nit

* draft pass at a native comfy implementation of Lotus-D depth and normal est

* fix model_sampling kludges

* fix ruff

---------

Co-authored-by: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com>

README.md

@@ -69,6 +69,8 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [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/)
+- 3D Models
+   - [Hunyuan3D 2.0](https://docs.comfy.org/tutorials/3d/hunyuan3D-2)
 - [Stable Audio](https://comfyanonymous.github.io/ComfyUI_examples/audio/)
 - Asynchronous Queue system
 - Many optimizations: Only re-executes the parts of the workflow that changes between executions.

app/frontend_management.py

@@ -22,13 +22,21 @@ 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:\n{sys.executable} {extra}-m pip install -r {req_path}\n\nThis error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead.\n\nIf you are on the portable package you can run: update\\update_comfyui.bat to solve this problem"
+    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():
@@ -43,7 +51,17 @@ def check_frontend_version():
         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("________________________________________________________________________\nWARNING WARNING WARNING WARNING WARNING\n\nInstalled frontend version {} is lower than the recommended version {}.\n\n{}\n________________________________________________________________________".format('.'.join(map(str, frontend_version)), '.'.join(map(str, required_frontend)), frontend_install_warning_message()))
+            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:
@@ -150,9 +168,20 @@ class FrontendManager:
     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"\n\n********** ERROR ***********\n\ncomfyui-frontend-package is not installed. {frontend_install_warning_message()}\n********** ERROR **********\n")
+            logging.error(
+                f"""
+********** ERROR ***********
+
+comfyui-frontend-package is not installed.
+
+{frontend_install_warning_message()}
+
+********** ERROR ***********
+""".strip()
+            )
             sys.exit(-1)
 
     @classmethod
@@ -175,7 +204,9 @@ 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.
 
@@ -197,12 +228,20 @@ class FrontendManager:
         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)

comfy/cli_args.py

@@ -49,7 +49,7 @@ parser.add_argument("--temp-directory", type=str, default=None, help="Set the Co
 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=str, default=None, metavar="DEVICE_ID", help="Set the ids of cuda devices this instance will use.")
+parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use.")
 cm_group = parser.add_mutually_exclusive_group()
 cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).")
 cm_group.add_argument("--disable-cuda-malloc", action="store_true", help="Disable cudaMallocAsync.")

comfy/controlnet.py

@@ -15,14 +15,13 @@
     You should have received a copy of the GNU General Public License
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 """
-from __future__ import annotations
+
 
 import torch
 from enum import Enum
 import math
 import os
 import logging
-import copy
 import comfy.utils
 import comfy.model_management
 import comfy.model_detection
@@ -37,7 +36,7 @@ import comfy.cldm.mmdit
 import comfy.ldm.hydit.controlnet
 import comfy.ldm.flux.controlnet
 import comfy.cldm.dit_embedder
-from typing import TYPE_CHECKING, Union
+from typing import TYPE_CHECKING
 if TYPE_CHECKING:
     from comfy.hooks import HookGroup
 
@@ -64,18 +63,6 @@ class StrengthType(Enum):
     CONSTANT = 1
     LINEAR_UP = 2
 
-class ControlIsolation:
-    '''Temporarily set a ControlBase object's previous_controlnet to None to prevent cascading calls.'''
-    def __init__(self, control: ControlBase):
-        self.control = control
-        self.orig_previous_controlnet = control.previous_controlnet
-
-    def __enter__(self):
-        self.control.previous_controlnet = None
-
-    def __exit__(self, *args):
-        self.control.previous_controlnet = self.orig_previous_controlnet
-
 class ControlBase:
     def __init__(self):
         self.cond_hint_original = None
@@ -89,7 +76,7 @@ class ControlBase:
         self.compression_ratio = 8
         self.upscale_algorithm = 'nearest-exact'
         self.extra_args = {}
-        self.previous_controlnet: Union[ControlBase, None] = None
+        self.previous_controlnet = None
         self.extra_conds = []
         self.strength_type = StrengthType.CONSTANT
         self.concat_mask = False
@@ -97,7 +84,6 @@ class ControlBase:
         self.extra_concat = None
         self.extra_hooks: HookGroup = None
         self.preprocess_image = lambda a: a
-        self.multigpu_clones: dict[torch.device, ControlBase] = {}
 
