ComfyUI version 0.3.39

* Added initial Flux.1 Kontext Pro Image node - recreated branch to save myself sanity from rebase crap after master got rebased

* Add safety filter to Kontext.

* Make safety = 2 and input image is optional.

* Add BFL kontext API nodes.

---------

Co-authored-by: Jedrzej Kosinski <kosinkadink1@gmail.com>

.ci/update_windows/update.py

@@ -63,6 +63,11 @@ except:
 print("checking out master branch")  # noqa: T201
 branch = repo.lookup_branch('master')
 if branch is None:
+    try:
+        ref = repo.lookup_reference('refs/remotes/origin/master')
+    except:
+        print("pulling.")  # noqa: T201
+        pull(repo)
         ref = repo.lookup_reference('refs/remotes/origin/master')
     repo.checkout(ref)
     branch = repo.lookup_branch('master')

.ci/windows_base_files/run_nvidia_gpu_fast_fp16_accumulation.bat

@@ -0,0 +1,2 @@
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --fast fp16_accumulation
+pause

.github/workflows/stable-release.yml

@@ -12,7 +12,7 @@ on:
         description: 'CUDA version'
         required: true
         type: string
-        default: "124"
+        default: "128"
       python_minor:
         description: 'Python minor version'
         required: true
@@ -22,7 +22,7 @@ on:
         description: 'Python patch version'
         required: true
         type: string
-        default: "8"
+        default: "10"
 
 
 jobs:
@@ -36,7 +36,7 @@ jobs:
       - uses: actions/checkout@v4
         with:
           ref: ${{ inputs.git_tag }}
-          fetch-depth: 0
+          fetch-depth: 150
           persist-credentials: false
       - uses: actions/cache/restore@v4
         id: cache
@@ -70,7 +70,7 @@ jobs:
             cd ..
 
           git clone --depth 1 https://github.com/comfyanonymous/taesd
-          cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
+          cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
 
           mkdir ComfyUI_windows_portable
           mv python_embeded ComfyUI_windows_portable
@@ -85,12 +85,14 @@ jobs:
 
           cd ..
 
-          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=8 -mfb=64 -md=32m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
+          "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
           mv ComfyUI_windows_portable.7z ComfyUI/ComfyUI_windows_portable_nvidia.7z
 
           cd ComfyUI_windows_portable
           python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu
 
+          python_embeded/python.exe -s ./update/update.py ComfyUI/
+
           ls
 
       - name: Upload binaries to release

.github/workflows/test-build.yml

@@ -18,7 +18,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12"]
+        python-version: ["3.9", "3.10", "3.11", "3.12", "3.13"]
     steps:
       - uses: actions/checkout@v4
       - name: Set up Python ${{ matrix.python-version }}

.github/workflows/test-launch.yml

@@ -17,7 +17,7 @@ jobs:
         path: "ComfyUI"
     - uses: actions/setup-python@v4
       with:
-        python-version: '3.9'
+        python-version: '3.10'
     - name: Install requirements
       run: |
         python -m pip install --upgrade pip

.github/workflows/test-unit.yml

@@ -18,7 +18,7 @@ jobs:
     - name: Set up Python      
       uses: actions/setup-python@v4
       with:
-        python-version: '3.10'
+        python-version: '3.12'
     - name: Install requirements
       run: |
         python -m pip install --upgrade pip

.github/workflows/update-api-stubs.yml

@@ -0,0 +1,56 @@
+name: Generate Pydantic Stubs from api.comfy.org
+
+on:
+  schedule:
+    - cron: '0 0 * * 1'
+  workflow_dispatch:
+
+jobs:
+  generate-models:
+    runs-on: ubuntu-latest
+    
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+      
+      - name: Set up Python
+        uses: actions/setup-python@v4
+        with:
+          python-version: '3.10'
+      
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install 'datamodel-code-generator[http]'
+          npm install @redocly/cli
+      
+      - name: Download OpenAPI spec
+        run: |
+          curl -o openapi.yaml https://api.comfy.org/openapi
+      
+      - name: Filter OpenAPI spec with Redocly
+        run: |
+          npx @redocly/cli bundle openapi.yaml --output filtered-openapi.yaml --config comfy_api_nodes/redocly.yaml --remove-unused-components
+      
+      - name: Generate API models
+        run: |
+          datamodel-codegen --use-subclass-enum --input filtered-openapi.yaml --output comfy_api_nodes/apis --output-model-type pydantic_v2.BaseModel
+      
+      - name: Check for changes
+        id: git-check
+        run: |
+          git diff --exit-code comfy_api_nodes/apis || echo "changes=true" >> $GITHUB_OUTPUT
+      
+      - name: Create Pull Request
+        if: steps.git-check.outputs.changes == 'true'
+        uses: peter-evans/create-pull-request@v5
+        with:
+          commit-message: 'chore: update API models from OpenAPI spec'
+          title: 'Update API models from api.comfy.org'
+          body: |
+            This PR updates the API models based on the latest api.comfy.org OpenAPI specification.
+            
+            Generated automatically by the a Github workflow.
+          branch: update-api-stubs
+          delete-branch: true
+          base: master

.github/workflows/update-frontend.yml

@@ -1,58 +0,0 @@
-name: Update Frontend Release
-
-on:
-  workflow_dispatch:
-    inputs:
-      version:
-        description: "Frontend version to update to (e.g., 1.0.0)"
-        required: true
-        type: string
-
-jobs:
-  update-frontend:
-    runs-on: ubuntu-latest
-    permissions:
-      contents: write
-      pull-requests: write
-
-    steps:
-      - name: Checkout ComfyUI
-        uses: actions/checkout@v4
-      - uses: actions/setup-python@v4
-        with:
-          python-version: '3.10'
-      - name: Install requirements
-        run: |
-          python -m pip install --upgrade pip
-          pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
-          pip install -r requirements.txt
-          pip install wait-for-it
-      # Frontend asset will be downloaded to ComfyUI/web_custom_versions/Comfy-Org_ComfyUI_frontend/{version}
-      - name: Start ComfyUI server
-        run: |
-          python main.py --cpu --front-end-version Comfy-Org/ComfyUI_frontend@${{ github.event.inputs.version }} 2>&1 | tee console_output.log &
-          wait-for-it --service 127.0.0.1:8188 -t 30
-      - name: Configure Git
-        run: |
-          git config --global user.name "GitHub Action"
-          git config --global user.email "action@github.com"
-      # Replace existing frontend content with the new version and remove .js.map files
-      # See https://github.com/Comfy-Org/ComfyUI_frontend/issues/2145 for why we remove .js.map files
-      - name: Update frontend content
-        run: |
-          rm -rf web/
-          cp -r web_custom_versions/Comfy-Org_ComfyUI_frontend/${{ github.event.inputs.version }} web/
-          rm web/**/*.js.map
-      - name: Create Pull Request
-        uses: peter-evans/create-pull-request@v7
-        with:
-          token: ${{ secrets.PR_BOT_PAT }}
-          commit-message: "Update frontend to v${{ github.event.inputs.version }}"
-          title: "Frontend Update: v${{ github.event.inputs.version }}"
-          body: |
-            Automated PR to update frontend content to version ${{ github.event.inputs.version }}
-
-            This PR was created automatically by the frontend update workflow.
-          branch: release-${{ github.event.inputs.version }}
-          base: master
-          labels: Frontend,dependencies

.github/workflows/windows_release_dependencies.yml

@@ -17,7 +17,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "124"
+        default: "128"
 
       python_minor:
         description: 'python minor version'
@@ -29,7 +29,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "8"
+        default: "10"
 #  push:
 #    branches:
 #      - master

.github/workflows/windows_release_nightly_pytorch.yml

@@ -7,7 +7,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "126"
+        default: "128"
 
       python_minor:
         description: 'python minor version'
@@ -19,7 +19,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "1"
+        default: "2"
 #  push:
 #    branches:
 #      - master
@@ -34,7 +34,7 @@ jobs:
     steps:
         - uses: actions/checkout@v4
           with:
-            fetch-depth: 0
+            fetch-depth: 30
             persist-credentials: false
         - uses: actions/setup-python@v5
           with:
@@ -56,7 +56,7 @@ jobs:
             cd ..
 
             git clone --depth 1 https://github.com/comfyanonymous/taesd
-            cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
+            cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
 
             mkdir ComfyUI_windows_portable_nightly_pytorch
             mv python_embeded ComfyUI_windows_portable_nightly_pytorch
@@ -74,7 +74,7 @@ jobs:
             pause" > ./update/update_comfyui_and_python_dependencies.bat
             cd ..
 
-            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=8 -mfb=64 -md=32m -ms=on -mf=BCJ2 ComfyUI_windows_portable_nightly_pytorch.7z ComfyUI_windows_portable_nightly_pytorch
+            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable_nightly_pytorch.7z ComfyUI_windows_portable_nightly_pytorch
             mv ComfyUI_windows_portable_nightly_pytorch.7z ComfyUI/ComfyUI_windows_portable_nvidia_or_cpu_nightly_pytorch.7z
 
             cd ComfyUI_windows_portable_nightly_pytorch

.github/workflows/windows_release_package.yml

@@ -7,7 +7,7 @@ on:
         description: 'cuda version'
         required: true
         type: string
-        default: "124"
+        default: "128"
 
       python_minor:
         description: 'python minor version'
@@ -19,7 +19,7 @@ on:
         description: 'python patch version'
         required: true
         type: string
-        default: "8"
+        default: "10"
 #  push:
 #    branches:
 #      - master
@@ -50,7 +50,7 @@ jobs:
 
         - uses: actions/checkout@v4
           with:
-            fetch-depth: 0
+            fetch-depth: 150
             persist-credentials: false
         - shell: bash
           run: |
@@ -67,7 +67,7 @@ jobs:
             cd ..
 
             git clone --depth 1 https://github.com/comfyanonymous/taesd
-            cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
+            cp taesd/*.safetensors ./ComfyUI_copy/models/vae_approx/
 
             mkdir ComfyUI_windows_portable
             mv python_embeded ComfyUI_windows_portable
@@ -82,12 +82,14 @@ jobs:
 
             cd ..
 
-            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=8 -mfb=64 -md=32m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
+            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=512m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
             mv ComfyUI_windows_portable.7z ComfyUI/new_ComfyUI_windows_portable_nvidia_cu${{ inputs.cu }}_or_cpu.7z
 
             cd ComfyUI_windows_portable
             python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu
 
+            python_embeded/python.exe -s ./update/update.py ComfyUI/
+
             ls
 
         - name: Upload binaries to release

.gitignore

@@ -21,3 +21,6 @@ venv/
 *.log
 web_custom_versions/
 .DS_Store
+openapi.yaml
+filtered-openapi.yaml
+uv.lock

CODEOWNERS

@@ -5,19 +5,20 @@
 # Inlined the team members for now.
 
 # Maintainers
-*.md @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink
-/tests/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink
-/tests-unit/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink
-/notebooks/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink
-/script_examples/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink
-/.github/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink
+*.md @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/tests/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/tests-unit/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/notebooks/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/script_examples/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/.github/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/requirements.txt @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
+/pyproject.toml @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @Kosinkadink @christian-byrne
 
 # Python web server
-/api_server/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata
-/app/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata
+/api_server/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @christian-byrne
+/app/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @christian-byrne
+/utils/ @yoland68 @robinjhuang @huchenlei @webfiltered @pythongosssss @ltdrdata @christian-byrne
 
-# Frontend assets
-/web/ @huchenlei @webfiltered @pythongosssss @yoland68 @robinjhuang
-
-# Extra nodes
-/comfy_extras/ @yoland68 @robinjhuang @huchenlei @pythongosssss @ltdrdata @Kosinkadink
+# Node developers
+/comfy_extras/ @yoland68 @robinjhuang @huchenlei @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne
+/comfy/comfy_types/ @yoland68 @robinjhuang @huchenlei @pythongosssss @ltdrdata @Kosinkadink @webfiltered @christian-byrne

README.md

@@ -1,7 +1,7 @@
 <div align="center">
 
 # ComfyUI
-**The most powerful and modular diffusion model GUI and backend.**
+**The most powerful and modular visual AI engine and application.**
 
 
 [![Website][website-shield]][website-url]
@@ -31,10 +31,23 @@
 ![ComfyUI Screenshot](https://github.com/user-attachments/assets/7ccaf2c1-9b72-41ae-9a89-5688c94b7abe)
 </div>
 
-This ui will let you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. For some workflow examples and see what ComfyUI can do you can check out:
-### [ComfyUI Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
+ComfyUI lets you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. Available on Windows, Linux, and macOS.
 
-### [Installing ComfyUI](#installing)
+## Get Started
+
+#### [Desktop Application](https://www.comfy.org/download)
+- The easiest way to get started. 
+- Available on Windows & macOS.
+
+#### [Windows Portable Package](#installing)
+- Get the latest commits and completely portable.
+- Available on Windows.
+
+#### [Manual Install](#manual-install-windows-linux)
+Supports all operating systems and GPU types (NVIDIA, AMD, Intel, Apple Silicon, Ascend).
+
+## [Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
+See what ComfyUI can do with the [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/).
 
 ## Features
 - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.
@@ -47,12 +60,20 @@ This ui will let you design and execute advanced stable diffusion pipelines usin
    - [AuraFlow](https://comfyanonymous.github.io/ComfyUI_examples/aura_flow/)
    - [HunyuanDiT](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_dit/)
    - [Flux](https://comfyanonymous.github.io/ComfyUI_examples/flux/)
+   - [Lumina Image 2.0](https://comfyanonymous.github.io/ComfyUI_examples/lumina2/)
+   - [HiDream](https://comfyanonymous.github.io/ComfyUI_examples/hidream/)
 - Video Models
    - [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/)
    - [Mochi](https://comfyanonymous.github.io/ComfyUI_examples/mochi/)
    - [LTX-Video](https://comfyanonymous.github.io/ComfyUI_examples/ltxv/)
    - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
+   - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos/)
+   - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/)
+- Audio Models
    - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
+   - [ACE Step](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
+- 3D Models
+   - [Hunyuan3D 2.0](https://docs.comfy.org/tutorials/3d/hunyuan3D-2)
 - Asynchronous Queue system
 - Many optimizations: Only re-executes the parts of the workflow that changes between executions.
 - Smart memory management: can automatically run models on GPUs with as low as 1GB vram.
@@ -79,6 +100,22 @@ This ui will let you design and execute advanced stable diffusion pipelines usin
 
 Workflow examples can be found on the [Examples page](https://comfyanonymous.github.io/ComfyUI_examples/)
 
+## Release Process
+
+ComfyUI follows a weekly release cycle every Friday, with three interconnected repositories:
+
+1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
+   - Releases a new stable version (e.g., v0.7.0)
+   - Serves as the foundation for the desktop release
+
+2. **[ComfyUI Desktop](https://github.com/Comfy-Org/desktop)**
+   - Builds a new release using the latest stable core version
+
+3. **[ComfyUI Frontend](https://github.com/Comfy-Org/ComfyUI_frontend)**
+   - Weekly frontend updates are merged into the core repository
+   - Features are frozen for the upcoming core release
+   - Development continues for the next release cycle
+
 ## Shortcuts
 
 | Keybind                            | Explanation                                                                                                        |
@@ -119,7 +156,7 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 
 # Installing
 
-## Windows
+## Windows Portable
 
 There is a portable standalone build for Windows that should work for running on Nvidia GPUs or for running on your CPU only on the [releases page](https://github.com/comfyanonymous/ComfyUI/releases).
 
@@ -137,9 +174,18 @@ See the [Config file](extra_model_paths.yaml.example) to set the search paths fo
 
 To run it on services like paperspace, kaggle or colab you can use my [Jupyter Notebook](notebooks/comfyui_colab.ipynb)
 
+
+## [comfy-cli](https://docs.comfy.org/comfy-cli/getting-started)
+
+You can install and start ComfyUI using comfy-cli:
+```bash
+pip install comfy-cli
+comfy install
+```
+
 ## Manual Install (Windows, Linux)
 
-Note that some dependencies do not yet support python 3.13 so using 3.12 is recommended.
+python 3.13 is supported but using 3.12 is recommended because some custom nodes and their dependencies might not support it yet.
 
 Git clone this repo.
 
@@ -151,11 +197,11 @@ Put your VAE in: models/vae
 ### AMD GPUs (Linux only)
 AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version:
 
-```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.2```
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.3```
 
-This is the command to install the nightly with ROCm 6.2 which might have some performance improvements:
+This is the command to install the nightly with ROCm 6.4 which might have some performance improvements:
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.2.4```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.4```
 
 ### Intel GPUs (Windows and Linux)
 
@@ -185,11 +231,11 @@ Additional discussion and help can be found [here](https://github.com/comfyanony
 
 Nvidia users should install stable pytorch using this command:
 
-```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu124```
+```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu128```
 
-This is the command to install pytorch nightly instead which might have performance improvements:
+This is the command to install pytorch nightly instead which might have performance improvements.
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu126```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128```
 
 #### Troubleshooting
 
@@ -233,6 +279,13 @@ For models compatible with Ascend Extension for PyTorch (torch_npu). To get star
 3. Next, install the necessary packages for torch-npu by adhering to the platform-specific instructions on the [Installation](https://ascend.github.io/docs/sources/pytorch/install.html#pytorch) page.
 4. Finally, adhere to the [ComfyUI manual installation](#manual-install-windows-linux) guide for Linux. Once all components are installed, you can run ComfyUI as described earlier.
 
+#### Cambricon MLUs
+
+For models compatible with Cambricon Extension for PyTorch (torch_mlu). Here's a step-by-step guide tailored to your platform and installation method:
+
+1. Install the Cambricon CNToolkit by adhering to the platform-specific instructions on the [Installation](https://www.cambricon.com/docs/sdk_1.15.0/cntoolkit_3.7.2/cntoolkit_install_3.7.2/index.html)
+2. Next, install the PyTorch(torch_mlu) following the instructions on the [Installation](https://www.cambricon.com/docs/sdk_1.15.0/cambricon_pytorch_1.17.0/user_guide_1.9/index.html)
+3. Launch ComfyUI by running `python main.py`
 
 # Running
 
@@ -248,7 +301,7 @@ For AMD 7600 and maybe other RDNA3 cards: ```HSA_OVERRIDE_GFX_VERSION=11.0.0 pyt
 