     def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0), vae=None, extra_concat=[]):
         self.cond_hint_original = cond_hint
@@ -124,38 +110,17 @@ class ControlBase:
     def cleanup(self):
         if self.previous_controlnet is not None:
             self.previous_controlnet.cleanup()
-        for device_cnet in self.multigpu_clones.values():
+
-            with ControlIsolation(device_cnet):
-                device_cnet.cleanup()
         self.cond_hint = None
         self.extra_concat = None
         self.timestep_range = None
 
     def get_models(self):
         out = []
-        for device_cnet in self.multigpu_clones.values():
-            out += device_cnet.get_models_only_self()
         if self.previous_controlnet is not None:
             out += self.previous_controlnet.get_models()
         return out
 
-    def get_models_only_self(self):
-        'Calls get_models, but temporarily sets previous_controlnet to None.'
-        with ControlIsolation(self):
-            return self.get_models()
-
-    def get_instance_for_device(self, device):
-        'Returns instance of this Control object intended for selected device.'
-        return self.multigpu_clones.get(device, self)
-
-    def deepclone_multigpu(self, load_device, autoregister=False):
-        '''
-        Create deep clone of Control object where model(s) is set to other devices.
-
-        When autoregister is set to True, the deep clone is also added to multigpu_clones dict.
-        '''
-        raise NotImplementedError("Classes inheriting from ControlBase should define their own deepclone_multigpu funtion.")
-
     def get_extra_hooks(self):
         out = []
         if self.extra_hooks is not None:
@@ -164,7 +129,7 @@ class ControlBase:
             out += self.previous_controlnet.get_extra_hooks()
         return out
 
-    def copy_to(self, c: ControlBase):
+    def copy_to(self, c):
         c.cond_hint_original = self.cond_hint_original
         c.strength = self.strength
         c.timestep_percent_range = self.timestep_percent_range
@@ -315,14 +280,6 @@ class ControlNet(ControlBase):
         self.copy_to(c)
         return c
 
-    def deepclone_multigpu(self, load_device, autoregister=False):
-        c = self.copy()
-        c.control_model = copy.deepcopy(c.control_model)
-        c.control_model_wrapped = comfy.model_patcher.ModelPatcher(c.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
-        if autoregister:
-            self.multigpu_clones[load_device] = c
-        return c
-
     def get_models(self):
         out = super().get_models()
         out.append(self.control_model_wrapped)
@@ -847,14 +804,6 @@ class T2IAdapter(ControlBase):
         self.copy_to(c)
         return c
 
-    def deepclone_multigpu(self, load_device, autoregister=False):
-        c = self.copy()
-        c.t2i_model = copy.deepcopy(c.t2i_model)
-        c.device = load_device
-        if autoregister:
-            self.multigpu_clones[load_device] = c
-        return c
-
 def load_t2i_adapter(t2i_data, model_options={}): #TODO: model_options
     compression_ratio = 8
     upscale_algorithm = 'nearest-exact'

comfy/latent_formats.py

@@ -456,3 +456,13 @@ class Wan21(LatentFormat):
         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

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/modules/attention.py

@@ -471,7 +471,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
 def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
     if skip_reshape:
         b, _, _, dim_head = q.shape
-        tensor_layout="HND"
+        tensor_layout = "HND"
     else:
         b, _, dim_head = q.shape
         dim_head //= heads
@@ -479,7 +479,7 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
             lambda t: t.view(b, -1, heads, dim_head),
             (q, k, v),
         )
-        tensor_layout="NHD"
+        tensor_layout = "NHD"
 
     if mask is not None:
         # add a batch dimension if there isn't already one
@@ -489,7 +489,17 @@ def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=
         if mask.ndim == 3:
             mask = mask.unsqueeze(1)
 