 ### AMD ROCm Tips
 
-You can enable experimental memory efficient attention on pytorch 2.5 in ComfyUI on RDNA3 and potentially other AMD GPUs using this command:
+You can enable experimental memory efficient attention on recent pytorch in ComfyUI on some AMD GPUs using this command, it should already be enabled by default on RDNA3. If this improves speed for you on latest pytorch on your GPU please report it so that I can enable it by default.
 
 ```TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1 python main.py --use-pytorch-cross-attention```
 
@@ -289,6 +342,8 @@ Use `--tls-keyfile key.pem --tls-certfile cert.pem` to enable TLS/SSL, the app w
 
 ## Support and dev channel
 
+[Discord](https://comfy.org/discord): Try the #help or #feedback channels.
+
 [Matrix space: #comfyui_space:matrix.org](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) (it's like discord but open source).
 
 See also: [https://www.comfy.org/](https://www.comfy.org/)
@@ -305,7 +360,7 @@ For any bugs, issues, or feature requests related to the frontend, please use th
 
 The new frontend is now the default for ComfyUI. However, please note:
 
-1. The frontend in the main ComfyUI repository is updated weekly.
+1. The frontend in the main ComfyUI repository is updated fortnightly.
 2. Daily releases are available in the separate frontend repository.
 
 To use the most up-to-date frontend version:
@@ -322,7 +377,7 @@ To use the most up-to-date frontend version:
    --front-end-version Comfy-Org/ComfyUI_frontend@1.2.2
    ```
 
-This approach allows you to easily switch between the stable weekly release and the cutting-edge daily updates, or even specific versions for testing purposes.
+This approach allows you to easily switch between the stable fortnightly release and the cutting-edge daily updates, or even specific versions for testing purposes.
 
 ### Accessing the Legacy Frontend
 

api_server/routes/internal/internal_routes.py

@@ -1,9 +1,9 @@
 from aiohttp import web
 from typing import Optional
-from folder_paths import models_dir, user_directory, output_directory, folder_names_and_paths
-from api_server.services.file_service import FileService
+from folder_paths import folder_names_and_paths, get_directory_by_type
 from api_server.services.terminal_service import TerminalService
 import app.logger
+import os
 
 class InternalRoutes:
     '''
@@ -15,26 +15,10 @@ class InternalRoutes:
     def __init__(self, prompt_server):
         self.routes: web.RouteTableDef = web.RouteTableDef()
         self._app: Optional[web.Application] = None
-        self.file_service = FileService({
-            "models": models_dir,
-            "user": user_directory,
-            "output": output_directory
-        })
         self.prompt_server = prompt_server
         self.terminal_service = TerminalService(prompt_server)
 
     def setup_routes(self):
-        @self.routes.get('/files')
-        async def list_files(request):
-            directory_key = request.query.get('directory', '')
-            try:
-                file_list = self.file_service.list_files(directory_key)
-                return web.json_response({"files": file_list})
-            except ValueError as e:
-                return web.json_response({"error": str(e)}, status=400)
-            except Exception as e:
-                return web.json_response({"error": str(e)}, status=500)
-
         @self.routes.get('/logs')
         async def get_logs(request):
             return web.json_response("".join([(l["t"] + " - " + l["m"]) for l in app.logger.get_logs()]))
@@ -67,6 +51,20 @@ class InternalRoutes:
                 response[key] = folder_names_and_paths[key][0]
             return web.json_response(response)
 
+        @self.routes.get('/files/{directory_type}')
+        async def get_files(request: web.Request) -> web.Response:
+            directory_type = request.match_info['directory_type']
+            if directory_type not in ("output", "input", "temp"):
+                return web.json_response({"error": "Invalid directory type"}, status=400)
+
+            directory = get_directory_by_type(directory_type)
+            sorted_files = sorted(
+                (entry for entry in os.scandir(directory) if entry.is_file()),
+                key=lambda entry: -entry.stat().st_mtime
+            )
+            return web.json_response([entry.name for entry in sorted_files], status=200)
+
+
     def get_app(self):
         if self._app is None:
             self._app = web.Application()

api_server/services/file_service.py

@@ -1,13 +0,0 @@
-from typing import Dict, List, Optional
-from api_server.utils.file_operations import FileSystemOperations, FileSystemItem
-
-class FileService:
-    def __init__(self, allowed_directories: Dict[str, str], file_system_ops: Optional[FileSystemOperations] = None):
-        self.allowed_directories: Dict[str, str] = allowed_directories
-        self.file_system_ops: FileSystemOperations = file_system_ops or FileSystemOperations()
-
-    def list_files(self, directory_key: str) -> List[FileSystemItem]:
-        if directory_key not in self.allowed_directories:
-            raise ValueError("Invalid directory key")
-        directory_path: str = self.allowed_directories[directory_key]
-        return self.file_system_ops.walk_directory(directory_path)

app/app_settings.py

@@ -9,8 +9,14 @@ class AppSettings():
         self.user_manager = user_manager
 
     def get_settings(self, request):
+        try:
             file = self.user_manager.get_request_user_filepath(
-            request, "comfy.settings.json")
+                request,
+                "comfy.settings.json"
+            )
+        except KeyError as e:
+            logging.error("User settings not found.")
+            raise web.HTTPUnauthorized() from e
         if os.path.isfile(file):
             try:
                 with open(file) as f:

app/custom_node_manager.py

@@ -4,31 +4,142 @@ import os
 import folder_paths
 import glob
 from aiohttp import web
+import json
+import logging
+from functools import lru_cache
+
+from utils.json_util import merge_json_recursive
+
+
+# Extra locale files to load into main.json
+EXTRA_LOCALE_FILES = [
+    "nodeDefs.json",
+    "commands.json",
+    "settings.json",
+]
+
+
+def safe_load_json_file(file_path: str) -> dict:
+    if not os.path.exists(file_path):
+        return {}
+
+    try:
+        with open(file_path, "r", encoding="utf-8") as f:
+            return json.load(f)
+    except json.JSONDecodeError:
+        logging.error(f"Error loading {file_path}")
+        return {}
+
 
 class CustomNodeManager:
+    @lru_cache(maxsize=1)
+    def build_translations(self):
+        """Load all custom nodes translations during initialization. Translations are
+        expected to be loaded from `locales/` folder.
+
+        The folder structure is expected to be the following:
+        - custom_nodes/
+            - custom_node_1/
+                - locales/
+                    - en/
+                        - main.json
+                        - commands.json
+                        - settings.json
+
+        returned translations are expected to be in the following format:
+        {
+            "en": {
+                "nodeDefs": {...},
+                "commands": {...},
+                "settings": {...},
+                ...{other main.json keys}
+            }
+        }
         """
-    Placeholder to refactor the custom node management features from ComfyUI-Manager.
-    Currently it only contains the custom workflow templates feature.
-    """
+
+        translations = {}
+
+        for folder in folder_paths.get_folder_paths("custom_nodes"):
+            # Sort glob results for deterministic ordering
+            for custom_node_dir in sorted(glob.glob(os.path.join(folder, "*/"))):
+                locales_dir = os.path.join(custom_node_dir, "locales")
+                if not os.path.exists(locales_dir):
+                    continue
+
+                for lang_dir in glob.glob(os.path.join(locales_dir, "*/")):
+                    lang_code = os.path.basename(os.path.dirname(lang_dir))
+
+                    if lang_code not in translations:
+                        translations[lang_code] = {}
+
+                    # Load main.json
+                    main_file = os.path.join(lang_dir, "main.json")
+                    node_translations = safe_load_json_file(main_file)
+
+                    # Load extra locale files
+                    for extra_file in EXTRA_LOCALE_FILES:
+                        extra_file_path = os.path.join(lang_dir, extra_file)
+                        key = extra_file.split(".")[0]
+                        json_data = safe_load_json_file(extra_file_path)
+                        if json_data:
+                            node_translations[key] = json_data
+
+                    if node_translations:
+                        translations[lang_code] = merge_json_recursive(
+                            translations[lang_code], node_translations
+                        )
+
+        return translations
+
     def add_routes(self, routes, webapp, loadedModules):
 
+        example_workflow_folder_names = ["example_workflows", "example", "examples", "workflow", "workflows"]
+
         @routes.get("/workflow_templates")
         async def get_workflow_templates(request):
             """Returns a web response that contains the map of custom_nodes names and their associated workflow templates. The ones without templates are omitted."""
-            files = [
-                file
-                for folder in folder_paths.get_folder_paths("custom_nodes")
-                for file in glob.glob(os.path.join(folder, '*/example_workflows/*.json'))
-            ]
-            workflow_templates_dict = {} # custom_nodes folder name -> example workflow names
+
+            files = []
+
+            for folder in folder_paths.get_folder_paths("custom_nodes"):
+                for folder_name in example_workflow_folder_names:
+                    pattern = os.path.join(folder, f"*/{folder_name}/*.json")
+                    matched_files = glob.glob(pattern)
+                    files.extend(matched_files)
+
+            workflow_templates_dict = (
+                {}
+            )  # custom_nodes folder name -> example workflow names
             for file in files:
-                custom_nodes_name = os.path.basename(os.path.dirname(os.path.dirname(file)))
+                custom_nodes_name = os.path.basename(
+                    os.path.dirname(os.path.dirname(file))
+                )
                 workflow_name = os.path.splitext(os.path.basename(file))[0]
-                workflow_templates_dict.setdefault(custom_nodes_name, []).append(workflow_name)
+                workflow_templates_dict.setdefault(custom_nodes_name, []).append(
+                    workflow_name
+                )
             return web.json_response(workflow_templates_dict)
 
         # Serve workflow templates from custom nodes.
         for module_name, module_dir in loadedModules:
-            workflows_dir = os.path.join(module_dir, 'example_workflows')
+            for folder_name in example_workflow_folder_names:
+                workflows_dir = os.path.join(module_dir, folder_name)
+
                 if os.path.exists(workflows_dir):
-                webapp.add_routes([web.static('/api/workflow_templates/' + module_name, workflows_dir)])
+                    if folder_name != "example_workflows":
+                        logging.debug(
+                            "Found example workflow folder '%s' for custom node '%s', consider renaming it to 'example_workflows'",
+                            folder_name, module_name)
+
+                    webapp.add_routes(
+                        [
+                            web.static(
+                                "/api/workflow_templates/" + module_name, workflows_dir
+                            )
+                        ]
+                    )
+
+        @routes.get("/i18n")
+        async def get_i18n(request):
+            """Returns translations from all custom nodes' locales folders."""
+            return web.json_response(self.build_translations())

app/frontend_management.py

@@ -3,16 +3,69 @@ import argparse
 import logging
 import os
 import re
+import sys
 import tempfile
 import zipfile
+import importlib
 from dataclasses import dataclass
 from functools import cached_property
 from pathlib import Path
 from typing import TypedDict, Optional
+from importlib.metadata import version
 
 import requests
 from typing_extensions import NotRequired
+
 from comfy.cli_args import DEFAULT_VERSION_STRING
+import app.logger
+
+# The path to the requirements.txt file
+req_path = Path(__file__).parents[1] / "requirements.txt"
+
+
+def frontend_install_warning_message():
+    """The warning message to display when the frontend version is not up to date."""
+
+    extra = ""
+    if sys.flags.no_user_site:
+        extra = "-s "
+    return f"""
+Please install the updated requirements.txt file by running:
+{sys.executable} {extra}-m pip install -r {req_path}
+
+This error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead.
+
+If you are on the portable package you can run: update\\update_comfyui.bat to solve this problem
+""".strip()
+
+
+def check_frontend_version():
+    """Check if the frontend version is up to date."""
+
+    def parse_version(version: str) -> tuple[int, int, int]:
+        return tuple(map(int, version.split(".")))
+
+    try:
+        frontend_version_str = version("comfyui-frontend-package")
+        frontend_version = parse_version(frontend_version_str)
+        with open(req_path, "r", encoding="utf-8") as f:
+            required_frontend = parse_version(f.readline().split("=")[-1])
+        if frontend_version < required_frontend:
+            app.logger.log_startup_warning(
+                f"""
+________________________________________________________________________
+WARNING WARNING WARNING WARNING WARNING
+
+Installed frontend version {".".join(map(str, frontend_version))} is lower than the recommended version {".".join(map(str, required_frontend))}.
+
+{frontend_install_warning_message()}
+________________________________________________________________________
+""".strip()
+            )
+        else:
+            logging.info("ComfyUI frontend version: {}".format(frontend_version_str))
+    except Exception as e:
+        logging.error(f"Failed to check frontend version: {e}")
 
 
 REQUEST_TIMEOUT = 10  # seconds
@@ -109,9 +162,49 @@ def download_release_asset_zip(release: Release, destination_path: str) -> None:
 
 
 class FrontendManager:
-    DEFAULT_FRONTEND_PATH = str(Path(__file__).parents[1] / "web")
     CUSTOM_FRONTENDS_ROOT = str(Path(__file__).parents[1] / "web_custom_versions")
 
+    @classmethod
+    def default_frontend_path(cls) -> str:
+        try:
+            import comfyui_frontend_package
+
+            return str(importlib.resources.files(comfyui_frontend_package) / "static")
+        except ImportError:
+            logging.error(
+                f"""
+********** ERROR ***********
+
+comfyui-frontend-package is not installed.
+
+{frontend_install_warning_message()}
+
+********** ERROR ***********
+""".strip()
+            )
+            sys.exit(-1)
+
+    @classmethod
+    def templates_path(cls) -> str:
+        try:
+            import comfyui_workflow_templates
+
+            return str(
+                importlib.resources.files(comfyui_workflow_templates) / "templates"
+            )
+        except ImportError:
+            logging.error(
+                f"""
+********** ERROR ***********
+
+comfyui-workflow-templates is not installed.
+
+{frontend_install_warning_message()}
+
+********** ERROR ***********
+""".strip()
+            )
+
     @classmethod
     def parse_version_string(cls, value: str) -> tuple[str, str, str]:
         """
@@ -132,7 +225,9 @@ class FrontendManager:
         return match_result.group(1), match_result.group(2), match_result.group(3)
 
     @classmethod
-    def init_frontend_unsafe(cls, version_string: str, provider: Optional[FrontEndProvider] = None) -> str:
+    def init_frontend_unsafe(
+        cls, version_string: str, provider: Optional[FrontEndProvider] = None
+    ) -> str:
         """
         Initializes the frontend for the specified version.
 
@@ -148,17 +243,26 @@ class FrontendManager:
             main error source might be request timeout or invalid URL.
         """
         if version_string == DEFAULT_VERSION_STRING:
-            return cls.DEFAULT_FRONTEND_PATH
+            check_frontend_version()
+            return cls.default_frontend_path()
 
         repo_owner, repo_name, version = cls.parse_version_string(version_string)
 
         if version.startswith("v"):
-            expected_path = str(Path(cls.CUSTOM_FRONTENDS_ROOT) / f"{repo_owner}_{repo_name}" / version.lstrip("v"))
+            expected_path = str(
+                Path(cls.CUSTOM_FRONTENDS_ROOT)
+                / f"{repo_owner}_{repo_name}"
+                / version.lstrip("v")
+            )
             if os.path.exists(expected_path):
-                logging.info(f"Using existing copy of specific frontend version tag: {repo_owner}/{repo_name}@{version}")
+                logging.info(
+                    f"Using existing copy of specific frontend version tag: {repo_owner}/{repo_name}@{version}"
+                )
                 return expected_path
 
-        logging.info(f"Initializing frontend: {repo_owner}/{repo_name}@{version}, requesting version details from GitHub...")
+        logging.info(
+            f"Initializing frontend: {repo_owner}/{repo_name}@{version}, requesting version details from GitHub..."
+        )
 
         provider = provider or FrontEndProvider(repo_owner, repo_name)
         release = provider.get_release(version)
@@ -201,4 +305,5 @@ class FrontendManager:
         except Exception as e:
             logging.error("Failed to initialize frontend: %s", e)
             logging.info("Falling back to the default frontend.")
-            return cls.DEFAULT_FRONTEND_PATH
+            check_frontend_version()
+            return cls.default_frontend_path()

app/logger.py

@@ -82,3 +82,17 @@ def setup_logger(log_level: str = 'INFO', capacity: int = 300, use_stdout: bool
         logger.addHandler(stdout_handler)
 
     logger.addHandler(stream_handler)
+
+
+STARTUP_WARNINGS = []
+
+
+def log_startup_warning(msg):
+    logging.warning(msg)
+    STARTUP_WARNINGS.append(msg)
+
+
+def print_startup_warnings():
+    for s in STARTUP_WARNINGS:
+        logging.warning(s)
+    STARTUP_WARNINGS.clear()

app/user_manager.py

@@ -197,6 +197,112 @@ class UserManager():
 
             return web.json_response(results)
 
+        @routes.get("/v2/userdata")
+        async def list_userdata_v2(request):
+            """
+            List files and directories in a user's data directory.
+
+            This endpoint provides a structured listing of contents within a specified
+            subdirectory of the user's data storage.
+
+            Query Parameters:
+            - path (optional): The relative path within the user's data directory
+                               to list. Defaults to the root ('').
+
+            Returns:
+            - 400: If the requested path is invalid, outside the user's data directory, or is not a directory.
+            - 404: If the requested path does not exist.
+            - 403: If the user is invalid.
+            - 500: If there is an error reading the directory contents.
+            - 200: JSON response containing a list of file and directory objects.
+                   Each object includes:
+                   - name: The name of the file or directory.
+                   - type: 'file' or 'directory'.
+                   - path: The relative path from the user's data root.
+                   - size (for files): The size in bytes.
+                   - modified (for files): The last modified timestamp (Unix epoch).
+            """
+            requested_rel_path = request.rel_url.query.get('path', '')
+
+            # URL-decode the path parameter
+            try:
+                requested_rel_path = parse.unquote(requested_rel_path)
+            except Exception as e:
+                logging.warning(f"Failed to decode path parameter: {requested_rel_path}, Error: {e}")
+                return web.Response(status=400, text="Invalid characters in path parameter")
+
+
+            # Check user validity and get the absolute path for the requested directory
+            try:
+                 base_user_path = self.get_request_user_filepath(request, None, create_dir=False)
+
+                 if requested_rel_path:
+                     target_abs_path = self.get_request_user_filepath(request, requested_rel_path, create_dir=False)
+                 else:
+                     target_abs_path = base_user_path
+
+            except KeyError as e:
+                 # Invalid user detected by get_request_user_id inside get_request_user_filepath
+                 logging.warning(f"Access denied for user: {e}")
+                 return web.Response(status=403, text="Invalid user specified in request")
+
+
+            if not target_abs_path:
+                 # Path traversal or other issue detected by get_request_user_filepath
+                 return web.Response(status=400, text="Invalid path requested")
+
+            # Handle cases where the user directory or target path doesn't exist
+            if not os.path.exists(target_abs_path):
+                # Check if it's the base user directory that's missing (new user case)
+                if target_abs_path == base_user_path:
+                    # It's okay if the base user directory doesn't exist yet, return empty list
+                     return web.json_response([])
+                else:
+                    # A specific subdirectory was requested but doesn't exist
+                     return web.Response(status=404, text="Requested path not found")
+
+            if not os.path.isdir(target_abs_path):
+                 return web.Response(status=400, text="Requested path is not a directory")
+
+            results = []
+            try:
+                for root, dirs, files in os.walk(target_abs_path, topdown=True):
+                    # Process directories
+                    for dir_name in dirs:
+                        dir_path = os.path.join(root, dir_name)
+                        rel_path = os.path.relpath(dir_path, base_user_path).replace(os.sep, '/')
+                        results.append({
+                            "name": dir_name,
+                            "path": rel_path,
+                            "type": "directory"
+                        })
+
+                    # Process files
+                    for file_name in files:
+                        file_path = os.path.join(root, file_name)
+                        rel_path = os.path.relpath(file_path, base_user_path).replace(os.sep, '/')
+                        entry_info = {
+                            "name": file_name,
+                            "path": rel_path,
+                            "type": "file"
+                        }
+                        try:
+                            stats = os.stat(file_path) # Use os.stat for potentially better performance with os.walk
+                            entry_info["size"] = stats.st_size
+                            entry_info["modified"] = stats.st_mtime
+                        except OSError as stat_error:
+                            logging.warning(f"Could not stat file {file_path}: {stat_error}")
+                            pass # Include file with available info
+                        results.append(entry_info)
+            except OSError as e:
+                logging.error(f"Error listing directory {target_abs_path}: {e}")
+                return web.Response(status=500, text="Error reading directory contents")
+
+            # Sort results alphabetically, directories first then files
+            results.sort(key=lambda x: (x['type'] != 'directory', x['name'].lower()))
+
+            return web.json_response(results)
+
         def get_user_data_path(request, check_exists = False, param = "file"):
             file = request.match_info.get(param, None)
             if not file:

comfy/cli_args.py

@@ -1,7 +1,6 @@
 import argparse
 import enum
 import os
-from typing import Optional
 import comfy.options
 
 
@@ -43,10 +42,11 @@ parser.add_argument("--tls-certfile", type=str, help="Path to TLS (SSL) certific
 parser.add_argument("--enable-cors-header", type=str, default=None, metavar="ORIGIN", nargs="?", const="*", help="Enable CORS (Cross-Origin Resource Sharing) with optional origin or allow all with default '*'.")
 parser.add_argument("--max-upload-size", type=float, default=100, help="Set the maximum upload size in MB.")
 
+parser.add_argument("--base-directory", type=str, default=None, help="Set the ComfyUI base directory for models, custom_nodes, input, output, temp, and user directories.")
 parser.add_argument("--extra-model-paths-config", type=str, default=None, metavar="PATH", nargs='+', action='append', help="Load one or more extra_model_paths.yaml files.")
-parser.add_argument("--output-directory", type=str, default=None, help="Set the ComfyUI output directory.")
-parser.add_argument("--temp-directory", type=str, default=None, help="Set the ComfyUI temp directory (default is in the ComfyUI directory).")
-parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory.")
+parser.add_argument("--output-directory", type=str, default=None, help="Set the ComfyUI output directory. Overrides --base-directory.")
+parser.add_argument("--temp-directory", type=str, default=None, help="Set the ComfyUI temp directory (default is in the ComfyUI directory). Overrides --base-directory.")
+parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.")
 parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.")
 parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.")
 parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use.")
@@ -66,6 +66,7 @@ fpunet_group.add_argument("--bf16-unet", action="store_true", help="Run the diff
 fpunet_group.add_argument("--fp16-unet", action="store_true", help="Run the diffusion model in fp16")
 fpunet_group.add_argument("--fp8_e4m3fn-unet", action="store_true", help="Store unet weights in fp8_e4m3fn.")
 fpunet_group.add_argument("--fp8_e5m2-unet", action="store_true", help="Store unet weights in fp8_e5m2.")
+fpunet_group.add_argument("--fp8_e8m0fnu-unet", action="store_true", help="Store unet weights in fp8_e8m0fnu.")
 
 fpvae_group = parser.add_mutually_exclusive_group()
 fpvae_group.add_argument("--fp16-vae", action="store_true", help="Run the VAE in fp16, might cause black images.")
@@ -79,6 +80,7 @@ fpte_group.add_argument("--fp8_e4m3fn-text-enc", action="store_true", help="Stor
 fpte_group.add_argument("--fp8_e5m2-text-enc", action="store_true", help="Store text encoder weights in fp8 (e5m2 variant).")
 fpte_group.add_argument("--fp16-text-enc", action="store_true", help="Store text encoder weights in fp16.")
 fpte_group.add_argument("--fp32-text-enc", action="store_true", help="Store text encoder weights in fp32.")
+fpte_group.add_argument("--bf16-text-enc", action="store_true", help="Store text encoder weights in bf16.")
 
 parser.add_argument("--force-channels-last", action="store_true", help="Force channels last format when inferencing the models.")
 
@@ -86,6 +88,7 @@ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE"
 
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
 parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.")
+parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"
@@ -100,12 +103,14 @@ parser.add_argument("--preview-size", type=int, default=512, help="Sets the maxi
 cache_group = parser.add_mutually_exclusive_group()
 cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
 cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
+cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.")
 
 attn_group = parser.add_mutually_exclusive_group()
 attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")
 attn_group.add_argument("--use-quad-cross-attention", action="store_true", help="Use the sub-quadratic cross attention optimization . Ignored when xformers is used.")
 attn_group.add_argument("--use-pytorch-cross-attention", action="store_true", help="Use the new pytorch 2.0 cross attention function.")
 attn_group.add_argument("--use-sage-attention", action="store_true", help="Use sage attention.")
+attn_group.add_argument("--use-flash-attention", action="store_true", help="Use FlashAttention.")
 
 parser.add_argument("--disable-xformers", action="store_true", help="Disable xformers.")
 
@@ -124,12 +129,21 @@ vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for e
 
 parser.add_argument("--reserve-vram", type=float, default=None, help="Set the amount of vram in GB you want to reserve for use by your OS/other software. By default some amount is reserved depending on your OS.")
 
+parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.")
 
 parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
 
 parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")
 parser.add_argument("--deterministic", action="store_true", help="Make pytorch use slower deterministic algorithms when it can. Note that this might not make images deterministic in all cases.")
-parser.add_argument("--fast", action="store_true", help="Enable some untested and potentially quality deteriorating optimizations.")
+
+class PerformanceFeature(enum.Enum):
+    Fp16Accumulation = "fp16_accumulation"
+    Fp8MatrixMultiplication = "fp8_matrix_mult"
+    CublasOps = "cublas_ops"
+
+parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: fp16_accumulation fp8_matrix_mult cublas_ops")
+
+parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.")
 
 parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")
 parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")
@@ -137,6 +151,7 @@ parser.add_argument("--windows-standalone-build", action="store_true", help="Win
 
 parser.add_argument("--disable-metadata", action="store_true", help="Disable saving prompt metadata in files.")
 parser.add_argument("--disable-all-custom-nodes", action="store_true", help="Disable loading all custom nodes.")
+parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.")
 
 parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.")
 
@@ -160,13 +175,14 @@ parser.add_argument(
     """,
 )
 
-def is_valid_directory(path: Optional[str]) -> Optional[str]:
-    """Validate if the given path is a directory."""
-    if path is None:
-        return None
-
+def is_valid_directory(path: str) -> str:
+    """Validate if the given path is a directory, and check permissions."""
+    if not os.path.exists(path):
+        raise argparse.ArgumentTypeError(f"The path '{path}' does not exist.")
     if not os.path.isdir(path):
-        raise argparse.ArgumentTypeError(f"{path} is not a valid directory.")
+        raise argparse.ArgumentTypeError(f"'{path}' is not a directory.")
+    if not os.access(path, os.R_OK):
+        raise argparse.ArgumentTypeError(f"You do not have read permissions for '{path}'.")
     return path
 
 parser.add_argument(
@@ -176,7 +192,16 @@ parser.add_argument(
     help="The local filesystem path to the directory where the frontend is located. Overrides --front-end-version.",
 )
 
-parser.add_argument("--user-directory", type=is_valid_directory, default=None, help="Set the ComfyUI user directory with an absolute path.")
+parser.add_argument("--user-directory", type=is_valid_directory, default=None, help="Set the ComfyUI user directory with an absolute path. Overrides --base-directory.")
+
+parser.add_argument("--enable-compress-response-body", action="store_true", help="Enable compressing response body.")
+
+parser.add_argument(
+    "--comfy-api-base",
+    type=str,
+    default="https://api.comfy.org",
+    help="Set the base URL for the ComfyUI API.  (default: https://api.comfy.org)",
+)
 
 if comfy.options.args_parsing:
     args = parser.parse_args()
@@ -188,3 +213,17 @@ if args.windows_standalone_build:
 
 if args.disable_auto_launch:
     args.auto_launch = False
+
+if args.force_fp16:
+    args.fp16_unet = True
+
+
+# '--fast' is not provided, use an empty set
+if args.fast is None:
+    args.fast = set()
+# '--fast' is provided with an empty list, enable all optimizations
+elif args.fast == []:
+    args.fast = set(PerformanceFeature)
+# '--fast' is provided with a list of performance features, use that list
+else:
+    args.fast = set(args.fast)

comfy/clip_model.py

@@ -97,14 +97,19 @@ class CLIPTextModel_(torch.nn.Module):
         self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
         self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
 
-    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32):
+    def forward(self, input_tokens=None, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32):
+        if embeds is not None:
+            x = embeds + comfy.ops.cast_to(self.embeddings.position_embedding.weight, dtype=dtype, device=embeds.device)
+        else:
             x = self.embeddings(input_tokens, dtype=dtype)
+
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
-            mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
+            mask = mask.masked_fill(mask.to(torch.bool), -torch.finfo(x.dtype).max)
+
+        causal_mask = torch.full((x.shape[1], x.shape[1]), -torch.finfo(x.dtype).max, dtype=x.dtype, device=x.device).triu_(1)
 
-        causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
         if mask is not None:
             mask += causal_mask
         else:
@@ -115,6 +120,9 @@ class CLIPTextModel_(torch.nn.Module):
         if i is not None and final_layer_norm_intermediate:
             i = self.final_layer_norm(i)
 
+        if num_tokens is not None:
+            pooled_output = x[list(range(x.shape[0])), list(map(lambda a: a - 1, num_tokens))]
+        else:
             pooled_output = x[torch.arange(x.shape[0], device=x.device), (torch.round(input_tokens).to(dtype=torch.int, device=x.device) == self.eos_token_id).int().argmax(dim=-1),]
         return x, i, pooled_output
 
@@ -203,6 +211,15 @@ class CLIPVision(torch.nn.Module):
             pooled_output = self.post_layernorm(x[:, 0, :])
         return x, i, pooled_output
 
+class LlavaProjector(torch.nn.Module):
+    def __init__(self, in_dim, out_dim, dtype, device, operations):
+        super().__init__()
+        self.linear_1 = operations.Linear(in_dim, out_dim, bias=True, device=device, dtype=dtype)
+        self.linear_2 = operations.Linear(out_dim, out_dim, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.linear_2(torch.nn.functional.gelu(self.linear_1(x[:, 1:])))
+
 class CLIPVisionModelProjection(torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
@@ -212,7 +229,16 @@ class CLIPVisionModelProjection(torch.nn.Module):
         else:
             self.visual_projection = lambda a: a
 
+        if "llava3" == config_dict.get("projector_type", None):
+            self.multi_modal_projector = LlavaProjector(config_dict["hidden_size"], 4096, dtype, device, operations)
+        else:
+            self.multi_modal_projector = None
+
     def forward(self, *args, **kwargs):
         x = self.vision_model(*args, **kwargs)
         out = self.visual_projection(x[2])
-        return (x[0], x[1], out)
+        projected = None
+        if self.multi_modal_projector is not None:
+            projected = self.multi_modal_projector(x[1])
+
+        return (x[0], x[1], out, projected)

comfy/clip_vision.py

@@ -9,6 +9,7 @@ import comfy.model_patcher
 import comfy.model_management
 import comfy.utils
 import comfy.clip_model
+import comfy.image_encoders.dino2
 
 class Output:
     def __getitem__(self, key):
@@ -17,6 +18,7 @@ class Output:
         setattr(self, key, item)
 
 def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True):
+    image = image[:, :, :, :3] if image.shape[3] > 3 else image
     mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
     std = torch.tensor(std, device=image.device, dtype=image.dtype)
     image = image.movedim(-1, 1)
@@ -34,6 +36,12 @@ def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], s
     image = torch.clip((255. * image), 0, 255).round() / 255.0
     return (image - mean.view([3,1,1])) / std.view([3,1,1])
 
+IMAGE_ENCODERS = {
+    "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
+    "siglip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
+    "dinov2": comfy.image_encoders.dino2.Dinov2Model,
+}
+
 class ClipVisionModel():
     def __init__(self, json_config):
         with open(json_config) as f:
@@ -42,10 +50,11 @@ class ClipVisionModel():
         self.image_size = config.get("image_size", 224)
         self.image_mean = config.get("image_mean", [0.48145466, 0.4578275, 0.40821073])
         self.image_std = config.get("image_std", [0.26862954, 0.26130258, 0.27577711])
+        model_class = IMAGE_ENCODERS.get(config.get("model_type", "clip_vision_model"))
         self.load_device = comfy.model_management.text_encoder_device()
         offload_device = comfy.model_management.text_encoder_offload_device()
         self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
-        self.model = comfy.clip_model.CLIPVisionModelProjection(config, self.dtype, offload_device, comfy.ops.manual_cast)
+        self.model = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast)
         self.model.eval()
 
         self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
@@ -65,6 +74,7 @@ class ClipVisionModel():
         outputs["last_hidden_state"] = out[0].to(comfy.model_management.intermediate_device())
         outputs["image_embeds"] = out[2].to(comfy.model_management.intermediate_device())
         outputs["penultimate_hidden_states"] = out[1].to(comfy.model_management.intermediate_device())
+        outputs["mm_projected"] = out[3]
         return outputs
 
 def convert_to_transformers(sd, prefix):
@@ -101,12 +111,21 @@ def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
     elif "vision_model.encoder.layers.30.layer_norm1.weight" in sd:
         json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_h.json")
     elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd:
+        embed_shape = sd["vision_model.embeddings.position_embedding.weight"].shape[0]
         if sd["vision_model.encoder.layers.0.layer_norm1.weight"].shape[0] == 1152:
+            if embed_shape == 729:
                 json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
-        elif sd["vision_model.embeddings.position_embedding.weight"].shape[0] == 577:
+            elif embed_shape == 1024:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json")
+        elif embed_shape == 577:
+            if "multi_modal_projector.linear_1.bias" in sd:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336_llava.json")
+            else:
                 json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336.json")
         else:
             json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl.json")
+    elif "embeddings.patch_embeddings.projection.weight" in sd:
+        json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "image_encoders"), "dino2_giant.json")
     else:
         return None
 

comfy/clip_vision_config_vitl_336_llava.json

@@ -0,0 +1,19 @@
+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 1024,
+  "image_size": 336,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-5,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 24,
+  "patch_size": 14,
+  "projection_dim": 768,
+  "projector_type": "llava3",
+  "torch_dtype": "float32"
+}

comfy/clip_vision_siglip_512.json

@@ -0,0 +1,13 @@
+{
+  "num_channels": 3,
+  "hidden_act": "gelu_pytorch_tanh",
+  "hidden_size": 1152,
+  "image_size": 512,
+  "intermediate_size": 4304,
+  "model_type": "siglip_vision_model",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 27,
+  "patch_size": 16,
+  "image_mean": [0.5, 0.5, 0.5],
+  "image_std": [0.5, 0.5, 0.5]
+}

comfy/comfy_types/__init__.py

@@ -1,6 +1,6 @@
 import torch
 from typing import Callable, Protocol, TypedDict, Optional, List
-from .node_typing import IO, InputTypeDict, ComfyNodeABC, CheckLazyMixin
+from .node_typing import IO, InputTypeDict, ComfyNodeABC, CheckLazyMixin, FileLocator
 
 
 class UnetApplyFunction(Protocol):
@@ -42,4 +42,5 @@ __all__ = [
     InputTypeDict.__name__,
     ComfyNodeABC.__name__,
     CheckLazyMixin.__name__,
+    FileLocator.__name__,
 ]

comfy/comfy_types/node_typing.py

@@ -1,7 +1,8 @@
 """Comfy-specific type hinting"""
 
 from __future__ import annotations
-from typing import Literal, TypedDict
+from typing import Literal, TypedDict, Optional
+from typing_extensions import NotRequired
 from abc import ABC, abstractmethod
 from enum import Enum
 
@@ -26,6 +27,7 @@ class IO(StrEnum):
     BOOLEAN = "BOOLEAN"
     INT = "INT"
     FLOAT = "FLOAT"
+    COMBO = "COMBO"
     CONDITIONING = "CONDITIONING"
     SAMPLER = "SAMPLER"
     SIGMAS = "SIGMAS"
@@ -46,6 +48,7 @@ class IO(StrEnum):
     FACE_ANALYSIS = "FACE_ANALYSIS"
     BBOX = "BBOX"
     SEGS = "SEGS"
+    VIDEO = "VIDEO"
 
     ANY = "*"
     """Always matches any type, but at a price.
@@ -67,90 +70,148 @@ class IO(StrEnum):
         return not (b.issubset(a) or a.issubset(b))
 
 
+class RemoteInputOptions(TypedDict):
+    route: str
+    """The route to the remote source."""
+    refresh_button: bool
+    """Specifies whether to show a refresh button in the UI below the widget."""
+    control_after_refresh: Literal["first", "last"]
+    """Specifies the control after the refresh button is clicked. If "first", the first item will be automatically selected, and so on."""
+    timeout: int
+    """The maximum amount of time to wait for a response from the remote source in milliseconds."""
+    max_retries: int
+    """The maximum number of retries before aborting the request."""
+    refresh: int
+    """The TTL of the remote input's value in milliseconds. Specifies the interval at which the remote input's value is refreshed."""
+
+
+class MultiSelectOptions(TypedDict):
+    placeholder: NotRequired[str]
+    """The placeholder text to display in the multi-select widget when no items are selected."""
+    chip: NotRequired[bool]
+    """Specifies whether to use chips instead of comma separated values for the multi-select widget."""
+
+
 class InputTypeOptions(TypedDict):
     """Provides type hinting for the return type of the INPUT_TYPES node function.
 
     Due to IDE limitations with unions, for now all options are available for all types (e.g. `label_on` is hinted even when the type is not `IO.BOOLEAN`).
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_datatypes
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/datatypes
     """
 
-    default: bool | str | float | int | list | tuple
+    default: NotRequired[bool | str | float | int | list | tuple]
     """The default value of the widget"""
-    defaultInput: bool
-    """Defaults to an input slot rather than a widget"""
-    forceInput: bool
-    """`defaultInput` and also don't allow converting to a widget"""
-    lazy: bool
+    defaultInput: NotRequired[bool]
+    """@deprecated in v1.16 frontend. v1.16 frontend allows input socket and widget to co-exist.
+    - defaultInput on required inputs should be dropped.
+    - defaultInput on optional inputs should be replaced with forceInput.
+    Ref: https://github.com/Comfy-Org/ComfyUI_frontend/pull/3364
+    """
+    forceInput: NotRequired[bool]
+    """Forces the input to be an input slot rather than a widget even a widget is available for the input type."""
+    lazy: NotRequired[bool]
     """Declares that this input uses lazy evaluation"""
-    rawLink: bool
+    rawLink: NotRequired[bool]
     """When a link exists, rather than receiving the evaluated value, you will receive the link (i.e. `["nodeId", <outputIndex>]`). Designed for node expansion."""
-    tooltip: str
+    tooltip: NotRequired[str]
     """Tooltip for the input (or widget), shown on pointer hover"""
+    socketless: NotRequired[bool]
+    """All inputs (including widgets) have an input socket to connect links. When ``true``, if there is a widget for this input, no socket will be created.
+    Available from frontend v1.17.5
+    Ref: https://github.com/Comfy-Org/ComfyUI_frontend/pull/3548
+    """
+    widgetType: NotRequired[str]
+    """Specifies a type to be used for widget initialization if different from the input type.
+    Available from frontend v1.18.0
+    https://github.com/Comfy-Org/ComfyUI_frontend/pull/3550"""
     # class InputTypeNumber(InputTypeOptions):
     # default: float | int
-    min: float
+    min: NotRequired[float]
     """The minimum value of a number (``FLOAT`` | ``INT``)"""
-    max: float
+    max: NotRequired[float]
     """The maximum value of a number (``FLOAT`` | ``INT``)"""
-    step: float
+    step: NotRequired[float]
     """The amount to increment or decrement a widget by when stepping up/down (``FLOAT`` | ``INT``)"""
-    round: float
+    round: NotRequired[float]
     """Floats are rounded by this value (``FLOAT``)"""
     # class InputTypeBoolean(InputTypeOptions):
     # default: bool
-    label_on: str
+    label_on: NotRequired[str]
     """The label to use in the UI when the bool is True (``BOOLEAN``)"""
-    label_on: str
+    label_off: NotRequired[str]
     """The label to use in the UI when the bool is False (``BOOLEAN``)"""
     # class InputTypeString(InputTypeOptions):
     # default: str
-    multiline: bool
+    multiline: NotRequired[bool]
     """Use a multiline text box (``STRING``)"""
-    placeholder: str
+    placeholder: NotRequired[str]
     """Placeholder text to display in the UI when empty (``STRING``)"""
     # Deprecated:
     # defaultVal: str
-    dynamicPrompts: bool
+    dynamicPrompts: NotRequired[bool]
     """Causes the front-end to evaluate dynamic prompts (``STRING``)"""
+    # class InputTypeCombo(InputTypeOptions):
+    image_upload: NotRequired[bool]
+    """Specifies whether the input should have an image upload button and image preview attached to it. Requires that the input's name is `image`."""
+    image_folder: NotRequired[Literal["input", "output", "temp"]]
+    """Specifies which folder to get preview images from if the input has the ``image_upload`` flag.
+    """
+    remote: NotRequired[RemoteInputOptions]
+    """Specifies the configuration for a remote input.
+    Available after ComfyUI frontend v1.9.7
+    https://github.com/Comfy-Org/ComfyUI_frontend/pull/2422"""
+    control_after_generate: NotRequired[bool]
+    """Specifies whether a control widget should be added to the input, adding options to automatically change the value after each prompt is queued. Currently only used for INT and COMBO types."""
+    options: NotRequired[list[str | int | float]]
+    """COMBO type only. Specifies the selectable options for the combo widget.
+    Prefer:
+    ["COMBO", {"options": ["Option 1", "Option 2", "Option 3"]}]
+    Over:
+    [["Option 1", "Option 2", "Option 3"]]
+    """
+    multi_select: NotRequired[MultiSelectOptions]
+    """COMBO type only. Specifies the configuration for a multi-select widget.
+    Available after ComfyUI frontend v1.13.4
+    https://github.com/Comfy-Org/ComfyUI_frontend/pull/2987"""
 
 
 class HiddenInputTypeDict(TypedDict):
     """Provides type hinting for the hidden entry of node INPUT_TYPES."""
 
-    node_id: Literal["UNIQUE_ID"]
+    node_id: NotRequired[Literal["UNIQUE_ID"]]
     """UNIQUE_ID is the unique identifier of the node, and matches the id property of the node on the client side. It is commonly used in client-server communications (see messages)."""
-    unique_id: Literal["UNIQUE_ID"]
+    unique_id: NotRequired[Literal["UNIQUE_ID"]]
     """UNIQUE_ID is the unique identifier of the node, and matches the id property of the node on the client side. It is commonly used in client-server communications (see messages)."""
-    prompt: Literal["PROMPT"]
+    prompt: NotRequired[Literal["PROMPT"]]
     """PROMPT is the complete prompt sent by the client to the server. See the prompt object for a full description."""
-    extra_pnginfo: Literal["EXTRA_PNGINFO"]
+    extra_pnginfo: NotRequired[Literal["EXTRA_PNGINFO"]]
     """EXTRA_PNGINFO is a dictionary that will be copied into the metadata of any .png files saved. Custom nodes can store additional information in this dictionary for saving (or as a way to communicate with a downstream node)."""
-    dynprompt: Literal["DYNPROMPT"]
+    dynprompt: NotRequired[Literal["DYNPROMPT"]]
     """DYNPROMPT is an instance of comfy_execution.graph.DynamicPrompt. It differs from PROMPT in that it may mutate during the course of execution in response to Node Expansion."""
 
 
 class InputTypeDict(TypedDict):
     """Provides type hinting for node INPUT_TYPES.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_more_on_inputs
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/more_on_inputs
     """
 
-    required: dict[str, tuple[IO, InputTypeOptions]]
+    required: NotRequired[dict[str, tuple[IO, InputTypeOptions]]]
     """Describes all inputs that must be connected for the node to execute."""
-    optional: dict[str, tuple[IO, InputTypeOptions]]
+    optional: NotRequired[dict[str, tuple[IO, InputTypeOptions]]]
     """Describes inputs which do not need to be connected."""
-    hidden: HiddenInputTypeDict
+    hidden: NotRequired[HiddenInputTypeDict]
     """Offers advanced functionality and server-client communication.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_more_on_inputs#hidden-inputs
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/more_on_inputs#hidden-inputs
     """
 
 
 class ComfyNodeABC(ABC):
     """Abstract base class for Comfy nodes.  Includes the names and expected types of attributes.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview
     """
 
     DESCRIPTION: str
@@ -167,12 +228,14 @@ class ComfyNodeABC(ABC):
     CATEGORY: str
     """The category of the node, as per the "Add Node" menu.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#category
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#category
     """
     EXPERIMENTAL: bool
     """Flags a node as experimental, informing users that it may change or not work as expected."""
     DEPRECATED: bool
     """Flags a node as deprecated, indicating to users that they should find alternatives to this node."""
+    API_NODE: Optional[bool]
+    """Flags a node as an API node. See: https://docs.comfy.org/tutorials/api-nodes/overview."""
 
     @classmethod
     @abstractmethod
@@ -181,9 +244,9 @@ class ComfyNodeABC(ABC):
 
         * Must include the ``required`` key, which describes all inputs that must be connected for the node to execute.
         * The ``optional`` key can be added to describe inputs which do not need to be connected.
-        * The ``hidden`` key offers some advanced functionality.  More info at: https://docs.comfy.org/essentials/custom_node_more_on_inputs#hidden-inputs
+        * The ``hidden`` key offers some advanced functionality.  More info at: https://docs.comfy.org/custom-nodes/backend/more_on_inputs#hidden-inputs
 
-        Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#input-types
+        Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#input-types
         """
         return {"required": {}}
 
@@ -198,7 +261,7 @@ class ComfyNodeABC(ABC):
 
     By default, a node is not considered an output. Set ``OUTPUT_NODE = True`` to specify that it is.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#output-node
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#output-node
     """
     INPUT_IS_LIST: bool
     """A flag indicating if this node implements the additional code necessary to deal with OUTPUT_IS_LIST nodes.
@@ -209,9 +272,9 @@ class ComfyNodeABC(ABC):
 
     A node can also override the default input behaviour and receive the whole list in a single call. This is done by setting a class attribute `INPUT_IS_LIST` to ``True``.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_lists#list-processing
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing
     """
-    OUTPUT_IS_LIST: tuple[bool]
+    OUTPUT_IS_LIST: tuple[bool, ...]
     """A tuple indicating which node outputs are lists, but will be connected to nodes that expect individual items.
 
     Connected nodes that do not implement `INPUT_IS_LIST` will be executed once for every item in the list.
@@ -227,29 +290,29 @@ class ComfyNodeABC(ABC):
     the node should provide a class attribute `OUTPUT_IS_LIST`, which is a ``tuple[bool]``, of the same length as `RETURN_TYPES`,
     specifying which outputs which should be so treated.
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_lists#list-processing
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lists#list-processing
     """
 
-    RETURN_TYPES: tuple[IO]
+    RETURN_TYPES: tuple[IO, ...]
     """A tuple representing the outputs of this node.
 
     Usage::
 
         RETURN_TYPES = (IO.INT, "INT", "CUSTOM_TYPE")
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#return-types
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-types
     """
-    RETURN_NAMES: tuple[str]
+    RETURN_NAMES: tuple[str, ...]
     """The output slot names for each item in `RETURN_TYPES`, e.g. ``RETURN_NAMES = ("count", "filter_string")``
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#return-names
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#return-names
     """
-    OUTPUT_TOOLTIPS: tuple[str]
+    OUTPUT_TOOLTIPS: tuple[str, ...]
     """A tuple of strings to use as tooltips for node outputs, one for each item in `RETURN_TYPES`."""
     FUNCTION: str
     """The name of the function to execute as a literal string, e.g. `FUNCTION = "execute"`
 
-    Comfy Docs: https://docs.comfy.org/essentials/custom_node_server_overview#function
+    Comfy Docs: https://docs.comfy.org/custom-nodes/backend/server_overview#function
     """
 
 
@@ -267,8 +330,19 @@ class CheckLazyMixin:
         Params should match the nodes execution ``FUNCTION`` (self, and all inputs by name).
         Will be executed repeatedly until it returns an empty list, or all requested items were already evaluated (and sent as params).
 
-        Comfy Docs: https://docs.comfy.org/essentials/custom_node_lazy_evaluation#defining-check-lazy-status
+        Comfy Docs: https://docs.comfy.org/custom-nodes/backend/lazy_evaluation#defining-check-lazy-status
         """
 
         need = [name for name in kwargs if kwargs[name] is None]
         return need
+
+
+class FileLocator(TypedDict):
+    """Provides type hinting for the file location"""
+
+    filename: str
+    """The filename of the file."""
+    subfolder: str
+    """The subfolder of the file."""
+    type: Literal["input", "output", "temp"]
+    """The root folder of the file."""

comfy/conds.py

@@ -3,9 +3,6 @@ import math
 import comfy.utils
 
 
-def lcm(a, b): #TODO: eventually replace by math.lcm (added in python3.9)
-    return abs(a*b) // math.gcd(a, b)
-
 class CONDRegular:
     def __init__(self, cond):
         self.cond = cond
@@ -27,6 +24,10 @@ class CONDRegular:
             conds.append(x.cond)
         return torch.cat(conds)
 
+    def size(self):
+        return list(self.cond.size())
+
+
 class CONDNoiseShape(CONDRegular):
     def process_cond(self, batch_size, device, area, **kwargs):
         data = self.cond
@@ -46,7 +47,7 @@ class CONDCrossAttn(CONDRegular):
             if s1[0] != s2[0] or s1[2] != s2[2]: #these 2 cases should not happen
                 return False
 
-            mult_min = lcm(s1[1], s2[1])
+            mult_min = math.lcm(s1[1], s2[1])
             diff = mult_min // min(s1[1], s2[1])
             if diff > 4: #arbitrary limit on the padding because it's probably going to impact performance negatively if it's too much
                 return False
@@ -57,7 +58,7 @@ class CONDCrossAttn(CONDRegular):
         crossattn_max_len = self.cond.shape[1]
         for x in others:
             c = x.cond
-            crossattn_max_len = lcm(crossattn_max_len, c.shape[1])
+            crossattn_max_len = math.lcm(crossattn_max_len, c.shape[1])
             conds.append(c)
 
         out = []
@@ -67,6 +68,7 @@ class CONDCrossAttn(CONDRegular):
             out.append(c)
         return torch.cat(out)
 
+
 class CONDConstant(CONDRegular):
     def __init__(self, cond):
         self.cond = cond
@@ -81,3 +83,6 @@ class CONDConstant(CONDRegular):
 
     def concat(self, others):
         return self.cond
+
+    def size(self):
+        return [1]

comfy/controlnet.py

@@ -418,10 +418,7 @@ def controlnet_config(sd, model_options={}):
         weight_dtype = comfy.utils.weight_dtype(sd)
 
         supported_inference_dtypes = list(model_config.supported_inference_dtypes)
-        if weight_dtype is not None:
-            supported_inference_dtypes.append(weight_dtype)
-
-        unet_dtype = comfy.model_management.unet_dtype(model_params=-1, supported_dtypes=supported_inference_dtypes)
+        unet_dtype = comfy.model_management.unet_dtype(model_params=-1, supported_dtypes=supported_inference_dtypes, weight_dtype=weight_dtype)
 
     load_device = comfy.model_management.get_torch_device()
     manual_cast_dtype = comfy.model_management.unet_manual_cast(unet_dtype, load_device)
@@ -689,10 +686,7 @@ def load_controlnet_state_dict(state_dict, model=None, model_options={}):
         if supported_inference_dtypes is None:
             supported_inference_dtypes = [comfy.model_management.unet_dtype()]
 
-        if weight_dtype is not None:
-            supported_inference_dtypes.append(weight_dtype)
-
-        unet_dtype = comfy.model_management.unet_dtype(model_params=-1, supported_dtypes=supported_inference_dtypes)
+        unet_dtype = comfy.model_management.unet_dtype(model_params=-1, supported_dtypes=supported_inference_dtypes, weight_dtype=weight_dtype)
 
     load_device = comfy.model_management.get_torch_device()
 
@@ -742,6 +736,7 @@ def load_controlnet_state_dict(state_dict, model=None, model_options={}):
     return control
 
 def load_controlnet(ckpt_path, model=None, model_options={}):
+    model_options = model_options.copy()
     if "global_average_pooling" not in model_options:
         filename = os.path.splitext(ckpt_path)[0]
         if filename.endswith("_shuffle") or filename.endswith("_shuffle_fp16"): #TODO: smarter way of enabling global_average_pooling

comfy/diffusers_convert.py

@@ -4,105 +4,6 @@ import logging
 
 # conversion code from https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_stable_diffusion.py
 
-# =================#
-# UNet Conversion #
-# =================#
-
-unet_conversion_map = [
-    # (stable-diffusion, HF Diffusers)
-    ("time_embed.0.weight", "time_embedding.linear_1.weight"),
-    ("time_embed.0.bias", "time_embedding.linear_1.bias"),
-    ("time_embed.2.weight", "time_embedding.linear_2.weight"),
-    ("time_embed.2.bias", "time_embedding.linear_2.bias"),
-    ("input_blocks.0.0.weight", "conv_in.weight"),
-    ("input_blocks.0.0.bias", "conv_in.bias"),
-    ("out.0.weight", "conv_norm_out.weight"),
-    ("out.0.bias", "conv_norm_out.bias"),
-    ("out.2.weight", "conv_out.weight"),
-    ("out.2.bias", "conv_out.bias"),
-]
-
-unet_conversion_map_resnet = [
-    # (stable-diffusion, HF Diffusers)
-    ("in_layers.0", "norm1"),
-    ("in_layers.2", "conv1"),
-    ("out_layers.0", "norm2"),
-    ("out_layers.3", "conv2"),
-    ("emb_layers.1", "time_emb_proj"),
-    ("skip_connection", "conv_shortcut"),
-]
-
-unet_conversion_map_layer = []
-# hardcoded number of downblocks and resnets/attentions...
-# would need smarter logic for other networks.
-for i in range(4):
-    # loop over downblocks/upblocks
-
-    for j in range(2):
-        # loop over resnets/attentions for downblocks
-        hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
-        sd_down_res_prefix = f"input_blocks.{3 * i + j + 1}.0."
-        unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
-
-        if i < 3:
-            # no attention layers in down_blocks.3
-            hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
-            sd_down_atn_prefix = f"input_blocks.{3 * i + j + 1}.1."
-            unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
-
-    for j in range(3):
-        # loop over resnets/attentions for upblocks
-        hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
-        sd_up_res_prefix = f"output_blocks.{3 * i + j}.0."
-        unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
-
-        if i > 0:
-            # no attention layers in up_blocks.0
-            hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
-            sd_up_atn_prefix = f"output_blocks.{3 * i + j}.1."
-            unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
-
-    if i < 3:
-        # no downsample in down_blocks.3
-        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
-        sd_downsample_prefix = f"input_blocks.{3 * (i + 1)}.0.op."
-        unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
-
-        # no upsample in up_blocks.3
-        hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
-        sd_upsample_prefix = f"output_blocks.{3 * i + 2}.{1 if i == 0 else 2}."
-        unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
-
-hf_mid_atn_prefix = "mid_block.attentions.0."
-sd_mid_atn_prefix = "middle_block.1."
-unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
-
-for j in range(2):
-    hf_mid_res_prefix = f"mid_block.resnets.{j}."
-    sd_mid_res_prefix = f"middle_block.{2 * j}."
-    unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
-
-
-def convert_unet_state_dict(unet_state_dict):
-    # buyer beware: this is a *brittle* function,
-    # and correct output requires that all of these pieces interact in
-    # the exact order in which I have arranged them.
-    mapping = {k: k for k in unet_state_dict.keys()}
-    for sd_name, hf_name in unet_conversion_map:
-        mapping[hf_name] = sd_name
-    for k, v in mapping.items():
-        if "resnets" in k:
-            for sd_part, hf_part in unet_conversion_map_resnet:
-                v = v.replace(hf_part, sd_part)
-            mapping[k] = v
-    for k, v in mapping.items():
-        for sd_part, hf_part in unet_conversion_map_layer:
-            v = v.replace(hf_part, sd_part)
-        mapping[k] = v
-    new_state_dict = {v: unet_state_dict[k] for k, v in mapping.items()}
-    return new_state_dict
-
-
 # ================#
 # VAE Conversion #
 # ================#
@@ -213,6 +114,7 @@ textenc_pattern = re.compile("|".join(protected.keys()))
 # Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp
 code2idx = {"q": 0, "k": 1, "v": 2}
 
+
 # This function exists because at the time of writing torch.cat can't do fp8 with cuda
 def cat_tensors(tensors):
     x = 0
@@ -229,6 +131,7 @@ def cat_tensors(tensors):
 
     return out
 
+
 def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""):
     new_state_dict = {}
     capture_qkv_weight = {}
@@ -284,5 +187,3 @@ def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""):
 
 def convert_text_enc_state_dict(text_enc_dict):
     return text_enc_dict
-
-

comfy/extra_samplers/uni_pc.py

@@ -661,7 +661,7 @@ class UniPC:
 
             if x_t is None:
                 if use_predictor:
-                    pred_res = torch.einsum('k,bkchw->bchw', rhos_p, D1s)
+                    pred_res = torch.tensordot(D1s, rhos_p, dims=([1], [0]))  # torch.einsum('k,bkchw->bchw', rhos_p, D1s)
                 else:
                     pred_res = 0
                 x_t = x_t_ - expand_dims(alpha_t * B_h, dims) * pred_res
@@ -669,7 +669,7 @@ class UniPC:
             if use_corrector:
                 model_t = self.model_fn(x_t, t)
                 if D1s is not None:
-                    corr_res = torch.einsum('k,bkchw->bchw', rhos_c[:-1], D1s)
+                    corr_res = torch.tensordot(D1s, rhos_c[:-1], dims=([1], [0]))  # torch.einsum('k,bkchw->bchw', rhos_c[:-1], D1s)
                 else:
                     corr_res = 0
                 D1_t = (model_t - model_prev_0)

comfy/image_encoders/dino2.py

@@ -0,0 +1,141 @@
+import torch
+from comfy.text_encoders.bert import BertAttention
+import comfy.model_management
+from comfy.ldm.modules.attention import optimized_attention_for_device
+
+
+class Dino2AttentionOutput(torch.nn.Module):
+    def __init__(self, input_dim, output_dim, layer_norm_eps, dtype, device, operations):
+        super().__init__()
+        self.dense = operations.Linear(input_dim, output_dim, dtype=dtype, device=device)
+
+    def forward(self, x):
+        return self.dense(x)
+
+
+class Dino2AttentionBlock(torch.nn.Module):
+    def __init__(self, embed_dim, heads, layer_norm_eps, dtype, device, operations):
+        super().__init__()
+        self.attention = BertAttention(embed_dim, heads, dtype, device, operations)
+        self.output = Dino2AttentionOutput(embed_dim, embed_dim, layer_norm_eps, dtype, device, operations)
+
+    def forward(self, x, mask, optimized_attention):
+        return self.output(self.attention(x, mask, optimized_attention))
+
+
+class LayerScale(torch.nn.Module):
+    def __init__(self, dim, dtype, device, operations):
+        super().__init__()
+        self.lambda1 = torch.nn.Parameter(torch.empty(dim, device=device, dtype=dtype))
+
+    def forward(self, x):
+        return x * comfy.model_management.cast_to_device(self.lambda1, x.device, x.dtype)
+
+
+class SwiGLUFFN(torch.nn.Module):
+    def __init__(self, dim, dtype, device, operations):
+        super().__init__()
+        in_features = out_features = dim
+        hidden_features = int(dim * 4)
+        hidden_features = (int(hidden_features * 2 / 3) + 7) // 8 * 8
+
+        self.weights_in = operations.Linear(in_features, 2 * hidden_features, bias=True, device=device, dtype=dtype)
+        self.weights_out = operations.Linear(hidden_features, out_features, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.weights_in(x)
+        x1, x2 = x.chunk(2, dim=-1)
+        x = torch.nn.functional.silu(x1) * x2
+        return self.weights_out(x)
+
+
+class Dino2Block(torch.nn.Module):
+    def __init__(self, dim, num_heads, layer_norm_eps, dtype, device, operations):
+        super().__init__()
+        self.attention = Dino2AttentionBlock(dim, num_heads, layer_norm_eps, dtype, device, operations)
+        self.layer_scale1 = LayerScale(dim, dtype, device, operations)
+        self.layer_scale2 = LayerScale(dim, dtype, device, operations)
+        self.mlp = SwiGLUFFN(dim, dtype, device, operations)
+        self.norm1 = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
+        self.norm2 = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
+
+    def forward(self, x, optimized_attention):
+        x = x + self.layer_scale1(self.attention(self.norm1(x), None, optimized_attention))
+        x = x + self.layer_scale2(self.mlp(self.norm2(x)))
+        return x
+
+
+class Dino2Encoder(torch.nn.Module):
+    def __init__(self, dim, num_heads, layer_norm_eps, num_layers, dtype, device, operations):
+        super().__init__()
+        self.layer = torch.nn.ModuleList([Dino2Block(dim, num_heads, layer_norm_eps, dtype, device, operations) for _ in range(num_layers)])
+
+    def forward(self, x, intermediate_output=None):
+        optimized_attention = optimized_attention_for_device(x.device, False, small_input=True)
+
+        if intermediate_output is not None:
+            if intermediate_output < 0:
+                intermediate_output = len(self.layer) + intermediate_output
+
+        intermediate = None
+        for i, l in enumerate(self.layer):
+            x = l(x, optimized_attention)
+            if i == intermediate_output:
+                intermediate = x.clone()
+        return x, intermediate
+
+
+class Dino2PatchEmbeddings(torch.nn.Module):
+    def __init__(self, dim, num_channels=3, patch_size=14, image_size=518, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.projection = operations.Conv2d(
+            in_channels=num_channels,
+            out_channels=dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=True,
+            dtype=dtype,
+            device=device
+        )
+
+    def forward(self, pixel_values):
+        return self.projection(pixel_values).flatten(2).transpose(1, 2)
+
+
+class Dino2Embeddings(torch.nn.Module):
+    def __init__(self, dim, dtype, device, operations):
+        super().__init__()
+        patch_size = 14
+        image_size = 518
+
+        self.patch_embeddings = Dino2PatchEmbeddings(dim, patch_size=patch_size, image_size=image_size, dtype=dtype, device=device, operations=operations)
+        self.position_embeddings = torch.nn.Parameter(torch.empty(1, (image_size // patch_size) ** 2 + 1, dim, dtype=dtype, device=device))
+        self.cls_token = torch.nn.Parameter(torch.empty(1, 1, dim, dtype=dtype, device=device))
+        self.mask_token = torch.nn.Parameter(torch.empty(1, dim, dtype=dtype, device=device))
+
+    def forward(self, pixel_values):
+        x = self.patch_embeddings(pixel_values)
+        # TODO: mask_token?
+        x = torch.cat((self.cls_token.to(device=x.device, dtype=x.dtype).expand(x.shape[0], -1, -1), x), dim=1)
+        x = x + comfy.model_management.cast_to_device(self.position_embeddings, x.device, x.dtype)
+        return x
+
+
+class Dinov2Model(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        num_layers = config_dict["num_hidden_layers"]
+        dim = config_dict["hidden_size"]
+        heads = config_dict["num_attention_heads"]
+        layer_norm_eps = config_dict["layer_norm_eps"]
+
+        self.embeddings = Dino2Embeddings(dim, dtype, device, operations)
+        self.encoder = Dino2Encoder(dim, heads, layer_norm_eps, num_layers, dtype, device, operations)
+        self.layernorm = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
+
+    def forward(self, pixel_values, attention_mask=None, intermediate_output=None):
+        x = self.embeddings(pixel_values)
+        x, i = self.encoder(x, intermediate_output=intermediate_output)
+        x = self.layernorm(x)
+        pooled_output = x[:, 0, :]
+        return x, i, pooled_output, None

comfy/image_encoders/dino2_giant.json

@@ -0,0 +1,21 @@
+{
+  "attention_probs_dropout_prob": 0.0,
+  "drop_path_rate": 0.0,
+  "hidden_act": "gelu",
+  "hidden_dropout_prob": 0.0,
+  "hidden_size": 1536,
+  "image_size": 518,
+  "initializer_range": 0.02,
+  "layer_norm_eps": 1e-06,
+  "layerscale_value": 1.0,
+  "mlp_ratio": 4,
+  "model_type": "dinov2",
+  "num_attention_heads": 24,
+  "num_channels": 3,
+  "num_hidden_layers": 40,
+  "patch_size": 14,
+  "qkv_bias": true,
+  "use_swiglu_ffn": true,
+  "image_mean": [0.485, 0.456, 0.406],
+  "image_std": [0.229, 0.224, 0.225]
+}

comfy/k_diffusion/sampling.py

@@ -40,7 +40,7 @@ def get_sigmas_polyexponential(n, sigma_min, sigma_max, rho=1., device='cpu'):
 def get_sigmas_vp(n, beta_d=19.9, beta_min=0.1, eps_s=1e-3, device='cpu'):
     """Constructs a continuous VP noise schedule."""
     t = torch.linspace(1, eps_s, n, device=device)
-    sigmas = torch.sqrt(torch.exp(beta_d * t ** 2 / 2 + beta_min * t) - 1)
+    sigmas = torch.sqrt(torch.special.expm1(beta_d * t ** 2 / 2 + beta_min * t))
     return append_zero(sigmas)
 
 
@@ -688,10 +688,10 @@ def sample_dpmpp_sde(model, x, sigmas, extra_args=None, callback=None, disable=N
     if len(sigmas) <= 1:
         return x
 
+    extra_args = {} if extra_args is None else extra_args
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     seed = extra_args.get("seed", None)
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
-    extra_args = {} if extra_args is None else extra_args
     s_in = x.new_ones([x.shape[0]])
     sigma_fn = lambda t: t.neg().exp()
     t_fn = lambda sigma: sigma.log().neg()
@@ -762,10 +762,10 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     if solver_type not in {'heun', 'midpoint'}:
         raise ValueError('solver_type must be \'heun\' or \'midpoint\'')
 
+    extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
-    extra_args = {} if extra_args is None else extra_args
     s_in = x.new_ones([x.shape[0]])
 
     old_denoised = None
@@ -808,10 +808,10 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     if len(sigmas) <= 1:
         return x
 
+    extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=seed, cpu=True) if noise_sampler is None else noise_sampler
-    extra_args = {} if extra_args is None else extra_args
     s_in = x.new_ones([x.shape[0]])
 
     denoised_1, denoised_2 = None, None
@@ -858,7 +858,7 @@ def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
 def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     if len(sigmas) <= 1:
         return x
-
+    extra_args = {} if extra_args is None else extra_args
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
     return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)
@@ -867,7 +867,7 @@ def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
 def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
     if len(sigmas) <= 1:
         return x
-
+    extra_args = {} if extra_args is None else extra_args
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
     return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
@@ -876,7 +876,7 @@ def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
 def sample_dpmpp_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=1 / 2):
     if len(sigmas) <= 1:
         return x
-
+    extra_args = {} if extra_args is None else extra_args
     sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
     noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
     return sample_dpmpp_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, r=r)
@@ -1267,7 +1267,7 @@ def sample_dpmpp_2m_cfg_pp(model, x, sigmas, extra_args=None, callback=None, dis
     return x
 
 @torch.no_grad()
-def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1., noise_sampler=None, cfg_pp=False):
+def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None, eta=1., cfg_pp=False):
     extra_args = {} if extra_args is None else extra_args
     seed = extra_args.get("seed", None)
     noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
@@ -1277,6 +1277,7 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None
     phi1_fn = lambda t: torch.expm1(t) / t
     phi2_fn = lambda t: (phi1_fn(t) - 1.0) / t
 
+    old_sigma_down = None
     old_denoised = None
     uncond_denoised = None
     def post_cfg_function(args):
@@ -1289,50 +1290,259 @@ def res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None
         extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
 
     for i in trange(len(sigmas) - 1, disable=disable):
-        if s_churn > 0:
-            gamma = min(s_churn / (len(sigmas) - 1), 2**0.5 - 1) if s_tmin <= sigmas[i] <= s_tmax else 0.0
-            sigma_hat = sigmas[i] * (gamma + 1)
-        else:
-            gamma = 0
-            sigma_hat = sigmas[i]
-
-        if gamma > 0:
-            eps = torch.randn_like(x) * s_noise
-            x = x + eps * (sigma_hat**2 - sigmas[i] ** 2) ** 0.5
-        denoised = model(x, sigma_hat * s_in, **extra_args)
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        sigma_down, sigma_up = get_ancestral_step(sigmas[i], sigmas[i + 1], eta=eta)
         if callback is not None:
-            callback({"x": x, "i": i, "sigma": sigmas[i], "sigma_hat": sigma_hat, "denoised": denoised})
-        if sigmas[i + 1] == 0 or old_denoised is None:
+            callback({"x": x, "i": i, "sigma": sigmas[i], "sigma_hat": sigmas[i], "denoised": denoised})
+        if sigma_down == 0 or old_denoised is None:
             # Euler method
             if cfg_pp:
-                d = to_d(x, sigma_hat, uncond_denoised)
-                x = denoised + d * sigmas[i + 1]
+                d = to_d(x, sigmas[i], uncond_denoised)
+                x = denoised + d * sigma_down
             else:
-                d = to_d(x, sigma_hat, denoised)
-                dt = sigmas[i + 1] - sigma_hat
+                d = to_d(x, sigmas[i], denoised)
+                dt = sigma_down - sigmas[i]
                 x = x + d * dt
         else:
             # Second order multistep method in https://arxiv.org/pdf/2308.02157
-            t, t_next, t_prev = t_fn(sigmas[i]), t_fn(sigmas[i + 1]), t_fn(sigmas[i - 1])
+            t, t_old, t_next, t_prev = t_fn(sigmas[i]), t_fn(old_sigma_down), t_fn(sigma_down), t_fn(sigmas[i - 1])
             h = t_next - t
-            c2 = (t_prev - t) / h
+            c2 = (t_prev - t_old) / h
 
             phi1_val, phi2_val = phi1_fn(-h), phi2_fn(-h)
-            b1 = torch.nan_to_num(phi1_val - 1.0 / c2 * phi2_val, nan=0.0)
-            b2 = torch.nan_to_num(1.0 / c2 * phi2_val, nan=0.0)
+            b1 = torch.nan_to_num(phi1_val - phi2_val / c2, nan=0.0)
+            b2 = torch.nan_to_num(phi2_val / c2, nan=0.0)
 
             if cfg_pp:
                 x = x + (denoised - uncond_denoised)
+                x = sigma_fn(h) * x + h * (b1 * uncond_denoised + b2 * old_denoised)
+            else:
+                x = sigma_fn(h) * x + h * (b1 * denoised + b2 * old_denoised)
 
-            x = (sigma_fn(t_next) / sigma_fn(t)) * x + h * (b1 * denoised + b2 * old_denoised)
+        # Noise addition
+        if sigmas[i + 1] > 0:
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * sigma_up
 
+        if cfg_pp:
+            old_denoised = uncond_denoised
+        else:
+            old_denoised = denoised
+        old_sigma_down = sigma_down
+    return x
+
+@torch.no_grad()
+def sample_res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None):
+    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=0., cfg_pp=False)
+
+@torch.no_grad()
+def sample_res_multistep_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None):
+    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=0., cfg_pp=True)
+
+@torch.no_grad()
+def sample_res_multistep_ancestral(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=False)
+
+@torch.no_grad()
+def sample_res_multistep_ancestral_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
+    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_noise=s_noise, noise_sampler=noise_sampler, eta=eta, cfg_pp=True)
+
+@torch.no_grad()
+def sample_gradient_estimation(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2., cfg_pp=False):
+    """Gradient-estimation sampler. Paper: https://openreview.net/pdf?id=o2ND9v0CeK"""
+    extra_args = {} if extra_args is None else extra_args
+    s_in = x.new_ones([x.shape[0]])
+    old_d = None
+
+    uncond_denoised = None
+    def post_cfg_function(args):
+        nonlocal uncond_denoised