-    out = sageattn(q, k, v, attn_mask=mask, is_causal=False, tensor_layout=tensor_layout)
+    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 = (

comfy/model_base.py

@@ -36,6 +36,7 @@ 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.model_management
 import comfy.patcher_extension
@@ -58,6 +59,7 @@ 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
@@ -88,6 +90,8 @@ 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
@@ -139,6 +143,7 @@ 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)
 
@@ -600,6 +605,19 @@ 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):
@@ -1013,3 +1031,18 @@ class WAN21(BaseModel):
         if clip_vision_output is not None:
             out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.penultimate_hidden_states)
         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

comfy/model_detection.py

@@ -154,7 +154,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         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
@@ -323,6 +323,21 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["model_type"] = "t2v"
         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 '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
@@ -667,8 +682,13 @@ 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

@@ -15,7 +15,6 @@
     You should have received a copy of the GNU General Public License
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 """
-from __future__ import annotations
 
 import psutil
 import logging
@@ -27,10 +26,6 @@ import platform
 import weakref
 import gc
 
-from typing import TYPE_CHECKING
-if TYPE_CHECKING:
-    from comfy.model_patcher import ModelPatcher
-
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
     NO_VRAM = 1     #Very low vram: enable all the options to save vram
@@ -51,6 +46,32 @@ 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:
@@ -150,25 +171,6 @@ def get_torch_device():
         else:
             return torch.device(torch.cuda.current_device())
 
-def get_all_torch_devices(exclude_current=False):
-    global cpu_state
-    devices = []
-    if cpu_state == CPUState.GPU:
-        if is_nvidia():
-            for i in range(torch.cuda.device_count()):
-                devices.append(torch.device(i))
-        elif is_intel_xpu():
-            for i in range(torch.xpu.device_count()):
-                devices.append(torch.device(i))
-        elif is_ascend_npu():
-            for i in range(torch.npu.device_count()):
-                devices.append(torch.device(i))
-    else:
-        devices.append(get_torch_device())
-    if exclude_current:
-        devices.remove(get_torch_device())
-    return devices
-
 def get_total_memory(dev=None, torch_total_too=False):
     global directml_enabled
     if dev is None:
@@ -380,13 +382,9 @@ try:
     logging.info("Device: {}".format(get_torch_device_name(get_torch_device())))
 except:
     logging.warning("Could not pick default device.")
-try:
-    for device in get_all_torch_devices(exclude_current=True):
-        logging.info("Device: {}".format(get_torch_device_name(device)))
-except:
-    pass
 
-current_loaded_models: list[LoadedModel] = []
+
+current_loaded_models = []
 
 def module_size(module):
     module_mem = 0
@@ -397,7 +395,7 @@ def module_size(module):
     return module_mem
 
 class LoadedModel:
-    def __init__(self, model: ModelPatcher):
+    def __init__(self, model):
         self._set_model(model)
         self.device = model.load_device
         self.real_model = None
@@ -405,7 +403,7 @@ class LoadedModel:
         self.model_finalizer = None
         self._patcher_finalizer = None
 
-    def _set_model(self, model: ModelPatcher):
+    def _set_model(self, model):
         self._model = weakref.ref(model)
         if model.parent is not None:
             self._parent_model = weakref.ref(model.parent)
@@ -729,11 +727,8 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         return torch.float8_e5m2
 
     fp8_dtype = None
-    try:
+    if weight_dtype in FLOAT8_TYPES:
-        if weight_dtype in [torch.float8_e4m3fn, torch.float8_e5m2]:
+        fp8_dtype = weight_dtype
-            fp8_dtype = weight_dtype
-    except:
-        pass
 
     if fp8_dtype is not None:
         if supports_fp8_compute(device): #if fp8 compute is supported the casting is most likely not expensive
@@ -1247,31 +1242,6 @@ def soft_empty_cache(force=False):
 def unload_all_models():
     free_memory(1e30, get_torch_device())
 
-def unload_model_and_clones(model: ModelPatcher, unload_additional_models=True, all_devices=False):
-    'Unload only model and its clones - primarily for multigpu cloning purposes.'
-    initial_keep_loaded: list[LoadedModel] = current_loaded_models.copy()
-    additional_models = []
-    if unload_additional_models:
-        additional_models = model.get_nested_additional_models()
-    keep_loaded = []
-    for loaded_model in initial_keep_loaded:
-        if loaded_model.model is not None:
-            if model.clone_base_uuid == loaded_model.model.clone_base_uuid:
-                continue
-            # check additional models if they are a match
-            skip = False
-            for add_model in additional_models:
-                if add_model.clone_base_uuid == loaded_model.model.clone_base_uuid:
-                    skip = True
-                    break
-            if skip:
-                continue
-        keep_loaded.append(loaded_model)
-    if not all_devices:
-        free_memory(1e30, get_torch_device(), keep_loaded)
-    else:
-        for device in get_all_torch_devices():
-            free_memory(1e30, device, keep_loaded)
 
 #TODO: might be cleaner to put this somewhere else
 import threading

comfy/model_patcher.py

@@ -84,15 +84,12 @@ def set_model_options_pre_cfg_function(model_options, pre_cfg_function, disable_
 def create_model_options_clone(orig_model_options: dict):
     return comfy.patcher_extension.copy_nested_dicts(orig_model_options)
 
-def create_hook_patches_clone(orig_hook_patches, copy_tuples=False):
+def create_hook_patches_clone(orig_hook_patches):
     new_hook_patches = {}
     for hook_ref in orig_hook_patches:
         new_hook_patches[hook_ref] = {}
         for k in orig_hook_patches[hook_ref]:
             new_hook_patches[hook_ref][k] = orig_hook_patches[hook_ref][k][:]
-            if copy_tuples:
-                for i in range(len(new_hook_patches[hook_ref][k])):
-                    new_hook_patches[hook_ref][k][i] = tuple(new_hook_patches[hook_ref][k][i])
     return new_hook_patches
 
 def wipe_lowvram_weight(m):
@@ -243,9 +240,6 @@ class ModelPatcher:
         self.is_clip = False
         self.hook_mode = comfy.hooks.EnumHookMode.MaxSpeed
 
-        self.is_multigpu_base_clone = False
-        self.clone_base_uuid = uuid.uuid4()
-
         if not hasattr(self.model, 'model_loaded_weight_memory'):
             self.model.model_loaded_weight_memory = 0
 
@@ -324,92 +318,18 @@ class ModelPatcher:
         n.is_clip = self.is_clip
         n.hook_mode = self.hook_mode
 
-        n.is_multigpu_base_clone = self.is_multigpu_base_clone
-        n.clone_base_uuid = self.clone_base_uuid
-
         for callback in self.get_all_callbacks(CallbacksMP.ON_CLONE):
             callback(self, n)
         return n
 