+        uncond_denoised = args["uncond_denoised"]
+        return args["denoised"]
+
+    if cfg_pp:
+        model_options = extra_args.get("model_options", {}).copy()
+        extra_args["model_options"] = comfy.model_patcher.set_model_options_post_cfg_function(model_options, post_cfg_function, disable_cfg1_optimization=True)
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if cfg_pp:
+            d = to_d(x, sigmas[i], uncond_denoised)
+        else:
+            d = to_d(x, sigmas[i], denoised)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        dt = sigmas[i + 1] - sigmas[i]
+        if i == 0:
+            # Euler method
+            if cfg_pp:
+                x = denoised + d * sigmas[i + 1]
+            else:
+                x = x + d * dt
+        else:
+            # Gradient estimation
+            if cfg_pp:
+                d_bar = (ge_gamma - 1) * (d - old_d)
+                x = denoised + d * sigmas[i + 1] + d_bar * dt
+            else:
+                d_bar = ge_gamma * d + (1 - ge_gamma) * old_d
+                x = x + d_bar * dt
+        old_d = d
+    return x
+
+@torch.no_grad()
+def sample_gradient_estimation_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, ge_gamma=2.):
+    return sample_gradient_estimation(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, ge_gamma=ge_gamma, cfg_pp=True)
+
+@torch.no_grad()
+def sample_er_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, s_noise=1., noise_sampler=None, noise_scaler=None, max_stage=3):
+    """
+    Extended Reverse-Time SDE solver (VE ER-SDE-Solver-3). Arxiv: https://arxiv.org/abs/2309.06169.
+    Code reference: https://github.com/QinpengCui/ER-SDE-Solver/blob/main/er_sde_solver.py.
+    """
+    extra_args = {} if extra_args is None else extra_args
+    seed = extra_args.get("seed", None)
+    noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+
+    def default_noise_scaler(sigma):
+        return sigma * ((sigma ** 0.3).exp() + 10.0)
+    noise_scaler = default_noise_scaler if noise_scaler is None else noise_scaler
+    num_integration_points = 200.0
+    point_indice = torch.arange(0, num_integration_points, dtype=torch.float32, device=x.device)
+
+    old_denoised = None
+    old_denoised_d = None
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        stage_used = min(max_stage, i + 1)
+        if sigmas[i + 1] == 0:
+            x = denoised
+        elif stage_used == 1:
+            r = noise_scaler(sigmas[i + 1]) / noise_scaler(sigmas[i])
+            x = r * x + (1 - r) * denoised
+        else:
+            r = noise_scaler(sigmas[i + 1]) / noise_scaler(sigmas[i])
+            x = r * x + (1 - r) * denoised
+
+            dt = sigmas[i + 1] - sigmas[i]
+            sigma_step_size = -dt / num_integration_points
+            sigma_pos = sigmas[i + 1] + point_indice * sigma_step_size
+            scaled_pos = noise_scaler(sigma_pos)
+
+            # Stage 2
+            s = torch.sum(1 / scaled_pos) * sigma_step_size
+            denoised_d = (denoised - old_denoised) / (sigmas[i] - sigmas[i - 1])
+            x = x + (dt + s * noise_scaler(sigmas[i + 1])) * denoised_d
+
+            if stage_used >= 3:
+                # Stage 3
+                s_u = torch.sum((sigma_pos - sigmas[i]) / scaled_pos) * sigma_step_size
+                denoised_u = (denoised_d - old_denoised_d) / ((sigmas[i] - sigmas[i - 2]) / 2)
+                x = x + ((dt ** 2) / 2 + s_u * noise_scaler(sigmas[i + 1])) * denoised_u
+            old_denoised_d = denoised_d
+
+        if s_noise != 0 and sigmas[i + 1] > 0:
+            x = x + noise_sampler(sigmas[i], sigmas[i + 1]) * s_noise * (sigmas[i + 1] ** 2 - sigmas[i] ** 2 * r ** 2).sqrt().nan_to_num(nan=0.0)
         old_denoised = denoised
     return x
 
 @torch.no_grad()
-def sample_res_multistep(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1., noise_sampler=None):
-    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_churn=s_churn, s_tmin=s_tmin, s_tmax=s_tmax, s_noise=s_noise, noise_sampler=noise_sampler, cfg_pp=False)
+def sample_seeds_2(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r=0.5):
+    '''
+    SEEDS-2 - Stochastic Explicit Exponential Derivative-free Solvers (VE Data Prediction) stage 2
+    Arxiv: https://arxiv.org/abs/2305.14267
+    '''
+    extra_args = {} if extra_args is None else extra_args
+    seed = extra_args.get("seed", None)
+    noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+
+    inject_noise = eta > 0 and s_noise > 0
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            x = denoised
+        else:
+            t, t_next = -sigmas[i].log(), -sigmas[i + 1].log()
+            h = t_next - t
+            h_eta = h * (eta + 1)
+            s = t + r * h
+            fac = 1 / (2 * r)
+            sigma_s = s.neg().exp()
+
+            coeff_1, coeff_2 = (-r * h_eta).expm1(), (-h_eta).expm1()
+            if inject_noise:
+                noise_coeff_1 = (-2 * r * h * eta).expm1().neg().sqrt()
+                noise_coeff_2 = ((-2 * r * h * eta).expm1() - (-2 * h * eta).expm1()).sqrt()
+                noise_1, noise_2 = noise_sampler(sigmas[i], sigma_s), noise_sampler(sigma_s, sigmas[i + 1])
+
+            # Step 1
+            x_2 = (coeff_1 + 1) * x - coeff_1 * denoised
+            if inject_noise:
+                x_2 = x_2 + sigma_s * (noise_coeff_1 * noise_1) * s_noise
+            denoised_2 = model(x_2, sigma_s * s_in, **extra_args)
+
+            # Step 2
+            denoised_d = (1 - fac) * denoised + fac * denoised_2
+            x = (coeff_2 + 1) * x - coeff_2 * denoised_d
+            if inject_noise:
+                x = x + sigmas[i + 1] * (noise_coeff_2 * noise_1 + noise_coeff_1 * noise_2) * s_noise
+    return x
 
 @torch.no_grad()
-def sample_res_multistep_cfg_pp(model, x, sigmas, extra_args=None, callback=None, disable=None, s_churn=0., s_tmin=0., s_tmax=float('inf'), s_noise=1., noise_sampler=None):
-    return res_multistep(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, s_churn=s_churn, s_tmin=s_tmin, s_tmax=s_tmax, s_noise=s_noise, noise_sampler=noise_sampler, cfg_pp=True)
+def sample_seeds_3(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, r_1=1./3, r_2=2./3):
+    '''
+    SEEDS-3 - Stochastic Explicit Exponential Derivative-free Solvers (VE Data Prediction) stage 3
+    Arxiv: https://arxiv.org/abs/2305.14267
+    '''
+    extra_args = {} if extra_args is None else extra_args
+    seed = extra_args.get("seed", None)
+    noise_sampler = default_noise_sampler(x, seed=seed) if noise_sampler is None else noise_sampler
+    s_in = x.new_ones([x.shape[0]])
+
+    inject_noise = eta > 0 and s_noise > 0
+
+    for i in trange(len(sigmas) - 1, disable=disable):
+        denoised = model(x, sigmas[i] * s_in, **extra_args)
+        if callback is not None:
+            callback({'x': x, 'i': i, 'sigma': sigmas[i], 'sigma_hat': sigmas[i], 'denoised': denoised})
+        if sigmas[i + 1] == 0:
+            x = denoised
+        else:
+            t, t_next = -sigmas[i].log(), -sigmas[i + 1].log()
+            h = t_next - t
+            h_eta = h * (eta + 1)
+            s_1 = t + r_1 * h
+            s_2 = t + r_2 * h
+            sigma_s_1, sigma_s_2 = s_1.neg().exp(), s_2.neg().exp()
+
+            coeff_1, coeff_2, coeff_3 = (-r_1 * h_eta).expm1(), (-r_2 * h_eta).expm1(), (-h_eta).expm1()
+            if inject_noise:
+                noise_coeff_1 = (-2 * r_1 * h * eta).expm1().neg().sqrt()
+                noise_coeff_2 = ((-2 * r_1 * h * eta).expm1() - (-2 * r_2 * h * eta).expm1()).sqrt()
+                noise_coeff_3 = ((-2 * r_2 * h * eta).expm1() - (-2 * h * eta).expm1()).sqrt()
+                noise_1, noise_2, noise_3 = noise_sampler(sigmas[i], sigma_s_1), noise_sampler(sigma_s_1, sigma_s_2), noise_sampler(sigma_s_2, sigmas[i + 1])
+
+            # Step 1
+            x_2 = (coeff_1 + 1) * x - coeff_1 * denoised
+            if inject_noise:
+                x_2 = x_2 + sigma_s_1 * (noise_coeff_1 * noise_1) * s_noise
+            denoised_2 = model(x_2, sigma_s_1 * s_in, **extra_args)
+
+            # Step 2
+            x_3 = (coeff_2 + 1) * x - coeff_2 * denoised + (r_2 / r_1) * (coeff_2 / (r_2 * h_eta) + 1) * (denoised_2 - denoised)
+            if inject_noise:
+                x_3 = x_3 + sigma_s_2 * (noise_coeff_2 * noise_1 + noise_coeff_1 * noise_2) * s_noise
+            denoised_3 = model(x_3, sigma_s_2 * s_in, **extra_args)
+
+            # Step 3
+            x = (coeff_3 + 1) * x - coeff_3 * denoised + (1. / r_2) * (coeff_3 / h_eta + 1) * (denoised_3 - denoised)
+            if inject_noise:
+                x = x + sigmas[i + 1] * (noise_coeff_3 * noise_1 + noise_coeff_2 * noise_2 + noise_coeff_1 * noise_3) * s_noise
+    return x

comfy/latent_formats.py

@@ -407,3 +407,66 @@ class Cosmos1CV8x8x8(LatentFormat):
     ]
 
     latent_rgb_factors_bias = [-0.1223, -0.1889, -0.1976]
+
+class Wan21(LatentFormat):
+    latent_channels = 16
+    latent_dimensions = 3
+
+    latent_rgb_factors = [
+            [-0.1299, -0.1692,  0.2932],
+            [ 0.0671,  0.0406,  0.0442],
+            [ 0.3568,  0.2548,  0.1747],
+            [ 0.0372,  0.2344,  0.1420],
+            [ 0.0313,  0.0189, -0.0328],
+            [ 0.0296, -0.0956, -0.0665],
+            [-0.3477, -0.4059, -0.2925],
+            [ 0.0166,  0.1902,  0.1975],
+            [-0.0412,  0.0267, -0.1364],
+            [-0.1293,  0.0740,  0.1636],
+            [ 0.0680,  0.3019,  0.1128],
+            [ 0.0032,  0.0581,  0.0639],
+            [-0.1251,  0.0927,  0.1699],
+            [ 0.0060, -0.0633,  0.0005],
+            [ 0.3477,  0.2275,  0.2950],
+            [ 0.1984,  0.0913,  0.1861]
+        ]
+
+    latent_rgb_factors_bias = [-0.1835, -0.0868, -0.3360]
+
+    def __init__(self):
+        self.scale_factor = 1.0
+        self.latents_mean = torch.tensor([
+            -0.7571, -0.7089, -0.9113, 0.1075, -0.1745, 0.9653, -0.1517, 1.5508,
+            0.4134, -0.0715, 0.5517, -0.3632, -0.1922, -0.9497, 0.2503, -0.2921
+        ]).view(1, self.latent_channels, 1, 1, 1)
+        self.latents_std = torch.tensor([
+            2.8184, 1.4541, 2.3275, 2.6558, 1.2196, 1.7708, 2.6052, 2.0743,
+            3.2687, 2.1526, 2.8652, 1.5579, 1.6382, 1.1253, 2.8251, 1.9160
+        ]).view(1, self.latent_channels, 1, 1, 1)
+
+
+        self.taesd_decoder_name = None #TODO
+
+    def process_in(self, latent):
+        latents_mean = self.latents_mean.to(latent.device, latent.dtype)
+        latents_std = self.latents_std.to(latent.device, latent.dtype)
+        return (latent - latents_mean) * self.scale_factor / latents_std
+
+    def process_out(self, latent):
+        latents_mean = self.latents_mean.to(latent.device, latent.dtype)
+        latents_std = self.latents_std.to(latent.device, latent.dtype)
+        return latent * latents_std / self.scale_factor + latents_mean
+
+class Hunyuan3Dv2(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+    scale_factor = 0.9990943042622529
+
+class Hunyuan3Dv2mini(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+    scale_factor = 1.0188137142395404
+
+class ACEAudio(LatentFormat):
+    latent_channels = 8
+    latent_dimensions = 2

comfy/ldm/ace/attention.py

@@ -0,0 +1,761 @@
+# Original from: https://github.com/ace-step/ACE-Step/blob/main/models/attention.py
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Tuple, Union, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+import comfy.model_management
+from comfy.ldm.modules.attention import optimized_attention
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        query_dim: int,
+        cross_attention_dim: Optional[int] = None,
+        heads: int = 8,
+        kv_heads: Optional[int] = None,
+        dim_head: int = 64,
+        dropout: float = 0.0,
+        bias: bool = False,
+        qk_norm: Optional[str] = None,
+        added_kv_proj_dim: Optional[int] = None,
+        added_proj_bias: Optional[bool] = True,
+        out_bias: bool = True,
+        scale_qk: bool = True,
+        only_cross_attention: bool = False,
+        eps: float = 1e-5,
+        rescale_output_factor: float = 1.0,
+        residual_connection: bool = False,
+        processor=None,
+        out_dim: int = None,
+        out_context_dim: int = None,
+        context_pre_only=None,
+        pre_only=False,
+        elementwise_affine: bool = True,
+        is_causal: bool = False,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads
+        self.query_dim = query_dim
+        self.use_bias = bias
+        self.is_cross_attention = cross_attention_dim is not None
+        self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
+        self.rescale_output_factor = rescale_output_factor
+        self.residual_connection = residual_connection
+        self.dropout = dropout
+        self.fused_projections = False
+        self.out_dim = out_dim if out_dim is not None else query_dim
+        self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim
+        self.context_pre_only = context_pre_only
+        self.pre_only = pre_only
+        self.is_causal = is_causal
+
+        self.scale_qk = scale_qk
+        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
+
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        # for slice_size > 0 the attention score computation
+        # is split across the batch axis to save memory
+        # You can set slice_size with `set_attention_slice`
+        self.sliceable_head_dim = heads
+
+        self.added_kv_proj_dim = added_kv_proj_dim
+        self.only_cross_attention = only_cross_attention
+
+        if self.added_kv_proj_dim is None and self.only_cross_attention:
+            raise ValueError(
+                "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."
+            )
+
+        self.group_norm = None
+        self.spatial_norm = None
+
+        self.norm_q = None
+        self.norm_k = None
+
+        self.norm_cross = None
+        self.to_q = operations.Linear(query_dim, self.inner_dim, bias=bias, dtype=dtype, device=device)
+
+        if not self.only_cross_attention:
+            # only relevant for the `AddedKVProcessor` classes
+            self.to_k = operations.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+            self.to_v = operations.Linear(self.cross_attention_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+        else:
+            self.to_k = None
+            self.to_v = None
+
+        self.added_proj_bias = added_proj_bias
+        if self.added_kv_proj_dim is not None:
+            self.add_k_proj = operations.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias, dtype=dtype, device=device)
+            self.add_v_proj = operations.Linear(added_kv_proj_dim, self.inner_kv_dim, bias=added_proj_bias, dtype=dtype, device=device)
+            if self.context_pre_only is not None:
+                self.add_q_proj = operations.Linear(added_kv_proj_dim, self.inner_dim, bias=added_proj_bias, dtype=dtype, device=device)
+        else:
+            self.add_q_proj = None
+            self.add_k_proj = None
+            self.add_v_proj = None
+
+        if not self.pre_only:
+            self.to_out = nn.ModuleList([])
+            self.to_out.append(operations.Linear(self.inner_dim, self.out_dim, bias=out_bias, dtype=dtype, device=device))
+            self.to_out.append(nn.Dropout(dropout))
+        else:
+            self.to_out = None
+
+        if self.context_pre_only is not None and not self.context_pre_only:
+            self.to_add_out = operations.Linear(self.inner_dim, self.out_context_dim, bias=out_bias, dtype=dtype, device=device)
+        else:
+            self.to_add_out = None
+
+        self.norm_added_q = None
+        self.norm_added_k = None
+        self.processor = processor
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        **cross_attention_kwargs,
+    ) -> torch.Tensor:
+        return self.processor(
+            self,
+            hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+
+class CustomLiteLAProcessor2_0:
+    """Attention processor used typically in processing the SD3-like self-attention projections. add rms norm for query and key and apply RoPE"""
+
+    def __init__(self):
+        self.kernel_func = nn.ReLU(inplace=False)
+        self.eps = 1e-15
+        self.pad_val = 1.0
+
+    def apply_rotary_emb(
+        self,
+        x: torch.Tensor,
+        freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
+        to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
+        reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
+        tensors contain rotary embeddings and are returned as real tensors.
+
+        Args:
+            x (`torch.Tensor`):
+                Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
+            freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
+        """
+        cos, sin = freqs_cis  # [S, D]
+        cos = cos[None, None]
+        sin = sin[None, None]
+        cos, sin = cos.to(x.device), sin.to(x.device)
+
+        x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+        x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+        return out
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        hidden_states_len = hidden_states.shape[1]
+
+        input_ndim = hidden_states.ndim
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+        if encoder_hidden_states is not None:
+            context_input_ndim = encoder_hidden_states.ndim
+            if context_input_ndim == 4:
+                batch_size, channel, height, width = encoder_hidden_states.shape
+                encoder_hidden_states = encoder_hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size = hidden_states.shape[0]
+
+        # `sample` projections.
+        dtype = hidden_states.dtype
+        query = attn.to_q(hidden_states)
+        key = attn.to_k(hidden_states)
+        value = attn.to_v(hidden_states)
+
+        # `context` projections.
+        has_encoder_hidden_state_proj = hasattr(attn, "add_q_proj") and hasattr(attn, "add_k_proj") and hasattr(attn, "add_v_proj")
+        if encoder_hidden_states is not None and has_encoder_hidden_state_proj:
+            encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
+            encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
+            encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
+
+            # attention
+            if not attn.is_cross_attention:
+                query = torch.cat([query, encoder_hidden_states_query_proj], dim=1)
+                key = torch.cat([key, encoder_hidden_states_key_proj], dim=1)
+                value = torch.cat([value, encoder_hidden_states_value_proj], dim=1)
+            else:
+                query = hidden_states
+                key = encoder_hidden_states
+                value = encoder_hidden_states
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.transpose(-1, -2).reshape(batch_size, attn.heads, head_dim, -1)
+        key = key.transpose(-1, -2).reshape(batch_size, attn.heads, head_dim, -1).transpose(-1, -2)
+        value = value.transpose(-1, -2).reshape(batch_size, attn.heads, head_dim, -1)
+
+        # RoPE需要 [B, H, S, D] 输入
+        # 此时 query是 [B, H, D, S], 需要转成 [B, H, S, D] 才能应用RoPE
+        query = query.permute(0, 1, 3, 2)  # [B, H, S, D]  (从 [B, H, D, S])
+
+        # Apply query and key normalization if needed
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if rotary_freqs_cis is not None:
+            query = self.apply_rotary_emb(query, rotary_freqs_cis)
+            if not attn.is_cross_attention:
+                key = self.apply_rotary_emb(key, rotary_freqs_cis)
+            elif rotary_freqs_cis_cross is not None and has_encoder_hidden_state_proj:
+                key = self.apply_rotary_emb(key, rotary_freqs_cis_cross)
+
+        # 此时 query是 [B, H, S, D]，需要还原成 [B, H, D, S]
+        query = query.permute(0, 1, 3, 2)  # [B, H, D, S]
+
+        if attention_mask is not None:
+            # attention_mask: [B, S] -> [B, 1, S, 1]
+            attention_mask = attention_mask[:, None, :, None].to(key.dtype)  # [B, 1, S, 1]
+            query = query * attention_mask.permute(0, 1, 3, 2)  # [B, H, S, D] * [B, 1, S, 1]
+            if not attn.is_cross_attention:
+                key = key * attention_mask  # key: [B, h, S, D] 与 mask [B, 1, S, 1] 相乘
+                value = value * attention_mask.permute(0, 1, 3, 2)  # 如果 value 是 [B, h, D, S]，那么需调整mask以匹配S维度
+
+        if attn.is_cross_attention and encoder_attention_mask is not None and has_encoder_hidden_state_proj:
+            encoder_attention_mask = encoder_attention_mask[:, None, :, None].to(key.dtype)  # [B, 1, S_enc, 1]
+            # 此时 key: [B, h, S_enc, D], value: [B, h, D, S_enc]
+            key = key * encoder_attention_mask  # [B, h, S_enc, D] * [B, 1, S_enc, 1]
+            value = value * encoder_attention_mask.permute(0, 1, 3, 2)  # [B, h, D, S_enc] * [B, 1, 1, S_enc]
+
+        query = self.kernel_func(query)
+        key = self.kernel_func(key)
+
+        query, key, value = query.float(), key.float(), value.float()
+
+        value = F.pad(value, (0, 0, 0, 1), mode="constant", value=self.pad_val)
+
+        vk = torch.matmul(value, key)
+
+        hidden_states = torch.matmul(vk, query)
+
+        if hidden_states.dtype in [torch.float16, torch.bfloat16]:
+            hidden_states = hidden_states.float()
+
+        hidden_states = hidden_states[:, :, :-1] / (hidden_states[:, :, -1:] + self.eps)
+
+        hidden_states = hidden_states.view(batch_size, attn.heads * head_dim, -1).permute(0, 2, 1)
+
+        hidden_states = hidden_states.to(dtype)
+        if encoder_hidden_states is not None:
+            encoder_hidden_states = encoder_hidden_states.to(dtype)
+
+        # Split the attention outputs.
+        if encoder_hidden_states is not None and not attn.is_cross_attention and has_encoder_hidden_state_proj:
+            hidden_states, encoder_hidden_states = (
+                hidden_states[:, : hidden_states_len],
+                hidden_states[:, hidden_states_len:],
+            )
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+        if encoder_hidden_states is not None and not attn.context_pre_only and not attn.is_cross_attention and hasattr(attn, "to_add_out"):
+            encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+        if encoder_hidden_states is not None and context_input_ndim == 4:
+            encoder_hidden_states = encoder_hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if torch.get_autocast_gpu_dtype() == torch.float16:
+            hidden_states = hidden_states.clip(-65504, 65504)
+            if encoder_hidden_states is not None:
+                encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
+
+        return hidden_states, encoder_hidden_states
+
+
+class CustomerAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def apply_rotary_emb(
+        self,
+        x: torch.Tensor,
+        freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
+        to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
+        reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
+        tensors contain rotary embeddings and are returned as real tensors.
+
+        Args:
+            x (`torch.Tensor`):
+                Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
+            freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
+        """
+        cos, sin = freqs_cis  # [S, D]
+        cos = cos[None, None]
+        sin = sin[None, None]
+        cos, sin = cos.to(x.device), sin.to(x.device)
+
+        x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)  # [B, S, H, D//2]
+        x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+        return out
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+
+        residual = hidden_states
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        has_encoder_hidden_state_proj = hasattr(attn, "add_q_proj") and hasattr(attn, "add_k_proj") and hasattr(attn, "add_v_proj")
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        if attn.norm_q is not None:
+            query = attn.norm_q(query)
+        if attn.norm_k is not None:
+            key = attn.norm_k(key)
+
+        # Apply RoPE if needed
+        if rotary_freqs_cis is not None:
+            query = self.apply_rotary_emb(query, rotary_freqs_cis)
+            if not attn.is_cross_attention:
+                key = self.apply_rotary_emb(key, rotary_freqs_cis)
+            elif rotary_freqs_cis_cross is not None and has_encoder_hidden_state_proj:
+                key = self.apply_rotary_emb(key, rotary_freqs_cis_cross)
+
+        if attn.is_cross_attention and encoder_attention_mask is not None and has_encoder_hidden_state_proj:
+            # attention_mask: N x S1
+            # encoder_attention_mask: N x S2
+            # cross attention 整合attention_mask和encoder_attention_mask
+            combined_mask = attention_mask[:, :, None] * encoder_attention_mask[:, None, :]
+            attention_mask = torch.where(combined_mask == 1, 0.0, -torch.inf)
+            attention_mask = attention_mask[:, None, :, :].expand(-1, attn.heads, -1, -1).to(query.dtype)
+
+        elif not attn.is_cross_attention and attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        hidden_states = optimized_attention(
+            query, key, value, heads=query.shape[1], mask=attention_mask, skip_reshape=True,
+        ).to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+def val2list(x: list or tuple or any, repeat_time=1) -> list:  # type: ignore
+    """Repeat `val` for `repeat_time` times and return the list or val if list/tuple."""
+    if isinstance(x, (list, tuple)):
+        return list(x)
+    return [x for _ in range(repeat_time)]
+
+
+def val2tuple(x: list or tuple or any, min_len: int = 1, idx_repeat: int = -1) -> tuple:  # type: ignore
+    """Return tuple with min_len by repeating element at idx_repeat."""
+    # convert to list first
+    x = val2list(x)
+
+    # repeat elements if necessary
+    if len(x) > 0:
+        x[idx_repeat:idx_repeat] = [x[idx_repeat] for _ in range(min_len - len(x))]
+
+    return tuple(x)
+
+
+def t2i_modulate(x, shift, scale):
+    return x * (1 + scale) + shift
+
+
+def get_same_padding(kernel_size: Union[int, Tuple[int, ...]]) -> Union[int, Tuple[int, ...]]:
+    if isinstance(kernel_size, tuple):
+        return tuple([get_same_padding(ks) for ks in kernel_size])
+    else:
+        assert kernel_size % 2 > 0, f"kernel size {kernel_size} should be odd number"
+        return kernel_size // 2
+
+class ConvLayer(nn.Module):
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: int,
+        kernel_size=3,
+        stride=1,
+        dilation=1,
+        groups=1,
+        padding: Union[int, None] = None,
+        use_bias=False,
+        norm=None,
+        act=None,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        if padding is None:
+            padding = get_same_padding(kernel_size)
+            padding *= dilation
+
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+        self.kernel_size = kernel_size
+        self.stride = stride
+        self.dilation = dilation
+        self.groups = groups
+        self.padding = padding
+        self.use_bias = use_bias
+
+        self.conv = operations.Conv1d(
+            in_dim,
+            out_dim,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=groups,
+            bias=use_bias,
+            device=device,
+            dtype=dtype
+        )
+        if norm is not None:
+            self.norm = operations.RMSNorm(out_dim, elementwise_affine=False, dtype=dtype, device=device)
+        else:
+            self.norm = None
+        if act is not None:
+            self.act = nn.SiLU(inplace=True)
+        else:
+            self.act = None
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.conv(x)
+        if self.norm:
+            x = self.norm(x)
+        if self.act:
+            x = self.act(x)
+        return x
+
+
+class GLUMBConv(nn.Module):
+    def __init__(
+        self,
+        in_features: int,
+        hidden_features: int,
+        out_feature=None,
+        kernel_size=3,
+        stride=1,
+        padding: Union[int, None] = None,
+        use_bias=False,
+        norm=(None, None, None),
+        act=("silu", "silu", None),
+        dilation=1,
+        dtype=None, device=None, operations=None
+    ):
+        out_feature = out_feature or in_features
+        super().__init__()
+        use_bias = val2tuple(use_bias, 3)
+        norm = val2tuple(norm, 3)
+        act = val2tuple(act, 3)
+
+        self.glu_act = nn.SiLU(inplace=False)
+        self.inverted_conv = ConvLayer(
+            in_features,
+            hidden_features * 2,
+            1,
+            use_bias=use_bias[0],
+            norm=norm[0],
+            act=act[0],
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.depth_conv = ConvLayer(
+            hidden_features * 2,
+            hidden_features * 2,
+            kernel_size,
+            stride=stride,
+            groups=hidden_features * 2,
+            padding=padding,
+            use_bias=use_bias[1],
+            norm=norm[1],
+            act=None,
+            dilation=dilation,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.point_conv = ConvLayer(
+            hidden_features,
+            out_feature,
+            1,
+            use_bias=use_bias[2],
+            norm=norm[2],
+            act=act[2],
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x.transpose(1, 2)
+        x = self.inverted_conv(x)
+        x = self.depth_conv(x)
+
+        x, gate = torch.chunk(x, 2, dim=1)
+        gate = self.glu_act(gate)
+        x = x * gate
+
+        x = self.point_conv(x)
+        x = x.transpose(1, 2)
+
+        return x
+
+
+class LinearTransformerBlock(nn.Module):
+    """
+    A Sana block with global shared adaptive layer norm (adaLN-single) conditioning.
+    """
+    def __init__(
+        self,
+        dim,
+        num_attention_heads,
+        attention_head_dim,
+        use_adaln_single=True,
+        cross_attention_dim=None,
+        added_kv_proj_dim=None,
+        context_pre_only=False,
+        mlp_ratio=4.0,
+        add_cross_attention=False,
+        add_cross_attention_dim=None,
+        qk_norm=None,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        self.norm1 = operations.RMSNorm(dim, elementwise_affine=False, eps=1e-6)
+        self.attn = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            added_kv_proj_dim=added_kv_proj_dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            bias=True,
+            qk_norm=qk_norm,
+            processor=CustomLiteLAProcessor2_0(),
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.add_cross_attention = add_cross_attention
+        self.context_pre_only = context_pre_only
+
+        if add_cross_attention and add_cross_attention_dim is not None:
+            self.cross_attn = Attention(
+                query_dim=dim,
+                cross_attention_dim=add_cross_attention_dim,
+                added_kv_proj_dim=add_cross_attention_dim,
+                dim_head=attention_head_dim,
+                heads=num_attention_heads,
+                out_dim=dim,
+                context_pre_only=context_pre_only,
+                bias=True,
+                qk_norm=qk_norm,
+                processor=CustomerAttnProcessor2_0(),
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            )
+
+        self.norm2 = operations.RMSNorm(dim, 1e-06, elementwise_affine=False)
+
+        self.ff = GLUMBConv(
+            in_features=dim,
+            hidden_features=int(dim * mlp_ratio),
+            use_bias=(True, True, False),
+            norm=(None, None, None),
+            act=("silu", "silu", None),
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.use_adaln_single = use_adaln_single
+        if use_adaln_single:
+            self.scale_shift_table = nn.Parameter(torch.empty(6, dim, dtype=dtype, device=device))
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor = None,
+        attention_mask: torch.FloatTensor = None,
+        encoder_attention_mask: torch.FloatTensor = None,
+        rotary_freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        rotary_freqs_cis_cross: Union[torch.Tensor, Tuple[torch.Tensor]] = None,
+        temb: torch.FloatTensor = None,
+    ):
+
+        N = hidden_states.shape[0]
+
+        # step 1: AdaLN single
+        if self.use_adaln_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                comfy.model_management.cast_to(self.scale_shift_table[None], dtype=temb.dtype, device=temb.device) + temb.reshape(N, 6, -1)
+            ).chunk(6, dim=1)
+
+        norm_hidden_states = self.norm1(hidden_states)
+        if self.use_adaln_single:
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+
+        # step 2: attention
+        if not self.add_cross_attention:
+            attn_output, encoder_hidden_states = self.attn(
+                hidden_states=norm_hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                rotary_freqs_cis=rotary_freqs_cis,
+                rotary_freqs_cis_cross=rotary_freqs_cis_cross,
+            )
+        else:
+            attn_output, _ = self.attn(
+                hidden_states=norm_hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=None,
+                encoder_attention_mask=None,
+                rotary_freqs_cis=rotary_freqs_cis,
+                rotary_freqs_cis_cross=None,
+            )
+
+        if self.use_adaln_single:
+            attn_output = gate_msa * attn_output
+        hidden_states = attn_output + hidden_states
+
+        if self.add_cross_attention:
+            attn_output = self.cross_attn(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                rotary_freqs_cis=rotary_freqs_cis,
+                rotary_freqs_cis_cross=rotary_freqs_cis_cross,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # step 3: add norm
+        norm_hidden_states = self.norm2(hidden_states)
+        if self.use_adaln_single:
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        # step 4: feed forward
+        ff_output = self.ff(norm_hidden_states)
+        if self.use_adaln_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = hidden_states + ff_output
+
+        return hidden_states

comfy/ldm/ace/model.py

@@ -0,0 +1,385 @@
+# Original from: https://github.com/ace-step/ACE-Step/blob/main/models/ace_step_transformer.py
+
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional, List, Union
+
+import torch
+from torch import nn
+
+import comfy.model_management
+
+from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
+from .attention import LinearTransformerBlock, t2i_modulate
+from .lyric_encoder import ConformerEncoder as LyricEncoder
+
+
+def cross_norm(hidden_states, controlnet_input):
+    # input N x T x c
+    mean_hidden_states, std_hidden_states = hidden_states.mean(dim=(1,2), keepdim=True), hidden_states.std(dim=(1,2), keepdim=True)
+    mean_controlnet_input, std_controlnet_input = controlnet_input.mean(dim=(1,2), keepdim=True), controlnet_input.std(dim=(1,2), keepdim=True)
+    controlnet_input = (controlnet_input - mean_controlnet_input) * (std_hidden_states / (std_controlnet_input + 1e-12)) + mean_hidden_states
+    return controlnet_input
+
+
+# Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->Qwen2
+class Qwen2RotaryEmbedding(nn.Module):
+    def __init__(self, dim, max_position_embeddings=2048, base=10000, dtype=None, device=None):
+        super().__init__()
+
+        self.dim = dim
+        self.max_position_embeddings = max_position_embeddings
+        self.base = base
+        inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=device).float() / self.dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+        # Build here to make `torch.jit.trace` work.
+        self._set_cos_sin_cache(
+            seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.float32
+        )
+
+    def _set_cos_sin_cache(self, seq_len, device, dtype):
+        self.max_seq_len_cached = seq_len
+        t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
+
+        freqs = torch.outer(t, self.inv_freq)
+        # Different from paper, but it uses a different permutation in order to obtain the same calculation
+        emb = torch.cat((freqs, freqs), dim=-1)
+        self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
+        self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
+
+    def forward(self, x, seq_len=None):
+        # x: [bs, num_attention_heads, seq_len, head_size]
+        if seq_len > self.max_seq_len_cached:
+            self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
+
+        return (
+            self.cos_cached[:seq_len].to(dtype=x.dtype),
+            self.sin_cached[:seq_len].to(dtype=x.dtype),
+        )
+
+
+class T2IFinalLayer(nn.Module):
+    """
+    The final layer of Sana.
+    """
+
+    def __init__(self, hidden_size, patch_size=[16, 1], out_channels=256, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.RMSNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, patch_size[0] * patch_size[1] * out_channels, bias=True, dtype=dtype, device=device)
+        self.scale_shift_table = nn.Parameter(torch.empty(2, hidden_size, dtype=dtype, device=device))
+        self.out_channels = out_channels
+        self.patch_size = patch_size
+
+    def unpatchfy(
+        self,
+        hidden_states: torch.Tensor,
+        width: int,
+    ):
+        # 4 unpatchify
+        new_height, new_width = 1, hidden_states.size(1)
+        hidden_states = hidden_states.reshape(
+            shape=(hidden_states.shape[0], new_height, new_width, self.patch_size[0], self.patch_size[1], self.out_channels)
+        ).contiguous()
+        hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(hidden_states.shape[0], self.out_channels, new_height * self.patch_size[0], new_width * self.patch_size[1])
+        ).contiguous()
+        if width > new_width:
+            output = torch.nn.functional.pad(output, (0, width - new_width, 0, 0), 'constant', 0)
+        elif width < new_width:
+            output = output[:, :, :, :width]
+        return output
+
+    def forward(self, x, t, output_length):
+        shift, scale = (comfy.model_management.cast_to(self.scale_shift_table[None], device=t.device, dtype=t.dtype) + t[:, None]).chunk(2, dim=1)
+        x = t2i_modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        # unpatchify
+        output = self.unpatchfy(x, output_length)
+        return output
+
+
+class PatchEmbed(nn.Module):
+    """2D Image to Patch Embedding"""
+
+    def __init__(
+        self,
+        height=16,
+        width=4096,
+        patch_size=(16, 1),
+        in_channels=8,
+        embed_dim=1152,
+        bias=True,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        patch_size_h, patch_size_w = patch_size
+        self.early_conv_layers = nn.Sequential(
+            operations.Conv2d(in_channels, in_channels*256, kernel_size=patch_size, stride=patch_size, padding=0, bias=bias, dtype=dtype, device=device),
+            operations.GroupNorm(num_groups=32, num_channels=in_channels*256, eps=1e-6, affine=True, dtype=dtype, device=device),
+            operations.Conv2d(in_channels*256, embed_dim, kernel_size=1, stride=1, padding=0, bias=bias, dtype=dtype, device=device)
+        )
+        self.patch_size = patch_size
+        self.height, self.width = height // patch_size_h, width // patch_size_w
+        self.base_size = self.width
+
+    def forward(self, latent):
+        # early convolutions, N x C x H x W -> N x 256 * sqrt(patch_size) x H/patch_size x W/patch_size
+        latent = self.early_conv_layers(latent)
+        latent = latent.flatten(2).transpose(1, 2)  # BCHW -> BNC
+        return latent
+
+
+class ACEStepTransformer2DModel(nn.Module):
+    # _supports_gradient_checkpointing = True
+
+    def __init__(
+        self,
+        in_channels: Optional[int] = 8,
+        num_layers: int = 28,
+        inner_dim: int = 1536,
+        attention_head_dim: int = 64,
+        num_attention_heads: int = 24,
+        mlp_ratio: float = 4.0,
+        out_channels: int = 8,
+        max_position: int = 32768,
+        rope_theta: float = 1000000.0,
+        speaker_embedding_dim: int = 512,
+        text_embedding_dim: int = 768,
+        ssl_encoder_depths: List[int] = [9, 9],
+        ssl_names: List[str] = ["mert", "m-hubert"],
+        ssl_latent_dims: List[int] = [1024, 768],
+        lyric_encoder_vocab_size: int = 6681,
+        lyric_hidden_size: int = 1024,
+        patch_size: List[int] = [16, 1],
+        max_height: int = 16,
+        max_width: int = 4096,
+        audio_model=None,
+        dtype=None, device=None, operations=None
+
+    ):
+        super().__init__()
+
+        self.dtype = dtype
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+        self.inner_dim = inner_dim
+        self.out_channels = out_channels
+        self.max_position = max_position
+        self.patch_size = patch_size
+
+        self.rope_theta = rope_theta
+
+        self.rotary_emb = Qwen2RotaryEmbedding(
+            dim=self.attention_head_dim,
+            max_position_embeddings=self.max_position,
+            base=self.rope_theta,
+            dtype=dtype,
+            device=device,
+        )
+
+        # 2. Define input layers
+        self.in_channels = in_channels
+
+        self.num_layers = num_layers
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                LinearTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=self.num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    mlp_ratio=mlp_ratio,
+                    add_cross_attention=True,
+                    add_cross_attention_dim=self.inner_dim,
+                    dtype=dtype,
+                    device=device,
+                    operations=operations,
+                )
+                for i in range(self.num_layers)
+            ]
+        )
+
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=self.inner_dim, dtype=dtype, device=device, operations=operations)
+        self.t_block = nn.Sequential(nn.SiLU(), operations.Linear(self.inner_dim, 6 * self.inner_dim, bias=True, dtype=dtype, device=device))
+
+        # speaker
+        self.speaker_embedder = operations.Linear(speaker_embedding_dim, self.inner_dim, dtype=dtype, device=device)
+
+        # genre
+        self.genre_embedder = operations.Linear(text_embedding_dim, self.inner_dim, dtype=dtype, device=device)
+
+        # lyric
+        self.lyric_embs = operations.Embedding(lyric_encoder_vocab_size, lyric_hidden_size, dtype=dtype, device=device)
+        self.lyric_encoder = LyricEncoder(input_size=lyric_hidden_size, static_chunk_size=0, dtype=dtype, device=device, operations=operations)
+        self.lyric_proj = operations.Linear(lyric_hidden_size, self.inner_dim, dtype=dtype, device=device)
+
+        projector_dim = 2 * self.inner_dim
+
+        self.projectors = nn.ModuleList([
+            nn.Sequential(
+                operations.Linear(self.inner_dim, projector_dim, dtype=dtype, device=device),
+                nn.SiLU(),
+                operations.Linear(projector_dim, projector_dim, dtype=dtype, device=device),
+                nn.SiLU(),
+                operations.Linear(projector_dim, ssl_dim, dtype=dtype, device=device),
+            ) for ssl_dim in ssl_latent_dims
+        ])
+
+        self.proj_in = PatchEmbed(
+            height=max_height,
+            width=max_width,
+            patch_size=patch_size,
+            embed_dim=self.inner_dim,
+            bias=True,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+        self.final_layer = T2IFinalLayer(self.inner_dim, patch_size=patch_size, out_channels=out_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward_lyric_encoder(
+        self,
+        lyric_token_idx: Optional[torch.LongTensor] = None,
+        lyric_mask: Optional[torch.LongTensor] = None,
+        out_dtype=None,
+    ):
+        # N x T x D
+        lyric_embs = self.lyric_embs(lyric_token_idx, out_dtype=out_dtype)
+        prompt_prenet_out, _mask = self.lyric_encoder(lyric_embs, lyric_mask, decoding_chunk_size=1, num_decoding_left_chunks=-1)
+        prompt_prenet_out = self.lyric_proj(prompt_prenet_out)
+        return prompt_prenet_out
+
+    def encode(
+        self,
+        encoder_text_hidden_states: Optional[torch.Tensor] = None,
+        text_attention_mask: Optional[torch.LongTensor] = None,
+        speaker_embeds: Optional[torch.FloatTensor] = None,
+        lyric_token_idx: Optional[torch.LongTensor] = None,
+        lyric_mask: Optional[torch.LongTensor] = None,
+        lyrics_strength=1.0,
+    ):
+
+        bs = encoder_text_hidden_states.shape[0]
+        device = encoder_text_hidden_states.device
+
+        # speaker embedding
+        encoder_spk_hidden_states = self.speaker_embedder(speaker_embeds).unsqueeze(1)
+
+        # genre embedding
+        encoder_text_hidden_states = self.genre_embedder(encoder_text_hidden_states)
+
+        # lyric
+        encoder_lyric_hidden_states = self.forward_lyric_encoder(
+            lyric_token_idx=lyric_token_idx,
+            lyric_mask=lyric_mask,
+            out_dtype=encoder_text_hidden_states.dtype,
+        )
+
+        encoder_lyric_hidden_states *= lyrics_strength
+
+        encoder_hidden_states = torch.cat([encoder_spk_hidden_states, encoder_text_hidden_states, encoder_lyric_hidden_states], dim=1)
+
+        encoder_hidden_mask = None
+        if text_attention_mask is not None:
+            speaker_mask = torch.ones(bs, 1, device=device)
+            encoder_hidden_mask = torch.cat([speaker_mask, text_attention_mask, lyric_mask], dim=1)
+
+        return encoder_hidden_states, encoder_hidden_mask
+
+    def decode(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_mask: torch.Tensor,
+        timestep: Optional[torch.Tensor],
+        output_length: int = 0,
+        block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None,
+        controlnet_scale: Union[float, torch.Tensor] = 1.0,
+    ):
+        embedded_timestep = self.timestep_embedder(self.time_proj(timestep).to(dtype=hidden_states.dtype))
+        temb = self.t_block(embedded_timestep)
+
+        hidden_states = self.proj_in(hidden_states)
+
+        # controlnet logic
+        if block_controlnet_hidden_states is not None:
+            control_condi = cross_norm(hidden_states, block_controlnet_hidden_states)
+            hidden_states = hidden_states + control_condi * controlnet_scale
+
+        # inner_hidden_states = []
+
+        rotary_freqs_cis = self.rotary_emb(hidden_states, seq_len=hidden_states.shape[1])
+        encoder_rotary_freqs_cis = self.rotary_emb(encoder_hidden_states, seq_len=encoder_hidden_states.shape[1])
+
+        for index_block, block in enumerate(self.transformer_blocks):
+            hidden_states = block(
+                hidden_states=hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_hidden_mask,
+                rotary_freqs_cis=rotary_freqs_cis,
+                rotary_freqs_cis_cross=encoder_rotary_freqs_cis,
+                temb=temb,
+            )
+
+        output = self.final_layer(hidden_states, embedded_timestep, output_length)
+        return output
+
+    def forward(
+        self,
+        x,
+        timestep,
+        attention_mask=None,
+        context: Optional[torch.Tensor] = None,
+        text_attention_mask: Optional[torch.LongTensor] = None,
+        speaker_embeds: Optional[torch.FloatTensor] = None,
+        lyric_token_idx: Optional[torch.LongTensor] = None,
+        lyric_mask: Optional[torch.LongTensor] = None,
+        block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None,
+        controlnet_scale: Union[float, torch.Tensor] = 1.0,
+        lyrics_strength=1.0,
+        **kwargs
+    ):
+        hidden_states = x
+        encoder_text_hidden_states = context
+        encoder_hidden_states, encoder_hidden_mask = self.encode(
+            encoder_text_hidden_states=encoder_text_hidden_states,
+            text_attention_mask=text_attention_mask,
+            speaker_embeds=speaker_embeds,
+            lyric_token_idx=lyric_token_idx,
+            lyric_mask=lyric_mask,
+            lyrics_strength=lyrics_strength,
+        )
+
+        output_length = hidden_states.shape[-1]
+
+        output = self.decode(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_hidden_mask=encoder_hidden_mask,
+            timestep=timestep,
+            output_length=output_length,
+            block_controlnet_hidden_states=block_controlnet_hidden_states,
+            controlnet_scale=controlnet_scale,
+        )
+
+        return output

comfy/ldm/ace/vae/autoencoder_dc.py

@@ -0,0 +1,644 @@
+# Rewritten from diffusers
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Tuple, Union
+
+import comfy.model_management
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+class RMSNorm(ops.RMSNorm):
+    def __init__(self, dim, eps=1e-5, elementwise_affine=True, bias=False):
+        super().__init__(dim, eps=eps, elementwise_affine=elementwise_affine)
+        if elementwise_affine:
+            self.bias = nn.Parameter(torch.empty(dim)) if bias else None
+
+    def forward(self, x):
+        x = super().forward(x)
+        if self.elementwise_affine:
+            if self.bias is not None:
+                x = x + comfy.model_management.cast_to(self.bias, dtype=x.dtype, device=x.device)
+        return x
+
+
+def get_normalization(norm_type, num_features, num_groups=32, eps=1e-5):
+    if norm_type == "batch_norm":
+        return nn.BatchNorm2d(num_features)
+    elif norm_type == "group_norm":
+        return ops.GroupNorm(num_groups, num_features)
+    elif norm_type == "layer_norm":
+        return ops.LayerNorm(num_features)
+    elif norm_type == "rms_norm":
+        return RMSNorm(num_features, eps=eps, elementwise_affine=True, bias=True)
+    else:
+        raise ValueError(f"Unknown normalization type: {norm_type}")
+
+
+def get_activation(activation_type):
+    if activation_type == "relu":
+        return nn.ReLU()
+    elif activation_type == "relu6":
+        return nn.ReLU6()
+    elif activation_type == "silu":
+        return nn.SiLU()
+    elif activation_type == "leaky_relu":
+        return nn.LeakyReLU(0.2)
+    else:
+        raise ValueError(f"Unknown activation type: {activation_type}")
+
+
+class ResBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        norm_type: str = "batch_norm",
+        act_fn: str = "relu6",
+    ) -> None:
+        super().__init__()
+
+        self.norm_type = norm_type
+        self.nonlinearity = get_activation(act_fn) if act_fn is not None else nn.Identity()
+        self.conv1 = ops.Conv2d(in_channels, in_channels, 3, 1, 1)
+        self.conv2 = ops.Conv2d(in_channels, out_channels, 3, 1, 1, bias=False)
+        self.norm = get_normalization(norm_type, out_channels)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        residual = hidden_states
+        hidden_states = self.conv1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.norm_type == "rms_norm":
+            # move channel to the last dimension so we apply RMSnorm across channel dimension
+            hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1)
+        else:
+            hidden_states = self.norm(hidden_states)
+
+        return hidden_states + residual
+
+class SanaMultiscaleAttentionProjection(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        num_attention_heads: int,
+        kernel_size: int,
+    ) -> None:
+        super().__init__()
+
+        channels = 3 * in_channels
+        self.proj_in = ops.Conv2d(
+            channels,
+            channels,
+            kernel_size,
+            padding=kernel_size // 2,
+            groups=channels,
+            bias=False,
+        )
+        self.proj_out = ops.Conv2d(channels, channels, 1, 1, 0, groups=3 * num_attention_heads, bias=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.proj_in(hidden_states)
+        hidden_states = self.proj_out(hidden_states)
+        return hidden_states
+
+class SanaMultiscaleLinearAttention(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        num_attention_heads: int = None,
+        attention_head_dim: int = 8,
+        mult: float = 1.0,
+        norm_type: str = "batch_norm",
+        kernel_sizes: tuple = (5,),
+        eps: float = 1e-15,
+        residual_connection: bool = False,
+    ):
+        super().__init__()
+
+        self.eps = eps
+        self.attention_head_dim = attention_head_dim
+        self.norm_type = norm_type
+        self.residual_connection = residual_connection
+
+        num_attention_heads = (
+            int(in_channels // attention_head_dim * mult)
+            if num_attention_heads is None
+            else num_attention_heads
+        )
+        inner_dim = num_attention_heads * attention_head_dim
+
+        self.to_q = ops.Linear(in_channels, inner_dim, bias=False)
+        self.to_k = ops.Linear(in_channels, inner_dim, bias=False)
+        self.to_v = ops.Linear(in_channels, inner_dim, bias=False)
+
+        self.to_qkv_multiscale = nn.ModuleList()
+        for kernel_size in kernel_sizes:
+            self.to_qkv_multiscale.append(
+                SanaMultiscaleAttentionProjection(inner_dim, num_attention_heads, kernel_size)
+            )
+
+        self.nonlinearity = nn.ReLU()
+        self.to_out = ops.Linear(inner_dim * (1 + len(kernel_sizes)), out_channels, bias=False)
+        self.norm_out = get_normalization(norm_type, out_channels)
+
+    def apply_linear_attention(self, query, key, value):
+        value = F.pad(value, (0, 0, 0, 1), mode="constant", value=1)
+        scores = torch.matmul(value, key.transpose(-1, -2))
+        hidden_states = torch.matmul(scores, query)
+
+        hidden_states = hidden_states.to(dtype=torch.float32)
+        hidden_states = hidden_states[:, :, :-1] / (hidden_states[:, :, -1:] + self.eps)
+        return hidden_states
+
+    def apply_quadratic_attention(self, query, key, value):
+        scores = torch.matmul(key.transpose(-1, -2), query)
+        scores = scores.to(dtype=torch.float32)
+        scores = scores / (torch.sum(scores, dim=2, keepdim=True) + self.eps)
+        hidden_states = torch.matmul(value, scores.to(value.dtype))
+        return hidden_states
+
+    def forward(self, hidden_states):
+        height, width = hidden_states.shape[-2:]
+        if height * width > self.attention_head_dim:
+            use_linear_attention = True
+        else:
+            use_linear_attention = False
+
+        residual = hidden_states
+
+        batch_size, _, height, width = list(hidden_states.size())
+        original_dtype = hidden_states.dtype
+
+        hidden_states = hidden_states.movedim(1, -1)
+        query = self.to_q(hidden_states)
+        key = self.to_k(hidden_states)
+        value = self.to_v(hidden_states)
+        hidden_states = torch.cat([query, key, value], dim=3)
+        hidden_states = hidden_states.movedim(-1, 1)
+
+        multi_scale_qkv = [hidden_states]
+        for block in self.to_qkv_multiscale:
+            multi_scale_qkv.append(block(hidden_states))
+
+        hidden_states = torch.cat(multi_scale_qkv, dim=1)
+
+        if use_linear_attention:
+            # for linear attention upcast hidden_states to float32
+            hidden_states = hidden_states.to(dtype=torch.float32)
+
+        hidden_states = hidden_states.reshape(batch_size, -1, 3 * self.attention_head_dim, height * width)
+
+        query, key, value = hidden_states.chunk(3, dim=2)
+        query = self.nonlinearity(query)
+        key = self.nonlinearity(key)
+
+        if use_linear_attention:
+            hidden_states = self.apply_linear_attention(query, key, value)
+            hidden_states = hidden_states.to(dtype=original_dtype)
+        else:
+            hidden_states = self.apply_quadratic_attention(query, key, value)
+
+        hidden_states = torch.reshape(hidden_states, (batch_size, -1, height, width))
+        hidden_states = self.to_out(hidden_states.movedim(1, -1)).movedim(-1, 1)
+
+        if self.norm_type == "rms_norm":
+            hidden_states = self.norm_out(hidden_states.movedim(1, -1)).movedim(-1, 1)
+        else:
+            hidden_states = self.norm_out(hidden_states)
+
+        if self.residual_connection:
+            hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
+class EfficientViTBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        mult: float = 1.0,
+        attention_head_dim: int = 32,
+        qkv_multiscales: tuple = (5,),
+        norm_type: str = "batch_norm",
+    ) -> None:
+        super().__init__()
+
+        self.attn = SanaMultiscaleLinearAttention(
+            in_channels=in_channels,
+            out_channels=in_channels,
+            mult=mult,
+            attention_head_dim=attention_head_dim,
+            norm_type=norm_type,
+            kernel_sizes=qkv_multiscales,
+            residual_connection=True,
+        )
+
+        self.conv_out = GLUMBConv(
+            in_channels=in_channels,
+            out_channels=in_channels,
+            norm_type="rms_norm",
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.attn(x)
+        x = self.conv_out(x)
+        return x
+
+
+class GLUMBConv(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        expand_ratio: float = 4,
+        norm_type: str = None,
+        residual_connection: bool = True,
+    ) -> None:
+        super().__init__()
+
+        hidden_channels = int(expand_ratio * in_channels)
+        self.norm_type = norm_type
+        self.residual_connection = residual_connection
+
+        self.nonlinearity = nn.SiLU()
+        self.conv_inverted = ops.Conv2d(in_channels, hidden_channels * 2, 1, 1, 0)
+        self.conv_depth = ops.Conv2d(hidden_channels * 2, hidden_channels * 2, 3, 1, 1, groups=hidden_channels * 2)
+        self.conv_point = ops.Conv2d(hidden_channels, out_channels, 1, 1, 0, bias=False)
+
+        self.norm = None
+        if norm_type == "rms_norm":
+            self.norm = RMSNorm(out_channels, eps=1e-5, elementwise_affine=True, bias=True)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.residual_connection:
+            residual = hidden_states
+
+        hidden_states = self.conv_inverted(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv_depth(hidden_states)
+        hidden_states, gate = torch.chunk(hidden_states, 2, dim=1)
+        hidden_states = hidden_states * self.nonlinearity(gate)
+
+        hidden_states = self.conv_point(hidden_states)
+
+        if self.norm_type == "rms_norm":
+            # move channel to the last dimension so we apply RMSnorm across channel dimension
+            hidden_states = self.norm(hidden_states.movedim(1, -1)).movedim(-1, 1)
+
+        if self.residual_connection:
+            hidden_states = hidden_states + residual
+
+        return hidden_states
+
+
+def get_block(
+    block_type: str,
+    in_channels: int,
+    out_channels: int,
+    attention_head_dim: int,
+    norm_type: str,
+    act_fn: str,
+    qkv_mutliscales: tuple = (),
+):
+    if block_type == "ResBlock":
+        block = ResBlock(in_channels, out_channels, norm_type, act_fn)
+    elif block_type == "EfficientViTBlock":
+        block = EfficientViTBlock(
+            in_channels,
+            attention_head_dim=attention_head_dim,
+            norm_type=norm_type,
+            qkv_multiscales=qkv_mutliscales
+        )
+    else:
+        raise ValueError(f"Block with {block_type=} is not supported.")
+
+    return block
+
+
+class DCDownBlock2d(nn.Module):
+    def __init__(self, in_channels: int, out_channels: int, downsample: bool = False, shortcut: bool = True) -> None:
+        super().__init__()
+
+        self.downsample = downsample
+        self.factor = 2
+        self.stride = 1 if downsample else 2
+        self.group_size = in_channels * self.factor**2 // out_channels
+        self.shortcut = shortcut
+
+        out_ratio = self.factor**2
+        if downsample:
+            assert out_channels % out_ratio == 0
+            out_channels = out_channels // out_ratio
+
+        self.conv = ops.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size=3,
+            stride=self.stride,
+            padding=1,
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        x = self.conv(hidden_states)
+        if self.downsample:
+            x = F.pixel_unshuffle(x, self.factor)
+
+        if self.shortcut:
+            y = F.pixel_unshuffle(hidden_states, self.factor)
+            y = y.unflatten(1, (-1, self.group_size))
+            y = y.mean(dim=2)
+            hidden_states = x + y
+        else:
+            hidden_states = x
+
+        return hidden_states
+
+
+class DCUpBlock2d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        interpolate: bool = False,
+        shortcut: bool = True,
+        interpolation_mode: str = "nearest",
+    ) -> None:
+        super().__init__()
+
+        self.interpolate = interpolate
+        self.interpolation_mode = interpolation_mode
+        self.shortcut = shortcut
+        self.factor = 2
+        self.repeats = out_channels * self.factor**2 // in_channels
+
+        out_ratio = self.factor**2
+        if not interpolate:
+            out_channels = out_channels * out_ratio
+
+        self.conv = ops.Conv2d(in_channels, out_channels, 3, 1, 1)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.interpolate:
+            x = F.interpolate(hidden_states, scale_factor=self.factor, mode=self.interpolation_mode)
+            x = self.conv(x)
+        else:
+            x = self.conv(hidden_states)
+            x = F.pixel_shuffle(x, self.factor)
+
+        if self.shortcut:
+            y = hidden_states.repeat_interleave(self.repeats, dim=1, output_size=hidden_states.shape[1] * self.repeats)
+            y = F.pixel_shuffle(y, self.factor)
+            hidden_states = x + y
+        else:
+            hidden_states = x
+
+        return hidden_states
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        latent_channels: int,
+        attention_head_dim: int = 32,
+        block_type: str or tuple = "ResBlock",
+        block_out_channels: tuple = (128, 256, 512, 512, 1024, 1024),
+        layers_per_block: tuple = (2, 2, 2, 2, 2, 2),
+        qkv_multiscales: tuple = ((), (), (), (5,), (5,), (5,)),
+        downsample_block_type: str = "pixel_unshuffle",
+        out_shortcut: bool = True,
+    ):
+        super().__init__()
+
+        num_blocks = len(block_out_channels)
+
+        if isinstance(block_type, str):
+            block_type = (block_type,) * num_blocks
+
+        if layers_per_block[0] > 0:
+            self.conv_in = ops.Conv2d(
+                in_channels,
+                block_out_channels[0] if layers_per_block[0] > 0 else block_out_channels[1],
+                kernel_size=3,
+                stride=1,
+                padding=1,
+            )
+        else:
+            self.conv_in = DCDownBlock2d(
+                in_channels=in_channels,
+                out_channels=block_out_channels[0] if layers_per_block[0] > 0 else block_out_channels[1],
+                downsample=downsample_block_type == "pixel_unshuffle",
+                shortcut=False,
+            )
+
+        down_blocks = []
+        for i, (out_channel, num_layers) in enumerate(zip(block_out_channels, layers_per_block)):
+            down_block_list = []
+
+            for _ in range(num_layers):
+                block = get_block(
+                    block_type[i],
+                    out_channel,
+                    out_channel,
+                    attention_head_dim=attention_head_dim,
+                    norm_type="rms_norm",
+                    act_fn="silu",
+                    qkv_mutliscales=qkv_multiscales[i],
+                )
+                down_block_list.append(block)
+
+            if i < num_blocks - 1 and num_layers > 0:
+                downsample_block = DCDownBlock2d(
+                    in_channels=out_channel,
+                    out_channels=block_out_channels[i + 1],
+                    downsample=downsample_block_type == "pixel_unshuffle",
+                    shortcut=True,
+                )
+                down_block_list.append(downsample_block)
+
+            down_blocks.append(nn.Sequential(*down_block_list))
+
+        self.down_blocks = nn.ModuleList(down_blocks)
+
+        self.conv_out = ops.Conv2d(block_out_channels[-1], latent_channels, 3, 1, 1)
+
+        self.out_shortcut = out_shortcut
+        if out_shortcut:
+            self.out_shortcut_average_group_size = block_out_channels[-1] // latent_channels
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.conv_in(hidden_states)
+        for down_block in self.down_blocks:
+            hidden_states = down_block(hidden_states)
+
+        if self.out_shortcut:
+            x = hidden_states.unflatten(1, (-1, self.out_shortcut_average_group_size))
+            x = x.mean(dim=2)
+            hidden_states = self.conv_out(hidden_states) + x
+        else:
+            hidden_states = self.conv_out(hidden_states)
+
+        return hidden_states
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        latent_channels: int,
+        attention_head_dim: int = 32,
+        block_type: str or tuple = "ResBlock",
+        block_out_channels: tuple = (128, 256, 512, 512, 1024, 1024),
+        layers_per_block: tuple = (2, 2, 2, 2, 2, 2),
+        qkv_multiscales: tuple = ((), (), (), (5,), (5,), (5,)),
+        norm_type: str or tuple = "rms_norm",
+        act_fn: str or tuple = "silu",
+        upsample_block_type: str = "pixel_shuffle",
+        in_shortcut: bool = True,
+    ):
+        super().__init__()
+
+        num_blocks = len(block_out_channels)
+
+        if isinstance(block_type, str):
+            block_type = (block_type,) * num_blocks
+        if isinstance(norm_type, str):
+            norm_type = (norm_type,) * num_blocks
+        if isinstance(act_fn, str):
+            act_fn = (act_fn,) * num_blocks
+
+        self.conv_in = ops.Conv2d(latent_channels, block_out_channels[-1], 3, 1, 1)
+
+        self.in_shortcut = in_shortcut
+        if in_shortcut:
+            self.in_shortcut_repeats = block_out_channels[-1] // latent_channels
+
+        up_blocks = []
+        for i, (out_channel, num_layers) in reversed(list(enumerate(zip(block_out_channels, layers_per_block)))):
+            up_block_list = []
+
+            if i < num_blocks - 1 and num_layers > 0:
+                upsample_block = DCUpBlock2d(
+                    block_out_channels[i + 1],
+                    out_channel,
+                    interpolate=upsample_block_type == "interpolate",
+                    shortcut=True,
+                )
+                up_block_list.append(upsample_block)
+
+            for _ in range(num_layers):
+                block = get_block(
+                    block_type[i],
+                    out_channel,
+                    out_channel,
+                    attention_head_dim=attention_head_dim,
+                    norm_type=norm_type[i],
+                    act_fn=act_fn[i],
+                    qkv_mutliscales=qkv_multiscales[i],
+                )
+                up_block_list.append(block)
+
+            up_blocks.insert(0, nn.Sequential(*up_block_list))
+
+        self.up_blocks = nn.ModuleList(up_blocks)
+
+        channels = block_out_channels[0] if layers_per_block[0] > 0 else block_out_channels[1]
+
+        self.norm_out = RMSNorm(channels, 1e-5, elementwise_affine=True, bias=True)
+        self.conv_act = nn.ReLU()
+        self.conv_out = None
+
+        if layers_per_block[0] > 0:
+            self.conv_out = ops.Conv2d(channels, in_channels, 3, 1, 1)
+        else:
+            self.conv_out = DCUpBlock2d(
+                channels, in_channels, interpolate=upsample_block_type == "interpolate", shortcut=False
+            )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self.in_shortcut:
+            x = hidden_states.repeat_interleave(
+                self.in_shortcut_repeats, dim=1, output_size=hidden_states.shape[1] * self.in_shortcut_repeats
+            )
+            hidden_states = self.conv_in(hidden_states) + x
+        else:
+            hidden_states = self.conv_in(hidden_states)
+
+        for up_block in reversed(self.up_blocks):
+            hidden_states = up_block(hidden_states)
+
+        hidden_states = self.norm_out(hidden_states.movedim(1, -1)).movedim(-1, 1)
+        hidden_states = self.conv_act(hidden_states)
+        hidden_states = self.conv_out(hidden_states)
+        return hidden_states
+
+
+class AutoencoderDC(nn.Module):
+    def __init__(
+        self,
+        in_channels: int = 2,
+        latent_channels: int = 8,
+        attention_head_dim: int = 32,
+        encoder_block_types: Union[str, Tuple[str]] = ["ResBlock", "ResBlock", "ResBlock", "EfficientViTBlock"],
+        decoder_block_types: Union[str, Tuple[str]] = ["ResBlock", "ResBlock", "ResBlock", "EfficientViTBlock"],
+        encoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 1024),
+        decoder_block_out_channels: Tuple[int, ...] = (128, 256, 512, 1024),
+        encoder_layers_per_block: Tuple[int] = (2, 2, 3, 3),
+        decoder_layers_per_block: Tuple[int] = (3, 3, 3, 3),
+        encoder_qkv_multiscales: Tuple[Tuple[int, ...], ...] = ((), (), (5,), (5,)),
+        decoder_qkv_multiscales: Tuple[Tuple[int, ...], ...] = ((), (), (5,), (5,)),
+        upsample_block_type: str = "interpolate",
+        downsample_block_type: str = "Conv",
+        decoder_norm_types: Union[str, Tuple[str]] = "rms_norm",
+        decoder_act_fns: Union[str, Tuple[str]] = "silu",
+        scaling_factor: float = 0.41407,
+    ) -> None:
+        super().__init__()
+
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            latent_channels=latent_channels,
+            attention_head_dim=attention_head_dim,
+            block_type=encoder_block_types,
+            block_out_channels=encoder_block_out_channels,
+            layers_per_block=encoder_layers_per_block,
+            qkv_multiscales=encoder_qkv_multiscales,
+            downsample_block_type=downsample_block_type,
+        )
+
+        self.decoder = Decoder(
+            in_channels=in_channels,
+            latent_channels=latent_channels,
+            attention_head_dim=attention_head_dim,
+            block_type=decoder_block_types,
+            block_out_channels=decoder_block_out_channels,
+            layers_per_block=decoder_layers_per_block,
+            qkv_multiscales=decoder_qkv_multiscales,
+            norm_type=decoder_norm_types,
+            act_fn=decoder_act_fns,
+            upsample_block_type=upsample_block_type,
+        )
+
+        self.scaling_factor = scaling_factor
+        self.spatial_compression_ratio = 2 ** (len(encoder_block_out_channels) - 1)
+
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
+        """Internal encoding function."""
+        encoded = self.encoder(x)
+        return encoded * self.scaling_factor
+
+    def decode(self, z: torch.Tensor) -> torch.Tensor:
+        # Scale the latents back
+        z = z / self.scaling_factor
+        decoded = self.decoder(z)
+        return decoded
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        z = self.encode(x)
+        return self.decode(z)
+

comfy/ldm/ace/vae/music_dcae_pipeline.py

@@ -0,0 +1,109 @@
+# Original from: https://github.com/ace-step/ACE-Step/blob/main/music_dcae/music_dcae_pipeline.py
+import torch
+from .autoencoder_dc import AutoencoderDC
+import logging
+try:
+    import torchaudio
+except:
+    logging.warning("torchaudio missing, ACE model will be broken")
+
+import torchvision.transforms as transforms
+from .music_vocoder import ADaMoSHiFiGANV1
+
+
+class MusicDCAE(torch.nn.Module):
+    def __init__(self, source_sample_rate=None, dcae_config={}, vocoder_config={}):
+        super(MusicDCAE, self).__init__()
+
+        self.dcae = AutoencoderDC(**dcae_config)
+        self.vocoder = ADaMoSHiFiGANV1(**vocoder_config)
+
+        if source_sample_rate is None:
+            self.source_sample_rate = 48000
+        else:
+            self.source_sample_rate = source_sample_rate
+
+        # self.resampler = torchaudio.transforms.Resample(source_sample_rate, 44100)
+
+        self.transform = transforms.Compose([
+            transforms.Normalize(0.5, 0.5),
+        ])
+        self.min_mel_value = -11.0
+        self.max_mel_value = 3.0
+        self.audio_chunk_size = int(round((1024 * 512 / 44100 * 48000)))
+        self.mel_chunk_size = 1024
+        self.time_dimention_multiple = 8
+        self.latent_chunk_size = self.mel_chunk_size // self.time_dimention_multiple
+        self.scale_factor = 0.1786
+        self.shift_factor = -1.9091
+
+    def load_audio(self, audio_path):
+        audio, sr = torchaudio.load(audio_path)
+        return audio, sr
+
+    def forward_mel(self, audios):
+        mels = []
+        for i in range(len(audios)):
+            image = self.vocoder.mel_transform(audios[i])
+            mels.append(image)
+        mels = torch.stack(mels)
+        return mels
+
+    @torch.no_grad()
+    def encode(self, audios, audio_lengths=None, sr=None):
+        if audio_lengths is None:
+            audio_lengths = torch.tensor([audios.shape[2]] * audios.shape[0])
+            audio_lengths = audio_lengths.to(audios.device)
+
+        if sr is None:
+            sr = self.source_sample_rate
+
+        if sr != 44100:
+            audios = torchaudio.functional.resample(audios, sr, 44100)
+
+        max_audio_len = audios.shape[-1]
+        if max_audio_len % (8 * 512) != 0:
+            audios = torch.nn.functional.pad(audios, (0, 8 * 512 - max_audio_len % (8 * 512)))
+
+        mels = self.forward_mel(audios)
+        mels = (mels - self.min_mel_value) / (self.max_mel_value - self.min_mel_value)
+        mels = self.transform(mels)
+        latents = []
+        for mel in mels:
+            latent = self.dcae.encoder(mel.unsqueeze(0))
+            latents.append(latent)
+        latents = torch.cat(latents, dim=0)
+        # latent_lengths = (audio_lengths / sr * 44100 / 512 / self.time_dimention_multiple).long()
+        latents = (latents - self.shift_factor) * self.scale_factor
+        return latents
+        # return latents, latent_lengths
+
+    @torch.no_grad()
+    def decode(self, latents, audio_lengths=None, sr=None):
+        latents = latents / self.scale_factor + self.shift_factor
+
+        pred_wavs = []
+
+        for latent in latents:
+            mels = self.dcae.decoder(latent.unsqueeze(0))
+            mels = mels * 0.5 + 0.5
+            mels = mels * (self.max_mel_value - self.min_mel_value) + self.min_mel_value
+            wav = self.vocoder.decode(mels[0]).squeeze(1)
+
+            if sr is not None:
+                # resampler = torchaudio.transforms.Resample(44100, sr).to(latents.device).to(latents.dtype)
+                wav = torchaudio.functional.resample(wav, 44100, sr)
+                # wav = resampler(wav)
+            else:
+                sr = 44100
+            pred_wavs.append(wav)
+
+        if audio_lengths is not None:
+            pred_wavs = [wav[:, :length].cpu() for wav, length in zip(pred_wavs, audio_lengths)]
+        return torch.stack(pred_wavs)
+        # return sr, pred_wavs
+
+    def forward(self, audios, audio_lengths=None, sr=None):
+        latents, latent_lengths = self.encode(audios=audios, audio_lengths=audio_lengths, sr=sr)
+        sr, pred_wavs = self.decode(latents=latents, audio_lengths=audio_lengths, sr=sr)
+        return sr, pred_wavs, latents, latent_lengths

comfy/ldm/ace/vae/music_log_mel.py

@@ -0,0 +1,113 @@
+# Original from: https://github.com/ace-step/ACE-Step/blob/main/music_dcae/music_log_mel.py
+import torch
+import torch.nn as nn
+from torch import Tensor
+import logging
+try:
+    from torchaudio.transforms import MelScale
+except:
+    logging.warning("torchaudio missing, ACE model will be broken")
+
+import comfy.model_management
+
+class LinearSpectrogram(nn.Module):
+    def __init__(
+        self,
+        n_fft=2048,
+        win_length=2048,
+        hop_length=512,
+        center=False,
+        mode="pow2_sqrt",
+    ):
+        super().__init__()
+
+        self.n_fft = n_fft
+        self.win_length = win_length
+        self.hop_length = hop_length
+        self.center = center
+        self.mode = mode
+
+        self.register_buffer("window", torch.hann_window(win_length))
+
+    def forward(self, y: Tensor) -> Tensor:
+        if y.ndim == 3:
+            y = y.squeeze(1)
+
+        y = torch.nn.functional.pad(
+            y.unsqueeze(1),
+            (
+                (self.win_length - self.hop_length) // 2,
+                (self.win_length - self.hop_length + 1) // 2,
+            ),
+            mode="reflect",
+        ).squeeze(1)
+        dtype = y.dtype
+        spec = torch.stft(
+            y.float(),
+            self.n_fft,
+            hop_length=self.hop_length,
+            win_length=self.win_length,
+            window=comfy.model_management.cast_to(self.window, dtype=torch.float32, device=y.device),
+            center=self.center,
+            pad_mode="reflect",
+            normalized=False,
+            onesided=True,
+            return_complex=True,
+        )
+        spec = torch.view_as_real(spec)
+
+        if self.mode == "pow2_sqrt":
+            spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
+        spec = spec.to(dtype)
+        return spec
+
+
+class LogMelSpectrogram(nn.Module):
+    def __init__(
+        self,
+        sample_rate=44100,
+        n_fft=2048,
+        win_length=2048,
+        hop_length=512,
+        n_mels=128,
+        center=False,
+        f_min=0.0,
+        f_max=None,
+    ):
+        super().__init__()
+
+        self.sample_rate = sample_rate
+        self.n_fft = n_fft
+        self.win_length = win_length
+        self.hop_length = hop_length
+        self.center = center
+        self.n_mels = n_mels
+        self.f_min = f_min
+        self.f_max = f_max or sample_rate // 2
+
+        self.spectrogram = LinearSpectrogram(n_fft, win_length, hop_length, center)
+        self.mel_scale = MelScale(
+            self.n_mels,
+            self.sample_rate,
+            self.f_min,
+            self.f_max,
+            self.n_fft // 2 + 1,
+            "slaney",
+            "slaney",
+        )
+
+    def compress(self, x: Tensor) -> Tensor:
+        return torch.log(torch.clamp(x, min=1e-5))
+
+    def decompress(self, x: Tensor) -> Tensor:
+        return torch.exp(x)
+
+    def forward(self, x: Tensor, return_linear: bool = False) -> Tensor:
+        linear = self.spectrogram(x)
+        x = self.mel_scale(linear)
+        x = self.compress(x)
+        # print(x.shape)
+        if return_linear:
+            return x, self.compress(linear)
+
+        return x

comfy/ldm/ace/vae/music_vocoder.py

@@ -0,0 +1,538 @@