-    def deepclone_multigpu(self, new_load_device=None, models_cache: dict[uuid.UUID,ModelPatcher]=None):
-        logging.info(f"Creating deepclone of {self.model.__class__.__name__} for {new_load_device if new_load_device else self.load_device}.")
-        comfy.model_management.unload_model_and_clones(self)
-        n = self.clone()
-        # set load device, if present
-        if new_load_device is not None:
-            n.load_device = new_load_device
-        # unlike for normal clone, backup dicts that shared same ref should not;
-        # otherwise, patchers that have deep copies of base models will erroneously influence each other.
-        n.backup = copy.deepcopy(n.backup)
-        n.object_patches_backup = copy.deepcopy(n.object_patches_backup)
-        n.hook_backup = copy.deepcopy(n.hook_backup)
-        n.model = copy.deepcopy(n.model)
-        # multigpu clone should not have multigpu additional_models entry
-        n.remove_additional_models("multigpu")
-        # multigpu_clone all stored additional_models; make sure circular references are properly handled
-        if models_cache is None:
-            models_cache = {}
-        for key, model_list in n.additional_models.items():
-            for i in range(len(model_list)):
-                add_model = n.additional_models[key][i]
-                if add_model.clone_base_uuid not in models_cache:
-                    models_cache[add_model.clone_base_uuid] = add_model.deepclone_multigpu(new_load_device=new_load_device, models_cache=models_cache)
-                n.additional_models[key][i] = models_cache[add_model.clone_base_uuid]
-        for callback in self.get_all_callbacks(CallbacksMP.ON_DEEPCLONE_MULTIGPU):
-            callback(self, n)
-        return n
-
-    def match_multigpu_clones(self):
-        multigpu_models = self.get_additional_models_with_key("multigpu")
-        if len(multigpu_models) > 0:
-            new_multigpu_models = []
-            for mm in multigpu_models:
-                # clone main model, but bring over relevant props from existing multigpu clone
-                n = self.clone()
-                n.load_device = mm.load_device
-                n.backup = mm.backup
-                n.object_patches_backup = mm.object_patches_backup
-                n.hook_backup = mm.hook_backup
-                n.model = mm.model
-                n.is_multigpu_base_clone = mm.is_multigpu_base_clone
-                n.remove_additional_models("multigpu")
-                orig_additional_models: dict[str, list[ModelPatcher]] = comfy.patcher_extension.copy_nested_dicts(n.additional_models)
-                n.additional_models = comfy.patcher_extension.copy_nested_dicts(mm.additional_models)
-                # figure out which additional models are not present in multigpu clone
-                models_cache = {}
-                for mm_add_model in mm.get_additional_models():
-                    models_cache[mm_add_model.clone_base_uuid] = mm_add_model
-                remove_models_uuids = set(list(models_cache.keys()))
-                for key, model_list in orig_additional_models.items():
-                    for orig_add_model in model_list:
-                        if orig_add_model.clone_base_uuid not in models_cache:
-                            models_cache[orig_add_model.clone_base_uuid] = orig_add_model.deepclone_multigpu(new_load_device=n.load_device, models_cache=models_cache)
-                            existing_list = n.get_additional_models_with_key(key)
-                            existing_list.append(models_cache[orig_add_model.clone_base_uuid])
-                            n.set_additional_models(key, existing_list)
-                        if orig_add_model.clone_base_uuid in remove_models_uuids:
-                            remove_models_uuids.remove(orig_add_model.clone_base_uuid)
-                # remove duplicate additional models
-                for key, model_list in n.additional_models.items():
-                    new_model_list = [x for x in model_list if x.clone_base_uuid not in remove_models_uuids]
-                    n.set_additional_models(key, new_model_list)
-                for callback in self.get_all_callbacks(CallbacksMP.ON_MATCH_MULTIGPU_CLONES):
-                    callback(self, n)
-                new_multigpu_models.append(n)
-            self.set_additional_models("multigpu", new_multigpu_models)
-
     def is_clone(self, other):
         if hasattr(other, 'model') and self.model is other.model:
             return True
         return False
 
-    def clone_has_same_weights(self, clone: ModelPatcher, allow_multigpu=False):
+    def clone_has_same_weights(self, clone: 'ModelPatcher'):
-        if allow_multigpu:
+        if not self.is_clone(clone):
-            if self.clone_base_uuid != clone.clone_base_uuid:
+            return False
-                return False
-        else:
-            if not self.is_clone(clone):
-                return False
 
         if self.current_hooks != clone.current_hooks:
             return False
@@ -1009,7 +929,7 @@ class ModelPatcher:
         return self.additional_models.get(key, [])
 
     def get_additional_models(self):
-        all_models: list[ModelPatcher] = []
+        all_models = []
         for models in self.additional_models.values():
             all_models.extend(models)
         return all_models
@@ -1063,13 +983,9 @@ class ModelPatcher:
         for callback in self.get_all_callbacks(CallbacksMP.ON_PRE_RUN):
             callback(self)
 