+# Original from: https://github.com/ace-step/ACE-Step/blob/main/music_dcae/music_vocoder.py
+import torch
+from torch import nn
+
+from functools import partial
+from math import prod
+from typing import Callable, Tuple, List
+
+import numpy as np
+import torch.nn.functional as F
+from torch.nn.utils.parametrize import remove_parametrizations as remove_weight_norm
+
+from .music_log_mel import LogMelSpectrogram
+
+import comfy.model_management
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+def drop_path(
+    x, drop_prob: float = 0.0, training: bool = False, scale_by_keep: bool = True
+):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+    'survival rate' as the argument.
+
+    """  # noqa: E501
+
+    if drop_prob == 0.0 or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0],) + (1,) * (
+        x.ndim - 1
+    )  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+    if keep_prob > 0.0 and scale_by_keep:
+        random_tensor.div_(keep_prob)
+    return x * random_tensor
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks)."""  # noqa: E501
+
+    def __init__(self, drop_prob: float = 0.0, scale_by_keep: bool = True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training, self.scale_by_keep)
+
+    def extra_repr(self):
+        return f"drop_prob={round(self.drop_prob,3):0.3f}"
+
+
+class LayerNorm(nn.Module):
+    r"""LayerNorm that supports two data formats: channels_last (default) or channels_first.
+    The ordering of the dimensions in the inputs. channels_last corresponds to inputs with
+    shape (batch_size, height, width, channels) while channels_first corresponds to inputs
+    with shape (batch_size, channels, height, width).
+    """  # noqa: E501
+
+    def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(normalized_shape))
+        self.bias = nn.Parameter(torch.zeros(normalized_shape))
+        self.eps = eps
+        self.data_format = data_format
+        if self.data_format not in ["channels_last", "channels_first"]:
+            raise NotImplementedError
+        self.normalized_shape = (normalized_shape,)
+
+    def forward(self, x):
+        if self.data_format == "channels_last":
+            return F.layer_norm(
+                x, self.normalized_shape, comfy.model_management.cast_to(self.weight, dtype=x.dtype, device=x.device), comfy.model_management.cast_to(self.bias, dtype=x.dtype, device=x.device), self.eps
+            )
+        elif self.data_format == "channels_first":
+            u = x.mean(1, keepdim=True)
+            s = (x - u).pow(2).mean(1, keepdim=True)
+            x = (x - u) / torch.sqrt(s + self.eps)
+            x = comfy.model_management.cast_to(self.weight[:, None], dtype=x.dtype, device=x.device) * x + comfy.model_management.cast_to(self.bias[:, None], dtype=x.dtype, device=x.device)
+            return x
+
+
+class ConvNeXtBlock(nn.Module):
+    r"""ConvNeXt Block. There are two equivalent implementations:
+    (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
+    (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
+    We use (2) as we find it slightly faster in PyTorch
+
+    Args:
+        dim (int): Number of input channels.
+        drop_path (float): Stochastic depth rate. Default: 0.0
+        layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.0.
+        kernel_size (int): Kernel size for depthwise conv. Default: 7.
+        dilation (int): Dilation for depthwise conv. Default: 1.
+    """  # noqa: E501
+
+    def __init__(
+        self,
+        dim: int,
+        drop_path: float = 0.0,
+        layer_scale_init_value: float = 1e-6,
+        mlp_ratio: float = 4.0,
+        kernel_size: int = 7,
+        dilation: int = 1,
+    ):
+        super().__init__()
+
+        self.dwconv = ops.Conv1d(
+            dim,
+            dim,
+            kernel_size=kernel_size,
+            padding=int(dilation * (kernel_size - 1) / 2),
+            groups=dim,
+        )  # depthwise conv
+        self.norm = LayerNorm(dim, eps=1e-6)
+        self.pwconv1 = ops.Linear(
+            dim, int(mlp_ratio * dim)
+        )  # pointwise/1x1 convs, implemented with linear layers
+        self.act = nn.GELU()
+        self.pwconv2 = ops.Linear(int(mlp_ratio * dim), dim)
+        self.gamma = (
+            nn.Parameter(torch.empty((dim)), requires_grad=False)
+            if layer_scale_init_value > 0
+            else None
+        )
+        self.drop_path = DropPath(
+            drop_path) if drop_path > 0.0 else nn.Identity()
+
+    def forward(self, x, apply_residual: bool = True):
+        input = x
+
+        x = self.dwconv(x)
+        x = x.permute(0, 2, 1)  # (N, C, L) -> (N, L, C)
+        x = self.norm(x)
+        x = self.pwconv1(x)
+        x = self.act(x)
+        x = self.pwconv2(x)
+
+        if self.gamma is not None:
+            x = comfy.model_management.cast_to(self.gamma, dtype=x.dtype, device=x.device) * x
+
+        x = x.permute(0, 2, 1)  # (N, L, C) -> (N, C, L)
+        x = self.drop_path(x)
+
+        if apply_residual:
+            x = input + x
+
+        return x
+
+
+class ParallelConvNeXtBlock(nn.Module):
+    def __init__(self, kernel_sizes: List[int], *args, **kwargs):
+        super().__init__()
+        self.blocks = nn.ModuleList(
+            [
+                ConvNeXtBlock(kernel_size=kernel_size, *args, **kwargs)
+                for kernel_size in kernel_sizes
+            ]
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return torch.stack(
+            [block(x, apply_residual=False) for block in self.blocks] + [x],
+            dim=1,
+        ).sum(dim=1)
+
+
+class ConvNeXtEncoder(nn.Module):
+    def __init__(
+        self,
+        input_channels=3,
+        depths=[3, 3, 9, 3],
+        dims=[96, 192, 384, 768],
+        drop_path_rate=0.0,
+        layer_scale_init_value=1e-6,
+        kernel_sizes: Tuple[int] = (7,),
+    ):
+        super().__init__()
+        assert len(depths) == len(dims)
+
+        self.channel_layers = nn.ModuleList()
+        stem = nn.Sequential(
+            ops.Conv1d(
+                input_channels,
+                dims[0],
+                kernel_size=7,
+                padding=3,
+                padding_mode="replicate",
+            ),
+            LayerNorm(dims[0], eps=1e-6, data_format="channels_first"),
+        )
+        self.channel_layers.append(stem)
+
+        for i in range(len(depths) - 1):
+            mid_layer = nn.Sequential(
+                LayerNorm(dims[i], eps=1e-6, data_format="channels_first"),
+                ops.Conv1d(dims[i], dims[i + 1], kernel_size=1),
+            )
+            self.channel_layers.append(mid_layer)
+
+        block_fn = (
+            partial(ConvNeXtBlock, kernel_size=kernel_sizes[0])
+            if len(kernel_sizes) == 1
+            else partial(ParallelConvNeXtBlock, kernel_sizes=kernel_sizes)
+        )
+
+        self.stages = nn.ModuleList()
+        drop_path_rates = [
+            x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))
+        ]
+
+        cur = 0
+        for i in range(len(depths)):
+            stage = nn.Sequential(
+                *[
+                    block_fn(
+                        dim=dims[i],
+                        drop_path=drop_path_rates[cur + j],
+                        layer_scale_init_value=layer_scale_init_value,
+                    )
+                    for j in range(depths[i])
+                ]
+            )
+            self.stages.append(stage)
+            cur += depths[i]
+
+        self.norm = LayerNorm(dims[-1], eps=1e-6, data_format="channels_first")
+
+    def forward(
+        self,
+        x: torch.Tensor,
+    ) -> torch.Tensor:
+        for channel_layer, stage in zip(self.channel_layers, self.stages):
+            x = channel_layer(x)
+            x = stage(x)
+
+        return self.norm(x)
+
+
+def get_padding(kernel_size, dilation=1):
+    return (kernel_size * dilation - dilation) // 2
+
+
+class ResBlock1(torch.nn.Module):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super().__init__()
+
+        self.convs1 = nn.ModuleList(
+            [
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+
+        self.convs2 = nn.ModuleList(
+            [
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.silu(x)
+            xt = c1(xt)
+            xt = F.silu(xt)
+            xt = c2(xt)
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for conv in self.convs1:
+            remove_weight_norm(conv)
+        for conv in self.convs2:
+            remove_weight_norm(conv)
+
+
+class HiFiGANGenerator(nn.Module):
+    def __init__(
+        self,
+        *,
+        hop_length: int = 512,
+        upsample_rates: Tuple[int] = (8, 8, 2, 2, 2),
+        upsample_kernel_sizes: Tuple[int] = (16, 16, 8, 2, 2),
+        resblock_kernel_sizes: Tuple[int] = (3, 7, 11),
+        resblock_dilation_sizes: Tuple[Tuple[int]] = (
+            (1, 3, 5), (1, 3, 5), (1, 3, 5)),
+        num_mels: int = 128,
+        upsample_initial_channel: int = 512,
+        use_template: bool = True,
+        pre_conv_kernel_size: int = 7,
+        post_conv_kernel_size: int = 7,
+        post_activation: Callable = partial(nn.SiLU, inplace=True),
+    ):
+        super().__init__()
+
+        assert (
+            prod(upsample_rates) == hop_length
+        ), f"hop_length must be {prod(upsample_rates)}"
+
+        self.conv_pre = torch.nn.utils.parametrizations.weight_norm(
+            ops.Conv1d(
+                num_mels,
+                upsample_initial_channel,
+                pre_conv_kernel_size,
+                1,
+                padding=get_padding(pre_conv_kernel_size),
+            )
+        )
+
+        self.num_upsamples = len(upsample_rates)
+        self.num_kernels = len(resblock_kernel_sizes)
+
+        self.noise_convs = nn.ModuleList()
+        self.use_template = use_template
+        self.ups = nn.ModuleList()
+
+        for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
+            c_cur = upsample_initial_channel // (2 ** (i + 1))
+            self.ups.append(
+                torch.nn.utils.parametrizations.weight_norm(
+                    ops.ConvTranspose1d(
+                        upsample_initial_channel // (2**i),
+                        upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+            if not use_template:
+                continue
+
+            if i + 1 < len(upsample_rates):
+                stride_f0 = np.prod(upsample_rates[i + 1:])
+                self.noise_convs.append(
+                    ops.Conv1d(
+                        1,
+                        c_cur,
+                        kernel_size=stride_f0 * 2,
+                        stride=stride_f0,
+                        padding=stride_f0 // 2,
+                    )
+                )
+            else:
+                self.noise_convs.append(ops.Conv1d(1, c_cur, kernel_size=1))
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = upsample_initial_channel // (2 ** (i + 1))
+            for k, d in zip(resblock_kernel_sizes, resblock_dilation_sizes):
+                self.resblocks.append(ResBlock1(ch, k, d))
+
+        self.activation_post = post_activation()
+        self.conv_post = torch.nn.utils.parametrizations.weight_norm(
+            ops.Conv1d(
+                ch,
+                1,
+                post_conv_kernel_size,
+                1,
+                padding=get_padding(post_conv_kernel_size),
+            )
+        )
+
+    def forward(self, x, template=None):
+        x = self.conv_pre(x)
+
+        for i in range(self.num_upsamples):
+            x = F.silu(x, inplace=True)
+            x = self.ups[i](x)
+
+            if self.use_template:
+                x = x + self.noise_convs[i](template)
+
+            xs = None
+
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+
+            x = xs / self.num_kernels
+
+        x = self.activation_post(x)
+        x = self.conv_post(x)
+        x = torch.tanh(x)
+
+        return x
+
+    def remove_weight_norm(self):
+        for up in self.ups:
+            remove_weight_norm(up)
+        for block in self.resblocks:
+            block.remove_weight_norm()
+        remove_weight_norm(self.conv_pre)
+        remove_weight_norm(self.conv_post)
+
+
+class ADaMoSHiFiGANV1(nn.Module):
+    def __init__(
+        self,
+        input_channels: int = 128,
+        depths: List[int] = [3, 3, 9, 3],
+        dims: List[int] = [128, 256, 384, 512],
+        drop_path_rate: float = 0.0,
+        kernel_sizes: Tuple[int] = (7,),
+        upsample_rates: Tuple[int] = (4, 4, 2, 2, 2, 2, 2),
+        upsample_kernel_sizes: Tuple[int] = (8, 8, 4, 4, 4, 4, 4),
+        resblock_kernel_sizes: Tuple[int] = (3, 7, 11, 13),
+        resblock_dilation_sizes: Tuple[Tuple[int]] = (
+            (1, 3, 5), (1, 3, 5), (1, 3, 5), (1, 3, 5)),
+        num_mels: int = 512,
+        upsample_initial_channel: int = 1024,
+        use_template: bool = False,
+        pre_conv_kernel_size: int = 13,
+        post_conv_kernel_size: int = 13,
+        sampling_rate: int = 44100,
+        n_fft: int = 2048,
+        win_length: int = 2048,
+        hop_length: int = 512,
+        f_min: int = 40,
+        f_max: int = 16000,
+        n_mels: int = 128,
+    ):
+        super().__init__()
+
+        self.backbone = ConvNeXtEncoder(
+            input_channels=input_channels,
+            depths=depths,
+            dims=dims,
+            drop_path_rate=drop_path_rate,
+            kernel_sizes=kernel_sizes,
+        )
+
+        self.head = HiFiGANGenerator(
+            hop_length=hop_length,
+            upsample_rates=upsample_rates,
+            upsample_kernel_sizes=upsample_kernel_sizes,
+            resblock_kernel_sizes=resblock_kernel_sizes,
+            resblock_dilation_sizes=resblock_dilation_sizes,
+            num_mels=num_mels,
+            upsample_initial_channel=upsample_initial_channel,
+            use_template=use_template,
+            pre_conv_kernel_size=pre_conv_kernel_size,
+            post_conv_kernel_size=post_conv_kernel_size,
+        )
+        self.sampling_rate = sampling_rate
+        self.mel_transform = LogMelSpectrogram(
+            sample_rate=sampling_rate,
+            n_fft=n_fft,
+            win_length=win_length,
+            hop_length=hop_length,
+            f_min=f_min,
+            f_max=f_max,
+            n_mels=n_mels,
+        )
+        self.eval()
+
+    @torch.no_grad()
+    def decode(self, mel):
+        y = self.backbone(mel)
+        y = self.head(y)
+        return y
+
+    @torch.no_grad()
+    def encode(self, x):
+        return self.mel_transform(x)
+
+    def forward(self, mel):
+        y = self.backbone(mel)
+        y = self.head(y)
+        return y

comfy/ldm/audio/autoencoder.py

@@ -75,16 +75,10 @@ class SnakeBeta(nn.Module):
         return x
 
 def WNConv1d(*args, **kwargs):
-    try:
     return torch.nn.utils.parametrizations.weight_norm(ops.Conv1d(*args, **kwargs))
-    except:
-        return torch.nn.utils.weight_norm(ops.Conv1d(*args, **kwargs)) #support pytorch 2.1 and older
 
 def WNConvTranspose1d(*args, **kwargs):
-    try:
     return torch.nn.utils.parametrizations.weight_norm(ops.ConvTranspose1d(*args, **kwargs))
-    except:
-        return torch.nn.utils.weight_norm(ops.ConvTranspose1d(*args, **kwargs)) #support pytorch 2.1 and older
 
 def get_activation(activation: Literal["elu", "snake", "none"], antialias=False, channels=None) -> nn.Module:
     if activation == "elu":

comfy/ldm/cascade/stage_a.py

@@ -19,6 +19,10 @@
 import torch
 from torch import nn
 from torch.autograd import Function
+import comfy.ops
+
+ops = comfy.ops.disable_weight_init
+
 
 class vector_quantize(Function):
     @staticmethod
@@ -121,15 +125,15 @@ class ResBlock(nn.Module):
         self.norm1 = nn.LayerNorm(c, elementwise_affine=False, eps=1e-6)
         self.depthwise = nn.Sequential(
             nn.ReplicationPad2d(1),
-            nn.Conv2d(c, c, kernel_size=3, groups=c)
+            ops.Conv2d(c, c, kernel_size=3, groups=c)
         )
 
         # channelwise
         self.norm2 = nn.LayerNorm(c, elementwise_affine=False, eps=1e-6)
         self.channelwise = nn.Sequential(
-            nn.Linear(c, c_hidden),
+            ops.Linear(c, c_hidden),
             nn.GELU(),
-            nn.Linear(c_hidden, c),
+            ops.Linear(c_hidden, c),
         )
 
         self.gammas = nn.Parameter(torch.zeros(6), requires_grad=True)
@@ -171,16 +175,16 @@ class StageA(nn.Module):
         # Encoder blocks
         self.in_block = nn.Sequential(
             nn.PixelUnshuffle(2),
-            nn.Conv2d(3 * 4, c_levels[0], kernel_size=1)
+            ops.Conv2d(3 * 4, c_levels[0], kernel_size=1)
         )
         down_blocks = []
         for i in range(levels):
             if i > 0:
-                down_blocks.append(nn.Conv2d(c_levels[i - 1], c_levels[i], kernel_size=4, stride=2, padding=1))
+                down_blocks.append(ops.Conv2d(c_levels[i - 1], c_levels[i], kernel_size=4, stride=2, padding=1))
             block = ResBlock(c_levels[i], c_levels[i] * 4)
             down_blocks.append(block)
         down_blocks.append(nn.Sequential(
-            nn.Conv2d(c_levels[-1], c_latent, kernel_size=1, bias=False),
+            ops.Conv2d(c_levels[-1], c_latent, kernel_size=1, bias=False),
             nn.BatchNorm2d(c_latent),  # then normalize them to have mean 0 and std 1
         ))
         self.down_blocks = nn.Sequential(*down_blocks)
@@ -191,7 +195,7 @@ class StageA(nn.Module):
 
         # Decoder blocks
         up_blocks = [nn.Sequential(
-            nn.Conv2d(c_latent, c_levels[-1], kernel_size=1)
+            ops.Conv2d(c_latent, c_levels[-1], kernel_size=1)
         )]
         for i in range(levels):
             for j in range(bottleneck_blocks if i == 0 else 1):
@@ -199,11 +203,11 @@ class StageA(nn.Module):
                 up_blocks.append(block)
             if i < levels - 1:
                 up_blocks.append(
-                    nn.ConvTranspose2d(c_levels[levels - 1 - i], c_levels[levels - 2 - i], kernel_size=4, stride=2,
+                    ops.ConvTranspose2d(c_levels[levels - 1 - i], c_levels[levels - 2 - i], kernel_size=4, stride=2,
                                        padding=1))
         self.up_blocks = nn.Sequential(*up_blocks)
         self.out_block = nn.Sequential(
-            nn.Conv2d(c_levels[0], 3 * 4, kernel_size=1),
+            ops.Conv2d(c_levels[0], 3 * 4, kernel_size=1),
             nn.PixelShuffle(2),
         )
 
@@ -232,17 +236,17 @@ class Discriminator(nn.Module):
         super().__init__()
         d = max(depth - 3, 3)
         layers = [
-            nn.utils.spectral_norm(nn.Conv2d(c_in, c_hidden // (2 ** d), kernel_size=3, stride=2, padding=1)),
+            nn.utils.spectral_norm(ops.Conv2d(c_in, c_hidden // (2 ** d), kernel_size=3, stride=2, padding=1)),
             nn.LeakyReLU(0.2),
         ]
         for i in range(depth - 1):
             c_in = c_hidden // (2 ** max((d - i), 0))
             c_out = c_hidden // (2 ** max((d - 1 - i), 0))
-            layers.append(nn.utils.spectral_norm(nn.Conv2d(c_in, c_out, kernel_size=3, stride=2, padding=1)))
+            layers.append(nn.utils.spectral_norm(ops.Conv2d(c_in, c_out, kernel_size=3, stride=2, padding=1)))
             layers.append(nn.InstanceNorm2d(c_out))
             layers.append(nn.LeakyReLU(0.2))
         self.encoder = nn.Sequential(*layers)
-        self.shuffle = nn.Conv2d((c_hidden + c_cond) if c_cond > 0 else c_hidden, 1, kernel_size=1)
+        self.shuffle = ops.Conv2d((c_hidden + c_cond) if c_cond > 0 else c_hidden, 1, kernel_size=1)
         self.logits = nn.Sigmoid()
 
     def forward(self, x, cond=None):

comfy/ldm/cascade/stage_c_coder.py

@@ -19,6 +19,9 @@ import torch
 import torchvision
 from torch import nn
 
+import comfy.ops
+
+ops = comfy.ops.disable_weight_init
 
 # EfficientNet
 class EfficientNetEncoder(nn.Module):
@@ -26,7 +29,7 @@ class EfficientNetEncoder(nn.Module):
         super().__init__()
         self.backbone = torchvision.models.efficientnet_v2_s().features.eval()
         self.mapper = nn.Sequential(
-            nn.Conv2d(1280, c_latent, kernel_size=1, bias=False),
+            ops.Conv2d(1280, c_latent, kernel_size=1, bias=False),
             nn.BatchNorm2d(c_latent, affine=False),  # then normalize them to have mean 0 and std 1
         )
         self.mean = nn.Parameter(torch.tensor([0.485, 0.456, 0.406]))
@@ -34,7 +37,7 @@ class EfficientNetEncoder(nn.Module):
 
     def forward(self, x):
         x = x * 0.5 + 0.5
-        x = (x - self.mean.view([3,1,1])) / self.std.view([3,1,1])
+        x = (x - self.mean.view([3,1,1]).to(device=x.device, dtype=x.dtype)) / self.std.view([3,1,1]).to(device=x.device, dtype=x.dtype)
         o = self.mapper(self.backbone(x))
         return o
 
@@ -44,39 +47,39 @@ class Previewer(nn.Module):
     def __init__(self, c_in=16, c_hidden=512, c_out=3):
         super().__init__()
         self.blocks = nn.Sequential(
-            nn.Conv2d(c_in, c_hidden, kernel_size=1),  # 16 channels to 512 channels
+            ops.Conv2d(c_in, c_hidden, kernel_size=1),  # 16 channels to 512 channels
             nn.GELU(),
             nn.BatchNorm2d(c_hidden),
 
-            nn.Conv2d(c_hidden, c_hidden, kernel_size=3, padding=1),
+            ops.Conv2d(c_hidden, c_hidden, kernel_size=3, padding=1),
             nn.GELU(),
             nn.BatchNorm2d(c_hidden),
 
-            nn.ConvTranspose2d(c_hidden, c_hidden // 2, kernel_size=2, stride=2),  # 16 -> 32
+            ops.ConvTranspose2d(c_hidden, c_hidden // 2, kernel_size=2, stride=2),  # 16 -> 32
             nn.GELU(),
             nn.BatchNorm2d(c_hidden // 2),
 
-            nn.Conv2d(c_hidden // 2, c_hidden // 2, kernel_size=3, padding=1),
+            ops.Conv2d(c_hidden // 2, c_hidden // 2, kernel_size=3, padding=1),
             nn.GELU(),
             nn.BatchNorm2d(c_hidden // 2),
 
-            nn.ConvTranspose2d(c_hidden // 2, c_hidden // 4, kernel_size=2, stride=2),  # 32 -> 64
+            ops.ConvTranspose2d(c_hidden // 2, c_hidden // 4, kernel_size=2, stride=2),  # 32 -> 64
             nn.GELU(),
             nn.BatchNorm2d(c_hidden // 4),
 
-            nn.Conv2d(c_hidden // 4, c_hidden // 4, kernel_size=3, padding=1),
+            ops.Conv2d(c_hidden // 4, c_hidden // 4, kernel_size=3, padding=1),
             nn.GELU(),
             nn.BatchNorm2d(c_hidden // 4),
 
-            nn.ConvTranspose2d(c_hidden // 4, c_hidden // 4, kernel_size=2, stride=2),  # 64 -> 128
+            ops.ConvTranspose2d(c_hidden // 4, c_hidden // 4, kernel_size=2, stride=2),  # 64 -> 128
             nn.GELU(),
             nn.BatchNorm2d(c_hidden // 4),
 
-            nn.Conv2d(c_hidden // 4, c_hidden // 4, kernel_size=3, padding=1),
+            ops.Conv2d(c_hidden // 4, c_hidden // 4, kernel_size=3, padding=1),
             nn.GELU(),
             nn.BatchNorm2d(c_hidden // 4),
 
-            nn.Conv2d(c_hidden // 4, c_out, kernel_size=1),
+            ops.Conv2d(c_hidden // 4, c_out, kernel_size=1),
         )
 
     def forward(self, x):

comfy/ldm/chroma/layers.py

@@ -0,0 +1,183 @@
+import torch
+from torch import Tensor, nn
+
+from comfy.ldm.flux.math import attention
+from comfy.ldm.flux.layers import (
+    MLPEmbedder,
+    RMSNorm,
+    QKNorm,
+    SelfAttention,
+    ModulationOut,
+)
+
+
+
+class ChromaModulationOut(ModulationOut):
+    @classmethod
+    def from_offset(cls, tensor: torch.Tensor, offset: int = 0) -> ModulationOut:
+        return cls(
+            shift=tensor[:, offset : offset + 1, :],
+            scale=tensor[:, offset + 1 : offset + 2, :],
+            gate=tensor[:, offset + 2 : offset + 3, :],
+        )
+
+
+
+
+class Approximator(nn.Module):
+    def __init__(self, in_dim: int, out_dim: int, hidden_dim: int, n_layers = 5, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.in_proj = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
+        self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
+        self.norms = nn.ModuleList([RMSNorm(hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
+        self.out_proj = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device)
+
+    @property
+    def device(self):
+        # Get the device of the module (assumes all parameters are on the same device)
+        return next(self.parameters()).device
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = self.in_proj(x)
+
+        for layer, norms in zip(self.layers, self.norms):
+            x = x + layer(norms(x))
+
+        x = self.out_proj(x)
+
+        return x
+
+
+class DoubleStreamBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.img_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.img_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.img_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+        self.txt_norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias, dtype=dtype, device=device, operations=operations)
+
+        self.txt_norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.txt_mlp = nn.Sequential(
+            operations.Linear(hidden_size, mlp_hidden_dim, bias=True, dtype=dtype, device=device),
+            nn.GELU(approximate="tanh"),
+            operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+        self.flipped_img_txt = flipped_img_txt
+
+    def forward(self, img: Tensor, txt: Tensor, pe: Tensor, vec: Tensor, attn_mask=None):
+        (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
+
+        # prepare image for attention
+        img_modulated = self.img_norm1(img)
+        img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+        img_qkv = self.img_attn.qkv(img_modulated)
+        img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
+
+        # prepare txt for attention
+        txt_modulated = self.txt_norm1(txt)
+        txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+        txt_qkv = self.txt_attn.qkv(txt_modulated)
+        txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
+
+        # run actual attention
+        attn = attention(torch.cat((txt_q, img_q), dim=2),
+                         torch.cat((txt_k, img_k), dim=2),
+                         torch.cat((txt_v, img_v), dim=2),
+                         pe=pe, mask=attn_mask)
+
+        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1] :]
+
+        # calculate the img bloks
+        img = img + img_mod1.gate * self.img_attn.proj(img_attn)
+        img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
+
+        # calculate the txt bloks
+        txt += txt_mod1.gate * self.txt_attn.proj(txt_attn)
+        txt += txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+
+        if txt.dtype == torch.float16:
+            txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
+
+        return img, txt
+
+
+class SingleStreamBlock(nn.Module):
+    """
+    A DiT block with parallel linear layers as described in
+    https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_scale: float = None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_size
+        self.num_heads = num_heads
+        head_dim = hidden_size // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        # qkv and mlp_in
+        self.linear1 = operations.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim, dtype=dtype, device=device)
+        # proj and mlp_out
+        self.linear2 = operations.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, dtype=dtype, device=device)
+
+        self.norm = QKNorm(head_dim, dtype=dtype, device=device, operations=operations)
+
+        self.hidden_size = hidden_size
+        self.pre_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+
+        self.mlp_act = nn.GELU(approximate="tanh")
+
+    def forward(self, x: Tensor, pe: Tensor, vec: Tensor, attn_mask=None) -> Tensor:
+        mod = vec
+        x_mod = (1 + mod.scale) * self.pre_norm(x) + mod.shift
+        qkv, mlp = torch.split(self.linear1(x_mod), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+
+        q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k = self.norm(q, k, v)
+
+        # compute attention
+        attn = attention(q, k, v, pe=pe, mask=attn_mask)
+        # compute activation in mlp stream, cat again and run second linear layer
+        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
+        x += mod.gate * output
+        if x.dtype == torch.float16:
+            x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
+        return x
+
+
+class LastLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, out_channels, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x: Tensor, vec: Tensor) -> Tensor:
+        shift, scale = vec
+        shift = shift.squeeze(1)
+        scale = scale.squeeze(1)
+        x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+        x = self.linear(x)
+        return x

comfy/ldm/chroma/model.py

@@ -0,0 +1,271 @@
+#Original code can be found on: https://github.com/black-forest-labs/flux
+
+from dataclasses import dataclass
+
+import torch
+from torch import Tensor, nn
+from einops import rearrange, repeat
+import comfy.ldm.common_dit
+
+from comfy.ldm.flux.layers import (
+    EmbedND,
+    timestep_embedding,
+)
+
+from .layers import (
+    DoubleStreamBlock,
+    LastLayer,
+    SingleStreamBlock,
+    Approximator,
+    ChromaModulationOut,
+)
+
+
+@dataclass
+class ChromaParams:
+    in_channels: int
+    out_channels: int
+    context_in_dim: int
+    hidden_size: int
+    mlp_ratio: float
+    num_heads: int
+    depth: int
+    depth_single_blocks: int
+    axes_dim: list
+    theta: int
+    patch_size: int
+    qkv_bias: bool
+    in_dim: int
+    out_dim: int
+    hidden_dim: int
+    n_layers: int
+
+
+
+
+class Chroma(nn.Module):
+    """
+    Transformer model for flow matching on sequences.
+    """
+
+    def __init__(self, image_model=None, final_layer=True, dtype=None, device=None, operations=None, **kwargs):
+        super().__init__()
+        self.dtype = dtype
+        params = ChromaParams(**kwargs)
+        self.params = params
+        self.patch_size = params.patch_size
+        self.in_channels = params.in_channels
+        self.out_channels = params.out_channels
+        if params.hidden_size % params.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}"
+            )
+        pe_dim = params.hidden_size // params.num_heads
+        if sum(params.axes_dim) != pe_dim:
+            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = params.hidden_size
+        self.num_heads = params.num_heads
+        self.in_dim = params.in_dim
+        self.out_dim = params.out_dim
+        self.hidden_dim = params.hidden_dim
+        self.n_layers = params.n_layers
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
+        self.img_in = operations.Linear(self.in_channels, self.hidden_size, bias=True, dtype=dtype, device=device)
+        self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, dtype=dtype, device=device)
+        # set as nn identity for now, will overwrite it later.
+        self.distilled_guidance_layer = Approximator(
+                    in_dim=self.in_dim,
+                    hidden_dim=self.hidden_dim,
+                    out_dim=self.out_dim,
+                    n_layers=self.n_layers,
+                    dtype=dtype, device=device, operations=operations
+                )
+
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=params.mlp_ratio,
+                    qkv_bias=params.qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(params.depth)
+            ]
+        )
+
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=params.mlp_ratio, dtype=dtype, device=device, operations=operations)
+                for _ in range(params.depth_single_blocks)
+            ]
+        )
+
+        if final_layer:
+            self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+        self.skip_mmdit = []
+        self.skip_dit = []
+        self.lite = False
+
+    def get_modulations(self, tensor: torch.Tensor, block_type: str, *, idx: int = 0):
+        # This function slices up the modulations tensor which has the following layout:
+        #   single     : num_single_blocks * 3 elements
+        #   double_img : num_double_blocks * 6 elements
+        #   double_txt : num_double_blocks * 6 elements
+        #   final      : 2 elements
+        if block_type == "final":
+            return (tensor[:, -2:-1, :], tensor[:, -1:, :])
+        single_block_count = self.params.depth_single_blocks
+        double_block_count = self.params.depth
+        offset = 3 * idx
+        if block_type == "single":
+            return ChromaModulationOut.from_offset(tensor, offset)
+        # Double block modulations are 6 elements so we double 3 * idx.
+        offset *= 2
+        if block_type in {"double_img", "double_txt"}:
+            # Advance past the single block modulations.
+            offset += 3 * single_block_count
+            if block_type == "double_txt":
+                # Advance past the double block img modulations.
+                offset += 6 * double_block_count
+            return (
+                ChromaModulationOut.from_offset(tensor, offset),
+                ChromaModulationOut.from_offset(tensor, offset + 3),
+            )
+        raise ValueError("Bad block_type")
+
+
+    def forward_orig(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        guidance: Tensor = None,
+        control = None,
+        transformer_options={},
+        attn_mask: Tensor = None,
+    ) -> Tensor:
+        patches_replace = transformer_options.get("patches_replace", {})
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img = self.img_in(img)
+
+        # distilled vector guidance
+        mod_index_length = 344
+        distill_timestep = timestep_embedding(timesteps.detach().clone(), 16).to(img.device, img.dtype)
+        # guidance = guidance *
+        distil_guidance = timestep_embedding(guidance.detach().clone(), 16).to(img.device, img.dtype)
+
+        # get all modulation index
+        modulation_index = timestep_embedding(torch.arange(mod_index_length, device=img.device), 32).to(img.device, img.dtype)
+        # we need to broadcast the modulation index here so each batch has all of the index
+        modulation_index = modulation_index.unsqueeze(0).repeat(img.shape[0], 1, 1).to(img.device, img.dtype)
+        # and we need to broadcast timestep and guidance along too
+        timestep_guidance = torch.cat([distill_timestep, distil_guidance], dim=1).unsqueeze(1).repeat(1, mod_index_length, 1).to(img.dtype).to(img.device, img.dtype)
+        # then and only then we could concatenate it together
+        input_vec = torch.cat([timestep_guidance, modulation_index], dim=-1).to(img.device, img.dtype)
+
+        mod_vectors = self.distilled_guidance_layer(input_vec)
+
+        txt = self.txt_in(txt)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.double_blocks):
+            if i not in self.skip_mmdit:
+                double_mod = (
+                    self.get_modulations(mod_vectors, "double_img", idx=i),
+                    self.get_modulations(mod_vectors, "double_txt", idx=i),
+                )
+                if ("double_block", i) in blocks_replace:
+                    def block_wrap(args):
+                        out = {}
+                        out["img"], out["txt"] = block(img=args["img"],
+                                                       txt=args["txt"],
+                                                       vec=args["vec"],
+                                                       pe=args["pe"],
+                                                       attn_mask=args.get("attn_mask"))
+                        return out
+
+                    out = blocks_replace[("double_block", i)]({"img": img,
+                                                               "txt": txt,
+                                                               "vec": double_mod,
+                                                               "pe": pe,
+                                                               "attn_mask": attn_mask},
+                                                              {"original_block": block_wrap})
+                    txt = out["txt"]
+                    img = out["img"]
+                else:
+                    img, txt = block(img=img,
+                                     txt=txt,
+                                     vec=double_mod,
+                                     pe=pe,
+                                     attn_mask=attn_mask)
+
+                if control is not None: # Controlnet
+                    control_i = control.get("input")
+                    if i < len(control_i):
+                        add = control_i[i]
+                        if add is not None:
+                            img += add
+
+        img = torch.cat((txt, img), 1)
+
+        for i, block in enumerate(self.single_blocks):
+            if i not in self.skip_dit:
+                single_mod = self.get_modulations(mod_vectors, "single", idx=i)
+                if ("single_block", i) in blocks_replace:
+                    def block_wrap(args):
+                        out = {}
+                        out["img"] = block(args["img"],
+                                           vec=args["vec"],
+                                           pe=args["pe"],
+                                           attn_mask=args.get("attn_mask"))
+                        return out
+
+                    out = blocks_replace[("single_block", i)]({"img": img,
+                                                               "vec": single_mod,
+                                                               "pe": pe,
+                                                               "attn_mask": attn_mask},
+                                                              {"original_block": block_wrap})
+                    img = out["img"]
+                else:
+                    img = block(img, vec=single_mod, pe=pe, attn_mask=attn_mask)
+
+                if control is not None: # Controlnet
+                    control_o = control.get("output")
+                    if i < len(control_o):
+                        add = control_o[i]
+                        if add is not None:
+                            img[:, txt.shape[1] :, ...] += add
+
+        img = img[:, txt.shape[1] :, ...]
+        final_mod = self.get_modulations(mod_vectors, "final")
+        img = self.final_layer(img, vec=final_mod)  # (N, T, patch_size ** 2 * out_channels)
+        return img
+
+    def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
+        bs, c, h, w = x.shape
+        patch_size = 2
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))
+
+        img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
+
+        h_len = ((h + (patch_size // 2)) // patch_size)
+        w_len = ((w + (patch_size // 2)) // patch_size)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+
+        txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
+        out = self.forward_orig(img, img_ids, context, txt_ids, timestep, guidance, control, transformer_options, attn_mask=kwargs.get("attention_mask", None))
+        return rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)[:,:,:h,:w]

comfy/ldm/common_dit.py

@@ -1,5 +1,6 @@
 import torch
-import comfy.ops
+import comfy.rmsnorm
+
 
 def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
     if padding_mode == "circular" and (torch.jit.is_tracing() or torch.jit.is_scripting()):
@@ -11,20 +12,5 @@ def pad_to_patch_size(img, patch_size=(2, 2), padding_mode="circular"):
 
     return torch.nn.functional.pad(img, pad, mode=padding_mode)
 
-try:
-    rms_norm_torch = torch.nn.functional.rms_norm
-except:
-    rms_norm_torch = None
 
-def rms_norm(x, weight=None, eps=1e-6):
-    if rms_norm_torch is not None and not (torch.jit.is_tracing() or torch.jit.is_scripting()):
-        if weight is None:
-            return rms_norm_torch(x, (x.shape[-1],), eps=eps)
-        else:
-            return rms_norm_torch(x, weight.shape, weight=comfy.ops.cast_to(weight, dtype=x.dtype, device=x.device), eps=eps)
-    else:
-        r = x * torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + eps)
-        if weight is None:
-            return r
-        else:
-            return r * comfy.ops.cast_to(weight, dtype=x.dtype, device=x.device)
+rms_norm = comfy.rmsnorm.rms_norm

comfy/ldm/cosmos/blocks.py

@@ -23,7 +23,6 @@ from einops import rearrange, repeat
 from einops.layers.torch import Rearrange
 from torch import nn
 
-from comfy.ldm.modules.diffusionmodules.mmdit import RMSNorm
 from comfy.ldm.modules.attention import optimized_attention
 
 
@@ -37,11 +36,11 @@ def apply_rotary_pos_emb(
     return t_out
 
 
-def get_normalization(name: str, channels: int, weight_args={}):
+def get_normalization(name: str, channels: int, weight_args={}, operations=None):
     if name == "I":
         return nn.Identity()
     elif name == "R":
-        return RMSNorm(channels, elementwise_affine=True, eps=1e-6, **weight_args)
+        return operations.RMSNorm(channels, elementwise_affine=True, eps=1e-6, **weight_args)
     else:
         raise ValueError(f"Normalization {name} not found")
 
@@ -120,15 +119,15 @@ class Attention(nn.Module):
 
         self.to_q = nn.Sequential(
             operations.Linear(query_dim, inner_dim, bias=qkv_bias, **weight_args),
-            get_normalization(qkv_norm[0], norm_dim),
+            get_normalization(qkv_norm[0], norm_dim, weight_args=weight_args, operations=operations),
         )
         self.to_k = nn.Sequential(
             operations.Linear(context_dim, inner_dim, bias=qkv_bias, **weight_args),
-            get_normalization(qkv_norm[1], norm_dim),
+            get_normalization(qkv_norm[1], norm_dim, weight_args=weight_args, operations=operations),
         )
         self.to_v = nn.Sequential(
             operations.Linear(context_dim, inner_dim, bias=qkv_bias, **weight_args),
-            get_normalization(qkv_norm[2], norm_dim),
+            get_normalization(qkv_norm[2], norm_dim, weight_args=weight_args, operations=operations),
         )
 
         self.to_out = nn.Sequential(

comfy/ldm/cosmos/model.py

@@ -27,8 +27,6 @@ from torchvision import transforms
 from enum import Enum
 import logging
 
-from comfy.ldm.modules.diffusionmodules.mmdit import RMSNorm
-
 from .blocks import (
     FinalLayer,
     GeneralDITTransformerBlock,
@@ -195,7 +193,7 @@ class GeneralDIT(nn.Module):
 
         if self.affline_emb_norm:
             logging.debug("Building affine embedding normalization layer")
-            self.affline_norm = RMSNorm(model_channels, elementwise_affine=True, eps=1e-6)
+            self.affline_norm = operations.RMSNorm(model_channels, elementwise_affine=True, eps=1e-6, device=device, dtype=dtype)
         else:
             self.affline_norm = nn.Identity()
 

comfy/ldm/flux/layers.py

@@ -105,7 +105,9 @@ class Modulation(nn.Module):
         self.lin = operations.Linear(dim, self.multiplier * dim, bias=True, dtype=dtype, device=device)
 
     def forward(self, vec: Tensor) -> tuple:
-        out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(self.multiplier, dim=-1)
+        if vec.ndim == 2:
+            vec = vec[:, None, :]
+        out = self.lin(nn.functional.silu(vec)).chunk(self.multiplier, dim=-1)
 
         return (
             ModulationOut(*out[:3]),
@@ -113,6 +115,20 @@ class Modulation(nn.Module):
         )
 
 
+def apply_mod(tensor, m_mult, m_add=None, modulation_dims=None):
+    if modulation_dims is None:
+        if m_add is not None:
+            return tensor * m_mult + m_add
+        else:
+            return tensor * m_mult
+    else:
+        for d in modulation_dims:
+            tensor[:, d[0]:d[1]] *= m_mult[:, d[2]]
+            if m_add is not None:
+                tensor[:, d[0]:d[1]] += m_add[:, d[2]]
+        return tensor
+
+
 class DoubleStreamBlock(nn.Module):
     def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, dtype=None, device=None, operations=None):
         super().__init__()
@@ -143,20 +159,20 @@ class DoubleStreamBlock(nn.Module):
         )
         self.flipped_img_txt = flipped_img_txt
 
-    def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None):
+    def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None):
         img_mod1, img_mod2 = self.img_mod(vec)
         txt_mod1, txt_mod2 = self.txt_mod(vec)
 
         # prepare image for attention
         img_modulated = self.img_norm1(img)
-        img_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+        img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img)
         img_qkv = self.img_attn.qkv(img_modulated)
         img_q, img_k, img_v = img_qkv.view(img_qkv.shape[0], img_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
         img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
 
         # prepare txt for attention
         txt_modulated = self.txt_norm1(txt)
-        txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+        txt_modulated = apply_mod(txt_modulated, (1 + txt_mod1.scale), txt_mod1.shift, modulation_dims_txt)
         txt_qkv = self.txt_attn.qkv(txt_modulated)
         txt_q, txt_k, txt_v = txt_qkv.view(txt_qkv.shape[0], txt_qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
         txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
@@ -179,12 +195,12 @@ class DoubleStreamBlock(nn.Module):
             txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
         # calculate the img bloks
-        img = img + img_mod1.gate * self.img_attn.proj(img_attn)
-        img = img + img_mod2.gate * self.img_mlp((1 + img_mod2.scale) * self.img_norm2(img) + img_mod2.shift)
+        img = img + apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
+        img = img + apply_mod(self.img_mlp(apply_mod(self.img_norm2(img), (1 + img_mod2.scale), img_mod2.shift, modulation_dims_img)), img_mod2.gate, None, modulation_dims_img)
 
         # calculate the txt bloks
-        txt += txt_mod1.gate * self.txt_attn.proj(txt_attn)
-        txt += txt_mod2.gate * self.txt_mlp((1 + txt_mod2.scale) * self.txt_norm2(txt) + txt_mod2.shift)
+        txt += apply_mod(self.txt_attn.proj(txt_attn), txt_mod1.gate, None, modulation_dims_txt)
+        txt += apply_mod(self.txt_mlp(apply_mod(self.txt_norm2(txt), (1 + txt_mod2.scale), txt_mod2.shift, modulation_dims_txt)), txt_mod2.gate, None, modulation_dims_txt)
 
         if txt.dtype == torch.float16:
             txt = torch.nan_to_num(txt, nan=0.0, posinf=65504, neginf=-65504)
@@ -228,9 +244,9 @@ class SingleStreamBlock(nn.Module):
         self.mlp_act = nn.GELU(approximate="tanh")
         self.modulation = Modulation(hidden_size, double=False, dtype=dtype, device=device, operations=operations)
 
-    def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None) -> Tensor:
+    def forward(self, x: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims=None) -> Tensor:
         mod, _ = self.modulation(vec)
-        qkv, mlp = torch.split(self.linear1((1 + mod.scale) * self.pre_norm(x) + mod.shift), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
+        qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim], dim=-1)
 
         q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
         q, k = self.norm(q, k, v)
@@ -239,7 +255,7 @@ class SingleStreamBlock(nn.Module):
         attn = attention(q, k, v, pe=pe, mask=attn_mask)
         # compute activation in mlp stream, cat again and run second linear layer
         output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
-        x += mod.gate * output
+        x += apply_mod(output, mod.gate, None, modulation_dims)
         if x.dtype == torch.float16:
             x = torch.nan_to_num(x, nan=0.0, posinf=65504, neginf=-65504)
         return x
@@ -252,8 +268,11 @@ class LastLayer(nn.Module):
         self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
         self.adaLN_modulation = nn.Sequential(nn.SiLU(), operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device))
 
-    def forward(self, x: Tensor, vec: Tensor) -> Tensor:
-        shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
-        x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+    def forward(self, x: Tensor, vec: Tensor, modulation_dims=None) -> Tensor:
+        if vec.ndim == 2:
+            vec = vec[:, None, :]
+
+        shift, scale = self.adaLN_modulation(vec).chunk(2, dim=-1)
+        x = apply_mod(self.norm_final(x), (1 + scale), shift, modulation_dims)
         x = self.linear(x)
         return x

comfy/ldm/flux/math.py

@@ -10,8 +10,9 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None) -> Tensor:
     q_shape = q.shape
     k_shape = k.shape
 
-    q = q.float().reshape(*q.shape[:-1], -1, 1, 2)
-    k = k.float().reshape(*k.shape[:-1], -1, 1, 2)
+    if pe is not None:
+        q = q.to(dtype=pe.dtype).reshape(*q.shape[:-1], -1, 1, 2)
+        k = k.to(dtype=pe.dtype).reshape(*k.shape[:-1], -1, 1, 2)
         q = (pe[..., 0] * q[..., 0] + pe[..., 1] * q[..., 1]).reshape(*q_shape).type_as(v)
         k = (pe[..., 0] * k[..., 0] + pe[..., 1] * k[..., 1]).reshape(*k_shape).type_as(v)
 
@@ -22,7 +23,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None) -> Tensor:
 
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
-    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu():
+    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled():
         device = torch.device("cpu")
     else:
         device = pos.device
@@ -36,8 +37,8 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
 
 
 def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
-    xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
-    xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
+    xq_ = xq.to(dtype=freqs_cis.dtype).reshape(*xq.shape[:-1], -1, 1, 2)
+    xk_ = xk.to(dtype=freqs_cis.dtype).reshape(*xk.shape[:-1], -1, 1, 2)
     xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
     xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
     return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)

comfy/ldm/flux/model.py

@@ -109,15 +109,17 @@ class Flux(nn.Module):
         img = self.img_in(img)
         vec = self.time_in(timestep_embedding(timesteps, 256).to(img.dtype))
         if self.params.guidance_embed:
-            if guidance is None:
-                raise ValueError("Didn't get guidance strength for guidance distilled model.")
+            if guidance is not None:
                 vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
 
         vec = vec + self.vector_in(y[:,:self.params.vec_in_dim])
         txt = self.txt_in(txt)
 
+        if img_ids is not None:
             ids = torch.cat((txt_ids, img_ids), dim=1)
             pe = self.pe_embedder(ids)
+        else:
+            pe = None
 
         blocks_replace = patches_replace.get("dit", {})
         for i, block in enumerate(self.double_blocks):
@@ -186,7 +188,7 @@ class Flux(nn.Module):
         img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
         return img
 
-    def forward(self, x, timestep, context, y, guidance, control=None, transformer_options={}, **kwargs):
+    def forward(self, x, timestep, context, y, guidance=None, control=None, transformer_options={}, **kwargs):
         bs, c, h, w = x.shape
         patch_size = self.patch_size
         x = comfy.ldm.common_dit.pad_to_patch_size(x, (patch_size, patch_size))

comfy/ldm/genmo/joint_model/asymm_models_joint.py

@@ -13,7 +13,6 @@ from comfy.ldm.modules.attention import optimized_attention
 from .layers import (
     FeedForward,
     PatchEmbed,
-    RMSNorm,
     TimestepEmbedder,
 )
 
@@ -90,10 +89,10 @@ class AsymmetricAttention(nn.Module):
 
         # Query and key normalization for stability.
         assert qk_norm
-        self.q_norm_x = RMSNorm(self.head_dim, device=device, dtype=dtype)
-        self.k_norm_x = RMSNorm(self.head_dim, device=device, dtype=dtype)
-        self.q_norm_y = RMSNorm(self.head_dim, device=device, dtype=dtype)
-        self.k_norm_y = RMSNorm(self.head_dim, device=device, dtype=dtype)
+        self.q_norm_x = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
+        self.k_norm_x = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
+        self.q_norm_y = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
+        self.k_norm_y = operations.RMSNorm(self.head_dim, eps=1e-5, device=device, dtype=dtype)
 
         # Output layers. y features go back down from dim_x -> dim_y.
         self.proj_x = operations.Linear(dim_x, dim_x, bias=out_bias, device=device, dtype=dtype)

comfy/ldm/genmo/joint_model/layers.py

@@ -151,14 +151,3 @@ class PatchEmbed(nn.Module):
 
         x = self.norm(x)
         return x
-
-
-class RMSNorm(torch.nn.Module):
-    def __init__(self, hidden_size, eps=1e-5, device=None, dtype=None):
-        super().__init__()
-        self.eps = eps
-        self.weight = torch.nn.Parameter(torch.empty(hidden_size, device=device, dtype=dtype))
-        self.register_parameter("bias", None)
-
-    def forward(self, x):
-        return comfy.ldm.common_dit.rms_norm(x, self.weight, self.eps)

comfy/ldm/hidream/model.py

@@ -0,0 +1,802 @@
+from typing import Optional, Tuple, List
+
+import torch
+import torch.nn as nn
+import einops
+from einops import repeat
+
+from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
+import torch.nn.functional as F
+
+from comfy.ldm.flux.math import apply_rope, rope
+from comfy.ldm.flux.layers import LastLayer
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+import comfy.ldm.common_dit
+
+
+# Copied from https://github.com/black-forest-labs/flux/blob/main/src/flux/modules/layers.py
+class EmbedND(nn.Module):
+    def __init__(self, theta: int, axes_dim: List[int]):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        n_axes = ids.shape[-1]
+        emb = torch.cat(
+            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
+            dim=-3,
+        )
+        return emb.unsqueeze(2)
+
+
+class PatchEmbed(nn.Module):
+    def __init__(
+        self,
+        patch_size=2,
+        in_channels=4,
+        out_channels=1024,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.out_channels = out_channels
+        self.proj = operations.Linear(in_channels * patch_size * patch_size, out_channels, bias=True, dtype=dtype, device=device)
+
+    def forward(self, latent):
+        latent = self.proj(latent)
+        return latent
+
+
+class PooledEmbed(nn.Module):
+    def __init__(self, text_emb_dim, hidden_size, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.pooled_embedder = TimestepEmbedding(in_channels=text_emb_dim, time_embed_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, pooled_embed):
+        return self.pooled_embedder(pooled_embed)
+
+
+class TimestepEmbed(nn.Module):
+    def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=frequency_embedding_size, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=frequency_embedding_size, time_embed_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, timesteps, wdtype):
+        t_emb = self.time_proj(timesteps).to(dtype=wdtype)
+        t_emb = self.timestep_embedder(t_emb)
+        return t_emb
+
+
+def attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor):
+    return optimized_attention(query.view(query.shape[0], -1, query.shape[-1] * query.shape[-2]), key.view(key.shape[0], -1, key.shape[-1] * key.shape[-2]), value.view(value.shape[0], -1, value.shape[-1] * value.shape[-2]), query.shape[2])
+
+
+class HiDreamAttnProcessor_flashattn:
+    """Attention processor used typically in processing the SD3-like self-attention projections."""
+
+    def __call__(
+        self,
+        attn,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+        *args,
+        **kwargs,
+    ) -> torch.FloatTensor:
+        dtype = image_tokens.dtype
+        batch_size = image_tokens.shape[0]
+
+        query_i = attn.q_rms_norm(attn.to_q(image_tokens)).to(dtype=dtype)
+        key_i = attn.k_rms_norm(attn.to_k(image_tokens)).to(dtype=dtype)
+        value_i = attn.to_v(image_tokens)
+
+        inner_dim = key_i.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query_i = query_i.view(batch_size, -1, attn.heads, head_dim)
+        key_i = key_i.view(batch_size, -1, attn.heads, head_dim)
+        value_i = value_i.view(batch_size, -1, attn.heads, head_dim)
+        if image_tokens_masks is not None:
+            key_i = key_i * image_tokens_masks.view(batch_size, -1, 1, 1)
+
+        if not attn.single:
+            query_t = attn.q_rms_norm_t(attn.to_q_t(text_tokens)).to(dtype=dtype)
+            key_t = attn.k_rms_norm_t(attn.to_k_t(text_tokens)).to(dtype=dtype)
+            value_t = attn.to_v_t(text_tokens)
+
+            query_t = query_t.view(batch_size, -1, attn.heads, head_dim)
+            key_t = key_t.view(batch_size, -1, attn.heads, head_dim)
+            value_t = value_t.view(batch_size, -1, attn.heads, head_dim)
+
+            num_image_tokens = query_i.shape[1]
+            num_text_tokens = query_t.shape[1]
+            query = torch.cat([query_i, query_t], dim=1)
+            key = torch.cat([key_i, key_t], dim=1)
+            value = torch.cat([value_i, value_t], dim=1)
+        else:
+            query = query_i
+            key = key_i
+            value = value_i
+
+        if query.shape[-1] == rope.shape[-3] * 2:
+            query, key = apply_rope(query, key, rope)
+        else:
+            query_1, query_2 = query.chunk(2, dim=-1)
+            key_1, key_2 = key.chunk(2, dim=-1)
+            query_1, key_1 = apply_rope(query_1, key_1, rope)
+            query = torch.cat([query_1, query_2], dim=-1)
+            key = torch.cat([key_1, key_2], dim=-1)
+
+        hidden_states = attention(query, key, value)
+
+        if not attn.single:
+            hidden_states_i, hidden_states_t = torch.split(hidden_states, [num_image_tokens, num_text_tokens], dim=1)
+            hidden_states_i = attn.to_out(hidden_states_i)
+            hidden_states_t = attn.to_out_t(hidden_states_t)
+            return hidden_states_i, hidden_states_t
+        else:
+            hidden_states = attn.to_out(hidden_states)
+            return hidden_states
+
+class HiDreamAttention(nn.Module):
+    def __init__(
+        self,
+        query_dim: int,
+        heads: int = 8,
+        dim_head: int = 64,
+        upcast_attention: bool = False,
+        upcast_softmax: bool = False,
+        scale_qk: bool = True,
+        eps: float = 1e-5,
+        processor = None,
+        out_dim: int = None,
+        single: bool = False,
+        dtype=None, device=None, operations=None
+    ):
+        # super(Attention, self).__init__()
+        super().__init__()
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.query_dim = query_dim
+        self.upcast_attention = upcast_attention
+        self.upcast_softmax = upcast_softmax
+        self.out_dim = out_dim if out_dim is not None else query_dim
+
+        self.scale_qk = scale_qk
+        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
+
+        self.heads = out_dim // dim_head if out_dim is not None else heads
+        self.sliceable_head_dim = heads
+        self.single = single
+
+        linear_cls = operations.Linear
+        self.linear_cls = linear_cls
+        self.to_q = linear_cls(query_dim, self.inner_dim, dtype=dtype, device=device)
+        self.to_k = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+        self.to_v = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+        self.to_out = linear_cls(self.inner_dim, self.out_dim, dtype=dtype, device=device)
+        self.q_rms_norm = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+        self.k_rms_norm = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+
+        if not single:
+            self.to_q_t = linear_cls(query_dim, self.inner_dim, dtype=dtype, device=device)
+            self.to_k_t = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+            self.to_v_t = linear_cls(self.inner_dim, self.inner_dim, dtype=dtype, device=device)
+            self.to_out_t = linear_cls(self.inner_dim, self.out_dim, dtype=dtype, device=device)
+            self.q_rms_norm_t = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+            self.k_rms_norm_t = operations.RMSNorm(self.inner_dim, eps, dtype=dtype, device=device)
+
+        self.processor = processor
+
+    def forward(
+        self,
+        norm_image_tokens: torch.FloatTensor,
+        image_tokens_masks: torch.FloatTensor = None,
+        norm_text_tokens: torch.FloatTensor = None,
+        rope: torch.FloatTensor = None,
+    ) -> torch.Tensor:
+        return self.processor(
+            self,
+            image_tokens = norm_image_tokens,
+            image_tokens_masks = image_tokens_masks,
+            text_tokens = norm_text_tokens,
+            rope = rope,
+        )
+
+
+class FeedForwardSwiGLU(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        hidden_dim = int(2 * hidden_dim / 3)
+        # custom dim factor multiplier
+        if ffn_dim_multiplier is not None:
+            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+        hidden_dim = multiple_of * (
+            (hidden_dim + multiple_of - 1) // multiple_of
+        )
+
+        self.w1 = operations.Linear(dim, hidden_dim, bias=False, dtype=dtype, device=device)
+        self.w2 = operations.Linear(hidden_dim, dim, bias=False, dtype=dtype, device=device)
+        self.w3 = operations.Linear(dim, hidden_dim, bias=False, dtype=dtype, device=device)
+
+    def forward(self, x):
+        return self.w2(torch.nn.functional.silu(self.w1(x)) * self.w3(x))
+
+
+# Modified from https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py
+class MoEGate(nn.Module):
+    def __init__(self, embed_dim, num_routed_experts=4, num_activated_experts=2, aux_loss_alpha=0.01, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.top_k = num_activated_experts
+        self.n_routed_experts = num_routed_experts
+
+        self.scoring_func = 'softmax'
+        self.alpha = aux_loss_alpha
+        self.seq_aux = False
+
+        # topk selection algorithm
+        self.norm_topk_prob = False
+        self.gating_dim = embed_dim
+        self.weight = nn.Parameter(torch.empty((self.n_routed_experts, self.gating_dim), dtype=dtype, device=device))
+        self.reset_parameters()
+
+    def reset_parameters(self) -> None:
+        pass
+        # import torch.nn.init  as init
+        # init.kaiming_uniform_(self.weight, a=math.sqrt(5))
+
+    def forward(self, hidden_states):
+        bsz, seq_len, h = hidden_states.shape
+
+        ### compute gating score
+        hidden_states = hidden_states.view(-1, h)
+        logits = F.linear(hidden_states, comfy.model_management.cast_to(self.weight, dtype=hidden_states.dtype, device=hidden_states.device), None)
+        if self.scoring_func == 'softmax':
+            scores = logits.softmax(dim=-1)
+        else:
+            raise NotImplementedError(f'insupportable scoring function for MoE gating: {self.scoring_func}')
+
+        ### select top-k experts
+        topk_weight, topk_idx = torch.topk(scores, k=self.top_k, dim=-1, sorted=False)
+
+        ### norm gate to sum 1
+        if self.top_k > 1 and self.norm_topk_prob:
+            denominator = topk_weight.sum(dim=-1, keepdim=True) + 1e-20
+            topk_weight = topk_weight / denominator
+
+        aux_loss = None
+        return topk_idx, topk_weight, aux_loss
+
+
+# Modified from https://github.com/deepseek-ai/DeepSeek-V3/blob/main/inference/model.py
+class MOEFeedForwardSwiGLU(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        num_routed_experts: int,
+        num_activated_experts: int,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.shared_experts = FeedForwardSwiGLU(dim, hidden_dim // 2, dtype=dtype, device=device, operations=operations)
+        self.experts = nn.ModuleList([FeedForwardSwiGLU(dim, hidden_dim, dtype=dtype, device=device, operations=operations) for i in range(num_routed_experts)])
+        self.gate = MoEGate(
+            embed_dim = dim,
+            num_routed_experts = num_routed_experts,
+            num_activated_experts = num_activated_experts,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.num_activated_experts = num_activated_experts
+
+    def forward(self, x):
+        wtype = x.dtype
+        identity = x
+        orig_shape = x.shape
+        topk_idx, topk_weight, aux_loss = self.gate(x)
+        x = x.view(-1, x.shape[-1])
+        flat_topk_idx = topk_idx.view(-1)
+        if True:  # self.training: # TODO: check which branch performs faster
+            x = x.repeat_interleave(self.num_activated_experts, dim=0)
+            y = torch.empty_like(x, dtype=wtype)
+            for i, expert in enumerate(self.experts):
+                y[flat_topk_idx == i] = expert(x[flat_topk_idx == i]).to(dtype=wtype)
+            y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
+            y =  y.view(*orig_shape).to(dtype=wtype)
+            #y = AddAuxiliaryLoss.apply(y, aux_loss)
+        else:
+            y = self.moe_infer(x, flat_topk_idx, topk_weight.view(-1, 1)).view(*orig_shape)
+        y = y + self.shared_experts(identity)
+        return y
+
+    @torch.no_grad()
+    def moe_infer(self, x, flat_expert_indices, flat_expert_weights):
+        expert_cache = torch.zeros_like(x)
+        idxs = flat_expert_indices.argsort()
+        tokens_per_expert = flat_expert_indices.bincount().cpu().numpy().cumsum(0)
+        token_idxs = idxs // self.num_activated_experts
+        for i, end_idx in enumerate(tokens_per_expert):
+            start_idx = 0 if i == 0 else tokens_per_expert[i-1]
+            if start_idx == end_idx:
+                continue
+            expert = self.experts[i]
+            exp_token_idx = token_idxs[start_idx:end_idx]
+            expert_tokens = x[exp_token_idx]
+            expert_out = expert(expert_tokens)
+            expert_out.mul_(flat_expert_weights[idxs[start_idx:end_idx]])
+
+            # for fp16 and other dtype
+            expert_cache = expert_cache.to(expert_out.dtype)
+            expert_cache.scatter_reduce_(0, exp_token_idx.view(-1, 1).repeat(1, x.shape[-1]), expert_out, reduce='sum')
+        return expert_cache
+
+
+class TextProjection(nn.Module):
+    def __init__(self, in_features, hidden_size, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.linear = operations.Linear(in_features=in_features, out_features=hidden_size, bias=False, dtype=dtype, device=device)
+
+    def forward(self, caption):
+        hidden_states = self.linear(caption)
+        return hidden_states
+
+
+class BlockType:
+    TransformerBlock = 1
+    SingleTransformerBlock = 2
+
+
+class HiDreamImageSingleTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(dim, 6 * dim, bias=True, dtype=dtype, device=device)
+        )
+
+        # 1. Attention
+        self.norm1_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        self.attn1 = HiDreamAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            processor = HiDreamAttnProcessor_flashattn(),
+            single = True,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        # 3. Feed-forward
+        self.norm3_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        if num_routed_experts > 0:
+            self.ff_i = MOEFeedForwardSwiGLU(
+                dim = dim,
+                hidden_dim = 4 * dim,
+                num_routed_experts = num_routed_experts,
+                num_activated_experts = num_activated_experts,
+                dtype=dtype, device=device, operations=operations
+            )
+        else:
+            self.ff_i = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim, dtype=dtype, device=device, operations=operations)
+
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+
+    ) -> torch.FloatTensor:
+        wtype = image_tokens.dtype
+        shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i = \
+            self.adaLN_modulation(adaln_input)[:,None].chunk(6, dim=-1)
+
+        # 1. MM-Attention
+        norm_image_tokens = self.norm1_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_msa_i) + shift_msa_i
+        attn_output_i = self.attn1(
+            norm_image_tokens,
+            image_tokens_masks,
+            rope = rope,
+        )
+        image_tokens = gate_msa_i * attn_output_i + image_tokens
+
+        # 2. Feed-forward
+        norm_image_tokens = self.norm3_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_mlp_i) + shift_mlp_i
+        ff_output_i = gate_mlp_i * self.ff_i(norm_image_tokens.to(dtype=wtype))
+        image_tokens = ff_output_i + image_tokens
+        return image_tokens
+
+
+class HiDreamImageTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.num_attention_heads = num_attention_heads
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(dim, 12 * dim, bias=True, dtype=dtype, device=device)
+        )
+        # nn.init.zeros_(self.adaLN_modulation[1].weight)
+        # nn.init.zeros_(self.adaLN_modulation[1].bias)
+
+        # 1. Attention
+        self.norm1_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        self.norm1_t = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        self.attn1 = HiDreamAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            processor = HiDreamAttnProcessor_flashattn(),
+            single = False,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        # 3. Feed-forward
+        self.norm3_i = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False, dtype=dtype, device=device)
+        if num_routed_experts > 0:
+            self.ff_i = MOEFeedForwardSwiGLU(
+                dim = dim,
+                hidden_dim = 4 * dim,
+                num_routed_experts = num_routed_experts,
+                num_activated_experts = num_activated_experts,
+                dtype=dtype, device=device, operations=operations
+            )
+        else:
+            self.ff_i = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim, dtype=dtype, device=device, operations=operations)
+        self.norm3_t = operations.LayerNorm(dim, eps = 1e-06, elementwise_affine = False)
+        self.ff_t = FeedForwardSwiGLU(dim = dim, hidden_dim = 4 * dim, dtype=dtype, device=device, operations=operations)
+
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: Optional[torch.FloatTensor] = None,
+        rope: torch.FloatTensor = None,
+    ) -> torch.FloatTensor:
+        wtype = image_tokens.dtype
+        shift_msa_i, scale_msa_i, gate_msa_i, shift_mlp_i, scale_mlp_i, gate_mlp_i, \
+        shift_msa_t, scale_msa_t, gate_msa_t, shift_mlp_t, scale_mlp_t, gate_mlp_t = \
+            self.adaLN_modulation(adaln_input)[:,None].chunk(12, dim=-1)
+
+        # 1. MM-Attention
+        norm_image_tokens = self.norm1_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_msa_i) + shift_msa_i
+        norm_text_tokens = self.norm1_t(text_tokens).to(dtype=wtype)
+        norm_text_tokens = norm_text_tokens * (1 + scale_msa_t) + shift_msa_t
+
+        attn_output_i, attn_output_t = self.attn1(
+            norm_image_tokens,
+            image_tokens_masks,
+            norm_text_tokens,
+            rope = rope,
+        )
+
+        image_tokens = gate_msa_i * attn_output_i + image_tokens
+        text_tokens = gate_msa_t * attn_output_t + text_tokens
+
+        # 2. Feed-forward
+        norm_image_tokens = self.norm3_i(image_tokens).to(dtype=wtype)
+        norm_image_tokens = norm_image_tokens * (1 + scale_mlp_i) + shift_mlp_i
+        norm_text_tokens = self.norm3_t(text_tokens).to(dtype=wtype)
+        norm_text_tokens = norm_text_tokens * (1 + scale_mlp_t) + shift_mlp_t
+
+        ff_output_i = gate_mlp_i * self.ff_i(norm_image_tokens)
+        ff_output_t = gate_mlp_t * self.ff_t(norm_text_tokens)
+        image_tokens = ff_output_i + image_tokens
+        text_tokens = ff_output_t + text_tokens
+        return image_tokens, text_tokens
+
+
+class HiDreamImageBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        block_type: BlockType = BlockType.TransformerBlock,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        block_classes = {
+            BlockType.TransformerBlock: HiDreamImageTransformerBlock,
+            BlockType.SingleTransformerBlock: HiDreamImageSingleTransformerBlock,
+        }
+        self.block = block_classes[block_type](
+            dim,
+            num_attention_heads,
+            attention_head_dim,
+            num_routed_experts,
+            num_activated_experts,
+            dtype=dtype, device=device, operations=operations
+        )
+
+    def forward(
+        self,
+        image_tokens: torch.FloatTensor,
+        image_tokens_masks: Optional[torch.FloatTensor] = None,
+        text_tokens: Optional[torch.FloatTensor] = None,
+        adaln_input: torch.FloatTensor = None,
+        rope: torch.FloatTensor = None,
+    ) -> torch.FloatTensor:
+        return self.block(
+            image_tokens,
+            image_tokens_masks,
+            text_tokens,
+            adaln_input,
+            rope,
+        )
+
+
+class HiDreamImageTransformer2DModel(nn.Module):
+    def __init__(
+        self,
+        patch_size: Optional[int] = None,
+        in_channels: int = 64,
+        out_channels: Optional[int] = None,
+        num_layers: int = 16,
+        num_single_layers: int = 32,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 20,
+        caption_channels: List[int] = None,
+        text_emb_dim: int = 2048,
+        num_routed_experts: int = 4,
+        num_activated_experts: int = 2,
+        axes_dims_rope: Tuple[int, int] = (32, 32),
+        max_resolution: Tuple[int, int] = (128, 128),
+        llama_layers: List[int] = None,
+        image_model=None,
+        dtype=None, device=None, operations=None
+    ):
+        self.patch_size = patch_size
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.num_layers = num_layers
+        self.num_single_layers = num_single_layers
+
+        self.gradient_checkpointing = False
+
+        super().__init__()
+        self.dtype = dtype
+        self.out_channels = out_channels or in_channels
+        self.inner_dim = self.num_attention_heads * self.attention_head_dim
+        self.llama_layers = llama_layers
+
+        self.t_embedder = TimestepEmbed(self.inner_dim, dtype=dtype, device=device, operations=operations)
+        self.p_embedder = PooledEmbed(text_emb_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
+        self.x_embedder = PatchEmbed(
+            patch_size = patch_size,
+            in_channels = in_channels,
+            out_channels = self.inner_dim,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.pe_embedder = EmbedND(theta=10000, axes_dim=axes_dims_rope)
+
+        self.double_stream_blocks = nn.ModuleList(
+            [
+                HiDreamImageBlock(
+                    dim = self.inner_dim,
+                    num_attention_heads = self.num_attention_heads,
+                    attention_head_dim = self.attention_head_dim,
+                    num_routed_experts = num_routed_experts,
+                    num_activated_experts = num_activated_experts,
+                    block_type = BlockType.TransformerBlock,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for i in range(self.num_layers)
+            ]
+        )
+
+        self.single_stream_blocks = nn.ModuleList(
+            [
+                HiDreamImageBlock(
+                    dim = self.inner_dim,
+                    num_attention_heads = self.num_attention_heads,
+                    attention_head_dim = self.attention_head_dim,
+                    num_routed_experts = num_routed_experts,
+                    num_activated_experts = num_activated_experts,
+                    block_type = BlockType.SingleTransformerBlock,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for i in range(self.num_single_layers)
+            ]
+        )
+
+        self.final_layer = LastLayer(self.inner_dim, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations)
+
+        caption_channels = [caption_channels[1], ] * (num_layers + num_single_layers) + [caption_channels[0], ]
+        caption_projection = []
+        for caption_channel in caption_channels:
+            caption_projection.append(TextProjection(in_features=caption_channel, hidden_size=self.inner_dim, dtype=dtype, device=device, operations=operations))
+        self.caption_projection = nn.ModuleList(caption_projection)
+        self.max_seq = max_resolution[0] * max_resolution[1] // (patch_size * patch_size)
+
+    def expand_timesteps(self, timesteps, batch_size, device):
+        if not torch.is_tensor(timesteps):
+            is_mps = device.type == "mps"
+            if isinstance(timesteps, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(device)
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(batch_size)
+        return timesteps
+
+    def unpatchify(self, x: torch.Tensor, img_sizes: List[Tuple[int, int]]) -> List[torch.Tensor]:
+        x_arr = []
+        for i, img_size in enumerate(img_sizes):
+            pH, pW = img_size
+            x_arr.append(
+                einops.rearrange(x[i, :pH*pW].reshape(1, pH, pW, -1), 'B H W (p1 p2 C) -> B C (H p1) (W p2)',
+                    p1=self.patch_size, p2=self.patch_size)
+            )
+        x = torch.cat(x_arr, dim=0)
+        return x
+
+    def patchify(self, x, max_seq, img_sizes=None):
+        pz2 = self.patch_size * self.patch_size
+        if isinstance(x, torch.Tensor):
+            B = x.shape[0]
+            device = x.device
+            dtype = x.dtype
+        else:
+            B = len(x)
+            device = x[0].device
+            dtype = x[0].dtype
+        x_masks = torch.zeros((B, max_seq), dtype=dtype, device=device)
+
+        if img_sizes is not None:
+            for i, img_size in enumerate(img_sizes):
+                x_masks[i, 0:img_size[0] * img_size[1]] = 1
+            x = einops.rearrange(x, 'B C S p -> B S (p C)', p=pz2)
+        elif isinstance(x, torch.Tensor):
+            pH, pW = x.shape[-2] // self.patch_size, x.shape[-1] // self.patch_size
+            x = einops.rearrange(x, 'B C (H p1) (W p2) -> B (H W) (p1 p2 C)', p1=self.patch_size, p2=self.patch_size)
+            img_sizes = [[pH, pW]] * B
+            x_masks = None
+        else:
+            raise NotImplementedError
+        return x, x_masks, img_sizes
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        t: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        encoder_hidden_states_llama3=None,
+        image_cond=None,
+        control = None,
+        transformer_options = {},
+    ) -> torch.Tensor:
+        bs, c, h, w = x.shape
+        if image_cond is not None:
+            x = torch.cat([x, image_cond], dim=-1)
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+        timesteps = t
+        pooled_embeds = y
+        T5_encoder_hidden_states = context
+
+        img_sizes = None
+
+        # spatial forward
+        batch_size = hidden_states.shape[0]
+        hidden_states_type = hidden_states.dtype
+
+        # 0. time
+        timesteps = self.expand_timesteps(timesteps, batch_size, hidden_states.device)
+        timesteps = self.t_embedder(timesteps, hidden_states_type)
+        p_embedder = self.p_embedder(pooled_embeds)
+        adaln_input = timesteps + p_embedder
+
+        hidden_states, image_tokens_masks, img_sizes = self.patchify(hidden_states, self.max_seq, img_sizes)
+        if image_tokens_masks is None:
+            pH, pW = img_sizes[0]
+            img_ids = torch.zeros(pH, pW, 3, device=hidden_states.device)
+            img_ids[..., 1] = img_ids[..., 1] + torch.arange(pH, device=hidden_states.device)[:, None]
+            img_ids[..., 2] = img_ids[..., 2] + torch.arange(pW, device=hidden_states.device)[None, :]
+            img_ids = repeat(img_ids, "h w c -> b (h w) c", b=batch_size)
+        hidden_states = self.x_embedder(hidden_states)
+
+        # T5_encoder_hidden_states = encoder_hidden_states[0]
+        encoder_hidden_states = encoder_hidden_states_llama3.movedim(1, 0)
+        encoder_hidden_states = [encoder_hidden_states[k] for k in self.llama_layers]
+
+        if self.caption_projection is not None:
+            new_encoder_hidden_states = []
+            for i, enc_hidden_state in enumerate(encoder_hidden_states):
+                enc_hidden_state = self.caption_projection[i](enc_hidden_state)
+                enc_hidden_state = enc_hidden_state.view(batch_size, -1, hidden_states.shape[-1])
+                new_encoder_hidden_states.append(enc_hidden_state)
+            encoder_hidden_states = new_encoder_hidden_states
+            T5_encoder_hidden_states = self.caption_projection[-1](T5_encoder_hidden_states)
+            T5_encoder_hidden_states = T5_encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+            encoder_hidden_states.append(T5_encoder_hidden_states)
+
+        txt_ids = torch.zeros(
+            batch_size,
+            encoder_hidden_states[-1].shape[1] + encoder_hidden_states[-2].shape[1] + encoder_hidden_states[0].shape[1],
+            3,
+            device=img_ids.device, dtype=img_ids.dtype
+        )
+        ids = torch.cat((img_ids, txt_ids), dim=1)
+        rope = self.pe_embedder(ids)
+
+        # 2. Blocks
+        block_id = 0
+        initial_encoder_hidden_states = torch.cat([encoder_hidden_states[-1], encoder_hidden_states[-2]], dim=1)
+        initial_encoder_hidden_states_seq_len = initial_encoder_hidden_states.shape[1]
+        for bid, block in enumerate(self.double_stream_blocks):
+            cur_llama31_encoder_hidden_states = encoder_hidden_states[block_id]
+            cur_encoder_hidden_states = torch.cat([initial_encoder_hidden_states, cur_llama31_encoder_hidden_states], dim=1)
+            hidden_states, initial_encoder_hidden_states = block(
+                image_tokens = hidden_states,
+                image_tokens_masks = image_tokens_masks,
+                text_tokens = cur_encoder_hidden_states,
+                adaln_input = adaln_input,
+                rope = rope,
+            )
+            initial_encoder_hidden_states = initial_encoder_hidden_states[:, :initial_encoder_hidden_states_seq_len]
+            block_id += 1
+
+        image_tokens_seq_len = hidden_states.shape[1]
+        hidden_states = torch.cat([hidden_states, initial_encoder_hidden_states], dim=1)
+        hidden_states_seq_len = hidden_states.shape[1]
+        if image_tokens_masks is not None:
+            encoder_attention_mask_ones = torch.ones(
+                (batch_size, initial_encoder_hidden_states.shape[1] + cur_llama31_encoder_hidden_states.shape[1]),
+                device=image_tokens_masks.device, dtype=image_tokens_masks.dtype
+            )
+            image_tokens_masks = torch.cat([image_tokens_masks, encoder_attention_mask_ones], dim=1)
+
+        for bid, block in enumerate(self.single_stream_blocks):
+            cur_llama31_encoder_hidden_states = encoder_hidden_states[block_id]
+            hidden_states = torch.cat([hidden_states, cur_llama31_encoder_hidden_states], dim=1)
+            hidden_states = block(
+                image_tokens=hidden_states,
+                image_tokens_masks=image_tokens_masks,
+                text_tokens=None,
+                adaln_input=adaln_input,
+                rope=rope,
+            )
+            hidden_states = hidden_states[:, :hidden_states_seq_len]
+            block_id += 1
+
+        hidden_states = hidden_states[:, :image_tokens_seq_len, ...]
+        output = self.final_layer(hidden_states, adaln_input)
+        output = self.unpatchify(output, img_sizes)
+        return -output[:, :, :h, :w]

comfy/ldm/hunyuan3d/model.py

@@ -0,0 +1,135 @@
+import torch
+from torch import nn
+from comfy.ldm.flux.layers import (
+    DoubleStreamBlock,
+    LastLayer,
+    MLPEmbedder,
+    SingleStreamBlock,
+    timestep_embedding,
+)
+
+
+class Hunyuan3Dv2(nn.Module):
+    def __init__(
+        self,
+        in_channels=64,
+        context_in_dim=1536,
+        hidden_size=1024,
+        mlp_ratio=4.0,
+        num_heads=16,
+        depth=16,
+        depth_single_blocks=32,
+        qkv_bias=True,
+        guidance_embed=False,
+        image_model=None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.dtype = dtype
+
+        if hidden_size % num_heads != 0:
+            raise ValueError(
+                f"Hidden size {hidden_size} must be divisible by num_heads {num_heads}"
+            )
+
+        self.max_period = 1000  # While reimplementing the model I noticed that they messed up. This 1000 value was meant to be the time_factor but they set the max_period instead
+        self.latent_in = operations.Linear(in_channels, hidden_size, bias=True, dtype=dtype, device=device)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=hidden_size, dtype=dtype, device=device, operations=operations)
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=hidden_size, dtype=dtype, device=device, operations=operations) if guidance_embed else None
+        )
+        self.cond_in = operations.Linear(context_in_dim, hidden_size, dtype=dtype, device=device)
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(
+                    hidden_size,
+                    num_heads,
+                    mlp_ratio=mlp_ratio,
+                    dtype=dtype, device=device, operations=operations
+                )
+                for _ in range(depth_single_blocks)
+            ]
+        )
+        self.final_layer = LastLayer(hidden_size, 1, in_channels, dtype=dtype, device=device, operations=operations)
+
+    def forward(self, x, timestep, context, guidance=None, transformer_options={}, **kwargs):
+        x = x.movedim(-1, -2)
+        timestep = 1.0 - timestep
+        txt = context
+        img = self.latent_in(x)
+
+        vec = self.time_in(timestep_embedding(timestep, 256, self.max_period).to(dtype=img.dtype))
+        if self.guidance_in is not None:
+            if guidance is not None:
+                vec = vec + self.guidance_in(timestep_embedding(guidance, 256, self.max_period).to(img.dtype))
+
+        txt = self.cond_in(txt)
+        pe = None
+        attn_mask = None
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.double_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"], out["txt"] = block(img=args["img"],
+                                                   txt=args["txt"],
+                                                   vec=args["vec"],
+                                                   pe=args["pe"],
+                                                   attn_mask=args.get("attn_mask"))
+                    return out
+
+                out = blocks_replace[("double_block", i)]({"img": img,
+                                                           "txt": txt,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask},
+                                                          {"original_block": block_wrap})
+                txt = out["txt"]
+                img = out["img"]
+            else:
+                img, txt = block(img=img,
+                                 txt=txt,
+                                 vec=vec,
+                                 pe=pe,
+                                 attn_mask=attn_mask)
+
+        img = torch.cat((txt, img), 1)
+
+        for i, block in enumerate(self.single_blocks):
+            if ("single_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"],
+                                       vec=args["vec"],
+                                       pe=args["pe"],
+                                       attn_mask=args.get("attn_mask"))
+                    return out
+
+                out = blocks_replace[("single_block", i)]({"img": img,
+                                                           "vec": vec,
+                                                           "pe": pe,
+                                                           "attn_mask": attn_mask},
+                                                          {"original_block": block_wrap})
+                img = out["img"]
+            else:
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
+
+        img = img[:, txt.shape[1]:, ...]
+        img = self.final_layer(img, vec)
+        return img.movedim(-2, -1) * (-1.0)

comfy/ldm/hunyuan3d/vae.py

@@ -0,0 +1,587 @@
+# Original: https://github.com/Tencent/Hunyuan3D-2/blob/main/hy3dgen/shapegen/models/autoencoders/model.py
+# Since the header on their VAE source file was a bit confusing we asked for permission to use this code from tencent under the GPL license used in ComfyUI.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+from typing import Union, Tuple, List, Callable, Optional
+
+import numpy as np
+from einops import repeat, rearrange
+from tqdm import tqdm
+import logging
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+def generate_dense_grid_points(
+    bbox_min: np.ndarray,
+    bbox_max: np.ndarray,
+    octree_resolution: int,
+    indexing: str = "ij",
+):
+    length = bbox_max - bbox_min
+    num_cells = octree_resolution
+
+    x = np.linspace(bbox_min[0], bbox_max[0], int(num_cells) + 1, dtype=np.float32)
+    y = np.linspace(bbox_min[1], bbox_max[1], int(num_cells) + 1, dtype=np.float32)
+    z = np.linspace(bbox_min[2], bbox_max[2], int(num_cells) + 1, dtype=np.float32)
+    [xs, ys, zs] = np.meshgrid(x, y, z, indexing=indexing)
+    xyz = np.stack((xs, ys, zs), axis=-1)
+    grid_size = [int(num_cells) + 1, int(num_cells) + 1, int(num_cells) + 1]
+
+    return xyz, grid_size, length
+
+
+class VanillaVolumeDecoder:
+    @torch.no_grad()
+    def __call__(
+        self,
+        latents: torch.FloatTensor,
+        geo_decoder: Callable,
+        bounds: Union[Tuple[float], List[float], float] = 1.01,
+        num_chunks: int = 10000,
+        octree_resolution: int = None,
+        enable_pbar: bool = True,
+        **kwargs,
+    ):
+        device = latents.device
+        dtype = latents.dtype
+        batch_size = latents.shape[0]
+
+        # 1. generate query points
+        if isinstance(bounds, float):
+            bounds = [-bounds, -bounds, -bounds, bounds, bounds, bounds]
+
+        bbox_min, bbox_max = np.array(bounds[0:3]), np.array(bounds[3:6])
+        xyz_samples, grid_size, length = generate_dense_grid_points(
+            bbox_min=bbox_min,
+            bbox_max=bbox_max,
+            octree_resolution=octree_resolution,
+            indexing="ij"
+        )
+        xyz_samples = torch.from_numpy(xyz_samples).to(device, dtype=dtype).contiguous().reshape(-1, 3)
+
+        # 2. latents to 3d volume
+        batch_logits = []
+        for start in tqdm(range(0, xyz_samples.shape[0], num_chunks), desc="Volume Decoding",
+                          disable=not enable_pbar):
+            chunk_queries = xyz_samples[start: start + num_chunks, :]
+            chunk_queries = repeat(chunk_queries, "p c -> b p c", b=batch_size)
+            logits = geo_decoder(queries=chunk_queries, latents=latents)
+            batch_logits.append(logits)
+
+        grid_logits = torch.cat(batch_logits, dim=1)
+        grid_logits = grid_logits.view((batch_size, *grid_size)).float()
+
+        return grid_logits
+
+
+class FourierEmbedder(nn.Module):
+    """The sin/cosine positional embedding. Given an input tensor `x` of shape [n_batch, ..., c_dim], it converts
+    each feature dimension of `x[..., i]` into:
+        [
+            sin(x[..., i]),
+            sin(f_1*x[..., i]),
+            sin(f_2*x[..., i]),
+            ...
+            sin(f_N * x[..., i]),
+            cos(x[..., i]),
+            cos(f_1*x[..., i]),
+            cos(f_2*x[..., i]),
+            ...
+            cos(f_N * x[..., i]),
+            x[..., i]     # only present if include_input is True.
+        ], here f_i is the frequency.
+
+    Denote the space is [0 / num_freqs, 1 / num_freqs, 2 / num_freqs, 3 / num_freqs, ..., (num_freqs - 1) / num_freqs].
+    If logspace is True, then the frequency f_i is [2^(0 / num_freqs), ..., 2^(i / num_freqs), ...];
+    Otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)].
+
+    Args:
+        num_freqs (int): the number of frequencies, default is 6;
+        logspace (bool): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+            otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)];
+        input_dim (int): the input dimension, default is 3;
+        include_input (bool): include the input tensor or not, default is True.
+
+    Attributes:
+        frequencies (torch.Tensor): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+                otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1);
+
+        out_dim (int): the embedding size, if include_input is True, it is input_dim * (num_freqs * 2 + 1),
+            otherwise, it is input_dim * num_freqs * 2.
+
+    """
+
+    def __init__(self,
+                 num_freqs: int = 6,
+                 logspace: bool = True,
+                 input_dim: int = 3,
+                 include_input: bool = True,
+                 include_pi: bool = True) -> None:
+
+        """The initialization"""
+
+        super().__init__()
+
+        if logspace:
+            frequencies = 2.0 ** torch.arange(
+                num_freqs,
+                dtype=torch.float32
+            )
+        else:
+            frequencies = torch.linspace(
+                1.0,
+                2.0 ** (num_freqs - 1),
+                num_freqs,
+                dtype=torch.float32
+            )
+
+        if include_pi:
+            frequencies *= torch.pi
+
+        self.register_buffer("frequencies", frequencies, persistent=False)
+        self.include_input = include_input
+        self.num_freqs = num_freqs
+
+        self.out_dim = self.get_dims(input_dim)
+
+    def get_dims(self, input_dim):
+        temp = 1 if self.include_input or self.num_freqs == 0 else 0
+        out_dim = input_dim * (self.num_freqs * 2 + temp)
+
+        return out_dim
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """ Forward process.
+
+        Args:
+            x: tensor of shape [..., dim]
+
+        Returns:
+            embedding: an embedding of `x` of shape [..., dim * (num_freqs * 2 + temp)]
+                where temp is 1 if include_input is True and 0 otherwise.
+        """
+
+        if self.num_freqs > 0:
+            embed = (x[..., None].contiguous() * self.frequencies.to(device=x.device, dtype=x.dtype)).view(*x.shape[:-1], -1)
+            if self.include_input:
+                return torch.cat((x, embed.sin(), embed.cos()), dim=-1)
+            else:
+                return torch.cat((embed.sin(), embed.cos()), dim=-1)
+        else:
+            return x
+
+
+class CrossAttentionProcessor:
+    def __call__(self, attn, q, k, v):
+        out = F.scaled_dot_product_attention(q, k, v)
+        return out
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+    """
+
+    def __init__(self, drop_prob: float = 0., scale_by_keep: bool = True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+        This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
+        the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+        See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
+        changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
+        'survival rate' as the argument.
+
+        """
+        if self.drop_prob == 0. or not self.training:
+            return x
+        keep_prob = 1 - self.drop_prob
+        shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+        if keep_prob > 0.0 and self.scale_by_keep:
+            random_tensor.div_(keep_prob)
+        return x * random_tensor
+
+    def extra_repr(self):
+        return f'drop_prob={round(self.drop_prob, 3):0.3f}'
+
+
+class MLP(nn.Module):
+    def __init__(
+        self, *,
+        width: int,
+        expand_ratio: int = 4,
+        output_width: int = None,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.c_fc = ops.Linear(width, width * expand_ratio)
+        self.c_proj = ops.Linear(width * expand_ratio, output_width if output_width is not None else width)
+        self.gelu = nn.GELU()
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(self, x):
+        return self.drop_path(self.c_proj(self.gelu(self.c_fc(x))))
+
+
+class QKVMultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        heads: int,
+        width=None,
+        qk_norm=False,
+        norm_layer=ops.LayerNorm
+    ):
+        super().__init__()
+        self.heads = heads
+        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+
+        self.attn_processor = CrossAttentionProcessor()
+
+    def forward(self, q, kv):
+        _, n_ctx, _ = q.shape
+        bs, n_data, width = kv.shape
+        attn_ch = width // self.heads // 2
+        q = q.view(bs, n_ctx, self.heads, -1)
+        kv = kv.view(bs, n_data, self.heads, -1)
+        k, v = torch.split(kv, attn_ch, dim=-1)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
+        out = self.attn_processor(self, q, k, v)
+        out = out.transpose(1, 2).reshape(bs, n_ctx, -1)
+        return out
+
+
+class MultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        data_width: Optional[int] = None,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        kv_cache: bool = False,
+    ):
+        super().__init__()
+        self.width = width
+        self.heads = heads
+        self.data_width = width if data_width is None else data_width
+        self.c_q = ops.Linear(width, width, bias=qkv_bias)
+        self.c_kv = ops.Linear(self.data_width, width * 2, bias=qkv_bias)
+        self.c_proj = ops.Linear(width, width)
+        self.attention = QKVMultiheadCrossAttention(
+            heads=heads,
+            width=width,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.kv_cache = kv_cache
+        self.data = None
+
+    def forward(self, x, data):
+        x = self.c_q(x)
+        if self.kv_cache:
+            if self.data is None:
+                self.data = self.c_kv(data)
+                logging.info('Save kv cache,this should be called only once for one mesh')
+            data = self.data
+        else:
+            data = self.c_kv(data)
+        x = self.attention(x, data)
+        x = self.c_proj(x)
+        return x
+
+
+class ResidualCrossAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        mlp_expand_ratio: int = 4,
+        data_width: Optional[int] = None,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False
+    ):
+        super().__init__()
+
+        if data_width is None:
+            data_width = width
+
+        self.attn = MultiheadCrossAttention(
+            width=width,
+            heads=heads,
+            data_width=data_width,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.ln_2 = norm_layer(data_width, elementwise_affine=True, eps=1e-6)
+        self.ln_3 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(width=width, expand_ratio=mlp_expand_ratio)
+
+    def forward(self, x: torch.Tensor, data: torch.Tensor):
+        x = x + self.attn(self.ln_1(x), self.ln_2(data))
+        x = x + self.mlp(self.ln_3(x))
+        return x
+
+
+class QKVMultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        heads: int,
+        width=None,
+        qk_norm=False,
+        norm_layer=ops.LayerNorm
+    ):
+        super().__init__()
+        self.heads = heads
+        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
+
+    def forward(self, qkv):
+        bs, n_ctx, width = qkv.shape
+        attn_ch = width // self.heads // 3
+        qkv = qkv.view(bs, n_ctx, self.heads, -1)
+        q, k, v = torch.split(qkv, attn_ch, dim=-1)
+
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
+        out = F.scaled_dot_product_attention(q, k, v).transpose(1, 2).reshape(bs, n_ctx, -1)
+        return out
+
+
+class MultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.heads = heads
+        self.c_qkv = ops.Linear(width, width * 3, bias=qkv_bias)
+        self.c_proj = ops.Linear(width, width)
+        self.attention = QKVMultiheadAttention(
+            heads=heads,
+            width=width,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm
+        )
+        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
+
+    def forward(self, x):
+        x = self.c_qkv(x)
+        x = self.attention(x)
+        x = self.drop_path(self.c_proj(x))
+        return x
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0,
+    ):
+        super().__init__()
+        self.attn = MultiheadAttention(
+            width=width,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            norm_layer=norm_layer,
+            qk_norm=qk_norm,
+            drop_path_rate=drop_path_rate
+        )
+        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+        self.mlp = MLP(width=width, drop_path_rate=drop_path_rate)
+        self.ln_2 = norm_layer(width, elementwise_affine=True, eps=1e-6)
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        *,
+        width: int,
+        layers: int,
+        heads: int,
+        qkv_bias: bool = True,
+        norm_layer=ops.LayerNorm,
+        qk_norm: bool = False,
+        drop_path_rate: float = 0.0
+    ):
+        super().__init__()
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    width=width,
+                    heads=heads,
+                    qkv_bias=qkv_bias,
+                    norm_layer=norm_layer,
+                    qk_norm=qk_norm,
+                    drop_path_rate=drop_path_rate
+                )
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor):
+        for block in self.resblocks:
+            x = block(x)
+        return x
+
+
+class CrossAttentionDecoder(nn.Module):
+
+    def __init__(
+        self,
+        *,
+        out_channels: int,
+        fourier_embedder: FourierEmbedder,
+        width: int,
+        heads: int,
+        mlp_expand_ratio: int = 4,
+        downsample_ratio: int = 1,
+        enable_ln_post: bool = True,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        label_type: str = "binary"
+    ):
+        super().__init__()
+
+        self.enable_ln_post = enable_ln_post
+        self.fourier_embedder = fourier_embedder
+        self.downsample_ratio = downsample_ratio
+        self.query_proj = ops.Linear(self.fourier_embedder.out_dim, width)
+        if self.downsample_ratio != 1:
+            self.latents_proj = ops.Linear(width * downsample_ratio, width)
+        if self.enable_ln_post == False:
+            qk_norm = False
+        self.cross_attn_decoder = ResidualCrossAttentionBlock(
+            width=width,
+            mlp_expand_ratio=mlp_expand_ratio,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm
+        )
+
+        if self.enable_ln_post:
+            self.ln_post = ops.LayerNorm(width)
+        self.output_proj = ops.Linear(width, out_channels)
+        self.label_type = label_type
+        self.count = 0
+
+    def forward(self, queries=None, query_embeddings=None, latents=None):
+        if query_embeddings is None:
+            query_embeddings = self.query_proj(self.fourier_embedder(queries).to(latents.dtype))
+        self.count += query_embeddings.shape[1]
+        if self.downsample_ratio != 1:
+            latents = self.latents_proj(latents)
+        x = self.cross_attn_decoder(query_embeddings, latents)
+        if self.enable_ln_post:
+            x = self.ln_post(x)
+        occ = self.output_proj(x)
+        return occ
+
+
+class ShapeVAE(nn.Module):
+    def __init__(
+        self,
+        *,
+        embed_dim: int,
+        width: int,
+        heads: int,
+        num_decoder_layers: int,
+        geo_decoder_downsample_ratio: int = 1,
+        geo_decoder_mlp_expand_ratio: int = 4,
+        geo_decoder_ln_post: bool = True,
+        num_freqs: int = 8,
+        include_pi: bool = True,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        label_type: str = "binary",
+        drop_path_rate: float = 0.0,
+        scale_factor: float = 1.0,
+    ):
+        super().__init__()
+        self.geo_decoder_ln_post = geo_decoder_ln_post
+
+        self.fourier_embedder = FourierEmbedder(num_freqs=num_freqs, include_pi=include_pi)
+
+        self.post_kl = ops.Linear(embed_dim, width)
+
+        self.transformer = Transformer(
+            width=width,
+            layers=num_decoder_layers,
+            heads=heads,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            drop_path_rate=drop_path_rate
+        )
+
+        self.geo_decoder = CrossAttentionDecoder(
+            fourier_embedder=self.fourier_embedder,
+            out_channels=1,
+            mlp_expand_ratio=geo_decoder_mlp_expand_ratio,
+            downsample_ratio=geo_decoder_downsample_ratio,
+            enable_ln_post=self.geo_decoder_ln_post,
+            width=width // geo_decoder_downsample_ratio,
+            heads=heads // geo_decoder_downsample_ratio,
+            qkv_bias=qkv_bias,
+            qk_norm=qk_norm,
+            label_type=label_type,
+        )
+
+        self.volume_decoder = VanillaVolumeDecoder()
+        self.scale_factor = scale_factor
+
+    def decode(self, latents, **kwargs):
+        latents = self.post_kl(latents.movedim(-2, -1))
+        latents = self.transformer(latents)
+
+        bounds = kwargs.get("bounds", 1.01)
+        num_chunks = kwargs.get("num_chunks", 8000)
+        octree_resolution = kwargs.get("octree_resolution", 256)
+        enable_pbar = kwargs.get("enable_pbar", True)
+
+        grid_logits = self.volume_decoder(latents, self.geo_decoder, bounds=bounds, num_chunks=num_chunks, octree_resolution=octree_resolution, enable_pbar=enable_pbar)
+        return grid_logits.movedim(-2, -1)
+
+    def encode(self, x):
+        return None

comfy/ldm/hunyuan_video/model.py

@@ -227,6 +227,8 @@ class HunyuanVideo(nn.Module):
         timesteps: Tensor,
         y: Tensor,
         guidance: Tensor = None,
+        guiding_frame_index=None,
+        ref_latent=None,
         control=None,
         transformer_options={},
     ) -> Tensor:
@@ -237,11 +239,28 @@ class HunyuanVideo(nn.Module):
         img = self.img_in(img)
         vec = self.time_in(timestep_embedding(timesteps, 256, time_factor=1.0).to(img.dtype))
 
+        if ref_latent is not None:
+            ref_latent_ids = self.img_ids(ref_latent)
+            ref_latent = self.img_in(ref_latent)
+            img = torch.cat([ref_latent, img], dim=-2)
+            ref_latent_ids[..., 0] = -1
+            ref_latent_ids[..., 2] += (initial_shape[-1] // self.patch_size[-1])
+            img_ids = torch.cat([ref_latent_ids, img_ids], dim=-2)
+
+        if guiding_frame_index is not None:
+            token_replace_vec = self.time_in(timestep_embedding(guiding_frame_index, 256, time_factor=1.0))
+            vec_ = self.vector_in(y[:, :self.params.vec_in_dim])
+            vec = torch.cat([(vec_ + token_replace_vec).unsqueeze(1), (vec_ + vec).unsqueeze(1)], dim=1)
+            frame_tokens = (initial_shape[-1] // self.patch_size[-1]) * (initial_shape[-2] // self.patch_size[-2])
+            modulation_dims = [(0, frame_tokens, 0), (frame_tokens, None, 1)]
+            modulation_dims_txt = [(0, None, 1)]
+        else:
             vec = vec + self.vector_in(y[:, :self.params.vec_in_dim])
+            modulation_dims = None
+            modulation_dims_txt = None
 
         if self.params.guidance_embed:
-            if guidance is None:
-                raise ValueError("Didn't get guidance strength for guidance distilled model.")
+            if guidance is not None:
                 vec = vec + self.guidance_in(timestep_embedding(guidance, 256).to(img.dtype))
 
         if txt_mask is not None and not torch.is_floating_point(txt_mask):
@@ -265,14 +284,14 @@ class HunyuanVideo(nn.Module):
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"], out["txt"] = block(img=args["img"], txt=args["txt"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"])
+                    out["img"], out["txt"] = block(img=args["img"], txt=args["txt"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"], modulation_dims_img=args["modulation_dims_img"], modulation_dims_txt=args["modulation_dims_txt"])
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": img, "txt": txt, "vec": vec, "pe": pe, "attention_mask": attn_mask}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": img, "txt": txt, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims_img': modulation_dims, 'modulation_dims_txt': modulation_dims_txt}, {"original_block": block_wrap})
                 txt = out["txt"]
                 img = out["img"]
             else:
-                img, txt = block(img=img, txt=txt, vec=vec, pe=pe, attn_mask=attn_mask)
+                img, txt = block(img=img, txt=txt, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims_img=modulation_dims, modulation_dims_txt=modulation_dims_txt)
 
             if control is not None: # Controlnet
                 control_i = control.get("input")
@@ -287,13 +306,13 @@ class HunyuanVideo(nn.Module):
             if ("single_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"])
+                    out["img"] = block(args["img"], vec=args["vec"], pe=args["pe"], attn_mask=args["attention_mask"], modulation_dims=args["modulation_dims"])
                     return out
 
-                out = blocks_replace[("single_block", i)]({"img": img, "vec": vec, "pe": pe, "attention_mask": attn_mask}, {"original_block": block_wrap})
+                out = blocks_replace[("single_block", i)]({"img": img, "vec": vec, "pe": pe, "attention_mask": attn_mask, 'modulation_dims': modulation_dims}, {"original_block": block_wrap})
                 img = out["img"]
             else:
-                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask)
+                img = block(img, vec=vec, pe=pe, attn_mask=attn_mask, modulation_dims=modulation_dims)
 
             if control is not None: # Controlnet
                 control_o = control.get("output")
@@ -303,18 +322,20 @@ class HunyuanVideo(nn.Module):
                         img[:, : img_len] += add
 
         img = img[:, : img_len]
+        if ref_latent is not None:
+            img = img[:, ref_latent.shape[1]:]
 
-        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+        img = self.final_layer(img, vec, modulation_dims=modulation_dims)  # (N, T, patch_size ** 2 * out_channels)
 
         shape = initial_shape[-3:]
         for i in range(len(shape)):
             shape[i] = shape[i] // self.patch_size[i]
         img = img.reshape([img.shape[0]] + shape + [self.out_channels] + self.patch_size)
         img = img.permute(0, 4, 1, 5, 2, 6, 3, 7)
-        img = img.reshape(initial_shape)
+        img = img.reshape(initial_shape[0], self.out_channels, initial_shape[2], initial_shape[3], initial_shape[4])
         return img
 
-    def forward(self, x, timestep, context, y, guidance, attention_mask=None, control=None, transformer_options={}, **kwargs):
+    def img_ids(self, x):
         bs, c, t, h, w = x.shape
         patch_size = self.patch_size
         t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
@@ -324,7 +345,11 @@ class HunyuanVideo(nn.Module):
         img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
         img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)
         img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
-        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+        return repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+
+    def forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, control=None, transformer_options={}, **kwargs):
+        bs, c, t, h, w = x.shape
+        img_ids = self.img_ids(x)
         txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)
-        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, guidance, control, transformer_options)
+        out = self.forward_orig(x, img_ids, context, txt_ids, attention_mask, timestep, y, guidance, guiding_frame_index, ref_latent, control=control, transformer_options=transformer_options)
         return out

comfy/ldm/hydit/models.py

@@ -3,7 +3,7 @@ import torch
 import torch.nn as nn
 
 import comfy.ops
-from comfy.ldm.modules.diffusionmodules.mmdit import Mlp, TimestepEmbedder, PatchEmbed, RMSNorm
+from comfy.ldm.modules.diffusionmodules.mmdit import Mlp, TimestepEmbedder, PatchEmbed
 from comfy.ldm.modules.diffusionmodules.util import timestep_embedding
 from torch.utils import checkpoint
 
@@ -51,7 +51,7 @@ class HunYuanDiTBlock(nn.Module):
         if norm_type == "layer":
             norm_layer = operations.LayerNorm
         elif norm_type == "rms":
-            norm_layer = RMSNorm
+            norm_layer = operations.RMSNorm
         else:
             raise ValueError(f"Unknown norm_type: {norm_type}")
 

comfy/ldm/lightricks/model.py

@@ -1,13 +1,12 @@
 import torch
 from torch import nn
 import comfy.ldm.modules.attention
-from comfy.ldm.genmo.joint_model.layers import RMSNorm
 import comfy.ldm.common_dit
 from einops import rearrange
 import math
 from typing import Dict, Optional, Tuple
 
-from .symmetric_patchifier import SymmetricPatchifier
+from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords
 
 
 def get_timestep_embedding(
@@ -262,8 +261,8 @@ class CrossAttention(nn.Module):
         self.heads = heads
         self.dim_head = dim_head
 
-        self.q_norm = RMSNorm(inner_dim, dtype=dtype, device=device)
-        self.k_norm = RMSNorm(inner_dim, dtype=dtype, device=device)
+        self.q_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device)
+        self.k_norm = operations.RMSNorm(inner_dim, dtype=dtype, device=device)
 
         self.to_q = operations.Linear(query_dim, inner_dim, bias=True, dtype=dtype, device=device)
         self.to_k = operations.Linear(context_dim, inner_dim, bias=True, dtype=dtype, device=device)
@@ -377,12 +376,16 @@ class LTXVModel(torch.nn.Module):
 
                  positional_embedding_theta=10000.0,
                  positional_embedding_max_pos=[20, 2048, 2048],
+                 causal_temporal_positioning=False,
+                 vae_scale_factors=(8, 32, 32),
                  dtype=None, device=None, operations=None, **kwargs):
         super().__init__()
         self.generator = None
+        self.vae_scale_factors = vae_scale_factors
         self.dtype = dtype
         self.out_channels = in_channels
         self.inner_dim = num_attention_heads * attention_head_dim
+        self.causal_temporal_positioning = causal_temporal_positioning
 
         self.patchify_proj = operations.Linear(in_channels, self.inner_dim, bias=True, dtype=dtype, device=device)
 
@@ -416,42 +419,23 @@ class LTXVModel(torch.nn.Module):
 
         self.patchifier = SymmetricPatchifier(1)
 
-    def forward(self, x, timestep, context, attention_mask, frame_rate=25, guiding_latent=None, guiding_latent_noise_scale=0, transformer_options={}, **kwargs):
+    def forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
         patches_replace = transformer_options.get("patches_replace", {})
 
-        indices_grid = self.patchifier.get_grid(
-            orig_num_frames=x.shape[2],
-            orig_height=x.shape[3],
-            orig_width=x.shape[4],
-            batch_size=x.shape[0],
-            scale_grid=((1 / frame_rate) * 8, 32, 32),
-            device=x.device,
-        )
-
-        if guiding_latent is not None:
-            ts = torch.ones([x.shape[0], 1, x.shape[2], x.shape[3], x.shape[4]], device=x.device, dtype=x.dtype)
-            input_ts = timestep.view([timestep.shape[0]] + [1] * (x.ndim - 1))
-            ts *= input_ts
-            ts[:, :, 0] = guiding_latent_noise_scale * (input_ts[:, :, 0] ** 2)
-            timestep = self.patchifier.patchify(ts)
-            input_x = x.clone()
-            x[:, :, 0] = guiding_latent[:, :, 0]
-            if guiding_latent_noise_scale > 0:
-                if self.generator is None:
-                    self.generator = torch.Generator(device=x.device).manual_seed(42)
-                elif self.generator.device != x.device:
-                    self.generator = torch.Generator(device=x.device).set_state(self.generator.get_state())
-
-                noise_shape = [guiding_latent.shape[0], guiding_latent.shape[1], 1, guiding_latent.shape[3], guiding_latent.shape[4]]
-                scale = guiding_latent_noise_scale * (input_ts ** 2)
-                guiding_noise = scale * torch.randn(size=noise_shape, device=x.device, generator=self.generator)
-
-                x[:, :, 0] = guiding_noise[:, :, 0] + x[:, :, 0] *  (1.0 - scale[:, :, 0])
-
-
         orig_shape = list(x.shape)
 
-        x = self.patchifier.patchify(x)
+        x, latent_coords = self.patchifier.patchify(x)
+        pixel_coords = latent_to_pixel_coords(
+            latent_coords=latent_coords,
+            scale_factors=self.vae_scale_factors,
+            causal_fix=self.causal_temporal_positioning,
+        )
+
+        if keyframe_idxs is not None:
+            pixel_coords[:, :, -keyframe_idxs.shape[2]:] = keyframe_idxs
+
+        fractional_coords = pixel_coords.to(torch.float32)
+        fractional_coords[:, 0] = fractional_coords[:, 0] * (1.0 / frame_rate)
 
         x = self.patchify_proj(x)
         timestep = timestep * 1000.0
@@ -459,7 +443,7 @@ class LTXVModel(torch.nn.Module):
         if attention_mask is not None and not torch.is_floating_point(attention_mask):
             attention_mask = (attention_mask - 1).to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])) * torch.finfo(x.dtype).max
 
-        pe = precompute_freqs_cis(indices_grid, dim=self.inner_dim, out_dtype=x.dtype)
+        pe = precompute_freqs_cis(fractional_coords, dim=self.inner_dim, out_dtype=x.dtype)
 
         batch_size = x.shape[0]
         timestep, embedded_timestep = self.adaln_single(
@@ -519,8 +503,4 @@ class LTXVModel(torch.nn.Module):
             out_channels=orig_shape[1] // math.prod(self.patchifier.patch_size),
         )
 
-        if guiding_latent is not None:
-            x[:, :, 0] = (input_x[:, :, 0] - guiding_latent[:, :, 0]) / input_ts[:, :, 0]
-
-        # print("res", x)
         return x

comfy/ldm/lightricks/symmetric_patchifier.py

@@ -6,16 +6,29 @@ from einops import rearrange
 from torch import Tensor
 
 
-def append_dims(x: torch.Tensor, target_dims: int) -> torch.Tensor:
-    """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
-    dims_to_append = target_dims - x.ndim
-    if dims_to_append < 0:
-        raise ValueError(
-            f"input has {x.ndim} dims but target_dims is {target_dims}, which is less"
+def latent_to_pixel_coords(
+    latent_coords: Tensor, scale_factors: Tuple[int, int, int], causal_fix: bool = False
+) -> Tensor:
+    """
+    Converts latent coordinates to pixel coordinates by scaling them according to the VAE's
+    configuration.
+    Args:
+        latent_coords (Tensor): A tensor of shape [batch_size, 3, num_latents]
+        containing the latent corner coordinates of each token.
+        scale_factors (Tuple[int, int, int]): The scale factors of the VAE's latent space.
+        causal_fix (bool): Whether to take into account the different temporal scale
+            of the first frame. Default = False for backwards compatibility.
+    Returns:
+        Tensor: A tensor of pixel coordinates corresponding to the input latent coordinates.
+    """
+    pixel_coords = (
+        latent_coords
+        * torch.tensor(scale_factors, device=latent_coords.device)[None, :, None]
     )
-    elif dims_to_append == 0:
-        return x
-    return x[(...,) + (None,) * dims_to_append]
+    if causal_fix:
+        # Fix temporal scale for first frame to 1 due to causality
+        pixel_coords[:, 0] = (pixel_coords[:, 0] + 1 - scale_factors[0]).clamp(min=0)
+    return pixel_coords
 
 
 class Patchifier(ABC):
@@ -44,29 +57,26 @@ class Patchifier(ABC):
     def patch_size(self):
         return self._patch_size
 
-    def get_grid(
-        self, orig_num_frames, orig_height, orig_width, batch_size, scale_grid, device
+    def get_latent_coords(
+        self, latent_num_frames, latent_height, latent_width, batch_size, device
     ):
-        f = orig_num_frames // self._patch_size[0]
-        h = orig_height // self._patch_size[1]
-        w = orig_width // self._patch_size[2]
-        grid_h = torch.arange(h, dtype=torch.float32, device=device)
-        grid_w = torch.arange(w, dtype=torch.float32, device=device)
-        grid_f = torch.arange(f, dtype=torch.float32, device=device)
-        grid = torch.meshgrid(grid_f, grid_h, grid_w, indexing='ij')
-        grid = torch.stack(grid, dim=0)
-        grid = grid.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
-
-        if scale_grid is not None:
-            for i in range(3):
-                if isinstance(scale_grid[i], Tensor):
-                    scale = append_dims(scale_grid[i], grid.ndim - 1)
-                else:
-                    scale = scale_grid[i]
-                grid[:, i, ...] = grid[:, i, ...] * scale * self._patch_size[i]
-
-        grid = rearrange(grid, "b c f h w -> b c (f h w)", b=batch_size)
-        return grid
+        """
+        Return a tensor of shape [batch_size, 3, num_patches] containing the
+            top-left corner latent coordinates of each latent patch.
+        The tensor is repeated for each batch element.
+        """
+        latent_sample_coords = torch.meshgrid(
+            torch.arange(0, latent_num_frames, self._patch_size[0], device=device),
+            torch.arange(0, latent_height, self._patch_size[1], device=device),
+            torch.arange(0, latent_width, self._patch_size[2], device=device),
+            indexing="ij",
+        )
+        latent_sample_coords = torch.stack(latent_sample_coords, dim=0)
+        latent_coords = latent_sample_coords.unsqueeze(0).repeat(batch_size, 1, 1, 1, 1)
+        latent_coords = rearrange(
+            latent_coords, "b c f h w -> b c (f h w)", b=batch_size
+        )
+        return latent_coords
 
 
 class SymmetricPatchifier(Patchifier):
@@ -74,6 +84,8 @@ class SymmetricPatchifier(Patchifier):
         self,
         latents: Tensor,
     ) -> Tuple[Tensor, Tensor]:
+        b, _, f, h, w = latents.shape
+        latent_coords = self.get_latent_coords(f, h, w, b, latents.device)
         latents = rearrange(
             latents,
             "b c (f p1) (h p2) (w p3) -> b (f h w) (c p1 p2 p3)",
@@ -81,7 +93,7 @@ class SymmetricPatchifier(Patchifier):
             p2=self._patch_size[1],
             p3=self._patch_size[2],
         )
-        return latents
+        return latents, latent_coords
 
     def unpatchify(
         self,

comfy/ldm/lightricks/vae/causal_conv3d.py

@@ -15,6 +15,7 @@ class CausalConv3d(nn.Module):
         stride: Union[int, Tuple[int]] = 1,
         dilation: int = 1,
         groups: int = 1,
+        spatial_padding_mode: str = "zeros",
         **kwargs,
     ):
         super().__init__()
@@ -38,7 +39,7 @@ class CausalConv3d(nn.Module):
             stride=stride,
             dilation=dilation,
             padding=padding,
-            padding_mode="zeros",
+            padding_mode=spatial_padding_mode,
             groups=groups,
         )
 

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -1,13 +1,15 @@
+from __future__ import annotations
 import torch
 from torch import nn
 from functools import partial
 import math
 from einops import rearrange
-from typing import Optional, Tuple, Union
+from typing import List, Optional, Tuple, Union
 from .conv_nd_factory import make_conv_nd, make_linear_nd
 from .pixel_norm import PixelNorm
 from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings
 import comfy.ops
+
 ops = comfy.ops.disable_weight_init
 
 class Encoder(nn.Module):
@@ -32,7 +34,7 @@ class Encoder(nn.Module):
         norm_layer (`str`, *optional*, defaults to `group_norm`):
             The normalization layer to use. Can be either `group_norm` or `pixel_norm`.
         latent_log_var (`str`, *optional*, defaults to `per_channel`):
-            The number of channels for the log variance. Can be either `per_channel`, `uniform`, or `none`.
+            The number of channels for the log variance. Can be either `per_channel`, `uniform`, `constant` or `none`.
     """
 
     def __init__(
@@ -40,12 +42,13 @@ class Encoder(nn.Module):
         dims: Union[int, Tuple[int, int]] = 3,
         in_channels: int = 3,
         out_channels: int = 3,
-        blocks=[("res_x", 1)],
+        blocks: List[Tuple[str, int | dict]] = [("res_x", 1)],
         base_channels: int = 128,
         norm_num_groups: int = 32,
         patch_size: Union[int, Tuple[int]] = 1,
         norm_layer: str = "group_norm",  # group_norm, pixel_norm
         latent_log_var: str = "per_channel",
+        spatial_padding_mode: str = "zeros",
     ):
         super().__init__()
         self.patch_size = patch_size
@@ -65,6 +68,7 @@ class Encoder(nn.Module):
             stride=1,
             padding=1,
             causal=True,
+            spatial_padding_mode=spatial_padding_mode,
         )
 
         self.down_blocks = nn.ModuleList([])
@@ -82,6 +86,7 @@ class Encoder(nn.Module):
                     resnet_eps=1e-6,
                     resnet_groups=norm_num_groups,
                     norm_layer=norm_layer,
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             elif block_name == "res_x_y":
                 output_channel = block_params.get("multiplier", 2) * output_channel
@@ -92,6 +97,7 @@ class Encoder(nn.Module):
                     eps=1e-6,
                     groups=norm_num_groups,
                     norm_layer=norm_layer,
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             elif block_name == "compress_time":
                 block = make_conv_nd(
@@ -101,6 +107,7 @@ class Encoder(nn.Module):
                     kernel_size=3,
                     stride=(2, 1, 1),
                     causal=True,
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             elif block_name == "compress_space":
                 block = make_conv_nd(
@@ -110,6 +117,7 @@ class Encoder(nn.Module):
                     kernel_size=3,
                     stride=(1, 2, 2),
                     causal=True,
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             elif block_name == "compress_all":
                 block = make_conv_nd(
@@ -119,6 +127,7 @@ class Encoder(nn.Module):
                     kernel_size=3,
                     stride=(2, 2, 2),
                     causal=True,
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             elif block_name == "compress_all_x_y":
                 output_channel = block_params.get("multiplier", 2) * output_channel
@@ -129,6 +138,34 @@ class Encoder(nn.Module):
                     kernel_size=3,
                     stride=(2, 2, 2),
                     causal=True,
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_all_res":
+                output_channel = block_params.get("multiplier", 2) * output_channel
+                block = SpaceToDepthDownsample(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    stride=(2, 2, 2),
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_space_res":
+                output_channel = block_params.get("multiplier", 2) * output_channel
+                block = SpaceToDepthDownsample(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    stride=(1, 2, 2),
+                    spatial_padding_mode=spatial_padding_mode,
+                )
+            elif block_name == "compress_time_res":
+                output_channel = block_params.get("multiplier", 2) * output_channel
+                block = SpaceToDepthDownsample(
+                    dims=dims,
+                    in_channels=input_channel,
+                    out_channels=output_channel,
+                    stride=(2, 1, 1),
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             else:
                 raise ValueError(f"unknown block: {block_name}")
@@ -152,10 +189,18 @@ class Encoder(nn.Module):
             conv_out_channels *= 2
         elif latent_log_var == "uniform":
             conv_out_channels += 1
+        elif latent_log_var == "constant":
+            conv_out_channels += 1
         elif latent_log_var != "none":
             raise ValueError(f"Invalid latent_log_var: {latent_log_var}")
         self.conv_out = make_conv_nd(
-            dims, output_channel, conv_out_channels, 3, padding=1, causal=True
+            dims,
+            output_channel,
+            conv_out_channels,
+            3,
+            padding=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
         )
 
         self.gradient_checkpointing = False
@@ -197,6 +242,15 @@ class Encoder(nn.Module):
                 sample = torch.cat([sample, repeated_last_channel], dim=1)
             else:
                 raise ValueError(f"Invalid input shape: {sample.shape}")
+        elif self.latent_log_var == "constant":
+            sample = sample[:, :-1, ...]
+            approx_ln_0 = (
+                -30
+            )  # this is the minimal clamp value in DiagonalGaussianDistribution objects
+            sample = torch.cat(
+                [sample, torch.ones_like(sample, device=sample.device) * approx_ln_0],
+                dim=1,
+            )
 
         return sample
 
@@ -231,7 +285,7 @@ class Decoder(nn.Module):
         dims,
         in_channels: int = 3,
         out_channels: int = 3,
-        blocks=[("res_x", 1)],
+        blocks: List[Tuple[str, int | dict]] = [("res_x", 1)],
         base_channels: int = 128,
         layers_per_block: int = 2,
         norm_num_groups: int = 32,
@@ -239,6 +293,7 @@ class Decoder(nn.Module):
         norm_layer: str = "group_norm",
         causal: bool = True,
         timestep_conditioning: bool = False,
+        spatial_padding_mode: str = "zeros",
     ):
         super().__init__()
         self.patch_size = patch_size
@@ -264,6 +319,7 @@ class Decoder(nn.Module):
             stride=1,
             padding=1,
             causal=True,
+            spatial_padding_mode=spatial_padding_mode,
         )
 
         self.up_blocks = nn.ModuleList([])
@@ -283,6 +339,7 @@ class Decoder(nn.Module):
                     norm_layer=norm_layer,
                     inject_noise=block_params.get("inject_noise", False),
                     timestep_conditioning=timestep_conditioning,
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             elif block_name == "attn_res_x":
                 block = UNetMidBlock3D(
@@ -294,6 +351,7 @@ class Decoder(nn.Module):
                     inject_noise=block_params.get("inject_noise", False),
                     timestep_conditioning=timestep_conditioning,
                     attention_head_dim=block_params["attention_head_dim"],
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             elif block_name == "res_x_y":
                 output_channel = output_channel // block_params.get("multiplier", 2)
@@ -306,14 +364,21 @@ class Decoder(nn.Module):
                     norm_layer=norm_layer,
                     inject_noise=block_params.get("inject_noise", False),
                     timestep_conditioning=False,
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             elif block_name == "compress_time":
                 block = DepthToSpaceUpsample(
-                    dims=dims, in_channels=input_channel, stride=(2, 1, 1)
+                    dims=dims,
+                    in_channels=input_channel,
+                    stride=(2, 1, 1),
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             elif block_name == "compress_space":
                 block = DepthToSpaceUpsample(
-                    dims=dims, in_channels=input_channel, stride=(1, 2, 2)
+                    dims=dims,
+                    in_channels=input_channel,
+                    stride=(1, 2, 2),
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             elif block_name == "compress_all":
                 output_channel = output_channel // block_params.get("multiplier", 1)
@@ -323,6 +388,7 @@ class Decoder(nn.Module):
                     stride=(2, 2, 2),
                     residual=block_params.get("residual", False),
                     out_channels_reduction_factor=block_params.get("multiplier", 1),
+                    spatial_padding_mode=spatial_padding_mode,
                 )
             else:
                 raise ValueError(f"unknown layer: {block_name}")
@@ -340,7 +406,13 @@ class Decoder(nn.Module):
 
         self.conv_act = nn.SiLU()
         self.conv_out = make_conv_nd(
-            dims, output_channel, out_channels, 3, padding=1, causal=True
+            dims,
+            output_channel,
+            out_channels,
+            3,
+            padding=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
         )
 
         self.gradient_checkpointing = False
@@ -433,6 +505,12 @@ class UNetMidBlock3D(nn.Module):
         resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks.
         resnet_groups (`int`, *optional*, defaults to 32):
             The number of groups to use in the group normalization layers of the resnet blocks.
+        norm_layer (`str`, *optional*, defaults to `group_norm`):
+            The normalization layer to use. Can be either `group_norm` or `pixel_norm`.
+        inject_noise (`bool`, *optional*, defaults to `False`):
+            Whether to inject noise into the hidden states.
+        timestep_conditioning (`bool`, *optional*, defaults to `False`):
+            Whether to condition the hidden states on the timestep.
 
     Returns:
         `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size,
@@ -451,6 +529,7 @@ class UNetMidBlock3D(nn.Module):
         norm_layer: str = "group_norm",
         inject_noise: bool = False,
         timestep_conditioning: bool = False,
+        spatial_padding_mode: str = "zeros",
     ):
         super().__init__()
         resnet_groups = (
@@ -476,13 +555,17 @@ class UNetMidBlock3D(nn.Module):
                     norm_layer=norm_layer,
                     inject_noise=inject_noise,
                     timestep_conditioning=timestep_conditioning,
+                    spatial_padding_mode=spatial_padding_mode,
                 )
                 for _ in range(num_layers)
             ]
         )
 
     def forward(
-        self, hidden_states: torch.FloatTensor, causal: bool = True, timestep: Optional[torch.Tensor] = None
+        self,
+        hidden_states: torch.FloatTensor,
+        causal: bool = True,
+        timestep: Optional[torch.Tensor] = None,
     ) -> torch.FloatTensor:
         timestep_embed = None
         if self.timestep_conditioning:
@@ -507,9 +590,62 @@ class UNetMidBlock3D(nn.Module):
         return hidden_states
 
 
+class SpaceToDepthDownsample(nn.Module):
+    def __init__(self, dims, in_channels, out_channels, stride, spatial_padding_mode):
+        super().__init__()
+        self.stride = stride
+        self.group_size = in_channels * math.prod(stride) // out_channels
+        self.conv = make_conv_nd(
+            dims=dims,
+            in_channels=in_channels,
+            out_channels=out_channels // math.prod(stride),
+            kernel_size=3,
+            stride=1,
+            causal=True,
+            spatial_padding_mode=spatial_padding_mode,
+        )
+
+    def forward(self, x, causal: bool = True):
+        if self.stride[0] == 2:
+            x = torch.cat(
+                [x[:, :, :1, :, :], x], dim=2
+            )  # duplicate first frames for padding
+
+        # skip connection
+        x_in = rearrange(
+            x,
+            "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
+            p1=self.stride[0],
+            p2=self.stride[1],
+            p3=self.stride[2],
+        )
+        x_in = rearrange(x_in, "b (c g) d h w -> b c g d h w", g=self.group_size)
+        x_in = x_in.mean(dim=2)
+
+        # conv
+        x = self.conv(x, causal=causal)
+        x = rearrange(
+            x,
+            "b c (d p1) (h p2) (w p3) -> b (c p1 p2 p3) d h w",
+            p1=self.stride[0],
+            p2=self.stride[1],
+            p3=self.stride[2],
+        )
+
+        x = x + x_in
+
+        return x
+
+
 class DepthToSpaceUpsample(nn.Module):
     def __init__(
-        self, dims, in_channels, stride, residual=False, out_channels_reduction_factor=1
+        self,
+        dims,
+        in_channels,
+        stride,
+        residual=False,
+        out_channels_reduction_factor=1,
+        spatial_padding_mode="zeros",
     ):
         super().__init__()
         self.stride = stride
@@ -523,6 +659,7 @@ class DepthToSpaceUpsample(nn.Module):
             kernel_size=3,
             stride=1,
             causal=True,
+            spatial_padding_mode=spatial_padding_mode,
         )
         self.residual = residual
         self.out_channels_reduction_factor = out_channels_reduction_factor
@@ -558,7 +695,7 @@ class DepthToSpaceUpsample(nn.Module):
 class LayerNorm(nn.Module):
     def __init__(self, dim, eps, elementwise_affine=True) -> None:
         super().__init__()
-        self.norm = nn.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine)
+        self.norm = ops.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine)
 
     def forward(self, x):
         x = rearrange(x, "b c d h w -> b d h w c")
@@ -591,6 +728,7 @@ class ResnetBlock3D(nn.Module):
         norm_layer: str = "group_norm",
         inject_noise: bool = False,
         timestep_conditioning: bool = False,
+        spatial_padding_mode: str = "zeros",
     ):
         super().__init__()
         self.in_channels = in_channels
@@ -617,6 +755,7 @@ class ResnetBlock3D(nn.Module):
             stride=1,
             padding=1,
             causal=True,
+            spatial_padding_mode=spatial_padding_mode,
         )
 
         if inject_noise:
@@ -641,6 +780,7 @@ class ResnetBlock3D(nn.Module):
             stride=1,
             padding=1,
             causal=True,
+            spatial_padding_mode=spatial_padding_mode,
         )
 
         if inject_noise:
@@ -801,9 +941,44 @@ class processor(nn.Module):
         return (x - self.get_buffer("mean-of-means").view(1, -1, 1, 1, 1).to(x)) / self.get_buffer("std-of-means").view(1, -1, 1, 1, 1).to(x)
 
 class VideoVAE(nn.Module):
-    def __init__(self, version=0):
+    def __init__(self, version=0, config=None):
         super().__init__()
 
+        if config is None:
+            config = self.guess_config(version)
+
+        self.timestep_conditioning = config.get("timestep_conditioning", False)
+        double_z = config.get("double_z", True)
+        latent_log_var = config.get(
+            "latent_log_var", "per_channel" if double_z else "none"
+        )
+
+        self.encoder = Encoder(
+            dims=config["dims"],
+            in_channels=config.get("in_channels", 3),
+            out_channels=config["latent_channels"],
+            blocks=config.get("encoder_blocks", config.get("encoder_blocks", config.get("blocks"))),
+            patch_size=config.get("patch_size", 1),
+            latent_log_var=latent_log_var,
+            norm_layer=config.get("norm_layer", "group_norm"),
+            spatial_padding_mode=config.get("spatial_padding_mode", "zeros"),
+        )
+
+        self.decoder = Decoder(
+            dims=config["dims"],
+            in_channels=config["latent_channels"],
+            out_channels=config.get("out_channels", 3),
+            blocks=config.get("decoder_blocks", config.get("decoder_blocks", config.get("blocks"))),
+            patch_size=config.get("patch_size", 1),
+            norm_layer=config.get("norm_layer", "group_norm"),
+            causal=config.get("causal_decoder", False),
+            timestep_conditioning=self.timestep_conditioning,
+            spatial_padding_mode=config.get("spatial_padding_mode", "zeros"),
+        )
+
+        self.per_channel_statistics = processor()
+
+    def guess_config(self, version):
         if version == 0:
             config = {
                 "_class_name": "CausalVideoAutoencoder",
@@ -830,7 +1005,7 @@ class VideoVAE(nn.Module):
                 "use_quant_conv": False,
                 "causal_decoder": False,
             }
-        else:
+        elif version == 1:
             config = {
                 "_class_name": "CausalVideoAutoencoder",
                 "dims": 3,
@@ -866,37 +1041,47 @@ class VideoVAE(nn.Module):
                 "causal_decoder": False,
                 "timestep_conditioning": True,
             }
-
-        double_z = config.get("double_z", True)
-        latent_log_var = config.get(
-            "latent_log_var", "per_channel" if double_z else "none"
-        )
-
-        self.encoder = Encoder(
-            dims=config["dims"],
-            in_channels=config.get("in_channels", 3),
-            out_channels=config["latent_channels"],
-            blocks=config.get("encoder_blocks", config.get("encoder_blocks", config.get("blocks"))),
-            patch_size=config.get("patch_size", 1),
-            latent_log_var=latent_log_var,
-            norm_layer=config.get("norm_layer", "group_norm"),
-        )
-
-        self.decoder = Decoder(
-            dims=config["dims"],
-            in_channels=config["latent_channels"],
-            out_channels=config.get("out_channels", 3),
-            blocks=config.get("decoder_blocks", config.get("decoder_blocks", config.get("blocks"))),
-            patch_size=config.get("patch_size", 1),
-            norm_layer=config.get("norm_layer", "group_norm"),
-            causal=config.get("causal_decoder", False),
-            timestep_conditioning=config.get("timestep_conditioning", False),
-        )
-
-        self.timestep_conditioning = config.get("timestep_conditioning", False)
-        self.per_channel_statistics = processor()
+        else:
+            config = {
+                "_class_name": "CausalVideoAutoencoder",
+                "dims": 3,
+                "in_channels": 3,
+                "out_channels": 3,
+                "latent_channels": 128,
+                "encoder_blocks": [
+                    ["res_x", {"num_layers": 4}],
+                    ["compress_space_res", {"multiplier": 2}],
+                    ["res_x", {"num_layers": 6}],
+                    ["compress_time_res", {"multiplier": 2}],
+                    ["res_x", {"num_layers": 6}],
+                    ["compress_all_res", {"multiplier": 2}],
+                    ["res_x", {"num_layers": 2}],
+                    ["compress_all_res", {"multiplier": 2}],
+                    ["res_x", {"num_layers": 2}]
+                ],
+                "decoder_blocks": [
+                    ["res_x", {"num_layers": 5, "inject_noise": False}],
+                    ["compress_all", {"residual": True, "multiplier": 2}],
+                    ["res_x", {"num_layers": 5, "inject_noise": False}],
+                    ["compress_all", {"residual": True, "multiplier": 2}],
+                    ["res_x", {"num_layers": 5, "inject_noise": False}],
+                    ["compress_all", {"residual": True, "multiplier": 2}],
+                    ["res_x", {"num_layers": 5, "inject_noise": False}]
+                ],
+                "scaling_factor": 1.0,
+                "norm_layer": "pixel_norm",
+                "patch_size": 4,
+                "latent_log_var": "uniform",
+                "use_quant_conv": False,
+                "causal_decoder": False,
+                "timestep_conditioning": True
+            }
+        return config
 
     def encode(self, x):
+        frames_count = x.shape[2]
+        if ((frames_count - 1) % 8) != 0:
+            raise ValueError("Invalid number of frames: Encode input must have 1 + 8 * x frames (e.g., 1, 9, 17, ...). Please check your input.")
         means, logvar = torch.chunk(self.encoder(x), 2, dim=1)
         return self.per_channel_statistics.normalize(means)
 

comfy/ldm/lightricks/vae/conv_nd_factory.py

@@ -17,7 +17,11 @@ def make_conv_nd(
     groups=1,
     bias=True,
     causal=False,
+    spatial_padding_mode="zeros",
+    temporal_padding_mode="zeros",
 ):
+    if not (spatial_padding_mode == temporal_padding_mode or causal):
+        raise NotImplementedError("spatial and temporal padding modes must be equal")
     if dims == 2:
         return ops.Conv2d(
             in_channels=in_channels,
@@ -28,6 +32,7 @@ def make_conv_nd(
             dilation=dilation,
             groups=groups,
             bias=bias,
+            padding_mode=spatial_padding_mode,
         )
     elif dims == 3:
         if causal:
@@ -40,6 +45,7 @@ def make_conv_nd(
                 dilation=dilation,
                 groups=groups,
                 bias=bias,
+                spatial_padding_mode=spatial_padding_mode,
             )
         return ops.Conv3d(
             in_channels=in_channels,
@@ -50,6 +56,7 @@ def make_conv_nd(
             dilation=dilation,
             groups=groups,
             bias=bias,
+            padding_mode=spatial_padding_mode,
         )
     elif dims == (2, 1):
         return DualConv3d(
@@ -59,6 +66,7 @@ def make_conv_nd(
             stride=stride,
             padding=padding,
             bias=bias,
+            padding_mode=spatial_padding_mode,
         )
     else:
         raise ValueError(f"unsupported dimensions: {dims}")

comfy/ldm/lightricks/vae/dual_conv3d.py

@@ -18,11 +18,13 @@ class DualConv3d(nn.Module):
         dilation: Union[int, Tuple[int, int, int]] = 1,
         groups=1,
         bias=True,
+        padding_mode="zeros",
     ):
         super(DualConv3d, self).__init__()
 
         self.in_channels = in_channels
         self.out_channels = out_channels
+        self.padding_mode = padding_mode
         # Ensure kernel_size, stride, padding, and dilation are tuples of length 3
         if isinstance(kernel_size, int):
             kernel_size = (kernel_size, kernel_size, kernel_size)
@@ -108,6 +110,7 @@ class DualConv3d(nn.Module):
             self.padding1,
             self.dilation1,
             self.groups,
+            padding_mode=self.padding_mode,
         )
 
         if skip_time_conv:
@@ -122,6 +125,7 @@ class DualConv3d(nn.Module):
             self.padding2,
             self.dilation2,
             self.groups,
+            padding_mode=self.padding_mode,
         )
 
         return x
@@ -137,7 +141,16 @@ class DualConv3d(nn.Module):
         stride1 = (self.stride1[1], self.stride1[2])
         padding1 = (self.padding1[1], self.padding1[2])
         dilation1 = (self.dilation1[1], self.dilation1[2])
-        x = F.conv2d(x, weight1, self.bias1, stride1, padding1, dilation1, self.groups)
+        x = F.conv2d(
+            x,
+            weight1,
+            self.bias1,
+            stride1,
+            padding1,
+            dilation1,
+            self.groups,
+            padding_mode=self.padding_mode,
+        )
 
         _, _, h, w = x.shape
 
@@ -154,7 +167,16 @@ class DualConv3d(nn.Module):
         stride2 = self.stride2[0]
         padding2 = self.padding2[0]
         dilation2 = self.dilation2[0]
-        x = F.conv1d(x, weight2, self.bias2, stride2, padding2, dilation2, self.groups)
+        x = F.conv1d(
+            x,
+            weight2,
+            self.bias2,
+            stride2,
+            padding2,
+            dilation2,
+            self.groups,
+            padding_mode=self.padding_mode,
+        )
         x = rearrange(x, "(b h w) c d -> b c d h w", b=b, h=h, w=w)
 
         return x

comfy/ldm/lumina/model.py

@@ -0,0 +1,622 @@
+# Code from: https://github.com/Alpha-VLLM/Lumina-Image-2.0/blob/main/models/model.py
+from __future__ import annotations
+
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import comfy.ldm.common_dit
+
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
+from comfy.ldm.modules.attention import optimized_attention_masked
+from comfy.ldm.flux.layers import EmbedND
+
+
+def modulate(x, scale):
+    return x * (1 + scale.unsqueeze(1))
+
+#############################################################################
+#                               Core NextDiT Model                              #
+#############################################################################
+
+
+class JointAttention(nn.Module):
+    """Multi-head attention module."""
+
+    def __init__(
+        self,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: Optional[int],
+        qk_norm: bool,
+        operation_settings={},
+    ):
+        """
+        Initialize the Attention module.
+
+        Args:
+            dim (int): Number of input dimensions.
+            n_heads (int): Number of heads.
+            n_kv_heads (Optional[int]): Number of kv heads, if using GQA.
+
+        """
+        super().__init__()
+        self.n_kv_heads = n_heads if n_kv_heads is None else n_kv_heads
+        self.n_local_heads = n_heads
+        self.n_local_kv_heads = self.n_kv_heads
+        self.n_rep = self.n_local_heads // self.n_local_kv_heads
+        self.head_dim = dim // n_heads
+
+        self.qkv = operation_settings.get("operations").Linear(
+            dim,
+            (n_heads + self.n_kv_heads + self.n_kv_heads) * self.head_dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.out = operation_settings.get("operations").Linear(
+            n_heads * self.head_dim,
+            dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+        if qk_norm:
+            self.q_norm = operation_settings.get("operations").RMSNorm(self.head_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+            self.k_norm = operation_settings.get("operations").RMSNorm(self.head_dim, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        else:
+            self.q_norm = self.k_norm = nn.Identity()
+
+    @staticmethod
+    def apply_rotary_emb(
+        x_in: torch.Tensor,
+        freqs_cis: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Apply rotary embeddings to input tensors using the given frequency
+        tensor.
+
+        This function applies rotary embeddings to the given query 'xq' and
+        key 'xk' tensors using the provided frequency tensor 'freqs_cis'. The
+        input tensors are reshaped as complex numbers, and the frequency tensor
+        is reshaped for broadcasting compatibility. The resulting tensors
+        contain rotary embeddings and are returned as real tensors.
+
+        Args:
+            x_in (torch.Tensor): Query or Key tensor to apply rotary embeddings.
+            freqs_cis (torch.Tensor): Precomputed frequency tensor for complex
+                exponentials.
+
+        Returns:
+            Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor
+                and key tensor with rotary embeddings.
+        """
+
+        t_ = x_in.reshape(*x_in.shape[:-1], -1, 1, 2)
+        t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
+        return t_out.reshape(*x_in.shape)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        freqs_cis: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+
+        Args:
+            x:
+            x_mask:
+            freqs_cis:
+
+        Returns:
+
+        """
+        bsz, seqlen, _ = x.shape
+
+        xq, xk, xv = torch.split(
+            self.qkv(x),
+            [
+                self.n_local_heads * self.head_dim,
+                self.n_local_kv_heads * self.head_dim,
+                self.n_local_kv_heads * self.head_dim,
+            ],
+            dim=-1,
+        )
+        xq = xq.view(bsz, seqlen, self.n_local_heads, self.head_dim)
+        xk = xk.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
+        xv = xv.view(bsz, seqlen, self.n_local_kv_heads, self.head_dim)
+
+        xq = self.q_norm(xq)
+        xk = self.k_norm(xk)
+
+        xq = JointAttention.apply_rotary_emb(xq, freqs_cis=freqs_cis)
+        xk = JointAttention.apply_rotary_emb(xk, freqs_cis=freqs_cis)
+
+        n_rep = self.n_local_heads // self.n_local_kv_heads
+        if n_rep >= 1:
+            xk = xk.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
+            xv = xv.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
+        output = optimized_attention_masked(xq.movedim(1, 2), xk.movedim(1, 2), xv.movedim(1, 2), self.n_local_heads, x_mask, skip_reshape=True)
+
+        return self.out(output)
+
+
+class FeedForward(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int,
+        ffn_dim_multiplier: Optional[float],
+        operation_settings={},
+    ):
+        """
+        Initialize the FeedForward module.
+
+        Args:
+            dim (int): Input dimension.
+            hidden_dim (int): Hidden dimension of the feedforward layer.
+            multiple_of (int): Value to ensure hidden dimension is a multiple
+                of this value.
+            ffn_dim_multiplier (float, optional): Custom multiplier for hidden
+                dimension. Defaults to None.
+
+        """
+        super().__init__()
+        # custom dim factor multiplier
+        if ffn_dim_multiplier is not None:
+            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+        self.w1 = operation_settings.get("operations").Linear(
+            dim,
+            hidden_dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.w2 = operation_settings.get("operations").Linear(
+            hidden_dim,
+            dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.w3 = operation_settings.get("operations").Linear(
+            dim,
+            hidden_dim,
+            bias=False,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+    # @torch.compile
+    def _forward_silu_gating(self, x1, x3):
+        return F.silu(x1) * x3
+
+    def forward(self, x):
+        return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x)))
+
+
+class JointTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        layer_id: int,
+        dim: int,
+        n_heads: int,
+        n_kv_heads: int,
+        multiple_of: int,
+        ffn_dim_multiplier: float,
+        norm_eps: float,
+        qk_norm: bool,
+        modulation=True,
+        operation_settings={},
+    ) -> None:
+        """
+        Initialize a TransformerBlock.
+
+        Args:
+            layer_id (int): Identifier for the layer.
+            dim (int): Embedding dimension of the input features.
+            n_heads (int): Number of attention heads.
+            n_kv_heads (Optional[int]): Number of attention heads in key and
+                value features (if using GQA), or set to None for the same as
+                query.
+            multiple_of (int):
+            ffn_dim_multiplier (float):
+            norm_eps (float):
+
+        """
+        super().__init__()
+        self.dim = dim
+        self.head_dim = dim // n_heads
+        self.attention = JointAttention(dim, n_heads, n_kv_heads, qk_norm, operation_settings=operation_settings)
+        self.feed_forward = FeedForward(
+            dim=dim,
+            hidden_dim=4 * dim,
+            multiple_of=multiple_of,
+            ffn_dim_multiplier=ffn_dim_multiplier,
+            operation_settings=operation_settings,
+        )
+        self.layer_id = layer_id
+        self.attention_norm1 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.ffn_norm1 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        self.attention_norm2 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.ffn_norm2 = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        self.modulation = modulation
+        if modulation:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                operation_settings.get("operations").Linear(
+                    min(dim, 1024),
+                    4 * dim,
+                    bias=True,
+                    device=operation_settings.get("device"),
+                    dtype=operation_settings.get("dtype"),
+                ),
+            )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        x_mask: torch.Tensor,
+        freqs_cis: torch.Tensor,
+        adaln_input: Optional[torch.Tensor]=None,
+    ):
+        """
+        Perform a forward pass through the TransformerBlock.
+
+        Args:
+            x (torch.Tensor): Input tensor.
+            freqs_cis (torch.Tensor): Precomputed cosine and sine frequencies.
+
+        Returns:
+            torch.Tensor: Output tensor after applying attention and
+                feedforward layers.
+
+        """
+        if self.modulation:
+            assert adaln_input is not None
+            scale_msa, gate_msa, scale_mlp, gate_mlp = self.adaLN_modulation(adaln_input).chunk(4, dim=1)
+
+            x = x + gate_msa.unsqueeze(1).tanh() * self.attention_norm2(
+                self.attention(
+                    modulate(self.attention_norm1(x), scale_msa),
+                    x_mask,
+                    freqs_cis,
+                )
+            )
+            x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(
+                self.feed_forward(
+                    modulate(self.ffn_norm1(x), scale_mlp),
+                )
+            )
+        else:
+            assert adaln_input is None
+            x = x + self.attention_norm2(
+                self.attention(
+                    self.attention_norm1(x),
+                    x_mask,
+                    freqs_cis,
+                )
+            )
+            x = x + self.ffn_norm2(
+                self.feed_forward(
+                    self.ffn_norm1(x),
+                )
+            )
+        return x
+
+
+class FinalLayer(nn.Module):
+    """
+    The final layer of NextDiT.
+    """
+
+    def __init__(self, hidden_size, patch_size, out_channels, operation_settings={}):
+        super().__init__()
+        self.norm_final = operation_settings.get("operations").LayerNorm(
+            hidden_size,
+            elementwise_affine=False,
+            eps=1e-6,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+        self.linear = operation_settings.get("operations").Linear(
+            hidden_size,
+            patch_size * patch_size * out_channels,
+            bias=True,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operation_settings.get("operations").Linear(
+                min(hidden_size, 1024),
+                hidden_size,
+                bias=True,
+                device=operation_settings.get("device"),
+                dtype=operation_settings.get("dtype"),
+            ),
+        )
+
+    def forward(self, x, c):
+        scale = self.adaLN_modulation(c)
+        x = modulate(self.norm_final(x), scale)
+        x = self.linear(x)
+        return x
+
+
+class NextDiT(nn.Module):
+    """
+    Diffusion model with a Transformer backbone.
+    """
+
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 4,
+        dim: int = 4096,
+        n_layers: int = 32,
+        n_refiner_layers: int = 2,
+        n_heads: int = 32,
+        n_kv_heads: Optional[int] = None,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+        norm_eps: float = 1e-5,
+        qk_norm: bool = False,
+        cap_feat_dim: int = 5120,
+        axes_dims: List[int] = (16, 56, 56),
+        axes_lens: List[int] = (1, 512, 512),
+        image_model=None,
+        device=None,
+        dtype=None,
+        operations=None,
+    ) -> None:
+        super().__init__()
+        self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+        self.in_channels = in_channels
+        self.out_channels = in_channels
+        self.patch_size = patch_size
+
+        self.x_embedder = operation_settings.get("operations").Linear(
+            in_features=patch_size * patch_size * in_channels,
+            out_features=dim,
+            bias=True,
+            device=operation_settings.get("device"),
+            dtype=operation_settings.get("dtype"),
+        )
+
+        self.noise_refiner = nn.ModuleList(
+            [
+                JointTransformerBlock(
+                    layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    modulation=True,
+                    operation_settings=operation_settings,
+                )
+                for layer_id in range(n_refiner_layers)
+            ]
+        )
+        self.context_refiner = nn.ModuleList(
+            [
+                JointTransformerBlock(
+                    layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    modulation=False,
+                    operation_settings=operation_settings,
+                )
+                for layer_id in range(n_refiner_layers)
+            ]
+        )
+
+        self.t_embedder = TimestepEmbedder(min(dim, 1024), **operation_settings)
+        self.cap_embedder = nn.Sequential(
+            operation_settings.get("operations").RMSNorm(cap_feat_dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+            operation_settings.get("operations").Linear(
+                cap_feat_dim,
+                dim,
+                bias=True,
+                device=operation_settings.get("device"),
+                dtype=operation_settings.get("dtype"),
+            ),
+        )
+
+        self.layers = nn.ModuleList(
+            [
+                JointTransformerBlock(
+                    layer_id,
+                    dim,
+                    n_heads,
+                    n_kv_heads,
+                    multiple_of,
+                    ffn_dim_multiplier,
+                    norm_eps,
+                    qk_norm,
+                    operation_settings=operation_settings,
+                )
+                for layer_id in range(n_layers)
+            ]
+        )
+        self.norm_final = operation_settings.get("operations").RMSNorm(dim, eps=norm_eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.final_layer = FinalLayer(dim, patch_size, self.out_channels, operation_settings=operation_settings)
+
+        assert (dim // n_heads) == sum(axes_dims)
+        self.axes_dims = axes_dims
+        self.axes_lens = axes_lens
+        self.rope_embedder = EmbedND(dim=dim // n_heads, theta=10000.0, axes_dim=axes_dims)
+        self.dim = dim
+        self.n_heads = n_heads
+
+    def unpatchify(
+        self, x: torch.Tensor, img_size: List[Tuple[int, int]], cap_size: List[int], return_tensor=False
+    ) -> List[torch.Tensor]:
+        """
+        x: (N, T, patch_size**2 * C)
+        imgs: (N, H, W, C)
+        """
+        pH = pW = self.patch_size
+        imgs = []
+        for i in range(x.size(0)):
+            H, W = img_size[i]
+            begin = cap_size[i]
+            end = begin + (H // pH) * (W // pW)
+            imgs.append(
+                x[i][begin:end]
+                .view(H // pH, W // pW, pH, pW, self.out_channels)
+                .permute(4, 0, 2, 1, 3)
+                .flatten(3, 4)
+                .flatten(1, 2)
+            )
+
+        if return_tensor:
+            imgs = torch.stack(imgs, dim=0)
+        return imgs
+
+    def patchify_and_embed(
+        self, x: List[torch.Tensor] | torch.Tensor, cap_feats: torch.Tensor, cap_mask: torch.Tensor, t: torch.Tensor, num_tokens
+    ) -> Tuple[torch.Tensor, torch.Tensor, List[Tuple[int, int]], List[int], torch.Tensor]:
+        bsz = len(x)
+        pH = pW = self.patch_size
+        device = x[0].device
+        dtype = x[0].dtype
+
+        if cap_mask is not None:
+            l_effective_cap_len = cap_mask.sum(dim=1).tolist()
+        else:
+            l_effective_cap_len = [num_tokens] * bsz
+
+        if cap_mask is not None and not torch.is_floating_point(cap_mask):
+            cap_mask = (cap_mask - 1).to(dtype) * torch.finfo(dtype).max
+
+        img_sizes = [(img.size(1), img.size(2)) for img in x]
+        l_effective_img_len = [(H // pH) * (W // pW) for (H, W) in img_sizes]
+
+        max_seq_len = max(
+            (cap_len+img_len for cap_len, img_len in zip(l_effective_cap_len, l_effective_img_len))
+        )
+        max_cap_len = max(l_effective_cap_len)
+        max_img_len = max(l_effective_img_len)
+
+        position_ids = torch.zeros(bsz, max_seq_len, 3, dtype=torch.int32, device=device)
+
+        for i in range(bsz):
+            cap_len = l_effective_cap_len[i]
+            img_len = l_effective_img_len[i]
+            H, W = img_sizes[i]
+            H_tokens, W_tokens = H // pH, W // pW
+            assert H_tokens * W_tokens == img_len
+
+            position_ids[i, :cap_len, 0] = torch.arange(cap_len, dtype=torch.int32, device=device)
+            position_ids[i, cap_len:cap_len+img_len, 0] = cap_len
+            row_ids = torch.arange(H_tokens, dtype=torch.int32, device=device).view(-1, 1).repeat(1, W_tokens).flatten()
+            col_ids = torch.arange(W_tokens, dtype=torch.int32, device=device).view(1, -1).repeat(H_tokens, 1).flatten()
+            position_ids[i, cap_len:cap_len+img_len, 1] = row_ids
+            position_ids[i, cap_len:cap_len+img_len, 2] = col_ids
+
+        freqs_cis = self.rope_embedder(position_ids).movedim(1, 2).to(dtype)
+
+        # build freqs_cis for cap and image individually
+        cap_freqs_cis_shape = list(freqs_cis.shape)
+        # cap_freqs_cis_shape[1] = max_cap_len
+        cap_freqs_cis_shape[1] = cap_feats.shape[1]
+        cap_freqs_cis = torch.zeros(*cap_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        img_freqs_cis_shape = list(freqs_cis.shape)
+        img_freqs_cis_shape[1] = max_img_len
+        img_freqs_cis = torch.zeros(*img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        for i in range(bsz):
+            cap_len = l_effective_cap_len[i]
+            img_len = l_effective_img_len[i]
+            cap_freqs_cis[i, :cap_len] = freqs_cis[i, :cap_len]
+            img_freqs_cis[i, :img_len] = freqs_cis[i, cap_len:cap_len+img_len]
+
+        # refine context
+        for layer in self.context_refiner:
+            cap_feats = layer(cap_feats, cap_mask, cap_freqs_cis)
+
+        # refine image
+        flat_x = []
+        for i in range(bsz):
+            img = x[i]
+            C, H, W = img.size()
+            img = img.view(C, H // pH, pH, W // pW, pW).permute(1, 3, 2, 4, 0).flatten(2).flatten(0, 1)
+            flat_x.append(img)
+        x = flat_x
+        padded_img_embed = torch.zeros(bsz, max_img_len, x[0].shape[-1], device=device, dtype=x[0].dtype)
+        padded_img_mask = torch.zeros(bsz, max_img_len, dtype=dtype, device=device)
+        for i in range(bsz):
+            padded_img_embed[i, :l_effective_img_len[i]] = x[i]
+            padded_img_mask[i, l_effective_img_len[i]:] = -torch.finfo(dtype).max
+
+        padded_img_embed = self.x_embedder(padded_img_embed)
+        padded_img_mask = padded_img_mask.unsqueeze(1)
+        for layer in self.noise_refiner:
+            padded_img_embed = layer(padded_img_embed, padded_img_mask, img_freqs_cis, t)
+
+        if cap_mask is not None:
+            mask = torch.zeros(bsz, max_seq_len, dtype=dtype, device=device)
+            mask[:, :max_cap_len] = cap_mask[:, :max_cap_len]
+        else:
+            mask = None
+
+        padded_full_embed = torch.zeros(bsz, max_seq_len, self.dim, device=device, dtype=x[0].dtype)
+        for i in range(bsz):
+            cap_len = l_effective_cap_len[i]
+            img_len = l_effective_img_len[i]
+
+            padded_full_embed[i, :cap_len] = cap_feats[i, :cap_len]
+            padded_full_embed[i, cap_len:cap_len+img_len] = padded_img_embed[i, :img_len]
+
+        return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis
+
+    # def forward(self, x, t, cap_feats, cap_mask):
+    def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
+        t = 1.0 - timesteps
+        cap_feats = context
+        cap_mask = attention_mask
+        bs, c, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+        """
+        Forward pass of NextDiT.
+        t: (N,) tensor of diffusion timesteps
+        y: (N,) tensor of text tokens/features
+        """
+
+        t = self.t_embedder(t, dtype=x.dtype)  # (N, D)
+        adaln_input = t
+
+        cap_feats = self.cap_embedder(cap_feats)  # (N, L, D)  # todo check if able to batchify w.o. redundant compute
+
+        x_is_tensor = isinstance(x, torch.Tensor)
+        x, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, t, num_tokens)
+        freqs_cis = freqs_cis.to(x.device)
+
+        for layer in self.layers:
+            x = layer(x, mask, freqs_cis, adaln_input)
+
+        x = self.final_layer(x, adaln_input)
+        x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)[:,:,:h,:w]
+
+        return -x
+

comfy/ldm/modules/attention.py

@@ -1,4 +1,6 @@
 import math
+import sys
+
 import torch
 import torch.nn.functional as F
 from torch import nn, einsum
@@ -16,7 +18,21 @@ if model_management.xformers_enabled():
     import xformers.ops
 
 if model_management.sage_attention_enabled():
+    try:
         from sageattention import sageattn
+    except ModuleNotFoundError as e:
+        if e.name == "sageattention":
+            logging.error(f"\n\nTo use the `--use-sage-attention` feature, the `sageattention` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install sageattention")
+        else:
+            raise e
+        exit(-1)
+
+if model_management.flash_attention_enabled():
+    try:
+        from flash_attn import flash_attn_func
+    except ModuleNotFoundError:
+        logging.error(f"\n\nTo use the `--use-flash-attention` feature, the `flash-attn` package must be installed first.\ncommand:\n\t{sys.executable} -m pip install flash-attn")
+        exit(-1)
 
 from comfy.cli_args import args
 import comfy.ops
@@ -24,38 +40,24 @@ ops = comfy.ops.disable_weight_init
 
 FORCE_UPCAST_ATTENTION_DTYPE = model_management.force_upcast_attention_dtype()
 
-def get_attn_precision(attn_precision):
+def get_attn_precision(attn_precision, current_dtype):
     if args.dont_upcast_attention:
         return None
-    if FORCE_UPCAST_ATTENTION_DTYPE is not None:
-        return FORCE_UPCAST_ATTENTION_DTYPE
+
+    if FORCE_UPCAST_ATTENTION_DTYPE is not None and current_dtype in FORCE_UPCAST_ATTENTION_DTYPE:
+        return FORCE_UPCAST_ATTENTION_DTYPE[current_dtype]
     return attn_precision
 
 def exists(val):
     return val is not None
 
 
-def uniq(arr):
-    return{el: True for el in arr}.keys()
-
-
 def default(val, d):
     if exists(val):
         return val
     return d
 
 
-def max_neg_value(t):
-    return -torch.finfo(t.dtype).max
-
-
-def init_(tensor):
-    dim = tensor.shape[-1]
-    std = 1 / math.sqrt(dim)
-    tensor.uniform_(-std, std)
-    return tensor
-
-
 # feedforward
 class GEGLU(nn.Module):
     def __init__(self, dim_in, dim_out, dtype=None, device=None, operations=ops):
@@ -90,7 +92,7 @@ def Normalize(in_channels, dtype=None, device=None):
     return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
 
 def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
-    attn_precision = get_attn_precision(attn_precision)
+    attn_precision = get_attn_precision(attn_precision, q.dtype)
 
     if skip_reshape:
         b, _, _, dim_head = q.shape
@@ -159,7 +161,7 @@ def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
 
 
 def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
-    attn_precision = get_attn_precision(attn_precision)
+    attn_precision = get_attn_precision(attn_precision, query.dtype)
 
     if skip_reshape:
         b, _, _, dim_head = query.shape
@@ -229,7 +231,7 @@ def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None,
     return hidden_states
 
 def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
-    attn_precision = get_attn_precision(attn_precision)
+    attn_precision = get_attn_precision(attn_precision, q.dtype)
 
     if skip_reshape:
         b, _, _, dim_head = q.shape
@@ -490,7 +492,17 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
         if mask.ndim == 3:
             mask = mask.unsqueeze(1)
 
+    try:
         out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
+    except Exception as e:
+        logging.error("Error running sage attention: {}, using pytorch attention instead.".format(e))
+        if tensor_layout == "NHD":
+            q, k, v = map(
+                lambda t: t.transpose(1, 2),
+                (q, k, v),
+            )
+        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=True, skip_output_reshape=skip_output_reshape)
+
     if tensor_layout == "HND":
         if not skip_output_reshape:
             out = (
@@ -504,6 +516,63 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
     return out
 
 
+try:
+    @torch.library.custom_op("flash_attention::flash_attn", mutates_args=())
+    def flash_attn_wrapper(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+                    dropout_p: float = 0.0, causal: bool = False) -> torch.Tensor:
+        return flash_attn_func(q, k, v, dropout_p=dropout_p, causal=causal)
+
+
+    @flash_attn_wrapper.register_fake
+    def flash_attn_fake(q, k, v, dropout_p=0.0, causal=False):
+        # Output shape is the same as q
+        return q.new_empty(q.shape)
+except AttributeError as error:
+    FLASH_ATTN_ERROR = error
+
+    def flash_attn_wrapper(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor,
+                    dropout_p: float = 0.0, causal: bool = False) -> torch.Tensor:
+        assert False, f"Could not define flash_attn_wrapper: {FLASH_ATTN_ERROR}"
+
+
+def attention_flash(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
+    if skip_reshape:
+        b, _, _, dim_head = q.shape
+    else:
+        b, _, dim_head = q.shape
+        dim_head //= heads
+        q, k, v = map(
+            lambda t: t.view(b, -1, heads, dim_head).transpose(1, 2),
+            (q, k, v),
+        )
+
+    if mask is not None:
+        # add a batch dimension if there isn't already one
+        if mask.ndim == 2:
+            mask = mask.unsqueeze(0)
+        # add a heads dimension if there isn't already one
+        if mask.ndim == 3:
+            mask = mask.unsqueeze(1)
+
+    try:
+        assert mask is None
+        out = flash_attn_wrapper(
+            q.transpose(1, 2),
+            k.transpose(1, 2),
+            v.transpose(1, 2),
+            dropout_p=0.0,
+            causal=False,
+        ).transpose(1, 2)
+    except Exception as e:
+        logging.warning(f"Flash Attention failed, using default SDPA: {e}")
+        out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+    if not skip_output_reshape:
+        out = (
+            out.transpose(1, 2).reshape(b, -1, heads * dim_head)
+        )
+    return out
+
+
 optimized_attention = attention_basic
 
 if model_management.sage_attention_enabled():
@@ -512,6 +581,9 @@ if model_management.sage_attention_enabled():
 elif model_management.xformers_enabled():
     logging.info("Using xformers attention")
     optimized_attention = attention_xformers
+elif model_management.flash_attention_enabled():
+    logging.info("Using Flash Attention")
+    optimized_attention = attention_flash
 elif model_management.pytorch_attention_enabled():
     logging.info("Using pytorch attention")
     optimized_attention = attention_pytorch
@@ -778,6 +850,7 @@ class SpatialTransformer(nn.Module):
         if not isinstance(context, list):
             context = [context] * len(self.transformer_blocks)
         b, c, h, w = x.shape
+        transformer_options["activations_shape"] = list(x.shape)
         x_in = x
         x = self.norm(x)
         if not self.use_linear:
@@ -893,6 +966,7 @@ class SpatialVideoTransformer(SpatialTransformer):
         transformer_options={}
     ) -> torch.Tensor:
         _, _, h, w = x.shape
+        transformer_options["activations_shape"] = list(x.shape)
         x_in = x
         spatial_context = None
         if exists(context):

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -321,7 +321,7 @@ class SelfAttention(nn.Module):
 
 class RMSNorm(torch.nn.Module):
     def __init__(
-        self, dim: int, elementwise_affine: bool = False, eps: float = 1e-6, device=None, dtype=None
+        self, dim: int, elementwise_affine: bool = False, eps: float = 1e-6, device=None, dtype=None, **kwargs
     ):
         """
         Initialize the RMSNorm normalization layer.

comfy/ldm/modules/diffusionmodules/model.py

@@ -297,7 +297,7 @@ def vae_attention():
     if model_management.xformers_enabled_vae():
         logging.info("Using xformers attention in VAE")
         return xformers_attention
-    elif model_management.pytorch_attention_enabled():
+    elif model_management.pytorch_attention_enabled_vae():
         logging.info("Using pytorch attention in VAE")
         return pytorch_attention
     else:
@@ -702,9 +702,6 @@ class Decoder(nn.Module):
                                         padding=1)
 
     def forward(self, z, **kwargs):
-        #assert z.shape[1:] == self.z_shape[1:]
-        self.last_z_shape = z.shape
-
         # timestep embedding
         temb = None
 

comfy/ldm/wan/model.py

@@ -0,0 +1,793 @@
+# original version: https://github.com/Wan-Video/Wan2.1/blob/main/wan/modules/model.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+import math
+
+import torch
+import torch.nn as nn
+from einops import repeat
+
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.flux.math import apply_rope
+import comfy.ldm.common_dit
+import comfy.model_management
+
+
+def sinusoidal_embedding_1d(dim, position):
+    # preprocess
+    assert dim % 2 == 0
+    half = dim // 2
+    position = position.type(torch.float32)
+
+    # calculation
+    sinusoid = torch.outer(
+        position, torch.pow(10000, -torch.arange(half).to(position).div(half)))
+    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+    return x
+
+
+class WanSelfAttention(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 eps=1e-6, operation_settings={}):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.eps = eps
+
+        # layers
+        self.q = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.k = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.v = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.o = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.norm_q = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+        self.norm_k = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+
+    def forward(self, x, freqs):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L, num_heads, C / num_heads]
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
+
+        # query, key, value function
+        def qkv_fn(x):
+            q = self.norm_q(self.q(x)).view(b, s, n, d)
+            k = self.norm_k(self.k(x)).view(b, s, n, d)
+            v = self.v(x).view(b, s, n * d)
+            return q, k, v
+
+        q, k, v = qkv_fn(x)
+        q, k = apply_rope(q, k, freqs)
+
+        x = optimized_attention(
+            q.view(b, s, n * d),
+            k.view(b, s, n * d),
+            v,
+            heads=self.num_heads,
+        )
+
+        x = self.o(x)
+        return x
+
+
+class WanT2VCrossAttention(WanSelfAttention):
+
+    def forward(self, x, context, **kwargs):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+        """
+        # compute query, key, value
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(context))
+        v = self.v(context)
+
+        # compute attention
+        x = optimized_attention(q, k, v, heads=self.num_heads)
+
+        x = self.o(x)
+        return x
+
+
+class WanI2VCrossAttention(WanSelfAttention):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 eps=1e-6, operation_settings={}):
+        super().__init__(dim, num_heads, window_size, qk_norm, eps, operation_settings=operation_settings)
+
+        self.k_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.v_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        # self.alpha = nn.Parameter(torch.zeros((1, )))
+        self.norm_k_img = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+
+    def forward(self, x, context, context_img_len):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+        """
+        context_img = context[:, :context_img_len]
+        context = context[:, context_img_len:]
+
+        # compute query, key, value
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(context))
+        v = self.v(context)
+        k_img = self.norm_k_img(self.k_img(context_img))
+        v_img = self.v_img(context_img)
+        img_x = optimized_attention(q, k_img, v_img, heads=self.num_heads)
+        # compute attention
+        x = optimized_attention(q, k, v, heads=self.num_heads)
+
+        # output
+        x = x + img_x
+        x = self.o(x)
+        return x
+
+
+WAN_CROSSATTENTION_CLASSES = {
+    't2v_cross_attn': WanT2VCrossAttention,
+    'i2v_cross_attn': WanI2VCrossAttention,
+}
+
+
+class WanAttentionBlock(nn.Module):
+
+    def __init__(self,
+                 cross_attn_type,
+                 dim,
+                 ffn_dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=False,
+                 eps=1e-6, operation_settings={}):
+        super().__init__()
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # layers
+        self.norm1 = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.self_attn = WanSelfAttention(dim, num_heads, window_size, qk_norm,
+                                          eps, operation_settings=operation_settings)
+        self.norm3 = operation_settings.get("operations").LayerNorm(
+            dim, eps,
+            elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if cross_attn_norm else nn.Identity()
+        self.cross_attn = WAN_CROSSATTENTION_CLASSES[cross_attn_type](dim,
+                                                                      num_heads,
+                                                                      (-1, -1),
+                                                                      qk_norm,
+                                                                      eps, operation_settings=operation_settings)
+        self.norm2 = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.ffn = nn.Sequential(
+            operation_settings.get("operations").Linear(dim, ffn_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.GELU(approximate='tanh'),
+            operation_settings.get("operations").Linear(ffn_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        # modulation
+        self.modulation = nn.Parameter(torch.empty(1, 6, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+    def forward(
+        self,
+        x,
+        e,
+        freqs,
+        context,
+        context_img_len=257,
+    ):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L, C]
+            e(Tensor): Shape [B, 6, C]
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        # assert e.dtype == torch.float32
+
+        e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        # assert e[0].dtype == torch.float32
+
+        # self-attention
+        y = self.self_attn(
+            self.norm1(x) * (1 + e[1]) + e[0],
+            freqs)
+
+        x = x + y * e[2]
+
+        # cross-attention & ffn
+        x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len)
+        y = self.ffn(self.norm2(x) * (1 + e[4]) + e[3])
+        x = x + y * e[5]
+        return x
+
+
+class VaceWanAttentionBlock(WanAttentionBlock):
+    def __init__(
+            self,
+            cross_attn_type,
+            dim,
+            ffn_dim,
+            num_heads,
+            window_size=(-1, -1),
+            qk_norm=True,
+            cross_attn_norm=False,
+            eps=1e-6,
+            block_id=0,
+            operation_settings={}
+    ):
+        super().__init__(cross_attn_type, dim, ffn_dim, num_heads, window_size, qk_norm, cross_attn_norm, eps, operation_settings=operation_settings)
+        self.block_id = block_id
+        if block_id == 0:
+            self.before_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.after_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, c, x, **kwargs):
+        if self.block_id == 0:
+            c = self.before_proj(c) + x
+        c = super().forward(c, **kwargs)
+        c_skip = self.after_proj(c)
+        return c_skip, c
+
+
+class WanCamAdapter(nn.Module):
+    def __init__(self, in_dim, out_dim, kernel_size, stride, num_residual_blocks=1, operation_settings={}):
+        super(WanCamAdapter, self).__init__()
+
+        # Pixel Unshuffle: reduce spatial dimensions by a factor of 8
+        self.pixel_unshuffle = nn.PixelUnshuffle(downscale_factor=8)
+
+        # Convolution: reduce spatial dimensions by a factor
+        #  of 2 (without overlap)
+        self.conv = operation_settings.get("operations").Conv2d(in_dim * 64, out_dim, kernel_size=kernel_size, stride=stride, padding=0, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        # Residual blocks for feature extraction
+        self.residual_blocks = nn.Sequential(
+            *[WanCamResidualBlock(out_dim, operation_settings = operation_settings) for _ in range(num_residual_blocks)]
+        )
+
+    def forward(self, x):
+        # Reshape to merge the frame dimension into batch
+        bs, c, f, h, w = x.size()
+        x = x.permute(0, 2, 1, 3, 4).contiguous().view(bs * f, c, h, w)
+
+        # Pixel Unshuffle operation
+        x_unshuffled = self.pixel_unshuffle(x)
+
+        # Convolution operation
+        x_conv = self.conv(x_unshuffled)
+
+        # Feature extraction with residual blocks
+        out = self.residual_blocks(x_conv)
+
+        # Reshape to restore original bf dimension
+        out = out.view(bs, f, out.size(1), out.size(2), out.size(3))
+
+        # Permute dimensions to reorder (if needed), e.g., swap channels and feature frames
+        out = out.permute(0, 2, 1, 3, 4)
+
+        return out
+
+
+class WanCamResidualBlock(nn.Module):
+    def __init__(self, dim, operation_settings={}):
+        super(WanCamResidualBlock, self).__init__()
+        self.conv1 = operation_settings.get("operations").Conv2d(dim, dim, kernel_size=3, padding=1, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = operation_settings.get("operations").Conv2d(dim, dim, kernel_size=3, padding=1, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, x):
+        residual = x
+        out = self.relu(self.conv1(x))
+        out = self.conv2(out)
+        out += residual
+        return out
+
+
+class Head(nn.Module):
+
+    def __init__(self, dim, out_dim, patch_size, eps=1e-6, operation_settings={}):
+        super().__init__()
+        self.dim = dim
+        self.out_dim = out_dim
+        self.patch_size = patch_size
+        self.eps = eps
+
+        # layers
+        out_dim = math.prod(patch_size) * out_dim
+        self.norm = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.head = operation_settings.get("operations").Linear(dim, out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        # modulation
+        self.modulation = nn.Parameter(torch.empty(1, 2, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+    def forward(self, x, e):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            e(Tensor): Shape [B, C]
+        """
+        # assert e.dtype == torch.float32
+        e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e.unsqueeze(1)).chunk(2, dim=1)
+        x = (self.head(self.norm(x) * (1 + e[1]) + e[0]))
+        return x
+
+
+class MLPProj(torch.nn.Module):
+
+    def __init__(self, in_dim, out_dim, flf_pos_embed_token_number=None, operation_settings={}):
+        super().__init__()
+
+        self.proj = torch.nn.Sequential(
+            operation_settings.get("operations").LayerNorm(in_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), operation_settings.get("operations").Linear(in_dim, in_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+            torch.nn.GELU(), operation_settings.get("operations").Linear(in_dim, out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+            operation_settings.get("operations").LayerNorm(out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        if flf_pos_embed_token_number is not None:
+            self.emb_pos = nn.Parameter(torch.empty((1, flf_pos_embed_token_number, in_dim), device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+        else:
+            self.emb_pos = None
+
+    def forward(self, image_embeds):
+        if self.emb_pos is not None:
+            image_embeds = image_embeds[:, :self.emb_pos.shape[1]] + comfy.model_management.cast_to(self.emb_pos[:, :image_embeds.shape[1]], dtype=image_embeds.dtype, device=image_embeds.device)
+
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class WanModel(torch.nn.Module):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='t2v',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 flf_pos_embed_token_number=None,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+        r"""
+        Initialize the diffusion model backbone.
+
+        Args:
+            model_type (`str`, *optional*, defaults to 't2v'):
+                Model variant - 't2v' (text-to-video) or 'i2v' (image-to-video)
+            patch_size (`tuple`, *optional*, defaults to (1, 2, 2)):
+                3D patch dimensions for video embedding (t_patch, h_patch, w_patch)
+            text_len (`int`, *optional*, defaults to 512):
+                Fixed length for text embeddings
+            in_dim (`int`, *optional*, defaults to 16):
+                Input video channels (C_in)
+            dim (`int`, *optional*, defaults to 2048):
+                Hidden dimension of the transformer
+            ffn_dim (`int`, *optional*, defaults to 8192):
+                Intermediate dimension in feed-forward network
+            freq_dim (`int`, *optional*, defaults to 256):
+                Dimension for sinusoidal time embeddings
+            text_dim (`int`, *optional*, defaults to 4096):
+                Input dimension for text embeddings
+            out_dim (`int`, *optional*, defaults to 16):
+                Output video channels (C_out)
+            num_heads (`int`, *optional*, defaults to 16):
+                Number of attention heads
+            num_layers (`int`, *optional*, defaults to 32):
+                Number of transformer blocks
+            window_size (`tuple`, *optional*, defaults to (-1, -1)):
+                Window size for local attention (-1 indicates global attention)
+            qk_norm (`bool`, *optional*, defaults to True):
+                Enable query/key normalization
+            cross_attn_norm (`bool`, *optional*, defaults to False):
+                Enable cross-attention normalization
+            eps (`float`, *optional*, defaults to 1e-6):
+                Epsilon value for normalization layers
+        """
+
+        super().__init__()
+        self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        assert model_type in ['t2v', 'i2v']
+        self.model_type = model_type
+
+        self.patch_size = patch_size
+        self.text_len = text_len
+        self.in_dim = in_dim
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.freq_dim = freq_dim
+        self.text_dim = text_dim
+        self.out_dim = out_dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # embeddings
+        self.patch_embedding = operations.Conv3d(
+            in_dim, dim, kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), dtype=torch.float32)
+        self.text_embedding = nn.Sequential(
+            operations.Linear(text_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.GELU(approximate='tanh'),
+            operations.Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        self.time_embedding = nn.Sequential(
+            operations.Linear(freq_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.SiLU(), operations.Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+        self.time_projection = nn.Sequential(nn.SiLU(), operations.Linear(dim, dim * 6, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        # blocks
+        cross_attn_type = 't2v_cross_attn' if model_type == 't2v' else 'i2v_cross_attn'
+        self.blocks = nn.ModuleList([
+            WanAttentionBlock(cross_attn_type, dim, ffn_dim, num_heads,
+                              window_size, qk_norm, cross_attn_norm, eps, operation_settings=operation_settings)
+            for _ in range(num_layers)
+        ])
+
+        # head
+        self.head = Head(dim, out_dim, patch_size, eps, operation_settings=operation_settings)
+
+        d = dim // num_heads
+        self.rope_embedder = EmbedND(dim=d, theta=10000.0, axes_dim=[d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)])
+
+        if model_type == 'i2v':
+            self.img_emb = MLPProj(1280, dim, flf_pos_embed_token_number=flf_pos_embed_token_number, operation_settings=operation_settings)
+        else:
+            self.img_emb = None
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        clip_fea=None,
+        freqs=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        r"""
+        Forward pass through the diffusion model
+
+        Args:
+            x (Tensor):
+                List of input video tensors with shape [B, C_in, F, H, W]
+            t (Tensor):
+                Diffusion timesteps tensor of shape [B]
+            context (List[Tensor]):
+                List of text embeddings each with shape [B, L, C]
+            seq_len (`int`):
+                Maximum sequence length for positional encoding
+            clip_fea (Tensor, *optional*):
+                CLIP image features for image-to-video mode
+            y (List[Tensor], *optional*):
+                Conditional video inputs for image-to-video mode, same shape as x
+
+        Returns:
+            List[Tensor]:
+                List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8]
+        """
+        # embeddings
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
+        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+
+    def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
+        bs, c, t, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+
+        if time_dim_concat is not None:
+            time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size)
+            x = torch.cat([x, time_dim_concat], dim=2)
+            t_len = ((x.shape[2] + (patch_size[0] // 2)) // patch_size[0])
+
+        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
+        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+
+        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w]
+
+    def unpatchify(self, x, grid_sizes):
+        r"""
+        Reconstruct video tensors from patch embeddings.
+
+        Args:
+            x (List[Tensor]):
+                List of patchified features, each with shape [L, C_out * prod(patch_size)]
+            grid_sizes (Tensor):
+                Original spatial-temporal grid dimensions before patching,
+                    shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches)
+
+        Returns:
+            List[Tensor]:
+                Reconstructed video tensors with shape [L, C_out, F, H / 8, W / 8]
+        """
+
+        c = self.out_dim
+        u = x
+        b = u.shape[0]
+        u = u[:, :math.prod(grid_sizes)].view(b, *grid_sizes, *self.patch_size, c)
+        u = torch.einsum('bfhwpqrc->bcfphqwr', u)
+        u = u.reshape(b, c, *[i * j for i, j in zip(grid_sizes, self.patch_size)])
+        return u
+
+
+class VaceWanModel(WanModel):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='vace',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 flf_pos_embed_token_number=None,
+                 image_model=None,
+                 vace_layers=None,
+                 vace_in_dim=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+
+        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        # Vace
+        if vace_layers is not None:
+            self.vace_layers = vace_layers
+            self.vace_in_dim = vace_in_dim
+            # vace blocks
+            self.vace_blocks = nn.ModuleList([
+                VaceWanAttentionBlock('t2v_cross_attn', self.dim, self.ffn_dim, self.num_heads, self.window_size, self.qk_norm, self.cross_attn_norm, self.eps, block_id=i, operation_settings=operation_settings)
+                for i in range(self.vace_layers)
+            ])
+
+            self.vace_layers_mapping = {i: n for n, i in enumerate(range(0, self.num_layers, self.num_layers // self.vace_layers))}
+            # vace patch embeddings
+            self.vace_patch_embedding = operations.Conv3d(
+                self.vace_in_dim, self.dim, kernel_size=self.patch_size, stride=self.patch_size, device=device, dtype=torch.float32
+            )
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        vace_context,
+        vace_strength,
+        clip_fea=None,
+        freqs=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        # embeddings
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
+        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        orig_shape = list(vace_context.shape)
+        vace_context = vace_context.movedim(0, 1).reshape([-1] + orig_shape[2:])
+        c = self.vace_patch_embedding(vace_context.float()).to(vace_context.dtype)
+        c = c.flatten(2).transpose(1, 2)
+        c = list(c.split(orig_shape[0], dim=0))
+
+        # arguments
+        x_orig = x
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+
+            ii = self.vace_layers_mapping.get(i, None)
+            if ii is not None:
+                for iii in range(len(c)):
+                    c_skip, c[iii] = self.vace_blocks[ii](c[iii], x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+                    x += c_skip * vace_strength[iii]
+                del c_skip
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+
+class CameraWanModel(WanModel):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='camera',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 flf_pos_embed_token_number=None,
+                 image_model=None,
+                 in_dim_control_adapter=24,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+
+        super().__init__(model_type='i2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        self.control_adapter = WanCamAdapter(in_dim_control_adapter, dim, kernel_size=patch_size[1:], stride=patch_size[1:], operation_settings=operation_settings)
+
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        clip_fea=None,
+        freqs=None,
+        camera_conditions = None,
+        transformer_options={},
+        **kwargs,
+    ):
+        # embeddings
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        if self.control_adapter is not None and camera_conditions is not None:
+            x_camera = self.control_adapter(camera_conditions).to(x.dtype)
+            x = x + x_camera
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
+        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x

comfy/ldm/wan/vae.py

@@ -0,0 +1,567 @@
+# original version: https://github.com/Wan-Video/Wan2.1/blob/main/wan/modules/vae.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from comfy.ldm.modules.diffusionmodules.model import vae_attention
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+CACHE_T = 2
+
+
+class CausalConv3d(ops.Conv3d):
+    """
+    Causal 3d convolusion.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._padding = (self.padding[2], self.padding[2], self.padding[1],
+                         self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
+
+    def forward(self, x, cache_x=None):
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+        x = F.pad(x, padding)
+
+        return super().forward(x)
+
+
+class RMS_norm(nn.Module):
+
+    def __init__(self, dim, channel_first=True, images=True, bias=False):
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else None
+
+    def forward(self, x):
+        return F.normalize(
+            x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma.to(x) + (self.bias.to(x) if self.bias is not None else 0)
+
+
+class Upsample(nn.Upsample):
+
+    def forward(self, x):
+        """
+        Fix bfloat16 support for nearest neighbor interpolation.
+        """
+        return super().forward(x.float()).type_as(x)
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in ('none', 'upsample2d', 'upsample3d', 'downsample2d',
+                        'downsample3d')
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == 'upsample2d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                ops.Conv2d(dim, dim // 2, 3, padding=1))
+        elif mode == 'upsample3d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                ops.Conv2d(dim, dim // 2, 3, padding=1))
+            self.time_conv = CausalConv3d(
+                dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+
+        elif mode == 'downsample2d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == 'downsample3d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == 'upsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = 'Rep'
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] != 'Rep':
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat([
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device), cache_x
+                        ],
+                                            dim=2)
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] == 'Rep':
+                        cache_x = torch.cat([
+                            torch.zeros_like(cache_x).to(cache_x.device),
+                            cache_x
+                        ],
+                                            dim=2)
+                    if feat_cache[idx] == 'Rep':
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.resample(x)
+        x = rearrange(x, '(b t) c h w -> b c t h w', t=t)
+
+        if self.mode == 'downsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    # if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx]!='Rep':
+                    #     # cache last frame of last two chunk
+                    #     cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+    def init_weight(self, conv):
+        conv_weight = conv.weight
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        one_matrix = torch.eye(c1, c2)
+        init_matrix = one_matrix
+        nn.init.zeros_(conv_weight)
+        #conv_weight.data[:,:,-1,1,1] = init_matrix * 0.5
+        conv_weight.data[:, :, 1, 0, 0] = init_matrix  #* 0.5
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+    def init_weight2(self, conv):
+        conv_weight = conv.weight.data
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        init_matrix = torch.eye(c1 // 2, c2)
+        #init_matrix = repeat(init_matrix, 'o ... -> (o 2) ...').permute(1,0,2).contiguous().reshape(c1,c2)
+        conv_weight[:c1 // 2, :, -1, 0, 0] = init_matrix
+        conv_weight[c1 // 2:, :, -1, 0, 0] = init_matrix
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False), nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False), nn.SiLU(), nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1))
+        self.shortcut = CausalConv3d(in_dim, out_dim, 1) \
+            if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        h = self.shortcut(x)
+        for layer in self.residual:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    Causal self-attention with a single head.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+        # layers
+        self.norm = RMS_norm(dim)
+        self.to_qkv = ops.Conv2d(dim, dim * 3, 1)
+        self.proj = ops.Conv2d(dim, dim, 1)
+        self.optimized_attention = vae_attention()
+
+    def forward(self, x):
+        identity = x
+        b, c, t, h, w = x.size()
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.norm(x)
+        # compute query, key, value
+
+        q, k, v = self.to_qkv(x).chunk(3, dim=1)
+        x = self.optimized_attention(q, k, v)
+
+        # output
+        x = self.proj(x)
+        x = rearrange(x, '(b t) c h w-> b c t h w', t=t)
+        return x + identity
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(3, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            for _ in range(num_res_blocks):
+                downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    downsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # downsample block
+            if i != len(dim_mult) - 1:
+                mode = 'downsample3d' if temperal_downsample[
+                    i] else 'downsample2d'
+                downsamples.append(Resample(out_dim, mode=mode))
+                scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout), AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout))
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False), nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2**(len(dim_mult) - 2)
+
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(dims[0], dims[0], dropout), AttentionBlock(dims[0]),
+            ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            if i == 1 or i == 2 or i == 3:
+                in_dim = in_dim // 2
+            for _ in range(num_res_blocks + 1):
+                upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    upsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # upsample block
+            if i != len(dim_mult) - 1:
+                mode = 'upsample3d' if temperal_upsample[i] else 'upsample2d'
+                upsamples.append(Resample(out_dim, mode=mode))
+                scale *= 2.0
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False), nn.SiLU(),
+            CausalConv3d(out_dim, 3, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        ## conv1
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class WanVAE(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_upsample, dropout)
+
+    def forward(self, x):
+        mu, log_var = self.encode(x)
+        z = self.reparameterize(mu, log_var)
+        x_recon = self.decode(z)
+        return x_recon, mu, log_var
+
+    def encode(self, x):
+        self.clear_cache()
+        ## cache
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        ## 对encode输入的x，按时间拆分为1、4、4、4....
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(
+                    x[:, :, :1, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        self.clear_cache()
+        return mu
+
+    def decode(self, z):
+        self.clear_cache()
+        # z: [b,c,t,h,w]
+
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx)
+            else:
+                out_ = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)
+        self.clear_cache()
+        return out
+
+    def reparameterize(self, mu, log_var):
+        std = torch.exp(0.5 * log_var)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample(self, imgs, deterministic=False):
+        mu, log_var = self.encode(imgs)
+        if deterministic:
+            return mu
+        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+        return mu + std * torch.randn_like(std)
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        #cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num

comfy/lora.py

@@ -20,6 +20,7 @@ from __future__ import annotations
 import comfy.utils
 import comfy.model_management
 import comfy.model_base
+import comfy.weight_adapter as weight_adapter
 import logging
 import torch
 
@@ -49,139 +50,12 @@ def load_lora(lora, to_load, log_missing=True):
             dora_scale = lora[dora_scale_name]
             loaded_keys.add(dora_scale_name)
 
-        reshape_name = "{}.reshape_weight".format(x)
-        reshape = None
-        if reshape_name in lora.keys():
-            try:
-                reshape = lora[reshape_name].tolist()
-                loaded_keys.add(reshape_name)
-            except:
-                pass
-
-        regular_lora = "{}.lora_up.weight".format(x)
-        diffusers_lora = "{}_lora.up.weight".format(x)
-        diffusers2_lora = "{}.lora_B.weight".format(x)
-        diffusers3_lora = "{}.lora.up.weight".format(x)
-        mochi_lora = "{}.lora_B".format(x)
-        transformers_lora = "{}.lora_linear_layer.up.weight".format(x)
-        A_name = None
-
-        if regular_lora in lora.keys():
-            A_name = regular_lora
-            B_name = "{}.lora_down.weight".format(x)
-            mid_name = "{}.lora_mid.weight".format(x)
-        elif diffusers_lora in lora.keys():
-            A_name = diffusers_lora
-            B_name = "{}_lora.down.weight".format(x)
-            mid_name = None
-        elif diffusers2_lora in lora.keys():
-            A_name = diffusers2_lora
-            B_name = "{}.lora_A.weight".format(x)
-            mid_name = None
-        elif diffusers3_lora in lora.keys():
-            A_name = diffusers3_lora
-            B_name = "{}.lora.down.weight".format(x)
-            mid_name = None
-        elif mochi_lora in lora.keys():
-            A_name = mochi_lora
-            B_name = "{}.lora_A".format(x)
-            mid_name = None
-        elif transformers_lora in lora.keys():
-            A_name = transformers_lora
-            B_name ="{}.lora_linear_layer.down.weight".format(x)
-            mid_name = None
-
-        if A_name is not None:
-            mid = None
-            if mid_name is not None and mid_name in lora.keys():
-                mid = lora[mid_name]
-                loaded_keys.add(mid_name)
-            patch_dict[to_load[x]] = ("lora", (lora[A_name], lora[B_name], alpha, mid, dora_scale, reshape))
-            loaded_keys.add(A_name)
-            loaded_keys.add(B_name)
-
-
-        ######## loha
-        hada_w1_a_name = "{}.hada_w1_a".format(x)
-        hada_w1_b_name = "{}.hada_w1_b".format(x)
-        hada_w2_a_name = "{}.hada_w2_a".format(x)
-        hada_w2_b_name = "{}.hada_w2_b".format(x)
-        hada_t1_name = "{}.hada_t1".format(x)
-        hada_t2_name = "{}.hada_t2".format(x)
-        if hada_w1_a_name in lora.keys():
-            hada_t1 = None
-            hada_t2 = None
-            if hada_t1_name in lora.keys():
-                hada_t1 = lora[hada_t1_name]
-                hada_t2 = lora[hada_t2_name]
-                loaded_keys.add(hada_t1_name)
-                loaded_keys.add(hada_t2_name)
-
-            patch_dict[to_load[x]] = ("loha", (lora[hada_w1_a_name], lora[hada_w1_b_name], alpha, lora[hada_w2_a_name], lora[hada_w2_b_name], hada_t1, hada_t2, dora_scale))
-            loaded_keys.add(hada_w1_a_name)
-            loaded_keys.add(hada_w1_b_name)
-            loaded_keys.add(hada_w2_a_name)
-            loaded_keys.add(hada_w2_b_name)
-
-
-        ######## lokr
-        lokr_w1_name = "{}.lokr_w1".format(x)
-        lokr_w2_name = "{}.lokr_w2".format(x)
-        lokr_w1_a_name = "{}.lokr_w1_a".format(x)
-        lokr_w1_b_name = "{}.lokr_w1_b".format(x)
-        lokr_t2_name = "{}.lokr_t2".format(x)
-        lokr_w2_a_name = "{}.lokr_w2_a".format(x)
-        lokr_w2_b_name = "{}.lokr_w2_b".format(x)
-
-        lokr_w1 = None
-        if lokr_w1_name in lora.keys():
-            lokr_w1 = lora[lokr_w1_name]
-            loaded_keys.add(lokr_w1_name)
-
-        lokr_w2 = None
-        if lokr_w2_name in lora.keys():
-            lokr_w2 = lora[lokr_w2_name]
-            loaded_keys.add(lokr_w2_name)
-
-        lokr_w1_a = None
-        if lokr_w1_a_name in lora.keys():
-            lokr_w1_a = lora[lokr_w1_a_name]
-            loaded_keys.add(lokr_w1_a_name)
-
-        lokr_w1_b = None
-        if lokr_w1_b_name in lora.keys():
-            lokr_w1_b = lora[lokr_w1_b_name]
-            loaded_keys.add(lokr_w1_b_name)
-
-        lokr_w2_a = None
-        if lokr_w2_a_name in lora.keys():
-            lokr_w2_a = lora[lokr_w2_a_name]
-            loaded_keys.add(lokr_w2_a_name)
-
-        lokr_w2_b = None
-        if lokr_w2_b_name in lora.keys():
-            lokr_w2_b = lora[lokr_w2_b_name]
-            loaded_keys.add(lokr_w2_b_name)
-
-        lokr_t2 = None
-        if lokr_t2_name in lora.keys():
-            lokr_t2 = lora[lokr_t2_name]
-            loaded_keys.add(lokr_t2_name)
-
-        if (lokr_w1 is not None) or (lokr_w2 is not None) or (lokr_w1_a is not None) or (lokr_w2_a is not None):
-            patch_dict[to_load[x]] = ("lokr", (lokr_w1, lokr_w2, alpha, lokr_w1_a, lokr_w1_b, lokr_w2_a, lokr_w2_b, lokr_t2, dora_scale))
-
-        #glora
-        a1_name = "{}.a1.weight".format(x)
-        a2_name = "{}.a2.weight".format(x)
-        b1_name = "{}.b1.weight".format(x)
-        b2_name = "{}.b2.weight".format(x)
-        if a1_name in lora:
-            patch_dict[to_load[x]] = ("glora", (lora[a1_name], lora[a2_name], lora[b1_name], lora[b2_name], alpha, dora_scale))
-            loaded_keys.add(a1_name)
-            loaded_keys.add(a2_name)
-            loaded_keys.add(b1_name)
-            loaded_keys.add(b2_name)
+        for adapter_cls in weight_adapter.adapters:
+            adapter = adapter_cls.load(x, lora, alpha, dora_scale, loaded_keys)
+            if adapter is not None:
+                patch_dict[to_load[x]] = adapter
+                loaded_keys.update(adapter.loaded_keys)
+                continue
 
         w_norm_name = "{}.w_norm".format(x)
         b_norm_name = "{}.b_norm".format(x)
@@ -307,7 +181,6 @@ def model_lora_keys_unet(model, key_map={}):
             if k.endswith(".weight"):
                 key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
                 key_map["lora_unet_{}".format(key_lora)] = k
-                key_map["lora_prior_unet_{}".format(key_lora)] = k #cascade lora: TODO put lora key prefix in the model config
                 key_map["{}".format(k[:-len(".weight")])] = k #generic lora format without any weird key names
             else:
                 key_map["{}".format(k)] = k #generic lora format for not .weight without any weird key names
@@ -327,6 +200,13 @@ def model_lora_keys_unet(model, key_map={}):
                     diffusers_lora_key = diffusers_lora_key[:-2]
                 key_map[diffusers_lora_key] = unet_key
 
+    if isinstance(model, comfy.model_base.StableCascade_C):
+        for k in sdk:
+            if k.startswith("diffusion_model."):
+                if k.endswith(".weight"):
+                    key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
+                    key_map["lora_prior_unet_{}".format(key_lora)] = k
+
     if isinstance(model, comfy.model_base.SD3): #Diffusers lora SD3
         diffusers_keys = comfy.utils.mmdit_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
         for k in diffusers_keys:
@@ -399,29 +279,23 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["transformer.{}".format(key_lora)] = k
                 key_map["diffusion_model.{}".format(key_lora)] = k  # Old loras
 
+    if isinstance(model, comfy.model_base.HiDream):
+        for k in sdk:
+            if k.startswith("diffusion_model."):
+                if k.endswith(".weight"):
+                    key_lora = k[len("diffusion_model."):-len(".weight")]
+                    key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format
+                    key_map["transformer.{}".format(key_lora)] = k #SimpleTuner regular format
+
+    if isinstance(model, comfy.model_base.ACEStep):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"): #Official ACE step lora format
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                key_map["{}".format(key_lora)] = k
+
     return key_map
 
 
-def weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function):
-    dora_scale = comfy.model_management.cast_to_device(dora_scale, weight.device, intermediate_dtype)
-    lora_diff *= alpha
-    weight_calc = weight + function(lora_diff).type(weight.dtype)
-    weight_norm = (
-        weight_calc.transpose(0, 1)
-        .reshape(weight_calc.shape[1], -1)
-        .norm(dim=1, keepdim=True)
-        .reshape(weight_calc.shape[1], *[1] * (weight_calc.dim() - 1))
-        .transpose(0, 1)
-    )
-
-    weight_calc *= (dora_scale / weight_norm).type(weight.dtype)
-    if strength != 1.0:
-        weight_calc -= weight
-        weight += strength * (weight_calc)
-    else:
-        weight[:] = weight_calc
-    return weight
-
 def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Tensor:
     """
     Pad a tensor to a new shape with zeros.
@@ -476,6 +350,16 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
         if isinstance(v, list):
             v = (calculate_weight(v[1:], v[0][1](comfy.model_management.cast_to_device(v[0][0], weight.device, intermediate_dtype, copy=True), inplace=True), key, intermediate_dtype=intermediate_dtype), )
 
+        if isinstance(v, weight_adapter.WeightAdapterBase):
+            output = v.calculate_weight(weight, key, strength, strength_model, offset, function, intermediate_dtype, original_weights)
+            if output is None:
+                logging.warning("Calculate Weight Failed: {} {}".format(v.name, key))
+            else:
+                weight = output
+                if old_weight is not None:
+                    weight = old_weight
+            continue
+
         if len(v) == 1:
             patch_type = "diff"
         elif len(v) == 2:
@@ -502,157 +386,6 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
             diff_weight = comfy.model_management.cast_to_device(target_weight, weight.device, intermediate_dtype) - \
                           comfy.model_management.cast_to_device(original_weights[key][0][0], weight.device, intermediate_dtype)
             weight += function(strength * comfy.model_management.cast_to_device(diff_weight, weight.device, weight.dtype))
-        elif patch_type == "lora": #lora/locon
-            mat1 = comfy.model_management.cast_to_device(v[0], weight.device, intermediate_dtype)
-            mat2 = comfy.model_management.cast_to_device(v[1], weight.device, intermediate_dtype)
-            dora_scale = v[4]
-            reshape = v[5]
-
-            if reshape is not None:
-                weight = pad_tensor_to_shape(weight, reshape)
-
-            if v[2] is not None:
-                alpha = v[2] / mat2.shape[0]
-            else:
-                alpha = 1.0
-
-            if v[3] is not None:
-                #locon mid weights, hopefully the math is fine because I didn't properly test it
-                mat3 = comfy.model_management.cast_to_device(v[3], weight.device, intermediate_dtype)
-                final_shape = [mat2.shape[1], mat2.shape[0], mat3.shape[2], mat3.shape[3]]
-                mat2 = torch.mm(mat2.transpose(0, 1).flatten(start_dim=1), mat3.transpose(0, 1).flatten(start_dim=1)).reshape(final_shape).transpose(0, 1)
-            try:
-                lora_diff = torch.mm(mat1.flatten(start_dim=1), mat2.flatten(start_dim=1)).reshape(weight.shape)
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
-        elif patch_type == "lokr":
-            w1 = v[0]
-            w2 = v[1]
-            w1_a = v[3]
-            w1_b = v[4]
-            w2_a = v[5]
-            w2_b = v[6]
-            t2 = v[7]
-            dora_scale = v[8]
-            dim = None
-
-            if w1 is None:
-                dim = w1_b.shape[0]
-                w1 = torch.mm(comfy.model_management.cast_to_device(w1_a, weight.device, intermediate_dtype),
-                                comfy.model_management.cast_to_device(w1_b, weight.device, intermediate_dtype))
-            else:
-                w1 = comfy.model_management.cast_to_device(w1, weight.device, intermediate_dtype)
-
-            if w2 is None:
-                dim = w2_b.shape[0]
-                if t2 is None:
-                    w2 = torch.mm(comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype))
-                else:
-                    w2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                        comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
-                                        comfy.model_management.cast_to_device(w2_b, weight.device, intermediate_dtype),
-                                        comfy.model_management.cast_to_device(w2_a, weight.device, intermediate_dtype))
-            else:
-                w2 = comfy.model_management.cast_to_device(w2, weight.device, intermediate_dtype)
-
-            if len(w2.shape) == 4:
-                w1 = w1.unsqueeze(2).unsqueeze(2)
-            if v[2] is not None and dim is not None:
-                alpha = v[2] / dim
-            else:
-                alpha = 1.0
-
-            try:
-                lora_diff = torch.kron(w1, w2).reshape(weight.shape)
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
-        elif patch_type == "loha":
-            w1a = v[0]
-            w1b = v[1]
-            if v[2] is not None:
-                alpha = v[2] / w1b.shape[0]
-            else:
-                alpha = 1.0
-
-            w2a = v[3]
-            w2b = v[4]
-            dora_scale = v[7]
-            if v[5] is not None: #cp decomposition
-                t1 = v[5]
-                t2 = v[6]
-                m1 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                    comfy.model_management.cast_to_device(t1, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype))
-
-                m2 = torch.einsum('i j k l, j r, i p -> p r k l',
-                                    comfy.model_management.cast_to_device(t2, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype),
-                                    comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype))
-            else:
-                m1 = torch.mm(comfy.model_management.cast_to_device(w1a, weight.device, intermediate_dtype),
-                                comfy.model_management.cast_to_device(w1b, weight.device, intermediate_dtype))
-                m2 = torch.mm(comfy.model_management.cast_to_device(w2a, weight.device, intermediate_dtype),
-                                comfy.model_management.cast_to_device(w2b, weight.device, intermediate_dtype))
-
-            try:
-                lora_diff = (m1 * m2).reshape(weight.shape)
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
-        elif patch_type == "glora":
-            dora_scale = v[5]
-
-            old_glora = False
-            if v[3].shape[1] == v[2].shape[0] == v[0].shape[0] == v[1].shape[1]:
-                rank = v[0].shape[0]
-                old_glora = True
-
-            if v[3].shape[0] == v[2].shape[1] == v[0].shape[1] == v[1].shape[0]:
-                if old_glora and v[1].shape[0] == weight.shape[0] and weight.shape[0] == weight.shape[1]:
-                    pass
-                else:
-                    old_glora = False
-                    rank = v[1].shape[0]
-
-            a1 = comfy.model_management.cast_to_device(v[0].flatten(start_dim=1), weight.device, intermediate_dtype)
-            a2 = comfy.model_management.cast_to_device(v[1].flatten(start_dim=1), weight.device, intermediate_dtype)
-            b1 = comfy.model_management.cast_to_device(v[2].flatten(start_dim=1), weight.device, intermediate_dtype)
-            b2 = comfy.model_management.cast_to_device(v[3].flatten(start_dim=1), weight.device, intermediate_dtype)
-
-            if v[4] is not None:
-                alpha = v[4] / rank
-            else:
-                alpha = 1.0
-
-            try:
-                if old_glora:
-                    lora_diff = (torch.mm(b2, b1) + torch.mm(torch.mm(weight.flatten(start_dim=1).to(dtype=intermediate_dtype), a2), a1)).reshape(weight.shape) #old lycoris glora
-                else:
-                    if weight.dim() > 2:
-                        lora_diff = torch.einsum("o i ..., i j -> o j ...", torch.einsum("o i ..., i j -> o j ...", weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
-                    else:
-                        lora_diff = torch.mm(torch.mm(weight.to(dtype=intermediate_dtype), a1), a2).reshape(weight.shape)
-                    lora_diff += torch.mm(b1, b2).reshape(weight.shape)
-
-                if dora_scale is not None:
-                    weight = weight_decompose(dora_scale, weight, lora_diff, alpha, strength, intermediate_dtype, function)
-                else:
-                    weight += function(((strength * alpha) * lora_diff).type(weight.dtype))
-            except Exception as e:
-                logging.error("ERROR {} {} {}".format(patch_type, key, e))
         else:
             logging.warning("patch type not recognized {} {}".format(patch_type, key))
 

comfy/lora_convert.py

@@ -1,4 +1,5 @@
 import torch
+import comfy.utils
 
 
 def convert_lora_bfl_control(sd): #BFL loras for Flux
@@ -11,7 +12,13 @@ def convert_lora_bfl_control(sd): #BFL loras for Flux
     return sd_out
 
 
+def convert_lora_wan_fun(sd): #Wan Fun loras
+    return comfy.utils.state_dict_prefix_replace(sd, {"lora_unet__": "lora_unet_"})
+
+
 def convert_lora(sd):
     if "img_in.lora_A.weight" in sd and "single_blocks.0.norm.key_norm.scale" in sd:
         return convert_lora_bfl_control(sd)
+    if "lora_unet__blocks_0_cross_attn_k.lora_down.weight" in sd:
+        return convert_lora_wan_fun(sd)
     return sd

comfy/model_base.py

@@ -34,6 +34,12 @@ import comfy.ldm.flux.model
 import comfy.ldm.lightricks.model
 import comfy.ldm.hunyuan_video.model
 import comfy.ldm.cosmos.model
+import comfy.ldm.lumina.model
+import comfy.ldm.wan.model
+import comfy.ldm.hunyuan3d.model
+import comfy.ldm.hidream.model
+import comfy.ldm.chroma.model
+import comfy.ldm.ace.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -56,6 +62,7 @@ class ModelType(Enum):
     FLOW = 6
     V_PREDICTION_CONTINUOUS = 7
     FLUX = 8
+    IMG_TO_IMG = 9
 
 
 from comfy.model_sampling import EPS, V_PREDICTION, EDM, ModelSamplingDiscrete, ModelSamplingContinuousEDM, StableCascadeSampling, ModelSamplingContinuousV
@@ -86,6 +93,8 @@ def model_sampling(model_config, model_type):
     elif model_type == ModelType.FLUX:
         c = comfy.model_sampling.CONST
         s = comfy.model_sampling.ModelSamplingFlux
+    elif model_type == ModelType.IMG_TO_IMG:
+        c = comfy.model_sampling.IMG_TO_IMG
 
     class ModelSampling(s, c):
         pass
@@ -106,7 +115,7 @@ class BaseModel(torch.nn.Module):
 
         if not unet_config.get("disable_unet_model_creation", False):
             if model_config.custom_operations is None:
-                fp8 = model_config.optimizations.get("fp8", model_config.scaled_fp8 is not None)
+                fp8 = model_config.optimizations.get("fp8", False)
                 operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8)
             else:
                 operations = model_config.custom_operations
@@ -126,6 +135,7 @@ class BaseModel(torch.nn.Module):
         logging.info("model_type {}".format(model_type.name))
         logging.debug("adm {}".format(self.adm_channels))
         self.memory_usage_factor = model_config.memory_usage_factor
+        self.memory_usage_factor_conds = ()
 
     def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
@@ -137,6 +147,7 @@ class BaseModel(torch.nn.Module):
     def _apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         sigma = t
         xc = self.model_sampling.calculate_input(sigma, x)
+
         if c_concat is not None:
             xc = torch.cat([xc] + [c_concat], dim=1)
 
@@ -148,7 +159,9 @@ class BaseModel(torch.nn.Module):
 
         xc = xc.to(dtype)
         t = self.model_sampling.timestep(t).float()
+        if context is not None:
             context = context.to(dtype)
+
         extra_conds = {}
         for o in kwargs:
             extra = kwargs[o]
@@ -157,15 +170,16 @@ class BaseModel(torch.nn.Module):
                     extra = extra.to(dtype)
             extra_conds[o] = extra
 
+        t = self.process_timestep(t, x=x, **extra_conds)
         model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float()
         return self.model_sampling.calculate_denoised(sigma, model_output, x)
 
+    def process_timestep(self, timestep, **kwargs):
+        return timestep
+
     def get_dtype(self):
         return self.diffusion_model.dtype
 
-    def is_adm(self):
-        return self.adm_channels > 0
-
     def encode_adm(self, **kwargs):
         return None
 
@@ -184,6 +198,11 @@ class BaseModel(torch.nn.Module):
 
             if concat_latent_image.shape[1:] != noise.shape[1:]:
                 concat_latent_image = utils.common_upscale(concat_latent_image, noise.shape[-1], noise.shape[-2], "bilinear", "center")
+                if noise.ndim == 5:
+                    if concat_latent_image.shape[-3] < noise.shape[-3]:
+                        concat_latent_image = torch.nn.functional.pad(concat_latent_image, (0, 0, 0, 0, 0, noise.shape[-3] - concat_latent_image.shape[-3]), "constant", 0)
+                    else:
+                        concat_latent_image = concat_latent_image[:, :, :noise.shape[-3]]
 
             concat_latent_image = utils.resize_to_batch_size(concat_latent_image, noise.shape[0])
 
@@ -212,6 +231,11 @@ class BaseModel(torch.nn.Module):
                         cond_concat.append(self.blank_inpaint_image_like(noise))
                     elif ck == "mask_inverted":
                         cond_concat.append(torch.zeros_like(noise)[:, :1])
+                if ck == "concat_image":
+                    if concat_latent_image is not None:
+                        cond_concat.append(concat_latent_image.to(device))
+                    else:
+                        cond_concat.append(torch.zeros_like(noise))
             data = torch.cat(cond_concat, dim=1)
             return data
         return None
@@ -302,19 +326,28 @@ class BaseModel(torch.nn.Module):
     def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
         return self.model_sampling.noise_scaling(sigma.reshape([sigma.shape[0]] + [1] * (len(noise.shape) - 1)), noise, latent_image)
 
-    def memory_required(self, input_shape):
+    def memory_required(self, input_shape, cond_shapes={}):
+        input_shapes = [input_shape]
+        for c in self.memory_usage_factor_conds:
+            shape = cond_shapes.get(c, None)
+            if shape is not None and len(shape) > 0:
+                input_shapes += shape
+
         if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
             dtype = self.get_dtype()
             if self.manual_cast_dtype is not None:
                 dtype = self.manual_cast_dtype
             #TODO: this needs to be tweaked
-            area = input_shape[0] * math.prod(input_shape[2:])
+            area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes))
             return (area * comfy.model_management.dtype_size(dtype) * 0.01 * self.memory_usage_factor) * (1024 * 1024)
         else:
             #TODO: this formula might be too aggressive since I tweaked the sub-quad and split algorithms to use less memory.
-            area = input_shape[0] * math.prod(input_shape[2:])
+            area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes))
             return (area * 0.15 * self.memory_usage_factor) * (1024 * 1024)
 
+    def extra_conds_shapes(self, **kwargs):
+        return {}
+
 
 def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0, seed=None):
     adm_inputs = []
@@ -549,6 +582,10 @@ class SD_X4Upscaler(BaseModel):
 
         out['c_concat'] = comfy.conds.CONDNoiseShape(image)
         out['y'] = comfy.conds.CONDRegular(noise_level)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
         return out
 
 class IP2P:
@@ -581,6 +618,19 @@ class SDXL_instructpix2pix(IP2P, SDXL):
         else:
             self.process_ip2p_image_in = lambda image: image #diffusers ip2p
 
+class Lotus(BaseModel):
+    def extra_conds(self, **kwargs):
+        out = {}
+        cross_attn = kwargs.get("cross_attn", None)
+        out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
+        device = kwargs["device"]
+        task_emb = torch.tensor([1, 0]).float().to(device)
+        task_emb = torch.cat([torch.sin(task_emb), torch.cos(task_emb)]).unsqueeze(0)
+        out['y'] = comfy.conds.CONDRegular(task_emb)
+        return out
+
+    def __init__(self, model_config, model_type=ModelType.IMG_TO_IMG, device=None):
+        super().__init__(model_config, model_type, device=device)
 
 class StableCascade_C(BaseModel):
     def __init__(self, model_config, model_type=ModelType.STABLE_CASCADE, device=None):
@@ -748,8 +798,8 @@ class PixArt(BaseModel):
         return out
 
 class Flux(BaseModel):
-    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
-        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.flux.model.Flux)
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None, unet_model=comfy.ldm.flux.model.Flux):
+        super().__init__(model_config, model_type, device=device, unet_model=unet_model)
 
     def concat_cond(self, **kwargs):
         try:
@@ -806,7 +856,10 @@ class Flux(BaseModel):
             (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size))
             attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok))
             out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
-        out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([kwargs.get("guidance", 3.5)]))
+
+        guidance = kwargs.get("guidance", 3.5)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
         return out
 
 class GenmoMochi(BaseModel):
@@ -837,17 +890,26 @@ class LTXV(BaseModel):
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
 
-        guiding_latent = kwargs.get("guiding_latent", None)
-        if guiding_latent is not None:
-            out['guiding_latent'] = comfy.conds.CONDRegular(guiding_latent)
-
-        guiding_latent_noise_scale = kwargs.get("guiding_latent_noise_scale", None)
-        if guiding_latent_noise_scale is not None:
-            out["guiding_latent_noise_scale"] = comfy.conds.CONDConstant(guiding_latent_noise_scale)
-
         out['frame_rate'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", 25))
+
+        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if denoise_mask is not None:
+            out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
+
+        keyframe_idxs = kwargs.get("keyframe_idxs", None)
+        if keyframe_idxs is not None:
+            out['keyframe_idxs'] = comfy.conds.CONDRegular(keyframe_idxs)
+
         return out
 
+    def process_timestep(self, timestep, x, denoise_mask=None, **kwargs):
+        if denoise_mask is None:
+            return timestep
+        return self.diffusion_model.patchifier.patchify(((denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1)))[:, :1])[0]
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        return latent_image
+
 class HunyuanVideo(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan_video.model.HunyuanVideo)
@@ -863,9 +925,40 @@ class HunyuanVideo(BaseModel):
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
-        out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([kwargs.get("guidance", 6.0)]))
+
+        guidance = kwargs.get("guidance", 6.0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+
+        guiding_frame_index = kwargs.get("guiding_frame_index", None)
+        if guiding_frame_index is not None:
+            out['guiding_frame_index'] = comfy.conds.CONDRegular(torch.FloatTensor([guiding_frame_index]))
+
+        ref_latent = kwargs.get("ref_latent", None)
+        if ref_latent is not None:
+            out['ref_latent'] = comfy.conds.CONDRegular(self.process_latent_in(ref_latent))
+
         return out
 
+    def scale_latent_inpaint(self, latent_image, **kwargs):
+        return latent_image
+
+class HunyuanVideoI2V(HunyuanVideo):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.concat_keys = ("concat_image", "mask_inverted")
+
+    def scale_latent_inpaint(self, latent_image, **kwargs):
+        return super().scale_latent_inpaint(latent_image=latent_image, **kwargs)
+
+class HunyuanVideoSkyreelsI2V(HunyuanVideo):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.concat_keys = ("concat_image",)
+
+    def scale_latent_inpaint(self, latent_image, **kwargs):
+        return super().scale_latent_inpaint(latent_image=latent_image, **kwargs)
+
 class CosmosVideo(BaseModel):
     def __init__(self, model_config, model_type=ModelType.EDM, image_to_video=False, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.cosmos.model.GeneralDIT)
@@ -892,3 +985,194 @@ class CosmosVideo(BaseModel):
             latent_image = latent_image + noise
         latent_image = self.model_sampling.calculate_input(torch.tensor([sigma_noise_augmentation], device=latent_image.device, dtype=latent_image.dtype), latent_image)
         return latent_image * ((sigma ** 2 + self.model_sampling.sigma_data ** 2) ** 0.5)
+
+class Lumina2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiT)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            if torch.numel(attention_mask) != attention_mask.sum():
+                out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+            out['num_tokens'] = comfy.conds.CONDConstant(max(1, torch.sum(attention_mask).item()))
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out
+
+class WAN21(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
+        self.image_to_video = image_to_video
+
+    def concat_cond(self, **kwargs):
+        noise = kwargs.get("noise", None)
+        extra_channels = self.diffusion_model.patch_embedding.weight.shape[1] - noise.shape[1]
+        if extra_channels == 0:
+            return None
+
+        image = kwargs.get("concat_latent_image", None)
+        device = kwargs["device"]
+
+        if image is None:
+            shape_image = list(noise.shape)
+            shape_image[1] = extra_channels
+            image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device)
+        else:
+            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            for i in range(0, image.shape[1], 16):
+                image[:, i: i + 16] = self.process_latent_in(image[:, i: i + 16])
+            image = utils.resize_to_batch_size(image, noise.shape[0])
+
+        if not self.image_to_video or extra_channels == image.shape[1]:
+            return image
+
+        if image.shape[1] > (extra_channels - 4):
+            image = image[:, :(extra_channels - 4)]
+
+        mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if mask is None:
+            mask = torch.zeros_like(noise)[:, :4]
+        else:
+            if mask.shape[1] != 4:
+                mask = torch.mean(mask, dim=1, keepdim=True)
+            mask = 1.0 - mask
+            mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            if mask.shape[-3] < noise.shape[-3]:
+                mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0)
+            if mask.shape[1] == 1:
+                mask = mask.repeat(1, 4, 1, 1, 1)
+            mask = utils.resize_to_batch_size(mask, noise.shape[0])
+
+        return torch.cat((mask, image), dim=1)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        clip_vision_output = kwargs.get("clip_vision_output", None)
+        if clip_vision_output is not None:
+            out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.penultimate_hidden_states)
+
+        time_dim_concat = kwargs.get("time_dim_concat", None)
+        if time_dim_concat is not None:
+            out['time_dim_concat'] = comfy.conds.CONDRegular(self.process_latent_in(time_dim_concat))
+
+        return out
+
+
+class WAN21_Vace(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.VaceWanModel)
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        noise = kwargs.get("noise", None)
+        noise_shape = list(noise.shape)
+        vace_frames = kwargs.get("vace_frames", None)
+        if vace_frames is None:
+            noise_shape[1] = 32
+            vace_frames = [torch.zeros(noise_shape, device=noise.device, dtype=noise.dtype)]
+
+        mask = kwargs.get("vace_mask", None)
+        if mask is None:
+            noise_shape[1] = 64
+            mask = [torch.ones(noise_shape, device=noise.device, dtype=noise.dtype)] * len(vace_frames)
+
+        vace_frames_out = []
+        for j in range(len(vace_frames)):
+            vf = vace_frames[j].clone()
+            for i in range(0, vf.shape[1], 16):
+                vf[:, i:i + 16] = self.process_latent_in(vf[:, i:i + 16])
+            vf = torch.cat([vf, mask[j]], dim=1)
+            vace_frames_out.append(vf)
+
+        vace_frames = torch.stack(vace_frames_out, dim=1)
+        out['vace_context'] = comfy.conds.CONDRegular(vace_frames)
+
+        vace_strength = kwargs.get("vace_strength", [1.0] * len(vace_frames_out))
+        out['vace_strength'] = comfy.conds.CONDConstant(vace_strength)
+        return out
+
+class WAN21_Camera(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.CameraWanModel)
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        camera_conditions = kwargs.get("camera_conditions", None)
+        if camera_conditions is not None:
+            out['camera_conditions'] = comfy.conds.CONDRegular(camera_conditions)
+        return out
+
+class Hunyuan3Dv2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan3d.model.Hunyuan3Dv2)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        guidance = kwargs.get("guidance", 5.0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+        return out
+
+class HiDream(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hidream.model.HiDreamImageTransformer2DModel)
+
+    def encode_adm(self, **kwargs):
+        return kwargs["pooled_output"]
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        conditioning_llama3 = kwargs.get("conditioning_llama3", None)
+        if conditioning_llama3 is not None:
+            out['encoder_hidden_states_llama3'] = comfy.conds.CONDRegular(conditioning_llama3)
+        image_cond = kwargs.get("concat_latent_image", None)
+        if image_cond is not None:
+            out['image_cond'] = comfy.conds.CONDNoiseShape(self.process_latent_in(image_cond))
+        return out
+
+class Chroma(Flux):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.chroma.model.Chroma)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+
+        guidance = kwargs.get("guidance", 0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+        return out
+
+class ACEStep(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ace.model.ACEStepTransformer2DModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        noise = kwargs.get("noise", None)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
+        if cross_attn is not None:
+            out['lyric_token_idx'] = comfy.conds.CONDRegular(conditioning_lyrics)
+        out['speaker_embeds'] = comfy.conds.CONDRegular(torch.zeros(noise.shape[0], 512, device=noise.device, dtype=noise.dtype))
+        out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0))
+        return out

comfy/model_detection.py

@@ -1,3 +1,4 @@
+import json
 import comfy.supported_models
 import comfy.supported_models_base
 import comfy.utils
@@ -33,7 +34,7 @@ def calculate_transformer_depth(prefix, state_dict_keys, state_dict):
         return last_transformer_depth, context_dim, use_linear_in_transformer, time_stack, time_stack_cross
     return None
 
-def detect_unet_config(state_dict, key_prefix):
+def detect_unet_config(state_dict, key_prefix, metadata=None):
     state_dict_keys = list(state_dict.keys())
 
     if '{}joint_blocks.0.context_block.attn.qkv.weight'.format(key_prefix) in state_dict_keys: #mmdit model
@@ -136,7 +137,7 @@ def detect_unet_config(state_dict, key_prefix):
     if '{}txt_in.individual_token_refiner.blocks.0.norm1.weight'.format(key_prefix) in state_dict_keys: #Hunyuan Video
         dit_config = {}
         dit_config["image_model"] = "hunyuan_video"
-        dit_config["in_channels"] = 16
+        dit_config["in_channels"] = state_dict['{}img_in.proj.weight'.format(key_prefix)].shape[1] #SkyReels img2video has 32 input channels
         dit_config["patch_size"] = [1, 2, 2]
         dit_config["out_channels"] = 16
         dit_config["vec_in_dim"] = 768
@@ -153,7 +154,7 @@ def detect_unet_config(state_dict, key_prefix):
         dit_config["guidance_embed"] = len(guidance_keys) > 0
         return dit_config
 
-    if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys: #Flux
+    if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and '{}img_in.weight'.format(key_prefix) in state_dict_keys: #Flux
         dit_config = {}
         dit_config["image_model"] = "flux"
         dit_config["in_channels"] = 16
@@ -163,7 +164,9 @@ def detect_unet_config(state_dict, key_prefix):
         if in_key in state_dict_keys:
             dit_config["in_channels"] = state_dict[in_key].shape[1] // (patch_size * patch_size)
         dit_config["out_channels"] = 16
-        dit_config["vec_in_dim"] = 768
+        vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix)
+        if vec_in_key in state_dict_keys:
+            dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1]
         dit_config["context_in_dim"] = 4096
         dit_config["hidden_size"] = 3072
         dit_config["mlp_ratio"] = 4.0
@@ -173,6 +176,15 @@ def detect_unet_config(state_dict, key_prefix):
         dit_config["axes_dim"] = [16, 56, 56]
         dit_config["theta"] = 10000
         dit_config["qkv_bias"] = True
+        if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma
+            dit_config["image_model"] = "chroma"
+            dit_config["in_channels"] = 64
+            dit_config["out_channels"] = 64
+            dit_config["in_dim"] = 64
+            dit_config["out_dim"] = 3072
+            dit_config["hidden_dim"] = 5120
+            dit_config["n_layers"] = 5
+        else:
             dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
         return dit_config
 
@@ -210,6 +222,37 @@ def detect_unet_config(state_dict, key_prefix):
     if '{}adaln_single.emb.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys: #Lightricks ltxv
         dit_config = {}
         dit_config["image_model"] = "ltxv"
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
+        shape = state_dict['{}transformer_blocks.0.attn2.to_k.weight'.format(key_prefix)].shape
+        dit_config["attention_head_dim"] = shape[0] // 32
+        dit_config["cross_attention_dim"] = shape[1]
+        if metadata is not None and "config" in metadata:
+            dit_config.update(json.loads(metadata["config"]).get("transformer", {}))
+        return dit_config
+
+    if '{}genre_embedder.weight'.format(key_prefix) in state_dict_keys: #ACE-Step model
+        dit_config = {}
+        dit_config["audio_model"] = "ace"
+        dit_config["attention_head_dim"] = 128
+        dit_config["in_channels"] = 8
+        dit_config["inner_dim"] = 2560
+        dit_config["max_height"] = 16
+        dit_config["max_position"] = 32768
+        dit_config["max_width"] = 32768
+        dit_config["mlp_ratio"] = 2.5
+        dit_config["num_attention_heads"] = 20
+        dit_config["num_layers"] = 24
+        dit_config["out_channels"] = 8
+        dit_config["patch_size"] = [16, 1]
+        dit_config["rope_theta"] = 1000000.0
+        dit_config["speaker_embedding_dim"] = 512
+        dit_config["text_embedding_dim"] = 768
+
+        dit_config["ssl_encoder_depths"] = [8, 8]
+        dit_config["ssl_latent_dims"] = [1024, 768]
+        dit_config["ssl_names"] = ["mert", "m-hubert"]
+        dit_config["lyric_encoder_vocab_size"] = 6693
+        dit_config["lyric_hidden_size"] = 1024
         return dit_config
 
     if '{}t_block.1.weight'.format(key_prefix) in state_dict_keys: # PixArt
@@ -239,7 +282,7 @@ def detect_unet_config(state_dict, key_prefix):
             dit_config["micro_condition"] = False
         return dit_config
 
-    if '{}blocks.block0.blocks.0.block.attn.to_q.0.weight'.format(key_prefix) in state_dict_keys:
+    if '{}blocks.block0.blocks.0.block.attn.to_q.0.weight'.format(key_prefix) in state_dict_keys:  # Cosmos
         dit_config = {}
         dit_config["image_model"] = "cosmos"
         dit_config["max_img_h"] = 240
@@ -284,6 +327,86 @@ def detect_unet_config(state_dict, key_prefix):
             dit_config["extra_per_block_abs_pos_emb_type"] = "learnable"
         return dit_config
 
+    if '{}cap_embedder.1.weight'.format(key_prefix) in state_dict_keys:  # Lumina 2
+        dit_config = {}
+        dit_config["image_model"] = "lumina2"
+        dit_config["patch_size"] = 2
+        dit_config["in_channels"] = 16
+        dit_config["dim"] = 2304
+        dit_config["cap_feat_dim"] = 2304
+        dit_config["n_layers"] = 26
+        dit_config["n_heads"] = 24
+        dit_config["n_kv_heads"] = 8
+        dit_config["qk_norm"] = True
+        dit_config["axes_dims"] = [32, 32, 32]
+        dit_config["axes_lens"] = [300, 512, 512]
+        return dit_config
+
+    if '{}head.modulation'.format(key_prefix) in state_dict_keys:  # Wan 2.1
+        dit_config = {}
+        dit_config["image_model"] = "wan2.1"
+        dim = state_dict['{}head.modulation'.format(key_prefix)].shape[-1]
+        dit_config["dim"] = dim
+        dit_config["num_heads"] = dim // 128
+        dit_config["ffn_dim"] = state_dict['{}blocks.0.ffn.0.weight'.format(key_prefix)].shape[0]
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')
+        dit_config["patch_size"] = (1, 2, 2)
+        dit_config["freq_dim"] = 256
+        dit_config["window_size"] = (-1, -1)
+        dit_config["qk_norm"] = True
+        dit_config["cross_attn_norm"] = True
+        dit_config["eps"] = 1e-6
+        dit_config["in_dim"] = state_dict['{}patch_embedding.weight'.format(key_prefix)].shape[1]
+        if '{}vace_patch_embedding.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "vace"
+            dit_config["vace_in_dim"] = state_dict['{}vace_patch_embedding.weight'.format(key_prefix)].shape[1]
+            dit_config["vace_layers"] = count_blocks(state_dict_keys, '{}vace_blocks.'.format(key_prefix) + '{}.')
+        elif '{}control_adapter.conv.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "camera"
+        else:
+            if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
+                dit_config["model_type"] = "i2v"
+            else:
+                dit_config["model_type"] = "t2v"
+        flf_weight = state_dict.get('{}img_emb.emb_pos'.format(key_prefix))
+        if flf_weight is not None:
+            dit_config["flf_pos_embed_token_number"] = flf_weight.shape[1]
+        return dit_config
+
+    if '{}latent_in.weight'.format(key_prefix) in state_dict_keys:  # Hunyuan 3D
+        in_shape = state_dict['{}latent_in.weight'.format(key_prefix)].shape
+        dit_config = {}
+        dit_config["image_model"] = "hunyuan3d2"
+        dit_config["in_channels"] = in_shape[1]
+        dit_config["context_in_dim"] = state_dict['{}cond_in.weight'.format(key_prefix)].shape[1]
+        dit_config["hidden_size"] = in_shape[0]
+        dit_config["mlp_ratio"] = 4.0
+        dit_config["num_heads"] = 16
+        dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
+        dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
+        dit_config["qkv_bias"] = True
+        dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
+        return dit_config
+
+    if '{}caption_projection.0.linear.weight'.format(key_prefix) in state_dict_keys:  # HiDream
+        dit_config = {}
+        dit_config["image_model"] = "hidream"
+        dit_config["attention_head_dim"] = 128
+        dit_config["axes_dims_rope"] = [64, 32, 32]
+        dit_config["caption_channels"] = [4096, 4096]
+        dit_config["max_resolution"] = [128, 128]
+        dit_config["in_channels"] = 16
+        dit_config["llama_layers"] = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31]
+        dit_config["num_attention_heads"] = 20
+        dit_config["num_routed_experts"] = 4
+        dit_config["num_activated_experts"] = 2
+        dit_config["num_layers"] = 16
+        dit_config["num_single_layers"] = 32
+        dit_config["out_channels"] = 16
+        dit_config["patch_size"] = 2
+        dit_config["text_emb_dim"] = 2048
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
@@ -418,8 +541,8 @@ def model_config_from_unet_config(unet_config, state_dict=None):
     logging.error("no match {}".format(unet_config))
     return None
 
-def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=False):
-    unet_config = detect_unet_config(state_dict, unet_key_prefix)
+def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=False, metadata=None):
+    unet_config = detect_unet_config(state_dict, unet_key_prefix, metadata=metadata)
     if unet_config is None:
         return None
     model_config = model_config_from_unet_config(unet_config, state_dict)
@@ -432,6 +555,10 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal
         model_config.scaled_fp8 = scaled_fp8_weight.dtype
         if model_config.scaled_fp8 == torch.float32:
             model_config.scaled_fp8 = torch.float8_e4m3fn
+        if scaled_fp8_weight.nelement() == 2:
+            model_config.optimizations["fp8"] = False
+        else:
+            model_config.optimizations["fp8"] = True
 
     return model_config
 
@@ -493,6 +620,9 @@ def convert_config(unet_config):
 
 
 def unet_config_from_diffusers_unet(state_dict, dtype=None):
+    if "conv_in.weight" not in state_dict:
+        return None
+
     match = {}
     transformer_depth = []
 
@@ -624,8 +754,13 @@ def unet_config_from_diffusers_unet(state_dict, dtype=None):
             'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
             'use_temporal_attention': False, 'use_temporal_resblock': False}
 
+    LotusD = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': 4,
+            'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024, 'num_heads': 8,
+            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
 
-    supported_models = [SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p, SD15_diffusers_inpaint]
+    supported_models = [LotusD, SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p, SD15_diffusers_inpaint]
 
     for unet_config in supported_models:
         matches = True

comfy/model_management.py

@@ -19,7 +19,7 @@
 import psutil
 import logging
 from enum import Enum
-from comfy.cli_args import args
+from comfy.cli_args import args, PerformanceFeature
 import torch
 import sys
 import platform
@@ -46,11 +46,39 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
+def get_supported_float8_types():
+    float8_types = []
+    try:
+        float8_types.append(torch.float8_e4m3fn)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e4m3fnuz)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e5m2)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e5m2fnuz)
+    except:
+        pass
+    try:
+        float8_types.append(torch.float8_e8m0fnu)
+    except:
+        pass
+    return float8_types
+
+FLOAT8_TYPES = get_supported_float8_types()
+
 xpu_available = False
 torch_version = ""
 try:
     torch_version = torch.version.__version__
-    xpu_available = (int(torch_version[0]) < 2 or (int(torch_version[0]) == 2 and int(torch_version[2]) <= 4)) and torch.xpu.is_available()
+    temp = torch_version.split(".")
+    torch_version_numeric = (int(temp[0]), int(temp[1]))
+    xpu_available = (torch_version_numeric[0] < 2 or (torch_version_numeric[0] == 2 and torch_version_numeric[1] <= 4)) and torch.xpu.is_available()
 except:
     pass
 
@@ -93,6 +121,13 @@ try:
 except:
     npu_available = False
 
+try:
+    import torch_mlu  # noqa: F401
+    _ = torch.mlu.device_count()
+    mlu_available = torch.mlu.is_available()
+except:
+    mlu_available = False
+
 if args.cpu:
     cpu_state = CPUState.CPU
 
@@ -110,6 +145,12 @@ def is_ascend_npu():
         return True
     return False
 
+def is_mlu():
+    global mlu_available
+    if mlu_available:
+        return True
+    return False
+
 def get_torch_device():
     global directml_enabled
     global cpu_state
@@ -125,6 +166,8 @@ def get_torch_device():
             return torch.device("xpu", torch.xpu.current_device())
         elif is_ascend_npu():
             return torch.device("npu", torch.npu.current_device())
+        elif is_mlu():
+            return torch.device("mlu", torch.mlu.current_device())
         else:
             return torch.device(torch.cuda.current_device())
 
@@ -151,6 +194,12 @@ def get_total_memory(dev=None, torch_total_too=False):
             _, mem_total_npu = torch.npu.mem_get_info(dev)
             mem_total_torch = mem_reserved
             mem_total = mem_total_npu
+        elif is_mlu():
+            stats = torch.mlu.memory_stats(dev)
+            mem_reserved = stats['reserved_bytes.all.current']
+            _, mem_total_mlu = torch.mlu.mem_get_info(dev)
+            mem_total_torch = mem_reserved
+            mem_total = mem_total_mlu
         else:
             stats = torch.cuda.memory_stats(dev)
             mem_reserved = stats['reserved_bytes.all.current']
@@ -163,12 +212,21 @@ def get_total_memory(dev=None, torch_total_too=False):
     else:
         return mem_total
 
+def mac_version():
+    try:
+        return tuple(int(n) for n in platform.mac_ver()[0].split("."))
+    except:
+        return None
+
 total_vram = get_total_memory(get_torch_device()) / (1024 * 1024)
 total_ram = psutil.virtual_memory().total / (1024 * 1024)
 logging.info("Total VRAM {:0.0f} MB, total RAM {:0.0f} MB".format(total_vram, total_ram))
 
 try:
     logging.info("pytorch version: {}".format(torch_version))
+    mac_ver = mac_version()
+    if mac_ver is not None:
+        logging.info("Mac Version {}".format(mac_ver))
 except:
     pass
 
@@ -218,7 +276,7 @@ def is_amd():
 
 MIN_WEIGHT_MEMORY_RATIO = 0.4
 if is_nvidia():
-    MIN_WEIGHT_MEMORY_RATIO = 0.2
+    MIN_WEIGHT_MEMORY_RATIO = 0.0
 
 ENABLE_PYTORCH_ATTENTION = False
 if args.use_pytorch_cross_attention:
@@ -227,22 +285,45 @@ if args.use_pytorch_cross_attention:
 
 try:
     if is_nvidia():
-        if int(torch_version[0]) >= 2:
+        if torch_version_numeric[0] >= 2:
             if ENABLE_PYTORCH_ATTENTION == False and args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
                 ENABLE_PYTORCH_ATTENTION = True
-    if is_intel_xpu() or is_ascend_npu():
+    if is_intel_xpu() or is_ascend_npu() or is_mlu():
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             ENABLE_PYTORCH_ATTENTION = True
 except:
     pass
 
+
+try:
+    if is_amd():
+        arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
+        logging.info("AMD arch: {}".format(arch))
+        if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
+            if torch_version_numeric[0] >= 2 and torch_version_numeric[1] >= 7:  # works on 2.6 but doesn't actually seem to improve much
+                if any((a in arch) for a in ["gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches
+                    ENABLE_PYTORCH_ATTENTION = True
+except:
+    pass
+
+
 if ENABLE_PYTORCH_ATTENTION:
     torch.backends.cuda.enable_math_sdp(True)
     torch.backends.cuda.enable_flash_sdp(True)
     torch.backends.cuda.enable_mem_efficient_sdp(True)
 
+
+PRIORITIZE_FP16 = False  # TODO: remove and replace with something that shows exactly which dtype is faster than the other
 try:
-    if int(torch_version[0]) == 2 and int(torch_version[2]) >= 5:
+    if is_nvidia() and PerformanceFeature.Fp16Accumulation in args.fast:
+        torch.backends.cuda.matmul.allow_fp16_accumulation = True
+        PRIORITIZE_FP16 = True  # TODO: limit to cards where it actually boosts performance
+        logging.info("Enabled fp16 accumulation.")
+except:
+    pass
+
+try:
+    if torch_version_numeric[0] == 2 and torch_version_numeric[1] >= 5:
         torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp(True)
 except:
     logging.warning("Warning, could not set allow_fp16_bf16_reduction_math_sdp")
@@ -256,15 +337,10 @@ elif args.highvram or args.gpu_only:
     vram_state = VRAMState.HIGH_VRAM
 
 FORCE_FP32 = False
-FORCE_FP16 = False
 if args.force_fp32:
     logging.info("Forcing FP32, if this improves things please report it.")
     FORCE_FP32 = True
 
-if args.force_fp16:
-    logging.info("Forcing FP16.")
-    FORCE_FP16 = True
-
 if lowvram_available:
     if set_vram_to in (VRAMState.LOW_VRAM, VRAMState.NO_VRAM):
         vram_state = set_vram_to
@@ -297,6 +373,8 @@ def get_torch_device_name(device):
         return "{} {}".format(device, torch.xpu.get_device_name(device))
     elif is_ascend_npu():
         return "{} {}".format(device, torch.npu.get_device_name(device))
+    elif is_mlu():
+        return "{} {}".format(device, torch.mlu.get_device_name(device))
     else:
         return "CUDA {}: {}".format(device, torch.cuda.get_device_name(device))
 
@@ -535,14 +613,11 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             vram_set_state = vram_state
         lowvram_model_memory = 0
         if lowvram_available and (vram_set_state == VRAMState.LOW_VRAM or vram_set_state == VRAMState.NORMAL_VRAM) and not force_full_load:
-            model_size = loaded_model.model_memory_required(torch_dev)
             loaded_memory = loaded_model.model_loaded_memory()
             current_free_mem = get_free_memory(torch_dev) + loaded_memory
 
-            lowvram_model_memory = max(64 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
+            lowvram_model_memory = max(128 * 1024 * 1024, (current_free_mem - minimum_memory_required), min(current_free_mem * MIN_WEIGHT_MEMORY_RATIO, current_free_mem - minimum_inference_memory()))
             lowvram_model_memory = max(0.1, lowvram_model_memory - loaded_memory)
-            if model_size <= lowvram_model_memory: #only switch to lowvram if really necessary
-                lowvram_model_memory = 0
 
         if vram_set_state == VRAMState.NO_VRAM:
             lowvram_model_memory = 0.1
@@ -620,7 +695,7 @@ def unet_inital_load_device(parameters, dtype):
         return torch_dev
 
     cpu_dev = torch.device("cpu")
-    if DISABLE_SMART_MEMORY:
+    if DISABLE_SMART_MEMORY or vram_state == VRAMState.NO_VRAM:
         return cpu_dev
 
     model_size = dtype_size(dtype) * parameters
@@ -635,7 +710,7 @@ def unet_inital_load_device(parameters, dtype):
 def maximum_vram_for_weights(device=None):
     return (get_total_memory(device) * 0.88 - minimum_inference_memory())
 
-def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, torch.bfloat16, torch.float32]):
+def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, torch.bfloat16, torch.float32], weight_dtype=None):
     if model_params < 0:
         model_params = 1000000000000000000000
     if args.fp32_unet:
@@ -650,15 +725,12 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         return torch.float8_e4m3fn
     if args.fp8_e5m2_unet:
         return torch.float8_e5m2
+    if args.fp8_e8m0fnu_unet:
+        return torch.float8_e8m0fnu
 
     fp8_dtype = None
-    try:
-        for dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
-            if dtype in supported_dtypes:
-                fp8_dtype = dtype
-                break
-    except:
-        pass
+    if weight_dtype in FLOAT8_TYPES:
+        fp8_dtype = weight_dtype
 
     if fp8_dtype is not None:
         if supports_fp8_compute(device): #if fp8 compute is supported the casting is most likely not expensive
@@ -668,6 +740,10 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         if model_params * 2 > free_model_memory:
             return fp8_dtype
 
+    if PRIORITIZE_FP16 or weight_dtype == torch.float16:
+        if torch.float16 in supported_dtypes and should_use_fp16(device=device, model_params=model_params):
+            return torch.float16
+
     for dt in supported_dtypes:
         if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params):
             if torch.float16 in supported_dtypes:
@@ -700,6 +776,9 @@ def unet_manual_cast(weight_dtype, inference_device, supported_dtypes=[torch.flo
         return None
 
     fp16_supported = should_use_fp16(inference_device, prioritize_performance=True)
+    if PRIORITIZE_FP16 and fp16_supported and torch.float16 in supported_dtypes:
+        return torch.float16
+
     for dt in supported_dtypes:
         if dt == torch.float16 and fp16_supported:
             return torch.float16
@@ -746,6 +825,8 @@ def text_encoder_dtype(device=None):
         return torch.float8_e5m2
     elif args.fp16_text_enc:
         return torch.float16
+    elif args.bf16_text_enc:
+        return torch.bfloat16
     elif args.fp32_text_enc:
         return torch.float32
 
@@ -858,13 +939,59 @@ def force_channels_last():
     #TODO
     return False
 
-def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False):
+
+STREAMS = {}
+NUM_STREAMS = 1
+if args.async_offload:
+    NUM_STREAMS = 2
+    logging.info("Using async weight offloading with {} streams".format(NUM_STREAMS))
+
+stream_counters = {}
+def get_offload_stream(device):
+    stream_counter = stream_counters.get(device, 0)
+    if NUM_STREAMS <= 1:
+        return None
+
+    if device in STREAMS:
+        ss = STREAMS[device]
+        s = ss[stream_counter]
+        stream_counter = (stream_counter + 1) % len(ss)
+        if is_device_cuda(device):
+            ss[stream_counter].wait_stream(torch.cuda.current_stream())
+        stream_counters[device] = stream_counter
+        return s
+    elif is_device_cuda(device):
+        ss = []
+        for k in range(NUM_STREAMS):
+            ss.append(torch.cuda.Stream(device=device, priority=0))
+        STREAMS[device] = ss
+        s = ss[stream_counter]
+        stream_counter = (stream_counter + 1) % len(ss)
+        stream_counters[device] = stream_counter
+        return s
+    return None
+
+def sync_stream(device, stream):
+    if stream is None:
+        return
+    if is_device_cuda(device):
+        torch.cuda.current_stream().wait_stream(stream)
+
+def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None):
     if device is None or weight.device == device:
         if not copy:
             if dtype is None or weight.dtype == dtype:
                 return weight
+        if stream is not None:
+            with stream:
+                return weight.to(dtype=dtype, copy=copy)
         return weight.to(dtype=dtype, copy=copy)
 
+    if stream is not None:
+        with stream:
+            r = torch.empty_like(weight, dtype=dtype, device=device)
+            r.copy_(weight, non_blocking=non_blocking)
+    else:
         r = torch.empty_like(weight, dtype=dtype, device=device)
         r.copy_(weight, non_blocking=non_blocking)
     return r
@@ -876,6 +1003,9 @@ def cast_to_device(tensor, device, dtype, copy=False):
 def sage_attention_enabled():
     return args.use_sage_attention
 
+def flash_attention_enabled():
+    return args.use_flash_attention
+
 def xformers_enabled():
     global directml_enabled
     global cpu_state
@@ -885,6 +1015,8 @@ def xformers_enabled():
         return False
     if is_ascend_npu():
         return False
+    if is_mlu():
+        return False
     if directml_enabled:
         return False
     return XFORMERS_IS_AVAILABLE
@@ -901,6 +1033,11 @@ def pytorch_attention_enabled():
     global ENABLE_PYTORCH_ATTENTION
     return ENABLE_PYTORCH_ATTENTION
 
+def pytorch_attention_enabled_vae():
+    if is_amd():
+        return False  # enabling pytorch attention on AMD currently causes crash when doing high res
+    return pytorch_attention_enabled()
+
 def pytorch_attention_flash_attention():
     global ENABLE_PYTORCH_ATTENTION
     if ENABLE_PYTORCH_ATTENTION:
@@ -911,23 +1048,21 @@ def pytorch_attention_flash_attention():
             return True
         if is_ascend_npu():
             return True
+        if is_mlu():
+            return True
+        if is_amd():
+            return True #if you have pytorch attention enabled on AMD it probably supports at least mem efficient attention
     return False
 
-def mac_version():
-    try:
-        return tuple(int(n) for n in platform.mac_ver()[0].split("."))
-    except:
-        return None
-
 def force_upcast_attention_dtype():
     upcast = args.force_upcast_attention
 
     macos_version = mac_version()
-    if macos_version is not None and ((14, 5) <= macos_version <= (15, 2)):  # black image bug on recent versions of macOS
+    if macos_version is not None and ((14, 5) <= macos_version < (16,)):  # black image bug on recent versions of macOS
         upcast = True
 
     if upcast:
-        return torch.float32
+        return {torch.float16: torch.float32}
     else:
         return None
 
@@ -957,6 +1092,13 @@ def get_free_memory(dev=None, torch_free_too=False):
             mem_free_npu, _ = torch.npu.mem_get_info(dev)
             mem_free_torch = mem_reserved - mem_active
             mem_free_total = mem_free_npu + mem_free_torch
+        elif is_mlu():
+            stats = torch.mlu.memory_stats(dev)
+            mem_active = stats['active_bytes.all.current']
+            mem_reserved = stats['reserved_bytes.all.current']
+            mem_free_mlu, _ = torch.mlu.mem_get_info(dev)
+            mem_free_torch = mem_reserved - mem_active
+            mem_free_total = mem_free_mlu + mem_free_torch
         else:
             stats = torch.cuda.memory_stats(dev)
             mem_active = stats['active_bytes.all.current']
@@ -993,21 +1135,26 @@ def is_device_mps(device):
 def is_device_cuda(device):
     return is_device_type(device, 'cuda')
 
-def should_use_fp16(device=None, model_params=0, prioritize_performance=True, manual_cast=False):
+def is_directml_enabled():
     global directml_enabled
+    if directml_enabled:
+        return True
 
+    return False
+
+def should_use_fp16(device=None, model_params=0, prioritize_performance=True, manual_cast=False):
     if device is not None:
         if is_device_cpu(device):
             return False
 
-    if FORCE_FP16:
+    if args.force_fp16:
         return True
 
     if FORCE_FP32:
         return False
 
-    if directml_enabled:
-        return False
+    if is_directml_enabled():
+        return True
 
     if (device is not None and is_device_mps(device)) or mps_mode():
         return True
@@ -1021,6 +1168,9 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     if is_ascend_npu():
         return True
 
+    if is_mlu():
+        return True
+
     if torch.version.hip:
         return True
 
@@ -1078,13 +1228,28 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     if is_intel_xpu():
         return True
 
+    if is_ascend_npu():
+        return True
+
+    if is_amd():
+        arch = torch.cuda.get_device_properties(device).gcnArchName
+        if any((a in arch) for a in ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]):  # RDNA2 and older don't support bf16
+            if manual_cast:
+                return True
+            return False
+
     props = torch.cuda.get_device_properties(device)
+
+    if is_mlu():
+        if props.major > 3:
+            return True
+
     if props.major >= 8:
         return True
 
     bf16_works = torch.cuda.is_bf16_supported()
 
-    if bf16_works or manual_cast:
+    if bf16_works and manual_cast:
         free_model_memory = maximum_vram_for_weights(device)
         if (not prioritize_performance) or model_params * 4 > free_model_memory:
             return True
@@ -1092,6 +1257,9 @@ def should_use_bf16(device=None, model_params=0, prioritize_performance=True, ma
     return False
 
 def supports_fp8_compute(device=None):
+    if args.supports_fp8_compute:
+        return True
+
     if not is_nvidia():
         return False
 
@@ -1103,11 +1271,11 @@ def supports_fp8_compute(device=None):
     if props.minor < 9:
         return False
 
-    if int(torch_version[0]) < 2 or (int(torch_version[0]) == 2 and int(torch_version[2]) < 3):
+    if torch_version_numeric[0] < 2 or (torch_version_numeric[0] == 2 and torch_version_numeric[1] < 3):
         return False
 
     if WINDOWS:
-        if (int(torch_version[0]) == 2 and int(torch_version[2]) < 4):
+        if (torch_version_numeric[0] == 2 and torch_version_numeric[1] < 4):
             return False
 
     return True
@@ -1120,6 +1288,8 @@ def soft_empty_cache(force=False):
         torch.xpu.empty_cache()
     elif is_ascend_npu():
         torch.npu.empty_cache()
+    elif is_mlu():
+        torch.mlu.empty_cache()
     elif torch.cuda.is_available():
         torch.cuda.empty_cache()
         torch.cuda.ipc_collect()

comfy/model_patcher.py

@@ -96,8 +96,28 @@ def wipe_lowvram_weight(m):
     if hasattr(m, "prev_comfy_cast_weights"):
         m.comfy_cast_weights = m.prev_comfy_cast_weights
         del m.prev_comfy_cast_weights
-    m.weight_function = None
-    m.bias_function = None
+
+    if hasattr(m, "weight_function"):
+        m.weight_function = []
+
+    if hasattr(m, "bias_function"):
+        m.bias_function = []
+
+def move_weight_functions(m, device):
+    if device is None:
+        return 0
+
+    memory = 0
+    if hasattr(m, "weight_function"):
+        for f in m.weight_function:
+            if hasattr(f, "move_to"):
+                memory += f.move_to(device=device)
+
+    if hasattr(m, "bias_function"):
+        for f in m.bias_function:
+            if hasattr(f, "move_to"):
+                memory += f.move_to(device=device)
+    return memory
 
 class LowVramPatch:
     def __init__(self, key, patches):
@@ -192,11 +212,13 @@ class ModelPatcher:
         self.backup = {}
         self.object_patches = {}
         self.object_patches_backup = {}
+        self.weight_wrapper_patches = {}
         self.model_options = {"transformer_options":{}}
         self.model_size()
         self.load_device = load_device
         self.offload_device = offload_device
         self.weight_inplace_update = weight_inplace_update
+        self.force_cast_weights = False
         self.patches_uuid = uuid.uuid4()
         self.parent = None
 
@@ -250,11 +272,14 @@ class ModelPatcher:
         n.patches_uuid = self.patches_uuid
 
         n.object_patches = self.object_patches.copy()
+        n.weight_wrapper_patches = self.weight_wrapper_patches.copy()
         n.model_options = copy.deepcopy(self.model_options)
         n.backup = self.backup
         n.object_patches_backup = self.object_patches_backup
         n.parent = self
 
+        n.force_cast_weights = self.force_cast_weights
+
         # attachments
         n.attachments = {}
         for k in self.attachments:
@@ -402,6 +427,16 @@ class ModelPatcher:
     def add_object_patch(self, name, obj):
         self.object_patches[name] = obj
 
+    def set_model_compute_dtype(self, dtype):
+        self.add_object_patch("manual_cast_dtype", dtype)
+        if dtype is not None:
+            self.force_cast_weights = True
+        self.patches_uuid = uuid.uuid4() #TODO: optimize by preventing a full model reload for this
+
+    def add_weight_wrapper(self, name, function):
+        self.weight_wrapper_patches[name] = self.weight_wrapper_patches.get(name, []) + [function]
+        self.patches_uuid = uuid.uuid4()
+
     def get_model_object(self, name: str) -> torch.nn.Module:
         """Retrieves a nested attribute from an object using dot notation considering
         object patches.
@@ -566,6 +601,9 @@ class ModelPatcher:
 
                 lowvram_weight = False
 
+                weight_key = "{}.weight".format(n)
+                bias_key = "{}.bias".format(n)
+
                 if not full_load and hasattr(m, "comfy_cast_weights"):
                     if mem_counter + module_mem >= lowvram_model_memory:
                         lowvram_weight = True
@@ -573,34 +611,46 @@ class ModelPatcher:
                         if hasattr(m, "prev_comfy_cast_weights"): #Already lowvramed
                             continue
 
-                weight_key = "{}.weight".format(n)
-                bias_key = "{}.bias".format(n)
-
+                cast_weight = self.force_cast_weights
                 if lowvram_weight:
+                    if hasattr(m, "comfy_cast_weights"):
+                        m.weight_function = []
+                        m.bias_function = []
+
                     if weight_key in self.patches:
                         if force_patch_weights:
                             self.patch_weight_to_device(weight_key)
                         else:
-                            m.weight_function = LowVramPatch(weight_key, self.patches)
+                            m.weight_function = [LowVramPatch(weight_key, self.patches)]
                             patch_counter += 1
                     if bias_key in self.patches:
                         if force_patch_weights:
                             self.patch_weight_to_device(bias_key)
                         else:
-                            m.bias_function = LowVramPatch(bias_key, self.patches)
+                            m.bias_function = [LowVramPatch(bias_key, self.patches)]
                             patch_counter += 1
 
-                    m.prev_comfy_cast_weights = m.comfy_cast_weights
-                    m.comfy_cast_weights = True
+                    cast_weight = True
                 else:
                     if hasattr(m, "comfy_cast_weights"):
-                        if m.comfy_cast_weights:
                         wipe_lowvram_weight(m)
 
                     if full_load or mem_counter + module_mem < lowvram_model_memory:
                         mem_counter += module_mem
                         load_completely.append((module_mem, n, m, params))
 
+                if cast_weight and hasattr(m, "comfy_cast_weights"):
+                    m.prev_comfy_cast_weights = m.comfy_cast_weights
+                    m.comfy_cast_weights = True
+
+                if weight_key in self.weight_wrapper_patches:
+                    m.weight_function.extend(self.weight_wrapper_patches[weight_key])
+
+                if bias_key in self.weight_wrapper_patches:
+                    m.bias_function.extend(self.weight_wrapper_patches[bias_key])
+
+                mem_counter += move_weight_functions(m, device_to)
+
             load_completely.sort(reverse=True)
             for x in load_completely:
                 n = x[1]
@@ -662,6 +712,7 @@ class ModelPatcher:
             self.unpatch_hooks()
             if self.model.model_lowvram:
                 for m in self.model.modules():
+                    move_weight_functions(m, device_to)
                     wipe_lowvram_weight(m)
 
                 self.model.model_lowvram = False
@@ -696,6 +747,7 @@ class ModelPatcher:
 
     def partially_unload(self, device_to, memory_to_free=0):
         with self.use_ejected():
+            hooks_unpatched = False
             memory_freed = 0
             patch_counter = 0
             unload_list = self._load_list()
@@ -719,6 +771,10 @@ class ModelPatcher:
                                 move_weight = False
                                 break
 
+                            if not hooks_unpatched:
+                                self.unpatch_hooks()
+                                hooks_unpatched = True
+
                             if bk.inplace_update:
                                 comfy.utils.copy_to_param(self.model, key, bk.weight)
                             else:
@@ -728,15 +784,19 @@ class ModelPatcher:
                     weight_key = "{}.weight".format(n)
                     bias_key = "{}.bias".format(n)
                     if move_weight:
+                        cast_weight = self.force_cast_weights
                         m.to(device_to)
+                        module_mem += move_weight_functions(m, device_to)
                         if lowvram_possible:
                             if weight_key in self.patches:
-                                m.weight_function = LowVramPatch(weight_key, self.patches)
+                                m.weight_function.append(LowVramPatch(weight_key, self.patches))
                                 patch_counter += 1
                             if bias_key in self.patches:
-                                m.bias_function = LowVramPatch(bias_key, self.patches)
+                                m.bias_function.append(LowVramPatch(bias_key, self.patches))
                                 patch_counter += 1
+                            cast_weight = True
 
+                        if cast_weight:
                             m.prev_comfy_cast_weights = m.comfy_cast_weights
                             m.comfy_cast_weights = True
                         m.comfy_patched_weights = False
@@ -1034,7 +1094,6 @@ class ModelPatcher:
 
     def patch_hooks(self, hooks: comfy.hooks.HookGroup):
         with self.use_ejected():
-            self.unpatch_hooks()
             if hooks is not None:
                 model_sd_keys = list(self.model_state_dict().keys())
                 memory_counter = None
@@ -1045,12 +1104,16 @@ class ModelPatcher:
                 # if have cached weights for hooks, use it
                 cached_weights = self.cached_hook_patches.get(hooks, None)
                 if cached_weights is not None:
+                    model_sd_keys_set = set(model_sd_keys)
                     for key in cached_weights:
                         if key not in model_sd_keys:
                             logging.warning(f"Cached hook could not patch. Key does not exist in model: {key}")
                             continue
                         self.patch_cached_hook_weights(cached_weights=cached_weights, key=key, memory_counter=memory_counter)
+                        model_sd_keys_set.remove(key)
+                    self.unpatch_hooks(model_sd_keys_set)
                 else:
+                    self.unpatch_hooks()
                     relevant_patches = self.get_combined_hook_patches(hooks=hooks)
                     original_weights = None
                     if len(relevant_patches) > 0:
@@ -1061,6 +1124,8 @@ class ModelPatcher:
                             continue
                         self.patch_hook_weight_to_device(hooks=hooks, combined_patches=relevant_patches, key=key, original_weights=original_weights,
                                                             memory_counter=memory_counter)
+            else:
+                self.unpatch_hooks()
             self.current_hooks = hooks
 
     def patch_cached_hook_weights(self, cached_weights: dict, key: str, memory_counter: MemoryCounter):
@@ -1117,12 +1182,18 @@ class ModelPatcher:
         del out_weight
         del weight
 
-    def unpatch_hooks(self) -> None:
+    def unpatch_hooks(self, whitelist_keys_set: set[str]=None) -> None:
         with self.use_ejected():
             if len(self.hook_backup) == 0:
                 self.current_hooks = None
                 return
             keys = list(self.hook_backup.keys())
+            if whitelist_keys_set:
+                for k in keys:
+                    if k in whitelist_keys_set:
+                        comfy.utils.copy_to_param(self.model, k, self.hook_backup[k][0].to(device=self.hook_backup[k][1]))
+                        self.hook_backup.pop(k)
+            else:
                 for k in keys:
                     comfy.utils.copy_to_param(self.model, k, self.hook_backup[k][0].to(device=self.hook_backup[k][1]))
 

comfy/model_sampling.py

@@ -31,6 +31,7 @@ class EPS:
         return model_input - model_output * sigma
 
     def noise_scaling(self, sigma, noise, latent_image, max_denoise=False):
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (noise.ndim - 1))
         if max_denoise:
             noise = noise * torch.sqrt(1.0 + sigma ** 2.0)
         else:
@@ -61,11 +62,22 @@ class CONST:
         return model_input - model_output * sigma
 
     def noise_scaling(self, sigma, noise, latent_image, max_denoise=False):
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (noise.ndim - 1))
         return sigma * noise + (1.0 - sigma) * latent_image
 
     def inverse_noise_scaling(self, sigma, latent):
+        sigma = sigma.view(sigma.shape[:1] + (1,) * (latent.ndim - 1))
         return latent / (1.0 - sigma)
 
+class X0(EPS):
+    def calculate_denoised(self, sigma, model_output, model_input):
+        return model_output
+
+class IMG_TO_IMG(X0):
+    def calculate_input(self, sigma, noise):
+        return noise
+
+
 class ModelSamplingDiscrete(torch.nn.Module):
     def __init__(self, model_config=None, zsnr=None):
         super().__init__()
@@ -99,13 +111,14 @@ class ModelSamplingDiscrete(torch.nn.Module):
         self.num_timesteps = int(timesteps)
         self.linear_start = linear_start
         self.linear_end = linear_end
+        self.zsnr = zsnr
 
         # self.register_buffer('betas', torch.tensor(betas, dtype=torch.float32))
         # self.register_buffer('alphas_cumprod', torch.tensor(alphas_cumprod, dtype=torch.float32))
         # self.register_buffer('alphas_cumprod_prev', torch.tensor(alphas_cumprod_prev, dtype=torch.float32))
 
         sigmas = ((1 - alphas_cumprod) / alphas_cumprod) ** 0.5
-        if zsnr:
+        if self.zsnr:
             sigmas = rescale_zero_terminal_snr_sigmas(sigmas)
 
         self.set_sigmas(sigmas)

comfy/ops.py

@@ -17,9 +17,12 @@
 """
 
 import torch
+import logging
 import comfy.model_management
-from comfy.cli_args import args
+from comfy.cli_args import args, PerformanceFeature
 import comfy.float
+import comfy.rmsnorm
+import contextlib
 
 cast_to = comfy.model_management.cast_to #TODO: remove once no more references
 
@@ -35,24 +38,37 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None):
         if device is None:
             device = input.device
 
+    offload_stream = comfy.model_management.get_offload_stream(device)
+    if offload_stream is not None:
+        wf_context = offload_stream
+    else:
+        wf_context = contextlib.nullcontext()
+
     bias = None
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
     if s.bias is not None:
-        has_function = s.bias_function is not None
-        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function)
-        if has_function:
-            bias = s.bias_function(bias)
+        has_function = len(s.bias_function) > 0
+        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream)
 
-    has_function = s.weight_function is not None
-    weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function)
         if has_function:
-        weight = s.weight_function(weight)
+            with wf_context:
+                for f in s.bias_function:
+                    bias = f(bias)
+
+    has_function = len(s.weight_function) > 0
+    weight = comfy.model_management.cast_to(s.weight, dtype, device, non_blocking=non_blocking, copy=has_function, stream=offload_stream)
+    if has_function:
+        with wf_context:
+            for f in s.weight_function:
+                weight = f(weight)
+
+    comfy.model_management.sync_stream(device, offload_stream)
     return weight, bias
 
 class CastWeightBiasOp:
     comfy_cast_weights = False
-    weight_function = None
-    bias_function = None
+    weight_function = []
+    bias_function = []
 
 class disable_weight_init:
     class Linear(torch.nn.Linear, CastWeightBiasOp):
@@ -64,7 +80,7 @@ class disable_weight_init:
             return torch.nn.functional.linear(input, weight, bias)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -78,7 +94,7 @@ class disable_weight_init:
             return self._conv_forward(input, weight, bias)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -92,7 +108,7 @@ class disable_weight_init:
             return self._conv_forward(input, weight, bias)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -106,7 +122,7 @@ class disable_weight_init:
             return self._conv_forward(input, weight, bias)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -120,12 +136,11 @@ class disable_weight_init:
             return torch.nn.functional.group_norm(input, self.num_groups, weight, bias, self.eps)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
 
-
     class LayerNorm(torch.nn.LayerNorm, CastWeightBiasOp):
         def reset_parameters(self):
             return None
@@ -139,7 +154,26 @@ class disable_weight_init:
             return torch.nn.functional.layer_norm(input, self.normalized_shape, weight, bias, self.eps)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
+                return self.forward_comfy_cast_weights(*args, **kwargs)
+            else:
+                return super().forward(*args, **kwargs)
+
+    class RMSNorm(comfy.rmsnorm.RMSNorm, CastWeightBiasOp):
+        def reset_parameters(self):
+            self.bias = None
+            return None
+
+        def forward_comfy_cast_weights(self, input):
+            if self.weight is not None:
+                weight, bias = cast_bias_weight(self, input)
+            else:
+                weight = None
+            return comfy.rmsnorm.rms_norm(input, weight, self.eps)  # TODO: switch to commented out line when old torch is deprecated
+            # return torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
+
+        def forward(self, *args, **kwargs):
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -160,7 +194,7 @@ class disable_weight_init:
                 output_padding, self.groups, self.dilation)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -181,7 +215,7 @@ class disable_weight_init:
                 output_padding, self.groups, self.dilation)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)
@@ -199,7 +233,7 @@ class disable_weight_init:
             return torch.nn.functional.embedding(input, weight, self.padding_idx, self.max_norm, self.norm_type, self.scale_grad_by_freq, self.sparse).to(dtype=output_dtype)
 
         def forward(self, *args, **kwargs):
-            if self.comfy_cast_weights:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 if "out_dtype" in kwargs:
@@ -241,6 +275,9 @@ class manual_cast(disable_weight_init):
     class ConvTranspose1d(disable_weight_init.ConvTranspose1d):
         comfy_cast_weights = True
 
+    class RMSNorm(disable_weight_init.RMSNorm):
+        comfy_cast_weights = True
+
     class Embedding(disable_weight_init.Embedding):
         comfy_cast_weights = True
 
@@ -271,10 +308,10 @@ def fp8_linear(self, input):
         if scale_input is None:
             scale_input = torch.ones((), device=input.device, dtype=torch.float32)
             input = torch.clamp(input, min=-448, max=448, out=input)
-            input = input.reshape(-1, input_shape[2]).to(dtype)
+            input = input.reshape(-1, input_shape[2]).to(dtype).contiguous()
         else:
             scale_input = scale_input.to(input.device)
-            input = (input * (1.0 / scale_input).to(input_dtype)).reshape(-1, input_shape[2]).to(dtype)
+            input = (input * (1.0 / scale_input).to(input_dtype)).reshape(-1, input_shape[2]).to(dtype).contiguous()
 
         if bias is not None:
             o = torch._scaled_mm(input, w, out_dtype=input_dtype, bias=bias, scale_a=scale_input, scale_b=scale_weight)
@@ -307,6 +344,7 @@ class fp8_ops(manual_cast):
             return torch.nn.functional.linear(input, weight, bias)
 
 def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None):
+    logging.info("Using scaled fp8: fp8 matrix mult: {}, scale input: {}".format(fp8_matrix_mult, scale_input))
     class scaled_fp8_op(manual_cast):
         class Linear(manual_cast.Linear):
             def __init__(self, *args, **kwargs):
@@ -354,14 +392,46 @@ def scaled_fp8_ops(fp8_matrix_mult=False, scale_input=False, override_dtype=None
 
     return scaled_fp8_op
 
+CUBLAS_IS_AVAILABLE = False
+try:
+    from cublas_ops import CublasLinear
+    CUBLAS_IS_AVAILABLE = True
+except ImportError:
+    pass
+
+if CUBLAS_IS_AVAILABLE:
+    class cublas_ops(disable_weight_init):
+        class Linear(CublasLinear, disable_weight_init.Linear):
+            def reset_parameters(self):
+                return None
+
+            def forward_comfy_cast_weights(self, input):
+                return super().forward(input)
+
+            def forward(self, *args, **kwargs):
+                return super().forward(*args, **kwargs)
+
 def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, scaled_fp8=None):
     fp8_compute = comfy.model_management.supports_fp8_compute(load_device)
     if scaled_fp8 is not None:
-        return scaled_fp8_ops(fp8_matrix_mult=fp8_compute, scale_input=True, override_dtype=scaled_fp8)
+        return scaled_fp8_ops(fp8_matrix_mult=fp8_compute and fp8_optimizations, scale_input=fp8_optimizations, override_dtype=scaled_fp8)
 
-    if fp8_compute and (fp8_optimizations or args.fast) and not disable_fast_fp8:
+    if (
+        fp8_compute and
+        (fp8_optimizations or PerformanceFeature.Fp8MatrixMultiplication in args.fast) and
+        not disable_fast_fp8
+    ):
         return fp8_ops
 
+    if (
+        PerformanceFeature.CublasOps in args.fast and
+        CUBLAS_IS_AVAILABLE and
+        weight_dtype == torch.float16 and
+        (compute_dtype == torch.float16 or compute_dtype is None)
+    ):
+        logging.info("Using cublas ops")
+        return cublas_ops
+
     if compute_dtype is None or weight_dtype == compute_dtype:
         return disable_weight_init
 

comfy/patcher_extension.py

@@ -48,6 +48,7 @@ def get_all_callbacks(call_type: str, transformer_options: dict, is_model_option
 
 class WrappersMP:
     OUTER_SAMPLE = "outer_sample"
+    PREPARE_SAMPLING = "prepare_sampling"
     SAMPLER_SAMPLE = "sampler_sample"
     CALC_COND_BATCH = "calc_cond_batch"
     APPLY_MODEL = "apply_model"

comfy/rmsnorm.py

@@ -0,0 +1,55 @@
+import torch
+import comfy.model_management
+import numbers
+
+RMSNorm = None
+
+try:
+    rms_norm_torch = torch.nn.functional.rms_norm
+    RMSNorm = torch.nn.RMSNorm
+except:
+    rms_norm_torch = None
+
+
+def rms_norm(x, weight=None, eps=1e-6):
+    if rms_norm_torch is not None and not (torch.jit.is_tracing() or torch.jit.is_scripting()):
+        if weight is None:
+            return rms_norm_torch(x, (x.shape[-1],), eps=eps)
+        else:
+            return rms_norm_torch(x, weight.shape, weight=comfy.model_management.cast_to(weight, dtype=x.dtype, device=x.device), eps=eps)
+    else:
+        r = x * torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + eps)
+        if weight is None:
+            return r
+        else:
+            return r * comfy.model_management.cast_to(weight, dtype=x.dtype, device=x.device)
+
+
+if RMSNorm is None:
+    class RMSNorm(torch.nn.Module):
+        def __init__(
+            self,
+            normalized_shape,
+            eps=1e-6,
+            elementwise_affine=True,
+            device=None,
+            dtype=None,
+        ):
+            factory_kwargs = {"device": device, "dtype": dtype}
+            super().__init__()
+            if isinstance(normalized_shape, numbers.Integral):
+                # mypy error: incompatible types in assignment
+                normalized_shape = (normalized_shape,)  # type: ignore[assignment]
+            self.normalized_shape = tuple(normalized_shape)  # type: ignore[arg-type]
+            self.eps = eps
+            self.elementwise_affine = elementwise_affine
+            if self.elementwise_affine:
+                self.weight = torch.nn.Parameter(
+                    torch.empty(self.normalized_shape, **factory_kwargs)
+                )
+            else:
+                self.register_parameter("weight", None)
+            self.bias = None
+
+        def forward(self, x):
+            return rms_norm(x, self.weight, self.eps)

comfy/sampler_helpers.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 import uuid
+import math
+import collections
 import comfy.model_management
 import comfy.conds
 import comfy.utils
@@ -58,7 +60,6 @@ def convert_cond(cond):
         temp = c[1].copy()
         model_conds = temp.get("model_conds", {})
         if c[0] is not None:
-            model_conds["c_crossattn"] = comfy.conds.CONDCrossAttn(c[0]) #TODO: remove
             temp["cross_attn"] = c[0]
         temp["model_conds"] = model_conds
         temp["uuid"] = uuid.uuid4()
@@ -105,15 +106,36 @@ def cleanup_additional_models(models):
         if hasattr(m, 'cleanup'):
             m.cleanup()
 
+def estimate_memory(model, noise_shape, conds):
+    cond_shapes = collections.defaultdict(list)
+    cond_shapes_min = {}
+    for _, cs in conds.items():
+        for cond in cs:
+            for k, v in model.model.extra_conds_shapes(**cond).items():
+                cond_shapes[k].append(v)
+                if cond_shapes_min.get(k, None) is None:
+                    cond_shapes_min[k] = [v]
+                elif math.prod(v) > math.prod(cond_shapes_min[k][0]):
+                    cond_shapes_min[k] = [v]
+
+    memory_required = model.model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:]), cond_shapes=cond_shapes)
+    minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min)
+    return memory_required, minimum_memory_required
 
 def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
+    executor = comfy.patcher_extension.WrapperExecutor.new_executor(
+        _prepare_sampling,
+        comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True)
+    )
+    return executor.execute(model, noise_shape, conds, model_options=model_options)
+
+def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
     real_model: BaseModel = None
     models, inference_memory = get_additional_models(conds, model.model_dtype())
     models += get_additional_models_from_model_options(model_options)
     models += model.get_nested_additional_models()  # TODO: does this require inference_memory update?
-    memory_required = model.memory_required([noise_shape[0] * 2] + list(noise_shape[1:])) + inference_memory
-    minimum_memory_required = model.memory_required([noise_shape[0]] + list(noise_shape[1:])) + inference_memory
-    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required)
+    memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
+    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory)
     real_model = model.model
 
     return real_model, conds, models

comfy/samplers.py

@@ -12,7 +12,6 @@ import collections
 from comfy import model_management
 import math
 import logging
-import comfy.samplers
 import comfy.sampler_helpers
 import comfy.model_patcher
 import comfy.patcher_extension
@@ -20,6 +19,12 @@ import comfy.hooks
 import scipy.stats
 import numpy
 
+
+def add_area_dims(area, num_dims):
+    while (len(area) // 2) < num_dims:
+        area = [2147483648] + area[:len(area) // 2] + [0] + area[len(area) // 2:]
+    return area
+
 def get_area_and_mult(conds, x_in, timestep_in):
     dims = tuple(x_in.shape[2:])
     area = None
@@ -35,6 +40,10 @@ def get_area_and_mult(conds, x_in, timestep_in):
             return None
     if 'area' in conds:
         area = list(conds['area'])
+        area = add_area_dims(area, len(dims))
+        if (len(area) // 2) > len(dims):
+            area = area[:len(dims)] + area[len(area) // 2:(len(area) // 2) + len(dims)]
+
     if 'strength' in conds:
         strength = conds['strength']
 
@@ -65,8 +74,9 @@ def get_area_and_mult(conds, x_in, timestep_in):
     mult = mask * strength
 
     if 'mask' not in conds and area is not None:
-        rr = 8
+        fuzz = 8
         for i in range(len(dims)):
+            rr = min(fuzz, mult.shape[2 + i] // 4)
             if area[len(dims) + i] != 0:
                 for t in range(rr):
                     m = mult.narrow(i + 2, t, 1)
@@ -178,7 +188,7 @@ def finalize_default_conds(model: 'BaseModel', hooked_to_run: dict[comfy.hooks.H
         cond = default_conds[i]
         for x in cond:
             # do get_area_and_mult to get all the expected values
-            p = comfy.samplers.get_area_and_mult(x, x_in, timestep)
+            p = get_area_and_mult(x, x_in, timestep)
             if p is None:
                 continue
             # replace p's mult with calculated mult
@@ -215,7 +225,7 @@ def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Te
                     default_c.append(x)
                     has_default_conds = True
                     continue
-                p = comfy.samplers.get_area_and_mult(x, x_in, timestep)
+                p = get_area_and_mult(x, x_in, timestep)
                 if p is None:
                     continue
                 if p.hooks is not None:
@@ -246,7 +256,13 @@ def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Te
             for i in range(1, len(to_batch_temp) + 1):
                 batch_amount = to_batch_temp[:len(to_batch_temp)//i]
                 input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]
-                if model.memory_required(input_shape) * 1.5 < free_memory:
+                cond_shapes = collections.defaultdict(list)
+                for tt in batch_amount:
+                    cond = {k: v.size() for k, v in to_run[tt][0].conditioning.items()}
+                    for k, v in to_run[tt][0].conditioning.items():
+                        cond_shapes[k].append(v.size())
+
+                if model.memory_required(input_shape, cond_shapes=cond_shapes) * 1.5 < free_memory:
                     to_batch = batch_amount
                     break
 
@@ -549,24 +565,36 @@ def resolve_areas_and_cond_masks(conditions, h, w, device):
     logging.warning("WARNING: The comfy.samplers.resolve_areas_and_cond_masks function is deprecated please use the resolve_areas_and_cond_masks_multidim one instead.")
     return resolve_areas_and_cond_masks_multidim(conditions, [h, w], device)
 
-def create_cond_with_same_area_if_none(conds, c): #TODO: handle dim != 2
+def create_cond_with_same_area_if_none(conds, c):
     if 'area' not in c:
         return
 
+    def area_inside(a, area_cmp):
+        a = add_area_dims(a, len(area_cmp) // 2)
+        area_cmp = add_area_dims(area_cmp, len(a) // 2)
+
+        a_l = len(a) // 2
+        area_cmp_l = len(area_cmp) // 2
+        for i in range(min(a_l, area_cmp_l)):
+            if a[a_l + i] < area_cmp[area_cmp_l + i]:
+                return False
+        for i in range(min(a_l, area_cmp_l)):
+            if (a[i] + a[a_l + i]) > (area_cmp[i] + area_cmp[area_cmp_l + i]):
+                return False
+        return True
+
     c_area = c['area']
     smallest = None
     for x in conds:
         if 'area' in x:
             a = x['area']
-            if c_area[2] >= a[2] and c_area[3] >= a[3]:
-                if a[0] + a[2] >= c_area[0] + c_area[2]:
-                    if a[1] + a[3] >= c_area[1] + c_area[3]:
+            if area_inside(c_area, a):
                 if smallest is None:
                     smallest = x
                 elif 'area' not in smallest:
                     smallest = x
                 else:
-                            if smallest['area'][0] * smallest['area'][1] > a[0] * a[1]:
+                    if math.prod(smallest['area'][:len(smallest['area']) // 2]) > math.prod(a[:len(a) // 2]):
                         smallest = x
         else:
             if smallest is None:
@@ -687,7 +715,8 @@ class Sampler:
 KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
                   "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
-                  "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp"]
+                  "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
+                  "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3"]
 
 class KSAMPLER(Sampler):
     def __init__(self, sampler_function, extra_options={}, inpaint_options={}):

comfy/sd.py

@@ -1,4 +1,5 @@
 from __future__ import annotations
+import json
 import torch
 from enum import Enum
 import logging
@@ -12,6 +13,9 @@ from .ldm.audio.autoencoder import AudioOobleckVAE
 import comfy.ldm.genmo.vae.model
 import comfy.ldm.lightricks.vae.causal_video_autoencoder
 import comfy.ldm.cosmos.vae
+import comfy.ldm.wan.vae
+import comfy.ldm.hunyuan3d.vae
+import comfy.ldm.ace.vae.music_dcae_pipeline
 import yaml
 import math
 
@@ -36,6 +40,10 @@ import comfy.text_encoders.genmo
 import comfy.text_encoders.lt
 import comfy.text_encoders.hunyuan_video
 import comfy.text_encoders.cosmos
+import comfy.text_encoders.lumina2
+import comfy.text_encoders.wan
+import comfy.text_encoders.hidream
+import comfy.text_encoders.ace
 
 import comfy.model_patcher
 import comfy.lora
@@ -114,6 +122,7 @@ class CLIP:
         self.layer_idx = None
         self.use_clip_schedule = False
         logging.info("CLIP/text encoder model load device: {}, offload device: {}, current: {}, dtype: {}".format(load_device, offload_device, params['device'], dtype))
+        self.tokenizer_options = {}
 
     def clone(self):
         n = CLIP(no_init=True)
@@ -121,6 +130,7 @@ class CLIP:
         n.cond_stage_model = self.cond_stage_model
         n.tokenizer = self.tokenizer
         n.layer_idx = self.layer_idx
+        n.tokenizer_options = self.tokenizer_options.copy()
         n.use_clip_schedule = self.use_clip_schedule
         n.apply_hooks_to_conds = self.apply_hooks_to_conds
         return n
@@ -128,11 +138,19 @@ class CLIP:
     def add_patches(self, patches, strength_patch=1.0, strength_model=1.0):
         return self.patcher.add_patches(patches, strength_patch, strength_model)
 
+    def set_tokenizer_option(self, option_name, value):
+        self.tokenizer_options[option_name] = value
+
     def clip_layer(self, layer_idx):
         self.layer_idx = layer_idx
 
-    def tokenize(self, text, return_word_ids=False):
-        return self.tokenizer.tokenize_with_weights(text, return_word_ids)
+    def tokenize(self, text, return_word_ids=False, **kwargs):
+        tokenizer_options = kwargs.get("tokenizer_options", {})
+        if len(self.tokenizer_options) > 0:
+            tokenizer_options = {**self.tokenizer_options, **tokenizer_options}
+        if len(tokenizer_options) > 0:
+            kwargs["tokenizer_options"] = tokenizer_options
+        return self.tokenizer.tokenize_with_weights(text, return_word_ids, **kwargs)
 
     def add_hooks_to_dict(self, pooled_dict: dict[str]):
         if self.apply_hooks_to_conds:
@@ -246,7 +264,7 @@ class CLIP:
         return self.patcher.get_key_patches()
 
 class VAE:
-    def __init__(self, sd=None, device=None, config=None, dtype=None):
+    def __init__(self, sd=None, device=None, config=None, dtype=None, metadata=None):
         if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format
             sd = diffusers_convert.convert_vae_state_dict(sd)
 
@@ -260,9 +278,11 @@ class VAE:
         self.process_input = lambda image: image * 2.0 - 1.0
         self.process_output = lambda image: torch.clamp((image + 1.0) / 2.0, min=0.0, max=1.0)
         self.working_dtypes = [torch.bfloat16, torch.float32]
+        self.disable_offload = False
 
         self.downscale_index_formula = None
         self.upscale_index_formula = None
+        self.extra_1d_channel = None
 
         if config is None:
             if "decoder.mid.block_1.mix_factor" in sd:
@@ -332,6 +352,7 @@ class VAE:
                 self.process_output = lambda audio: audio
                 self.process_input = lambda audio: audio
                 self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+                self.disable_offload = True
             elif "blocks.2.blocks.3.stack.5.weight" in sd or "decoder.blocks.2.blocks.3.stack.5.weight" in sd or "layers.4.layers.1.attn_block.attn.qkv.weight" in sd or "encoder.layers.4.layers.1.attn_block.attn.qkv.weight" in sd: #genmo mochi vae
                 if "blocks.2.blocks.3.stack.5.weight" in sd:
                     sd = comfy.utils.state_dict_prefix_replace(sd, {"": "decoder."})
@@ -354,7 +375,12 @@ class VAE:
                     version = 0
                 elif tensor_conv1.shape[0] == 1024:
                     version = 1
-                self.first_stage_model = comfy.ldm.lightricks.vae.causal_video_autoencoder.VideoVAE(version=version)
+                    if "encoder.down_blocks.1.conv.conv.bias" in sd:
+                        version = 2
+                vae_config = None
+                if metadata is not None and "config" in metadata:
+                    vae_config = json.loads(metadata["config"]).get("vae", None)
+                self.first_stage_model = comfy.ldm.lightricks.vae.causal_video_autoencoder.VideoVAE(version=version, config=vae_config)
                 self.latent_channels = 128
                 self.latent_dim = 3
                 self.memory_used_decode = lambda shape, dtype: (900 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
@@ -391,6 +417,43 @@ class VAE:
                 self.memory_used_decode = lambda shape, dtype: (50 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
                 self.memory_used_encode = lambda shape, dtype: (50 * (round((shape[2] + 7) / 8) * 8) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
                 self.working_dtypes = [torch.bfloat16, torch.float32]
+            elif "decoder.middle.0.residual.0.gamma" in sd:
+                self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
+                self.upscale_index_formula = (4, 8, 8)
+                self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8)
+                self.downscale_index_formula = (4, 8, 8)
+                self.latent_dim = 3
+                self.latent_channels = 16
+                ddconfig = {"dim": 96, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
+                self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig)
+                self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+                self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype)
+            elif "geo_decoder.cross_attn_decoder.ln_1.bias" in sd:
+                self.latent_dim = 1
+                ln_post = "geo_decoder.ln_post.weight" in sd
+                inner_size = sd["geo_decoder.output_proj.weight"].shape[1]
+                downsample_ratio = sd["post_kl.weight"].shape[0] // inner_size
+                mlp_expand = sd["geo_decoder.cross_attn_decoder.mlp.c_fc.weight"].shape[0] // inner_size
+                self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype)  # TODO
+                self.memory_used_decode = lambda shape, dtype: (1024 * 1024 * 1024 * 2.0) * model_management.dtype_size(dtype)  # TODO
+                ddconfig = {"embed_dim": 64, "num_freqs": 8, "include_pi": False, "heads": 16, "width": 1024, "num_decoder_layers": 16, "qkv_bias": False, "qk_norm": True, "geo_decoder_mlp_expand_ratio": mlp_expand, "geo_decoder_downsample_ratio": downsample_ratio, "geo_decoder_ln_post": ln_post}
+                self.first_stage_model = comfy.ldm.hunyuan3d.vae.ShapeVAE(**ddconfig)
+                self.working_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+            elif "vocoder.backbone.channel_layers.0.0.bias" in sd: #Ace Step Audio
+                self.first_stage_model = comfy.ldm.ace.vae.music_dcae_pipeline.MusicDCAE(source_sample_rate=44100)
+                self.memory_used_encode = lambda shape, dtype: (shape[2] * 330) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (shape[2] * shape[3] * 87000) * model_management.dtype_size(dtype)
+                self.latent_channels = 8
+                self.output_channels = 2
+                self.upscale_ratio = 4096
+                self.downscale_ratio = 4096
+                self.latent_dim = 2
+                self.process_output = lambda audio: audio
+                self.process_input = lambda audio: audio
+                self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+                self.disable_offload = True
+                self.extra_1d_channel = 16
             else:
                 logging.warning("WARNING: No VAE weights detected, VAE not initalized.")
                 self.first_stage_model = None
@@ -419,6 +482,10 @@ class VAE:
         self.patcher = comfy.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device)
         logging.info("VAE load device: {}, offload device: {}, dtype: {}".format(self.device, offload_device, self.vae_dtype))
 
+    def throw_exception_if_invalid(self):
+        if self.first_stage_model is None:
+            raise RuntimeError("ERROR: VAE is invalid: None\n\nIf the VAE is from a checkpoint loader node your checkpoint does not contain a valid VAE.")
+
     def vae_encode_crop_pixels(self, pixels):
         downscale_ratio = self.spacial_compression_encode()
 
@@ -445,7 +512,13 @@ class VAE:
         return output
 
     def decode_tiled_1d(self, samples, tile_x=128, overlap=32):
+        if samples.ndim == 3:
             decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        else:
+            og_shape = samples.shape
+            samples = samples.reshape((og_shape[0], og_shape[1] * og_shape[2], -1))
+            decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).float()
+
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, output_device=self.output_device))
 
     def decode_tiled_3d(self, samples, tile_t=999, tile_x=32, tile_y=32, overlap=(1, 8, 8)):
@@ -465,33 +538,49 @@ class VAE:
         samples /= 3.0
         return samples
 
-    def encode_tiled_1d(self, samples, tile_x=128 * 2048, overlap=32 * 2048):
+    def encode_tiled_1d(self, samples, tile_x=256 * 2048, overlap=64 * 2048):
+        if self.latent_dim == 1:
             encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
-        return comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=(1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device)
+            out_channels = self.latent_channels
+            upscale_amount = 1 / self.downscale_ratio
+        else:
+            extra_channel_size = self.extra_1d_channel
+            out_channels = self.latent_channels * extra_channel_size
+            tile_x = tile_x // extra_channel_size
+            overlap = overlap // extra_channel_size
+            upscale_amount = 1 / self.downscale_ratio
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).float()
+
+        out = comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=upscale_amount, out_channels=out_channels, output_device=self.output_device)
+        if self.latent_dim == 1:
+            return out
+        else:
+            return out.reshape(samples.shape[0], self.latent_channels, extra_channel_size, -1)
 
     def encode_tiled_3d(self, samples, tile_t=9999, tile_x=512, tile_y=512, overlap=(1, 64, 64)):
         encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
         return comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.downscale_ratio, out_channels=self.latent_channels, downscale=True, index_formulas=self.downscale_index_formula, output_device=self.output_device)
 
-    def decode(self, samples_in):
+    def decode(self, samples_in, vae_options={}):
+        self.throw_exception_if_invalid()
         pixel_samples = None
         try:
             memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
-            model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+            model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
             free_memory = model_management.get_free_memory(self.device)
             batch_number = int(free_memory / memory_used)
             batch_number = max(1, batch_number)
 
             for x in range(0, samples_in.shape[0], batch_number):
                 samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device)
-                out = self.process_output(self.first_stage_model.decode(samples).to(self.output_device).float())
+                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).float())
                 if pixel_samples is None:
                     pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
                 pixel_samples[x:x+batch_number] = out
         except model_management.OOM_EXCEPTION:
             logging.warning("Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.")
             dims = samples_in.ndim - 2
-            if dims == 1:
+            if dims == 1 or self.extra_1d_channel is not None:
                 pixel_samples = self.decode_tiled_1d(samples_in)
             elif dims == 2:
                 pixel_samples = self.decode_tiled_(samples_in)
@@ -504,8 +593,9 @@ class VAE:
         return pixel_samples
 
     def decode_tiled(self, samples, tile_x=None, tile_y=None, overlap=None, tile_t=None, overlap_t=None):
+        self.throw_exception_if_invalid()
         memory_used = self.memory_used_decode(samples.shape, self.vae_dtype) #TODO: calculate mem required for tile
-        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+        model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
         dims = samples.ndim - 2
         args = {}
         if tile_x is not None:
@@ -532,13 +622,14 @@ class VAE:
         return output.movedim(1, -1)
 
     def encode(self, pixel_samples):
+        self.throw_exception_if_invalid()
         pixel_samples = self.vae_encode_crop_pixels(pixel_samples)
         pixel_samples = pixel_samples.movedim(-1, 1)
         if self.latent_dim == 3 and pixel_samples.ndim < 5:
             pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0)
         try:
             memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)
-            model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+            model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
             free_memory = model_management.get_free_memory(self.device)
             batch_number = int(free_memory / max(1, memory_used))
             batch_number = max(1, batch_number)
@@ -556,7 +647,7 @@ class VAE:
                 tile = 256
                 overlap = tile // 4
                 samples = self.encode_tiled_3d(pixel_samples, tile_x=tile, tile_y=tile, overlap=(1, overlap, overlap))
-            elif self.latent_dim == 1:
+            elif self.latent_dim == 1 or self.extra_1d_channel is not None:
                 samples = self.encode_tiled_1d(pixel_samples)
             else:
                 samples = self.encode_tiled_(pixel_samples)
@@ -564,6 +655,7 @@ class VAE:
         return samples
 
     def encode_tiled(self, pixel_samples, tile_x=None, tile_y=None, overlap=None, tile_t=None, overlap_t=None):
+        self.throw_exception_if_invalid()
         pixel_samples = self.vae_encode_crop_pixels(pixel_samples)
         dims = self.latent_dim
         pixel_samples = pixel_samples.movedim(-1, 1)
@@ -571,7 +663,7 @@ class VAE:
             pixel_samples = pixel_samples.movedim(1, 0).unsqueeze(0)
 
         memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)  # TODO: calculate mem required for tile
-        model_management.load_models_gpu([self.patcher], memory_required=memory_used)
+        model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
 
         args = {}
         if tile_x is not None:
@@ -657,6 +749,11 @@ class CLIPType(Enum):
     HUNYUAN_VIDEO = 9
     PIXART = 10
     COSMOS = 11
+    LUMINA2 = 12
+    WAN = 13
+    HIDREAM = 14
+    CHROMA = 15
+    ACE = 16
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -675,6 +772,7 @@ class TEModel(Enum):
     T5_BASE = 6
     LLAMA3_8 = 7
     T5_XXL_OLD = 8
+    GEMMA_2_2B = 9
 
 def detect_te_model(sd):
     if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
@@ -693,6 +791,8 @@ def detect_te_model(sd):
         return TEModel.T5_XXL_OLD
     if "encoder.block.0.layer.0.SelfAttention.k.weight" in sd:
         return TEModel.T5_BASE
+    if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
+        return TEModel.GEMMA_2_2B
     if "model.layers.0.post_attention_layernorm.weight" in sd:
         return TEModel.LLAMA3_8
     return None
@@ -730,6 +830,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             if "text_projection" in clip_data[i]:
                 clip_data[i]["text_projection.weight"] = clip_data[i]["text_projection"].transpose(0, 1) #old models saved with the CLIPSave node
 
+    tokenizer_data = {}
     clip_target = EmptyClass()
     clip_target.params = {}
     if len(clip_data) == 1:
@@ -741,6 +842,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             elif clip_type == CLIPType.SD3:
                 clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=False, clip_g=True, t5=False)
                 clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
+            elif clip_type == CLIPType.HIDREAM:
+                clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=False, clip_g=True, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None)
+                clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
             else:
                 clip_target.clip = sdxl_clip.SDXLRefinerClipModel
                 clip_target.tokenizer = sdxl_clip.SDXLTokenizer
@@ -754,9 +858,17 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             elif clip_type == CLIPType.LTXV:
                 clip_target.clip = comfy.text_encoders.lt.ltxv_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.lt.LTXVT5Tokenizer
-            elif clip_type == CLIPType.PIXART:
+            elif clip_type == CLIPType.PIXART or clip_type == CLIPType.CHROMA:
                 clip_target.clip = comfy.text_encoders.pixart_t5.pixart_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.pixart_t5.PixArtTokenizer
+            elif clip_type == CLIPType.WAN:
+                clip_target.clip = comfy.text_encoders.wan.te(**t5xxl_detect(clip_data))
+                clip_target.tokenizer = comfy.text_encoders.wan.WanT5Tokenizer
+                tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
+            elif clip_type == CLIPType.HIDREAM:
+                clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data),
+                                                                        clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None, llama_scaled_fp8=None)
+                clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
             else: #CLIPType.MOCHI
                 clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer
@@ -767,12 +879,29 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             clip_target.clip = comfy.text_encoders.aura_t5.AuraT5Model
             clip_target.tokenizer = comfy.text_encoders.aura_t5.AuraT5Tokenizer
         elif te_model == TEModel.T5_BASE:
+            if clip_type == CLIPType.ACE or "spiece_model" in clip_data[0]:
+                clip_target.clip = comfy.text_encoders.ace.AceT5Model
+                clip_target.tokenizer = comfy.text_encoders.ace.AceT5Tokenizer
+                tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
+            else:
                 clip_target.clip = comfy.text_encoders.sa_t5.SAT5Model
                 clip_target.tokenizer = comfy.text_encoders.sa_t5.SAT5Tokenizer
+        elif te_model == TEModel.GEMMA_2_2B:
+            clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer
+            tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
+        elif te_model == TEModel.LLAMA3_8:
+            clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**llama_detect(clip_data),
+                                                                        clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None, t5xxl_scaled_fp8=None)
+            clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
         else:
+            # clip_l
             if clip_type == CLIPType.SD3:
                 clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(clip_l=True, clip_g=False, t5=False)
                 clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
+            elif clip_type == CLIPType.HIDREAM:
+                clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=True, clip_g=False, t5=False, llama=False, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None)
+                clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
             else:
                 clip_target.clip = sd1_clip.SD1ClipModel
                 clip_target.tokenizer = sd1_clip.SD1Tokenizer
@@ -790,15 +919,35 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif clip_type == CLIPType.HUNYUAN_VIDEO:
             clip_target.clip = comfy.text_encoders.hunyuan_video.hunyuan_video_clip(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.hunyuan_video.HunyuanVideoTokenizer
+        elif clip_type == CLIPType.HIDREAM:
+            # Detect
+            hidream_dualclip_classes = []
+            for hidream_te in clip_data:
+                te_model = detect_te_model(hidream_te)
+                hidream_dualclip_classes.append(te_model)
+
+            clip_l = TEModel.CLIP_L in hidream_dualclip_classes
+            clip_g = TEModel.CLIP_G in hidream_dualclip_classes
+            t5 = TEModel.T5_XXL in hidream_dualclip_classes
+            llama = TEModel.LLAMA3_8 in hidream_dualclip_classes
+
+            # Initialize t5xxl_detect and llama_detect kwargs if needed
+            t5_kwargs = t5xxl_detect(clip_data) if t5 else {}
+            llama_kwargs = llama_detect(clip_data) if llama else {}
+
+            clip_target.clip = comfy.text_encoders.hidream.hidream_clip(clip_l=clip_l, clip_g=clip_g, t5=t5, llama=llama, **t5_kwargs, **llama_kwargs)
+            clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer
     elif len(clip_data) == 3:
         clip_target.clip = comfy.text_encoders.sd3_clip.sd3_clip(**t5xxl_detect(clip_data))
         clip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer
+    elif len(clip_data) == 4:
+        clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data), **llama_detect(clip_data))
+        clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
 
     parameters = 0
-    tokenizer_data = {}
     for c in clip_data:
         parameters += comfy.utils.calculate_parameters(c)
         tokenizer_data, model_options = comfy.text_encoders.long_clipl.model_options_long_clip(c, tokenizer_data, model_options)
@@ -845,13 +994,13 @@ def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_cl
     return (model, clip, vae)
 
 def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
-    sd = comfy.utils.load_torch_file(ckpt_path)
-    out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options)
+    sd, metadata = comfy.utils.load_torch_file(ckpt_path, return_metadata=True)
+    out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options, metadata=metadata)
     if out is None:
         raise RuntimeError("ERROR: Could not detect model type of: {}".format(ckpt_path))
     return out
 
-def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
+def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}, metadata=None):
     clip = None
     clipvision = None
     vae = None
@@ -863,19 +1012,24 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
     weight_dtype = comfy.utils.weight_dtype(sd, diffusion_model_prefix)
     load_device = model_management.get_torch_device()
 
-    model_config = model_detection.model_config_from_unet(sd, diffusion_model_prefix)
+    model_config = model_detection.model_config_from_unet(sd, diffusion_model_prefix, metadata=metadata)
     if model_config is None:
+        logging.warning("Warning, This is not a checkpoint file, trying to load it as a diffusion model only.")
+        diffusion_model = load_diffusion_model_state_dict(sd, model_options={})
+        if diffusion_model is None:
             return None
+        return (diffusion_model, None, VAE(sd={}), None)  # The VAE object is there to throw an exception if it's actually used'
+
 
     unet_weight_dtype = list(model_config.supported_inference_dtypes)
-    if weight_dtype is not None and model_config.scaled_fp8 is None:
-        unet_weight_dtype.append(weight_dtype)
+    if model_config.scaled_fp8 is not None:
+        weight_dtype = None
 
     model_config.custom_operations = model_options.get("custom_operations", None)
     unet_dtype = model_options.get("dtype", model_options.get("weight_dtype", None))
 
     if unet_dtype is None:
-        unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=unet_weight_dtype)
+        unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=unet_weight_dtype, weight_dtype=weight_dtype)
 
     manual_cast_dtype = model_management.unet_manual_cast(unet_dtype, load_device, model_config.supported_inference_dtypes)
     model_config.set_inference_dtype(unet_dtype, manual_cast_dtype)
@@ -892,7 +1046,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
     if output_vae:
         vae_sd = comfy.utils.state_dict_prefix_replace(sd, {k: "" for k in model_config.vae_key_prefix}, filter_keys=True)
         vae_sd = model_config.process_vae_state_dict(vae_sd)
-        vae = VAE(sd=vae_sd)
+        vae = VAE(sd=vae_sd, metadata=metadata)
 
     if output_clip:
         clip_target = model_config.clip_target(state_dict=sd)
@@ -966,11 +1120,11 @@ def load_diffusion_model_state_dict(sd, model_options={}): #load unet in diffuse
 
     offload_device = model_management.unet_offload_device()
     unet_weight_dtype = list(model_config.supported_inference_dtypes)
-    if weight_dtype is not None and model_config.scaled_fp8 is None:
-        unet_weight_dtype.append(weight_dtype)
+    if model_config.scaled_fp8 is not None:
+        weight_dtype = None
 
     if dtype is None:
-        unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=unet_weight_dtype)
+        unet_dtype = model_management.unet_dtype(model_params=parameters, supported_dtypes=unet_weight_dtype, weight_dtype=weight_dtype)
     else:
         unet_dtype = dtype
 

comfy/sd1_clip.py

@@ -82,7 +82,8 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
     LAYERS = [
         "last",
         "pooled",
-        "hidden"
+        "hidden",
+        "all"
     ]
     def __init__(self, device="cpu", max_length=77,
                  freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, dtype=None, model_class=comfy.clip_model.CLIPTextModel,
@@ -93,6 +94,8 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
 
         if textmodel_json_config is None:
             textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_clip_config.json")
+            if "model_name" not in model_options:
+                model_options = {**model_options, "model_name": "clip_l"}
 
         if isinstance(textmodel_json_config, dict):
             config = textmodel_json_config
@@ -100,6 +103,10 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
             with open(textmodel_json_config) as f:
                 config = json.load(f)
 
+        te_model_options = model_options.get("{}_model_config".format(model_options.get("model_name", "")), {})
+        for k, v in te_model_options.items():
+            config[k] = v
+
         operations = model_options.get("custom_operations", None)
         scaled_fp8 = None
 
@@ -147,7 +154,9 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
     def set_clip_options(self, options):
         layer_idx = options.get("layer", self.layer_idx)
         self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled)
-        if layer_idx is None or abs(layer_idx) > self.num_layers:
+        if self.layer == "all":
+            pass
+        elif layer_idx is None or abs(layer_idx) > self.num_layers:
             self.layer = "last"
         else:
             self.layer = "hidden"
@@ -158,71 +167,98 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         self.layer_idx = self.options_default[1]
         self.return_projected_pooled = self.options_default[2]
 
-    def set_up_textual_embeddings(self, tokens, current_embeds):
-        out_tokens = []
-        next_new_token = token_dict_size = current_embeds.weight.shape[0]
-        embedding_weights = []
-
-        for x in tokens:
-            tokens_temp = []
-            for y in x:
-                if isinstance(y, numbers.Integral):
-                    tokens_temp += [int(y)]
-                else:
-                    if y.shape[0] == current_embeds.weight.shape[1]:
-                        embedding_weights += [y]
-                        tokens_temp += [next_new_token]
-                        next_new_token += 1
-                    else:
-                        logging.warning("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored {} != {}".format(y.shape[0], current_embeds.weight.shape[1]))
-            while len(tokens_temp) < len(x):
-                tokens_temp += [self.special_tokens["pad"]]
-            out_tokens += [tokens_temp]
-
-        n = token_dict_size
-        if len(embedding_weights) > 0:
-            new_embedding = self.operations.Embedding(next_new_token + 1, current_embeds.weight.shape[1], device=current_embeds.weight.device, dtype=current_embeds.weight.dtype)
-            new_embedding.weight[:token_dict_size] = current_embeds.weight
-            for x in embedding_weights:
-                new_embedding.weight[n] = x
-                n += 1
-            self.transformer.set_input_embeddings(new_embedding)
-
-        processed_tokens = []
-        for x in out_tokens:
-            processed_tokens += [list(map(lambda a: n if a == -1 else a, x))] #The EOS token should always be the largest one
-
-        return processed_tokens
-
-    def forward(self, tokens):
-        backup_embeds = self.transformer.get_input_embeddings()
-        device = backup_embeds.weight.device
-        tokens = self.set_up_textual_embeddings(tokens, backup_embeds)
-        tokens = torch.LongTensor(tokens).to(device)
-
-        attention_mask = None
-        if self.enable_attention_masks or self.zero_out_masked or self.return_attention_masks:
-            attention_mask = torch.zeros_like(tokens)
+    def process_tokens(self, tokens, device):
         end_token = self.special_tokens.get("end", None)
         if end_token is None:
             cmp_token = self.special_tokens.get("pad", -1)
         else:
             cmp_token = end_token
 
-            for x in range(attention_mask.shape[0]):
-                for y in range(attention_mask.shape[1]):
-                    attention_mask[x, y] = 1
-                    if tokens[x, y] == cmp_token:
+        embeds_out = []
+        attention_masks = []
+        num_tokens = []
+
+        for x in tokens:
+            attention_mask = []
+            tokens_temp = []
+            other_embeds = []
+            eos = False
+            index = 0
+            for y in x:
+                if isinstance(y, numbers.Integral):
+                    if eos:
+                        attention_mask.append(0)
+                    else:
+                        attention_mask.append(1)
+                    token = int(y)
+                    tokens_temp += [token]
+                    if not eos and token == cmp_token:
                         if end_token is None:
-                            attention_mask[x, y] = 0
-                        break
+                            attention_mask[-1] = 0
+                        eos = True
+                else:
+                    other_embeds.append((index, y))
+                index += 1
+
+            tokens_embed = torch.tensor([tokens_temp], device=device, dtype=torch.long)
+            tokens_embed = self.transformer.get_input_embeddings()(tokens_embed, out_dtype=torch.float32)
+            index = 0
+            pad_extra = 0
+            for o in other_embeds:
+                emb = o[1]
+                if torch.is_tensor(emb):
+                    emb = {"type": "embedding", "data": emb}
+
+                emb_type = emb.get("type", None)
+                if emb_type == "embedding":
+                    emb = emb.get("data", None)
+                else:
+                    if hasattr(self.transformer, "preprocess_embed"):
+                        emb = self.transformer.preprocess_embed(emb, device=device)
+                    else:
+                        emb = None
+
+                if emb is None:
+                    index += -1
+                    continue
+
+                ind = index + o[0]
+                emb = emb.view(1, -1, emb.shape[-1]).to(device=device, dtype=torch.float32)
+                emb_shape = emb.shape[1]
+                if emb.shape[-1] == tokens_embed.shape[-1]:
+                    tokens_embed = torch.cat([tokens_embed[:, :ind], emb, tokens_embed[:, ind:]], dim=1)
+                    attention_mask = attention_mask[:ind] + [1] * emb_shape + attention_mask[ind:]
+                    index += emb_shape - 1
+                else:
+                    index += -1
+                    pad_extra += emb_shape
+                    logging.warning("WARNING: shape mismatch when trying to apply embedding, embedding will be ignored {} != {}".format(emb.shape[-1], tokens_embed.shape[-1]))
+
+            if pad_extra > 0:
+                padd_embed = self.transformer.get_input_embeddings()(torch.tensor([[self.special_tokens["pad"]] * pad_extra], device=device, dtype=torch.long), out_dtype=torch.float32)
+                tokens_embed = torch.cat([tokens_embed, padd_embed], dim=1)
+                attention_mask = attention_mask + [0] * pad_extra
+
+            embeds_out.append(tokens_embed)
+            attention_masks.append(attention_mask)
+            num_tokens.append(sum(attention_mask))
+
+        return torch.cat(embeds_out), torch.tensor(attention_masks, device=device, dtype=torch.long), num_tokens
+
+    def forward(self, tokens):
+        device = self.transformer.get_input_embeddings().weight.device
+        embeds, attention_mask, num_tokens = self.process_tokens(tokens, device)
 
         attention_mask_model = None
         if self.enable_attention_masks:
             attention_mask_model = attention_mask
 
-        outputs = self.transformer(tokens, attention_mask_model, intermediate_output=self.layer_idx, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32)
-        self.transformer.set_input_embeddings(backup_embeds)
+        if self.layer == "all":
+            intermediate_output = "all"
+        else:
+            intermediate_output = self.layer_idx
+
+        outputs = self.transformer(None, attention_mask_model, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=self.layer_norm_hidden_state, dtype=torch.float32)
 
         if self.layer == "last":
             z = outputs[0].float()
@@ -421,13 +457,14 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No
     return embed_out
 
 class SDTokenizer:
-    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, tokenizer_data={}):
+    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, tokenizer_data={}, tokenizer_args={}):
         if tokenizer_path is None:
             tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer")
-        self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path)
-        self.max_length = max_length
+        self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args)
+        self.max_length = tokenizer_data.get("{}_max_length".format(embedding_key), max_length)
         self.min_length = min_length
         self.end_token = None
+        self.min_padding = min_padding
 
         empty = self.tokenizer('')["input_ids"]
         self.tokenizer_adds_end_token = has_end_token
@@ -482,13 +519,15 @@ class SDTokenizer:
         return (embed, leftover)
 
 
-    def tokenize_with_weights(self, text:str, return_word_ids=False):
+    def tokenize_with_weights(self, text:str, return_word_ids=False, tokenizer_options={}, **kwargs):
         '''
         Takes a prompt and converts it to a list of (token, weight, word id) elements.
         Tokens can both be integer tokens and pre computed CLIP tensors.
         Word id values are unique per word and embedding, where the id 0 is reserved for non word tokens.
         Returned list has the dimensions NxM where M is the input size of CLIP
         '''
+        min_length = tokenizer_options.get("{}_min_length".format(self.embedding_key), self.min_length)
+        min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding)
 
         text = escape_important(text)
         parsed_weights = token_weights(text, 1.0)
@@ -567,10 +606,12 @@ class SDTokenizer:
         #fill last batch
         if self.end_token is not None:
             batch.append((self.end_token, 1.0, 0))
-        if self.pad_to_max_length:
+        if min_padding is not None:
+            batch.extend([(self.pad_token, 1.0, 0)] * min_padding)
+        if self.pad_to_max_length and len(batch) < self.max_length:
             batch.extend([(self.pad_token, 1.0, 0)] * (self.max_length - len(batch)))
-        if self.min_length is not None and len(batch) < self.min_length:
-            batch.extend([(self.pad_token, 1.0, 0)] * (self.min_length - len(batch)))
+        if min_length is not None and len(batch) < min_length:
+            batch.extend([(self.pad_token, 1.0, 0)] * (min_length - len(batch)))
 
         if not return_word_ids:
             batched_tokens = [[(t, w) for t, w,_ in x] for x in batched_tokens]
@@ -585,22 +626,27 @@ class SDTokenizer:
         return {}
 
 class SD1Tokenizer:
-    def __init__(self, embedding_directory=None, tokenizer_data={}, clip_name="l", tokenizer=SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}, clip_name="l", tokenizer=SDTokenizer, name=None):
+        if name is not None:
+            self.clip_name = name
+            self.clip = "{}".format(self.clip_name)
+        else:
             self.clip_name = clip_name
             self.clip = "clip_{}".format(self.clip_name)
+
         tokenizer = tokenizer_data.get("{}_tokenizer_class".format(self.clip), tokenizer)
         setattr(self, self.clip, tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data))
 
-    def tokenize_with_weights(self, text:str, return_word_ids=False):
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
-        out[self.clip_name] = getattr(self, self.clip).tokenize_with_weights(text, return_word_ids)
+        out[self.clip_name] = getattr(self, self.clip).tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
     def untokenize(self, token_weight_pair):
         return getattr(self, self.clip).untokenize(token_weight_pair)
 
     def state_dict(self):
-        return {}
+        return getattr(self, self.clip).state_dict()
 
 class SD1CheckpointClipModel(SDClipModel):
     def __init__(self, device="cpu", dtype=None, model_options={}):
@@ -618,6 +664,7 @@ class SD1ClipModel(torch.nn.Module):
             self.clip = "clip_{}".format(self.clip_name)
 
         clip_model = model_options.get("{}_class".format(self.clip), clip_model)
+        model_options = {**model_options, "model_name": self.clip}
         setattr(self, self.clip, clip_model(device=device, dtype=dtype, model_options=model_options, **kwargs))
 
         self.dtypes = set()

comfy/sdxl_clip.py

@@ -9,6 +9,7 @@ class SDXLClipG(sd1_clip.SDClipModel):
             layer_idx=-2
 
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
+        model_options = {**model_options, "model_name": "clip_g"}
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
                          special_tokens={"start": 49406, "end": 49407, "pad": 0}, layer_norm_hidden_state=False, return_projected_pooled=True, model_options=model_options)
 
@@ -17,19 +18,18 @@ class SDXLClipG(sd1_clip.SDClipModel):
 
 class SDXLClipGTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, tokenizer_path=None, embedding_directory=None, tokenizer_data={}):
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g')
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g', tokenizer_data=tokenizer_data)
 
 
 class SDXLTokenizer:
     def __init__(self, embedding_directory=None, tokenizer_data={}):
-        clip_l_tokenizer_class = tokenizer_data.get("clip_l_tokenizer_class", sd1_clip.SDTokenizer)
-        self.clip_l = clip_l_tokenizer_class(embedding_directory=embedding_directory)
-        self.clip_g = SDXLClipGTokenizer(embedding_directory=embedding_directory)
+        self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+        self.clip_g = SDXLClipGTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
 
-    def tokenize_with_weights(self, text:str, return_word_ids=False):
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
-        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids)
-        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
+        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
     def untokenize(self, token_weight_pair):
@@ -41,8 +41,7 @@ class SDXLTokenizer:
 class SDXLClipModel(torch.nn.Module):
     def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__()
-        clip_l_class = model_options.get("clip_l_class", sd1_clip.SDClipModel)
-        self.clip_l = clip_l_class(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, model_options=model_options)
+        self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False, model_options=model_options)
         self.clip_g = SDXLClipG(device=device, dtype=dtype, model_options=model_options)
         self.dtypes = set([dtype])
 
@@ -75,7 +74,7 @@ class SDXLRefinerClipModel(sd1_clip.SD1ClipModel):
 
 class StableCascadeClipGTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, tokenizer_path=None, embedding_directory=None, tokenizer_data={}):
-        super().__init__(tokenizer_path, pad_with_end=True, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g')
+        super().__init__(tokenizer_path, pad_with_end=True, embedding_directory=embedding_directory, embedding_size=1280, embedding_key='clip_g', tokenizer_data=tokenizer_data)
 
 class StableCascadeTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -84,6 +83,7 @@ class StableCascadeTokenizer(sd1_clip.SD1Tokenizer):
 class StableCascadeClipG(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", max_length=77, freeze=True, layer="hidden", layer_idx=-1, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_config_bigg.json")
+        model_options = {**model_options, "model_name": "clip_g"}
         super().__init__(device=device, freeze=freeze, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype,
                          special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=False, enable_attention_masks=True, return_projected_pooled=True, model_options=model_options)
 

comfy/supported_models.py

@@ -15,6 +15,9 @@ import comfy.text_encoders.genmo
 import comfy.text_encoders.lt
 import comfy.text_encoders.hunyuan_video
 import comfy.text_encoders.cosmos
+import comfy.text_encoders.lumina2
+import comfy.text_encoders.wan
+import comfy.text_encoders.ace
 
 from . import supported_models_base
 from . import latent_formats
@@ -504,6 +507,22 @@ class SDXL_instructpix2pix(SDXL):
     def get_model(self, state_dict, prefix="", device=None):
         return model_base.SDXL_instructpix2pix(self, model_type=self.model_type(state_dict, prefix), device=device)
 
+class LotusD(SD20):
+    unet_config = {
+        "model_channels": 320,
+        "use_linear_in_transformer": True,
+        "use_temporal_attention": False,
+        "adm_in_channels": 4,
+        "in_channels": 4,
+    }
+
+    unet_extra_config = {
+        "num_classes": 'sequential'
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        return model_base.Lotus(self, device=device)
+
 class SD3(supported_models_base.BASE):
     unet_config = {
         "in_channels": 16,
@@ -760,13 +779,17 @@ class LTXV(supported_models_base.BASE):
     unet_extra_config = {}
     latent_format = latent_formats.LTXV
 
-    memory_usage_factor = 2.7
+    memory_usage_factor = 5.5 # TODO: img2vid is about 2x vs txt2vid
 
     supported_inference_dtypes = [torch.bfloat16, torch.float32]
 
     vae_key_prefix = ["vae."]
     text_encoder_key_prefix = ["text_encoders."]
 
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = (unet_config.get("cross_attention_dim", 2048) / 2048) * 5.5
+
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.LTXV(self, device=device)
         return out
@@ -824,6 +847,26 @@ class HunyuanVideo(supported_models_base.BASE):
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}llama.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.hunyuan_video.HunyuanVideoTokenizer, comfy.text_encoders.hunyuan_video.hunyuan_video_clip(**hunyuan_detect))
 
+class HunyuanVideoI2V(HunyuanVideo):
+    unet_config = {
+        "image_model": "hunyuan_video",
+        "in_channels": 33,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.HunyuanVideoI2V(self, device=device)
+        return out
+
+class HunyuanVideoSkyreelsI2V(HunyuanVideo):
+    unet_config = {
+        "image_model": "hunyuan_video",
+        "in_channels": 32,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.HunyuanVideoSkyreelsI2V(self, device=device)
+        return out
+
 class CosmosT2V(supported_models_base.BASE):
     unet_config = {
         "image_model": "cosmos",
@@ -865,6 +908,237 @@ class CosmosI2V(CosmosT2V):
         out = model_base.CosmosVideo(self, image_to_video=True, device=device)
         return out
 
-models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo, CosmosT2V, CosmosI2V]
+class Lumina2(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "lumina2",
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 6.0,
+    }
+
+    memory_usage_factor = 1.2
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Flux
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Lumina2(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}gemma2_2b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.lumina2.LuminaTokenizer, comfy.text_encoders.lumina2.te(**hunyuan_detect))
+
+class WAN21_T2V(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "t2v",
+    }
+
+    sampling_settings = {
+        "shift": 8.0,
+    }
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Wan21
+
+    memory_usage_factor = 1.0
+
+    supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = self.unet_config.get("dim", 2000) / 2000
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}umt5xxl.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.wan.WanT5Tokenizer, comfy.text_encoders.wan.te(**t5_detect))
+
+class WAN21_I2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "i2v",
+        "in_dim": 36,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21(self, image_to_video=True, device=device)
+        return out
+
+class WAN21_FunControl2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "i2v",
+        "in_dim": 48,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21(self, image_to_video=False, device=device)
+        return out
+
+class WAN21_Camera(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "camera",
+        "in_dim": 32,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21_Camera(self, image_to_video=False, device=device)
+        return out
+class WAN21_Vace(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "vace",
+    }
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = 1.2 * self.memory_usage_factor
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21_Vace(self, image_to_video=False, device=device)
+        return out
+
+class Hunyuan3Dv2(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "hunyuan3d2",
+    }
+
+    unet_extra_config = {}
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 1.0,
+    }
+
+    memory_usage_factor = 3.5
+
+    clip_vision_prefix = "conditioner.main_image_encoder.model."
+    vae_key_prefix = ["vae."]
+
+    latent_format = latent_formats.Hunyuan3Dv2
+
+    def process_unet_state_dict_for_saving(self, state_dict):
+        replace_prefix = {"": "model."}
+        return utils.state_dict_prefix_replace(state_dict, replace_prefix)
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Hunyuan3Dv2(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return None
+
+class Hunyuan3Dv2mini(Hunyuan3Dv2):
+    unet_config = {
+        "image_model": "hunyuan3d2",
+        "depth": 8,
+    }
+
+    latent_format = latent_formats.Hunyuan3Dv2mini
+
+class HiDream(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "hidream",
+    }
+
+    sampling_settings = {
+        "shift": 3.0,
+    }
+
+    sampling_settings = {
+    }
+
+    # memory_usage_factor = 1.2 # TODO
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Flux
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.HiDream(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return None #  TODO
+
+class Chroma(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "chroma",
+    }
+
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+    }
+
+    latent_format = comfy.latent_formats.Flux
+
+    memory_usage_factor = 3.2
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.Chroma(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.pixart_t5.PixArtTokenizer, comfy.text_encoders.pixart_t5.pixart_te(**t5_detect))
+
+class ACEStep(supported_models_base.BASE):
+    unet_config = {
+        "audio_model": "ace",
+    }
+
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "shift": 3.0,
+    }
+
+    latent_format = comfy.latent_formats.ACEAudio
+
+    memory_usage_factor = 0.5
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.ACEStep(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return supported_models_base.ClipTarget(comfy.text_encoders.ace.AceT5Tokenizer, comfy.text_encoders.ace.AceT5Model)
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep]
 
 models += [SVD_img2vid]

comfy/text_encoders/ace.py

@@ -0,0 +1,153 @@
+from comfy import sd1_clip
+from .spiece_tokenizer import SPieceTokenizer
+import comfy.text_encoders.t5
+import os
+import re
+import torch
+import logging
+
+from tokenizers import Tokenizer
+from .ace_text_cleaners import multilingual_cleaners, japanese_to_romaji
+
+SUPPORT_LANGUAGES = {
+    "en": 259, "de": 260, "fr": 262, "es": 284, "it": 285,
+    "pt": 286, "pl": 294, "tr": 295, "ru": 267, "cs": 293,
+    "nl": 297, "ar": 5022, "zh": 5023, "ja": 5412, "hu": 5753,
+    "ko": 6152, "hi": 6680
+}
+
+structure_pattern = re.compile(r"\[.*?\]")
+
+DEFAULT_VOCAB_FILE = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "ace_lyrics_tokenizer"), "vocab.json")
+
+
+class VoiceBpeTokenizer:
+    def __init__(self, vocab_file=DEFAULT_VOCAB_FILE):
+        self.tokenizer = None
+        if vocab_file is not None:
+            self.tokenizer = Tokenizer.from_file(vocab_file)
+
+    def preprocess_text(self, txt, lang):
+        txt = multilingual_cleaners(txt, lang)
+        return txt
+
+    def encode(self, txt, lang='en'):
+        # lang = lang.split("-")[0]  # remove the region
+        # self.check_input_length(txt, lang)
+        txt = self.preprocess_text(txt, lang)
+        lang = "zh-cn" if lang == "zh" else lang
+        txt = f"[{lang}]{txt}"
+        txt = txt.replace(" ", "[SPACE]")
+        return self.tokenizer.encode(txt).ids
+
+    def get_lang(self, line):
+        if line.startswith("[") and line[3:4] == ']':
+            lang = line[1:3].lower()
+            if lang in SUPPORT_LANGUAGES:
+                return lang, line[4:]
+        return "en", line
+
+    def __call__(self, string):
+        lines = string.split("\n")
+        lyric_token_idx = [261]
+        for line in lines:
+            line = line.strip()
+            if not line:
+                lyric_token_idx += [2]
+                continue
+
+            lang, line = self.get_lang(line)
+
+            if lang not in SUPPORT_LANGUAGES:
+                lang = "en"
+            if "zh" in lang:
+                lang = "zh"
+            if "spa" in lang:
+                lang = "es"
+
+            try:
+                line_out = japanese_to_romaji(line)
+                if line_out != line:
+                    lang = "ja"
+                line = line_out
+            except:
+                pass
+
+            try:
+                if structure_pattern.match(line):
+                    token_idx = self.encode(line, "en")
+                else:
+                    token_idx = self.encode(line, lang)
+                lyric_token_idx = lyric_token_idx + token_idx + [2]
+            except Exception as e:
+                logging.warning("tokenize error {} for line {} major_language {}".format(e, line, lang))
+        return {"input_ids": lyric_token_idx}
+
+    @staticmethod
+    def from_pretrained(path, **kwargs):
+        return VoiceBpeTokenizer(path, **kwargs)
+
+    def get_vocab(self):
+        return {}
+
+
+class UMT5BaseModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "umt5_config_base.json")
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=False, model_options=model_options)
+
+class UMT5BaseTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = tokenizer_data.get("spiece_model", None)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=768, embedding_key='umt5base', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=0, tokenizer_data=tokenizer_data)
+
+    def state_dict(self):
+        return {"spiece_model": self.tokenizer.serialize_model()}
+
+class LyricsTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "ace_lyrics_tokenizer"), "vocab.json")
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=1024, embedding_key='lyrics', tokenizer_class=VoiceBpeTokenizer, has_start_token=True, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=2, has_end_token=False, tokenizer_data=tokenizer_data)
+
+class AceT5Tokenizer:
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        self.voicebpe = LyricsTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+        self.umt5base = UMT5BaseTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
+        out = {}
+        out["lyrics"] = self.voicebpe.tokenize_with_weights(kwargs.get("lyrics", ""), return_word_ids, **kwargs)
+        out["umt5base"] = self.umt5base.tokenize_with_weights(text, return_word_ids, **kwargs)
+        return out
+
+    def untokenize(self, token_weight_pair):
+        return self.umt5base.untokenize(token_weight_pair)
+
+    def state_dict(self):
+        return self.umt5base.state_dict()
+
+class AceT5Model(torch.nn.Module):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__()
+        self.umt5base = UMT5BaseModel(device=device, dtype=dtype, model_options=model_options)
+        self.dtypes = set()
+        if dtype is not None:
+            self.dtypes.add(dtype)
+
+    def set_clip_options(self, options):
+        self.umt5base.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.umt5base.reset_clip_options()
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_umt5base = token_weight_pairs["umt5base"]
+        token_weight_pairs_lyrics = token_weight_pairs["lyrics"]
+
+        t5_out, t5_pooled = self.umt5base.encode_token_weights(token_weight_pairs_umt5base)
+
+        lyrics_embeds = torch.tensor(list(map(lambda a: a[0], token_weight_pairs_lyrics[0]))).unsqueeze(0)
+        return t5_out, None, {"conditioning_lyrics": lyrics_embeds}
+
+    def load_sd(self, sd):
+        return self.umt5base.load_sd(sd)

comfy/text_encoders/ace_text_cleaners.py

@@ -0,0 +1,395 @@
+# basic text cleaners for the ACE step model
+# I didn't copy the ones from the reference code because I didn't want to deal with the dependencies
+# TODO: more languages than english?
+
+import re
+
+def japanese_to_romaji(japanese_text):
+    """
+    Convert Japanese hiragana and katakana to romaji (Latin alphabet representation).
+
+    Args:
+        japanese_text (str): Text containing hiragana and/or katakana characters
+
+    Returns:
+        str: The romaji (Latin alphabet) equivalent
+    """
+    # Dictionary mapping kana characters to their romaji equivalents
+    kana_map = {
+        # Katakana characters
+        'ア': 'a', 'イ': 'i', 'ウ': 'u', 'エ': 'e', 'オ': 'o',
+        'カ': 'ka', 'キ': 'ki', 'ク': 'ku', 'ケ': 'ke', 'コ': 'ko',
+        'サ': 'sa', 'シ': 'shi', 'ス': 'su', 'セ': 'se', 'ソ': 'so',
+        'タ': 'ta', 'チ': 'chi', 'ツ': 'tsu', 'テ': 'te', 'ト': 'to',
+        'ナ': 'na', 'ニ': 'ni', 'ヌ': 'nu', 'ネ': 'ne', 'ノ': 'no',
+        'ハ': 'ha', 'ヒ': 'hi', 'フ': 'fu', 'ヘ': 'he', 'ホ': 'ho',
+        'マ': 'ma', 'ミ': 'mi', 'ム': 'mu', 'メ': 'me', 'モ': 'mo',
+        'ヤ': 'ya', 'ユ': 'yu', 'ヨ': 'yo',
+        'ラ': 'ra', 'リ': 'ri', 'ル': 'ru', 'レ': 're', 'ロ': 'ro',
+        'ワ': 'wa', 'ヲ': 'wo', 'ン': 'n',
+
+        # Katakana voiced consonants
+        'ガ': 'ga', 'ギ': 'gi', 'グ': 'gu', 'ゲ': 'ge', 'ゴ': 'go',
+        'ザ': 'za', 'ジ': 'ji', 'ズ': 'zu', 'ゼ': 'ze', 'ゾ': 'zo',
+        'ダ': 'da', 'ヂ': 'ji', 'ヅ': 'zu', 'デ': 'de', 'ド': 'do',
+        'バ': 'ba', 'ビ': 'bi', 'ブ': 'bu', 'ベ': 'be', 'ボ': 'bo',
+        'パ': 'pa', 'ピ': 'pi', 'プ': 'pu', 'ペ': 'pe', 'ポ': 'po',
+
+        # Katakana combinations
+        'キャ': 'kya', 'キュ': 'kyu', 'キョ': 'kyo',
+        'シャ': 'sha', 'シュ': 'shu', 'ショ': 'sho',
+        'チャ': 'cha', 'チュ': 'chu', 'チョ': 'cho',
+        'ニャ': 'nya', 'ニュ': 'nyu', 'ニョ': 'nyo',
+        'ヒャ': 'hya', 'ヒュ': 'hyu', 'ヒョ': 'hyo',
+        'ミャ': 'mya', 'ミュ': 'myu', 'ミョ': 'myo',
+        'リャ': 'rya', 'リュ': 'ryu', 'リョ': 'ryo',
+        'ギャ': 'gya', 'ギュ': 'gyu', 'ギョ': 'gyo',
+        'ジャ': 'ja', 'ジュ': 'ju', 'ジョ': 'jo',
+        'ビャ': 'bya', 'ビュ': 'byu', 'ビョ': 'byo',
+        'ピャ': 'pya', 'ピュ': 'pyu', 'ピョ': 'pyo',
+
+        # Katakana small characters and special cases
+        'ッ': '', # Small tsu (doubles the following consonant)
+        'ャ': 'ya', 'ュ': 'yu', 'ョ': 'yo',
+
+        # Katakana extras
+        'ヴ': 'vu', 'ファ': 'fa', 'フィ': 'fi', 'フェ': 'fe', 'フォ': 'fo',
+        'ウィ': 'wi', 'ウェ': 'we', 'ウォ': 'wo',
+
+        # Hiragana characters
+        'あ': 'a', 'い': 'i', 'う': 'u', 'え': 'e', 'お': 'o',
+        'か': 'ka', 'き': 'ki', 'く': 'ku', 'け': 'ke', 'こ': 'ko',
+        'さ': 'sa', 'し': 'shi', 'す': 'su', 'せ': 'se', 'そ': 'so',
+        'た': 'ta', 'ち': 'chi', 'つ': 'tsu', 'て': 'te', 'と': 'to',
+        'な': 'na', 'に': 'ni', 'ぬ': 'nu', 'ね': 'ne', 'の': 'no',
+        'は': 'ha', 'ひ': 'hi', 'ふ': 'fu', 'へ': 'he', 'ほ': 'ho',
+        'ま': 'ma', 'み': 'mi', 'む': 'mu', 'め': 'me', 'も': 'mo',
+        'や': 'ya', 'ゆ': 'yu', 'よ': 'yo',
+        'ら': 'ra', 'り': 'ri', 'る': 'ru', 'れ': 're', 'ろ': 'ro',
+        'わ': 'wa', 'を': 'wo', 'ん': 'n',
+
+        # Hiragana voiced consonants
+        'が': 'ga', 'ぎ': 'gi', 'ぐ': 'gu', 'げ': 'ge', 'ご': 'go',
+        'ざ': 'za', 'じ': 'ji', 'ず': 'zu', 'ぜ': 'ze', 'ぞ': 'zo',
+        'だ': 'da', 'ぢ': 'ji', 'づ': 'zu', 'で': 'de', 'ど': 'do',
+        'ば': 'ba', 'び': 'bi', 'ぶ': 'bu', 'べ': 'be', 'ぼ': 'bo',
+        'ぱ': 'pa', 'ぴ': 'pi', 'ぷ': 'pu', 'ぺ': 'pe', 'ぽ': 'po',
+
+        # Hiragana combinations
+        'きゃ': 'kya', 'きゅ': 'kyu', 'きょ': 'kyo',
+        'しゃ': 'sha', 'しゅ': 'shu', 'しょ': 'sho',
+        'ちゃ': 'cha', 'ちゅ': 'chu', 'ちょ': 'cho',
+        'にゃ': 'nya', 'にゅ': 'nyu', 'にょ': 'nyo',
+        'ひゃ': 'hya', 'ひゅ': 'hyu', 'ひょ': 'hyo',
+        'みゃ': 'mya', 'みゅ': 'myu', 'みょ': 'myo',
+        'りゃ': 'rya', 'りゅ': 'ryu', 'りょ': 'ryo',
+        'ぎゃ': 'gya', 'ぎゅ': 'gyu', 'ぎょ': 'gyo',
+        'じゃ': 'ja', 'じゅ': 'ju', 'じょ': 'jo',
+        'びゃ': 'bya', 'びゅ': 'byu', 'びょ': 'byo',
+        'ぴゃ': 'pya', 'ぴゅ': 'pyu', 'ぴょ': 'pyo',
+
+        # Hiragana small characters and special cases
+        'っ': '', # Small tsu (doubles the following consonant)
+        'ゃ': 'ya', 'ゅ': 'yu', 'ょ': 'yo',
+
+        # Common punctuation and spaces
+        '　': ' ', # Japanese space
+        '、': ', ', '。': '. ',
+    }
+
+    result = []
+    i = 0
+
+    while i < len(japanese_text):
+        # Check for small tsu (doubling the following consonant)
+        if i < len(japanese_text) - 1 and (japanese_text[i] == 'っ' or japanese_text[i] == 'ッ'):
+            if i < len(japanese_text) - 1 and japanese_text[i+1] in kana_map:
+                next_romaji = kana_map[japanese_text[i+1]]
+                if next_romaji and next_romaji[0] not in 'aiueon':
+                    result.append(next_romaji[0])  # Double the consonant
+            i += 1
+            continue
+
+        # Check for combinations with small ya, yu, yo
+        if i < len(japanese_text) - 1 and japanese_text[i+1] in ('ゃ', 'ゅ', 'ょ', 'ャ', 'ュ', 'ョ'):
+            combo = japanese_text[i:i+2]
+            if combo in kana_map:
+                result.append(kana_map[combo])
+                i += 2
+                continue
+
+        # Regular character
+        if japanese_text[i] in kana_map:
+            result.append(kana_map[japanese_text[i]])
+        else:
+            # If it's not in our map, keep it as is (might be kanji, romaji, etc.)
+            result.append(japanese_text[i])
+
+        i += 1
+
+    return ''.join(result)
+
+def number_to_text(num, ordinal=False):
+    """
+    Convert a number (int or float) to its text representation.
+
+    Args:
+        num: The number to convert
+
+    Returns:
+        str: Text representation of the number
+    """
+
+    if not isinstance(num, (int, float)):
+        return "Input must be a number"
+
+    # Handle special case of zero
+    if num == 0:
+        return "zero"
+
+    # Handle negative numbers
+    negative = num < 0
+    num = abs(num)
+
+    # Handle floats
+    if isinstance(num, float):
+        # Split into integer and decimal parts
+        int_part = int(num)
+
+        # Convert both parts
+        int_text = _int_to_text(int_part)
+
+        # Handle decimal part (convert to string and remove '0.')
+        decimal_str = str(num).split('.')[1]
+        decimal_text = " point " + " ".join(_digit_to_text(int(digit)) for digit in decimal_str)
+
+        result = int_text + decimal_text
+    else:
+        # Handle integers
+        result = _int_to_text(num)
+
+    # Add 'negative' prefix for negative numbers
+    if negative:
+        result = "negative " + result
+
+    return result
+
+
+def _int_to_text(num):
+    """Helper function to convert an integer to text"""
+
+    ones = ["", "one", "two", "three", "four", "five", "six", "seven", "eight", "nine",
+            "ten", "eleven", "twelve", "thirteen", "fourteen", "fifteen", "sixteen",
+            "seventeen", "eighteen", "nineteen"]
+
+    tens = ["", "", "twenty", "thirty", "forty", "fifty", "sixty", "seventy", "eighty", "ninety"]
+
+    if num < 20:
+        return ones[num]
+
+    if num < 100:
+        return tens[num // 10] + (" " + ones[num % 10] if num % 10 != 0 else "")
+
+    if num < 1000:
+        return ones[num // 100] + " hundred" + (" " + _int_to_text(num % 100) if num % 100 != 0 else "")
+
+    if num < 1000000:
+        return _int_to_text(num // 1000) + " thousand" + (" " + _int_to_text(num % 1000) if num % 1000 != 0 else "")
+
+    if num < 1000000000:
+        return _int_to_text(num // 1000000) + " million" + (" " + _int_to_text(num % 1000000) if num % 1000000 != 0 else "")
+
+    return _int_to_text(num // 1000000000) + " billion" + (" " + _int_to_text(num % 1000000000) if num % 1000000000 != 0 else "")
+
+
+def _digit_to_text(digit):
+    """Convert a single digit to text"""
+    digits = ["zero", "one", "two", "three", "four", "five", "six", "seven", "eight", "nine"]
+    return digits[digit]
+
+
+_whitespace_re = re.compile(r"\s+")
+
+
+# List of (regular expression, replacement) pairs for abbreviations:
+_abbreviations = {
+    "en": [
+        (re.compile("\\b%s\\." % x[0], re.IGNORECASE), x[1])
+        for x in [
+            ("mrs", "misess"),
+            ("mr", "mister"),
+            ("dr", "doctor"),
+            ("st", "saint"),
+            ("co", "company"),
+            ("jr", "junior"),
+            ("maj", "major"),
+            ("gen", "general"),
+            ("drs", "doctors"),
+            ("rev", "reverend"),
+            ("lt", "lieutenant"),
+            ("hon", "honorable"),
+            ("sgt", "sergeant"),
+            ("capt", "captain"),
+            ("esq", "esquire"),
+            ("ltd", "limited"),
+            ("col", "colonel"),
+            ("ft", "fort"),
+        ]
+    ],
+}
+
+
+def expand_abbreviations_multilingual(text, lang="en"):
+    for regex, replacement in _abbreviations[lang]:
+        text = re.sub(regex, replacement, text)
+    return text
+
+
+_symbols_multilingual = {
+    "en": [
+        (re.compile(r"%s" % re.escape(x[0]), re.IGNORECASE), x[1])
+        for x in [
+            ("&", " and "),
+            ("@", " at "),
+            ("%", " percent "),
+            ("#", " hash "),
+            ("$", " dollar "),
+            ("£", " pound "),
+            ("°", " degree "),
+        ]
+    ],
+}
+
+
+def expand_symbols_multilingual(text, lang="en"):
+    for regex, replacement in _symbols_multilingual[lang]:
+        text = re.sub(regex, replacement, text)
+        text = text.replace("  ", " ")  # Ensure there are no double spaces
+    return text.strip()
+
+
+_ordinal_re = {
+    "en": re.compile(r"([0-9]+)(st|nd|rd|th)"),
+}
+_number_re = re.compile(r"[0-9]+")
+_currency_re = {
+    "USD": re.compile(r"((\$[0-9\.\,]*[0-9]+)|([0-9\.\,]*[0-9]+\$))"),
+    "GBP": re.compile(r"((£[0-9\.\,]*[0-9]+)|([0-9\.\,]*[0-9]+£))"),
+    "EUR": re.compile(r"(([0-9\.\,]*[0-9]+€)|((€[0-9\.\,]*[0-9]+)))"),
+}
+
+_comma_number_re = re.compile(r"\b\d{1,3}(,\d{3})*(\.\d+)?\b")
+_dot_number_re = re.compile(r"\b\d{1,3}(.\d{3})*(\,\d+)?\b")
+_decimal_number_re = re.compile(r"([0-9]+[.,][0-9]+)")
+
+
+def _remove_commas(m):
+    text = m.group(0)
+    if "," in text:
+        text = text.replace(",", "")
+    return text
+
+
+def _remove_dots(m):
+    text = m.group(0)
+    if "." in text:
+        text = text.replace(".", "")
+    return text
+
+
+def _expand_decimal_point(m, lang="en"):
+    amount = m.group(1).replace(",", ".")
+    return number_to_text(float(amount))
+
+
+def _expand_currency(m, lang="en", currency="USD"):
+    amount = float((re.sub(r"[^\d.]", "", m.group(0).replace(",", "."))))
+    full_amount = number_to_text(amount)
+
+    and_equivalents = {
+        "en": ", ",
+        "es": " con ",
+        "fr": " et ",
+        "de": " und ",
+        "pt": " e ",
+        "it": " e ",
+        "pl": ", ",
+        "cs": ", ",
+        "ru": ", ",
+        "nl": ", ",
+        "ar": ", ",
+        "tr": ", ",
+        "hu": ", ",
+        "ko": ", ",
+    }
+
+    if amount.is_integer():
+        last_and = full_amount.rfind(and_equivalents[lang])
+        if last_and != -1:
+            full_amount = full_amount[:last_and]
+
+    return full_amount
+
+
+def _expand_ordinal(m, lang="en"):
+    return number_to_text(int(m.group(1)), ordinal=True)
+
+
+def _expand_number(m, lang="en"):
+    return number_to_text(int(m.group(0)))
+
+
+def expand_numbers_multilingual(text, lang="en"):
+    if lang in ["en", "ru"]:
+        text = re.sub(_comma_number_re, _remove_commas, text)
+    else:
+        text = re.sub(_dot_number_re, _remove_dots, text)
+    try:
+        text = re.sub(_currency_re["GBP"], lambda m: _expand_currency(m, lang, "GBP"), text)
+        text = re.sub(_currency_re["USD"], lambda m: _expand_currency(m, lang, "USD"), text)
+        text = re.sub(_currency_re["EUR"], lambda m: _expand_currency(m, lang, "EUR"), text)
+    except:
+        pass
+
+    text = re.sub(_decimal_number_re, lambda m: _expand_decimal_point(m, lang), text)
+    text = re.sub(_ordinal_re[lang], lambda m: _expand_ordinal(m, lang), text)
+    text = re.sub(_number_re, lambda m: _expand_number(m, lang), text)
+    return text
+
+
+def lowercase(text):
+    return text.lower()
+
+
+def collapse_whitespace(text):
+    return re.sub(_whitespace_re, " ", text)
+
+
+def multilingual_cleaners(text, lang):
+    text = text.replace('"', "")
+    if lang == "tr":
+        text = text.replace("İ", "i")
+        text = text.replace("Ö", "ö")
+        text = text.replace("Ü", "ü")
+    text = lowercase(text)
+    try:
+        text = expand_numbers_multilingual(text, lang)
+    except:
+        pass
+    try:
+        text = expand_abbreviations_multilingual(text, lang)
+    except:
+        pass
+    try:
+        text = expand_symbols_multilingual(text, lang=lang)
+    except:
+        pass
+    text = collapse_whitespace(text)
+    return text
+
+
+def basic_cleaners(text):
+    """Basic pipeline that lowercases and collapses whitespace without transliteration."""
+    text = lowercase(text)
+    text = collapse_whitespace(text)
+    return text

comfy/text_encoders/aura_t5.py

@@ -11,7 +11,7 @@ class PT5XlModel(sd1_clip.SDClipModel):
 class PT5XlTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_pile_tokenizer"), "tokenizer.model")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2048, embedding_key='pile_t5xl', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256, pad_token=1)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2048, embedding_key='pile_t5xl', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256, pad_token=1, tokenizer_data=tokenizer_data)
 
 class AuraT5Tokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):

comfy/text_encoders/bert.py

@@ -93,7 +93,10 @@ class BertEmbeddings(torch.nn.Module):
 
         self.LayerNorm = operations.LayerNorm(embed_dim, eps=layer_norm_eps, dtype=dtype, device=device)
 
-    def forward(self, input_tokens, token_type_ids=None, dtype=None):
+    def forward(self, input_tokens, embeds=None, token_type_ids=None, dtype=None):
+        if embeds is not None:
+            x = embeds
+        else:
             x = self.word_embeddings(input_tokens, out_dtype=dtype)
         x += comfy.ops.cast_to_input(self.position_embeddings.weight[:x.shape[1]], x)
         if token_type_ids is not None:
@@ -113,12 +116,12 @@ class BertModel_(torch.nn.Module):
         self.embeddings = BertEmbeddings(config_dict["vocab_size"], config_dict["max_position_embeddings"], config_dict["type_vocab_size"], config_dict["pad_token_id"], embed_dim, layer_norm_eps, dtype, device, operations)
         self.encoder = BertEncoder(config_dict["num_hidden_layers"], embed_dim, config_dict["intermediate_size"], config_dict["num_attention_heads"], layer_norm_eps, dtype, device, operations)
 
-    def forward(self, input_tokens, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
-        x = self.embeddings(input_tokens, dtype=dtype)
+    def forward(self, input_tokens, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
+        x = self.embeddings(input_tokens, embeds=embeds, dtype=dtype)
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
-            mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
+            mask = mask.masked_fill(mask.to(torch.bool), -torch.finfo(x.dtype).max)
 
         x, i = self.encoder(x, mask, intermediate_output)
         return x, i

comfy/text_encoders/cosmos.py

@@ -22,7 +22,7 @@ class CosmosT5XXL(sd1_clip.SD1ClipModel):
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
-        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=1024, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=512)
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=1024, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, tokenizer_data=tokenizer_data)
 
 
 class CosmosT5Tokenizer(sd1_clip.SD1Tokenizer):

comfy/text_encoders/flux.py

@@ -9,19 +9,18 @@ import os
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
-        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256)
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256, tokenizer_data=tokenizer_data)
 
 
 class FluxTokenizer:
     def __init__(self, embedding_directory=None, tokenizer_data={}):
-        clip_l_tokenizer_class = tokenizer_data.get("clip_l_tokenizer_class", sd1_clip.SDTokenizer)
-        self.clip_l = clip_l_tokenizer_class(embedding_directory=embedding_directory)
-        self.t5xxl = T5XXLTokenizer(embedding_directory=embedding_directory)
+        self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+        self.t5xxl = T5XXLTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
 
-    def tokenize_with_weights(self, text:str, return_word_ids=False):
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
-        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
-        out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
     def untokenize(self, token_weight_pair):
@@ -35,8 +34,7 @@ class FluxClipModel(torch.nn.Module):
     def __init__(self, dtype_t5=None, device="cpu", dtype=None, model_options={}):
         super().__init__()
         dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
-        clip_l_class = model_options.get("clip_l_class", sd1_clip.SDClipModel)
-        self.clip_l = clip_l_class(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options)
+        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options)
         self.t5xxl = comfy.text_encoders.sd3_clip.T5XXLModel(device=device, dtype=dtype_t5, model_options=model_options)
         self.dtypes = set([dtype, dtype_t5])
 

comfy/text_encoders/genmo.py

@@ -18,7 +18,7 @@ class MochiT5XXL(sd1_clip.SD1ClipModel):
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "t5_tokenizer")
-        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256)
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='t5xxl', tokenizer_class=T5TokenizerFast, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256, tokenizer_data=tokenizer_data)
 
 
 class MochiT5Tokenizer(sd1_clip.SD1Tokenizer):

comfy/text_encoders/hidream.py

@@ -0,0 +1,155 @@
+from . import hunyuan_video
+from . import sd3_clip
+from comfy import sd1_clip
+from comfy import sdxl_clip
+import comfy.model_management
+import torch
+import logging
+
+
+class HiDreamTokenizer:
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+        self.clip_g = sdxl_clip.SDXLClipGTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+        self.t5xxl = sd3_clip.T5XXLTokenizer(embedding_directory=embedding_directory, min_length=128, max_length=128, tokenizer_data=tokenizer_data)
+        self.llama = hunyuan_video.LLAMA3Tokenizer(embedding_directory=embedding_directory, min_length=128, pad_token=128009, tokenizer_data=tokenizer_data)
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
+        out = {}
+        out["g"] = self.clip_g.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
+        t5xxl = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["t5xxl"] = [t5xxl[0]]  # Use only first 128 tokens
+        out["llama"] = self.llama.tokenize_with_weights(text, return_word_ids, **kwargs)
+        return out
+
+    def untokenize(self, token_weight_pair):
+        return self.clip_g.untokenize(token_weight_pair)
+
+    def state_dict(self):
+        return {}
+
+
+class HiDreamTEModel(torch.nn.Module):
+    def __init__(self, clip_l=True, clip_g=True, t5=True, llama=True, dtype_t5=None, dtype_llama=None, device="cpu", dtype=None, model_options={}):
+        super().__init__()
+        self.dtypes = set()
+        if clip_l:
+            self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=True, model_options=model_options)
+            self.dtypes.add(dtype)
+        else:
+            self.clip_l = None
+
+        if clip_g:
+            self.clip_g = sdxl_clip.SDXLClipG(device=device, dtype=dtype, model_options=model_options)
+            self.dtypes.add(dtype)
+        else:
+            self.clip_g = None
+
+        if t5:
+            dtype_t5 = comfy.model_management.pick_weight_dtype(dtype_t5, dtype, device)
+            self.t5xxl = sd3_clip.T5XXLModel(device=device, dtype=dtype_t5, model_options=model_options, attention_mask=True)
+            self.dtypes.add(dtype_t5)
+        else:
+            self.t5xxl = None
+
+        if llama:
+            dtype_llama = comfy.model_management.pick_weight_dtype(dtype_llama, dtype, device)
+            if "vocab_size" not in model_options:
+                model_options["vocab_size"] = 128256
+            self.llama = hunyuan_video.LLAMAModel(device=device, dtype=dtype_llama, model_options=model_options, layer="all", layer_idx=None, special_tokens={"start": 128000, "pad": 128009})
+            self.dtypes.add(dtype_llama)
+        else:
+            self.llama = None
+
+        logging.debug("Created HiDream text encoder with: clip_l {}, clip_g {}, t5xxl {}:{}, llama {}:{}".format(clip_l, clip_g, t5, dtype_t5, llama, dtype_llama))
+
+    def set_clip_options(self, options):
+        if self.clip_l is not None:
+            self.clip_l.set_clip_options(options)
+        if self.clip_g is not None:
+            self.clip_g.set_clip_options(options)
+        if self.t5xxl is not None:
+            self.t5xxl.set_clip_options(options)
+        if self.llama is not None:
+            self.llama.set_clip_options(options)
+
+    def reset_clip_options(self):
+        if self.clip_l is not None:
+            self.clip_l.reset_clip_options()
+        if self.clip_g is not None:
+            self.clip_g.reset_clip_options()
+        if self.t5xxl is not None:
+            self.t5xxl.reset_clip_options()
+        if self.llama is not None:
+            self.llama.reset_clip_options()
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_l = token_weight_pairs["l"]
+        token_weight_pairs_g = token_weight_pairs["g"]
+        token_weight_pairs_t5 = token_weight_pairs["t5xxl"]
+        token_weight_pairs_llama = token_weight_pairs["llama"]
+        lg_out = None
+        pooled = None
+        extra = {}
+
+        if len(token_weight_pairs_g) > 0 or len(token_weight_pairs_l) > 0:
+            if self.clip_l is not None:
+                lg_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs_l)
+            else:
+                l_pooled = torch.zeros((1, 768), device=comfy.model_management.intermediate_device())
+
+            if self.clip_g is not None:
+                g_out, g_pooled = self.clip_g.encode_token_weights(token_weight_pairs_g)
+            else:
+                g_pooled = torch.zeros((1, 1280), device=comfy.model_management.intermediate_device())
+
+            pooled = torch.cat((l_pooled, g_pooled), dim=-1)
+
+        if self.t5xxl is not None:
+            t5_output = self.t5xxl.encode_token_weights(token_weight_pairs_t5)
+            t5_out, t5_pooled = t5_output[:2]
+        else:
+            t5_out = None
+
+        if self.llama is not None:
+            ll_output = self.llama.encode_token_weights(token_weight_pairs_llama)
+            ll_out, ll_pooled = ll_output[:2]
+            ll_out = ll_out[:, 1:]
+        else:
+            ll_out = None
+
+        if t5_out is None:
+            t5_out = torch.zeros((1, 128, 4096), device=comfy.model_management.intermediate_device())
+
+        if ll_out is None:
+            ll_out = torch.zeros((1, 32, 1, 4096), device=comfy.model_management.intermediate_device())
+
+        if pooled is None:
+            pooled = torch.zeros((1, 768 + 1280), device=comfy.model_management.intermediate_device())
+
+        extra["conditioning_llama3"] = ll_out
+        return t5_out, pooled, extra
+
+    def load_sd(self, sd):
+        if "text_model.encoder.layers.30.mlp.fc1.weight" in sd:
+            return self.clip_g.load_sd(sd)
+        elif "text_model.encoder.layers.1.mlp.fc1.weight" in sd:
+            return self.clip_l.load_sd(sd)
+        elif "encoder.block.23.layer.1.DenseReluDense.wi_1.weight" in sd:
+            return self.t5xxl.load_sd(sd)
+        else:
+            return self.llama.load_sd(sd)
+
+
+def hidream_clip(clip_l=True, clip_g=True, t5=True, llama=True, dtype_t5=None, dtype_llama=None, t5xxl_scaled_fp8=None, llama_scaled_fp8=None):
+    class HiDreamTEModel_(HiDreamTEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if t5xxl_scaled_fp8 is not None and "t5xxl_scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["t5xxl_scaled_fp8"] = t5xxl_scaled_fp8
+            if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["llama_scaled_fp8"] = llama_scaled_fp8
+            super().__init__(clip_l=clip_l, clip_g=clip_g, t5=t5, llama=llama, dtype_t5=dtype_t5, dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)
+    return HiDreamTEModel_

comfy/text_encoders/hunyuan_video.py

@@ -4,6 +4,7 @@ import comfy.text_encoders.llama
 from transformers import LlamaTokenizerFast
 import torch
 import os
+import numbers
 
 
 def llama_detect(state_dict, prefix=""):
@@ -20,33 +21,49 @@ def llama_detect(state_dict, prefix=""):
 
 
 class LLAMA3Tokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}, min_length=256):
+    def __init__(self, embedding_directory=None, tokenizer_data={}, min_length=256, pad_token=128258):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "llama_tokenizer")
-        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='llama', tokenizer_class=LlamaTokenizerFast, has_start_token=True, has_end_token=False, pad_to_max_length=False, max_length=99999999, pad_token=128258, end_token=128009, min_length=min_length)
+        super().__init__(tokenizer_path, embedding_directory=embedding_directory, pad_with_end=False, embedding_size=4096, embedding_key='llama', tokenizer_class=LlamaTokenizerFast, has_start_token=True, has_end_token=False, pad_to_max_length=False, max_length=99999999, pad_token=pad_token, min_length=min_length, tokenizer_data=tokenizer_data)
 
 class LLAMAModel(sd1_clip.SDClipModel):
-    def __init__(self, device="cpu", layer="hidden", layer_idx=-3, dtype=None, attention_mask=True, model_options={}):
+    def __init__(self, device="cpu", layer="hidden", layer_idx=-3, dtype=None, attention_mask=True, model_options={}, special_tokens={"start": 128000, "pad": 128258}):
         llama_scaled_fp8 = model_options.get("llama_scaled_fp8", None)
         if llama_scaled_fp8 is not None:
             model_options = model_options.copy()
             model_options["scaled_fp8"] = llama_scaled_fp8
 
-        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 128000, "pad": 128258}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Llama2, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+        textmodel_json_config = {}
+        vocab_size = model_options.get("vocab_size", None)
+        if vocab_size is not None:
+            textmodel_json_config["vocab_size"] = vocab_size
+
+        model_options = {**model_options, "model_name": "llama"}
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens=special_tokens, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Llama2, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
 
 class HunyuanVideoTokenizer:
     def __init__(self, embedding_directory=None, tokenizer_data={}):
-        clip_l_tokenizer_class = tokenizer_data.get("clip_l_tokenizer_class", sd1_clip.SDTokenizer)
-        self.clip_l = clip_l_tokenizer_class(embedding_directory=embedding_directory)
-        self.llama_template = """<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: 1. The main content and theme of the video.2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects.3. Actions, events, behaviors temporal relationships, physical movement changes of the objects.4. background environment, light, style and atmosphere.5. camera angles, movements, and transitions used in the video:<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n"""  # 95 tokens
-        self.llama = LLAMA3Tokenizer(embedding_directory=embedding_directory, min_length=1)
+        self.clip_l = sd1_clip.SDTokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
+        self.llama_template = """<|start_header_id|>system<|end_header_id|>\n\nDescribe the video by detailing the following aspects: 1. The main content and theme of the video.2. The color, shape, size, texture, quantity, text, and spatial relationships of the objects.3. Actions, events, behaviors temporal relationships, physical movement changes of the objects.4. background environment, light, style and atmosphere.5. camera angles, movements, and transitions used in the video:<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{}<|eot_id|>"""  # 95 tokens
+        self.llama = LLAMA3Tokenizer(embedding_directory=embedding_directory, min_length=1, tokenizer_data=tokenizer_data)
 
-    def tokenize_with_weights(self, text:str, return_word_ids=False):
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, image_embeds=None, image_interleave=1, **kwargs):
         out = {}
-        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids)
+        out["l"] = self.clip_l.tokenize_with_weights(text, return_word_ids, **kwargs)
 
-        llama_text = "{}{}".format(self.llama_template, text)
-        out["llama"] = self.llama.tokenize_with_weights(llama_text, return_word_ids)
+        if llama_template is None:
+            llama_text = self.llama_template.format(text)
+        else:
+            llama_text = llama_template.format(text)
+        llama_text_tokens = self.llama.tokenize_with_weights(llama_text, return_word_ids, **kwargs)
+        embed_count = 0
+        for r in llama_text_tokens:
+            for i in range(len(r)):
+                if r[i][0] == 128257:
+                    if image_embeds is not None and embed_count < image_embeds.shape[0]:
+                        r[i] = ({"type": "embedding", "data": image_embeds[embed_count], "original_type": "image", "image_interleave": image_interleave},) + r[i][1:]
+                        embed_count += 1
+        out["llama"] = llama_text_tokens
         return out
 
     def untokenize(self, token_weight_pair):
@@ -60,8 +77,7 @@ class HunyuanVideoClipModel(torch.nn.Module):
     def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}):
         super().__init__()
         dtype_llama = comfy.model_management.pick_weight_dtype(dtype_llama, dtype, device)
-        clip_l_class = model_options.get("clip_l_class", sd1_clip.SDClipModel)
-        self.clip_l = clip_l_class(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options)
+        self.clip_l = sd1_clip.SDClipModel(device=device, dtype=dtype, return_projected_pooled=False, model_options=model_options)
         self.llama = LLAMAModel(device=device, dtype=dtype_llama, model_options=model_options)
         self.dtypes = set([dtype, dtype_llama])
 
@@ -80,20 +96,51 @@ class HunyuanVideoClipModel(torch.nn.Module):
         llama_out, llama_pooled, llama_extra_out = self.llama.encode_token_weights(token_weight_pairs_llama)
 
         template_end = 0
-        for i, v in enumerate(token_weight_pairs_llama[0]):
-            if v[0] == 128007:  # <|end_header_id|>
-                template_end = i
+        extra_template_end = 0
+        extra_sizes = 0
+        user_end = 9999999999999
+        images = []
+
+        tok_pairs = token_weight_pairs_llama[0]
+        for i, v in enumerate(tok_pairs):
+            elem = v[0]
+            if not torch.is_tensor(elem):
+                if isinstance(elem, numbers.Integral):
+                    if elem == 128006:
+                        if tok_pairs[i + 1][0] == 882:
+                            if tok_pairs[i + 2][0] == 128007:
+                                template_end = i + 2
+                                user_end = -1
+                    if elem == 128009 and user_end == -1:
+                        user_end = i + 1
+                else:
+                    if elem.get("original_type") == "image":
+                        elem_size = elem.get("data").shape[0]
+                        if template_end > 0:
+                            if user_end == -1:
+                                extra_template_end += elem_size - 1
+                        else:
+                            image_start = i + extra_sizes
+                            image_end = i + elem_size + extra_sizes
+                            images.append((image_start, image_end, elem.get("image_interleave", 1)))
+                            extra_sizes += elem_size - 1
 
         if llama_out.shape[1] > (template_end + 2):
-            if token_weight_pairs_llama[0][template_end + 1][0] == 271:
+            if tok_pairs[template_end + 1][0] == 271:
                 template_end += 2
-        llama_out = llama_out[:, template_end:]
-        llama_extra_out["attention_mask"] = llama_extra_out["attention_mask"][:, template_end:]
+        llama_output = llama_out[:, template_end + extra_sizes:user_end + extra_sizes + extra_template_end]
+        llama_extra_out["attention_mask"] = llama_extra_out["attention_mask"][:, template_end + extra_sizes:user_end + extra_sizes + extra_template_end]
         if llama_extra_out["attention_mask"].sum() == torch.numel(llama_extra_out["attention_mask"]):
             llama_extra_out.pop("attention_mask")  # attention mask is useless if no masked elements
 
+        if len(images) > 0:
+            out = []
+            for i in images:
+                out.append(llama_out[:, i[0]: i[1]: i[2]])
+            llama_output = torch.cat(out + [llama_output], dim=1)
+
         l_out, l_pooled = self.clip_l.encode_token_weights(token_weight_pairs_l)
-        return llama_out, l_pooled, llama_extra_out
+        return llama_output, l_pooled, llama_extra_out
 
     def load_sd(self, sd):
         if "text_model.encoder.layers.1.mlp.fc1.weight" in sd:

comfy/text_encoders/hydit.py

@@ -9,24 +9,26 @@ import torch
 class HyditBertModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "hydit_clip.json")
+        model_options = {**model_options, "model_name": "hydit_clip"}
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 101, "end": 102, "pad": 0}, model_class=BertModel, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
 
 class HyditBertTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "hydit_clip_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='chinese_roberta', tokenizer_class=BertTokenizer, pad_to_max_length=False, max_length=512, min_length=77)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='chinese_roberta', tokenizer_class=BertTokenizer, pad_to_max_length=False, max_length=512, min_length=77, tokenizer_data=tokenizer_data)
 
 
 class MT5XLModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
         textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "mt5_config_xl.json")
+        model_options = {**model_options, "model_name": "mt5xl"}
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
 
 class MT5XLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         #tokenizer_path = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "mt5_tokenizer"), "spiece.model")
         tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=2048, embedding_key='mt5xl', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=2048, embedding_key='mt5xl', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=256, tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}
@@ -35,12 +37,12 @@ class HyditTokenizer:
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         mt5_tokenizer_data = tokenizer_data.get("mt5xl.spiece_model", None)
         self.hydit_clip = HyditBertTokenizer(embedding_directory=embedding_directory)
-        self.mt5xl = MT5XLTokenizer(tokenizer_data={"spiece_model": mt5_tokenizer_data}, embedding_directory=embedding_directory)
+        self.mt5xl = MT5XLTokenizer(tokenizer_data={**tokenizer_data, "spiece_model": mt5_tokenizer_data}, embedding_directory=embedding_directory)
 
-    def tokenize_with_weights(self, text:str, return_word_ids=False):
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
-        out["hydit_clip"] = self.hydit_clip.tokenize_with_weights(text, return_word_ids)
-        out["mt5xl"] = self.mt5xl.tokenize_with_weights(text, return_word_ids)
+        out["hydit_clip"] = self.hydit_clip.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["mt5xl"] = self.mt5xl.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
     def untokenize(self, token_weight_pair):