-    def prepare_state(self, timestep, model_options, ignore_multigpu=False):
+    def prepare_state(self, timestep):
         for callback in self.get_all_callbacks(CallbacksMP.ON_PREPARE_STATE):
-            callback(self, timestep, model_options, ignore_multigpu)
+            callback(self, timestep)
-        if not ignore_multigpu and "multigpu_clones" in model_options:
-            for p in model_options["multigpu_clones"].values():
-                p: ModelPatcher
-                p.prepare_state(timestep, model_options, ignore_multigpu=True)
 
     def restore_hook_patches(self):
         if self.hook_patches_backup is not None:
@@ -1082,18 +998,12 @@ class ModelPatcher:
     def prepare_hook_patches_current_keyframe(self, t: torch.Tensor, hook_group: comfy.hooks.HookGroup, model_options: dict[str]):
         curr_t = t[0]
         reset_current_hooks = False
-        multigpu_kf_changed_cache = None
         transformer_options = model_options.get("transformer_options", {})
         for hook in hook_group.hooks:
             changed = hook.hook_keyframe.prepare_current_keyframe(curr_t=curr_t, transformer_options=transformer_options)
             # if keyframe changed, remove any cached HookGroups that contain hook with the same hook_ref;
             # this will cause the weights to be recalculated when sampling
             if changed:
-                # cache changed for multigpu usage
-                if "multigpu_clones" in model_options:
-                    if multigpu_kf_changed_cache is None:
-                        multigpu_kf_changed_cache = []
-                    multigpu_kf_changed_cache.append(hook)
                 # reset current_hooks if contains hook that changed
                 if self.current_hooks is not None:
                     for current_hook in self.current_hooks.hooks:
@@ -1105,28 +1015,6 @@ class ModelPatcher:
                         self.cached_hook_patches.pop(cached_group)
         if reset_current_hooks:
             self.patch_hooks(None)
-        if "multigpu_clones" in model_options:
-            for p in model_options["multigpu_clones"].values():
-                p: ModelPatcher
-                p._handle_changed_hook_keyframes(multigpu_kf_changed_cache)
-
-    def _handle_changed_hook_keyframes(self, kf_changed_cache: list[comfy.hooks.Hook]):
-        'Used to handle multigpu behavior inside prepare_hook_patches_current_keyframe.'
-        if kf_changed_cache is None:
-            return
-        reset_current_hooks = False
-        # reset current_hooks if contains hook that changed
-        for hook in kf_changed_cache:
-            if self.current_hooks is not None:
-                for current_hook in self.current_hooks.hooks:
-                    if current_hook == hook:
-                        reset_current_hooks = True
-                        break
-            for cached_group in list(self.cached_hook_patches.keys()):
-                if cached_group.contains(hook):
-                    self.cached_hook_patches.pop(cached_group)
-        if reset_current_hooks:
-            self.patch_hooks(None)
 
     def register_all_hook_patches(self, hooks: comfy.hooks.HookGroup, target_dict: dict[str], model_options: dict=None,
                                   registered: comfy.hooks.HookGroup = None):

comfy/model_sampling.py

@@ -69,6 +69,15 @@ class CONST:
         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__()

comfy/multigpu.py

@@ -1,176 +0,0 @@
-from __future__ import annotations
-import torch
-import logging
-
-from collections import namedtuple
-from typing import TYPE_CHECKING
-if TYPE_CHECKING:
-    from comfy.model_patcher import ModelPatcher
-import comfy.utils
-import comfy.patcher_extension
-import comfy.model_management
-
-
-class GPUOptions:
-    def __init__(self, device_index: int, relative_speed: float):
-        self.device_index = device_index
-        self.relative_speed = relative_speed
-
-    def clone(self):
-        return GPUOptions(self.device_index, self.relative_speed)
-
-    def create_dict(self):
-        return {
-            "relative_speed": self.relative_speed
-        }
-
-class GPUOptionsGroup:
-    def __init__(self):
-        self.options: dict[int, GPUOptions] = {}
-
-    def add(self, info: GPUOptions):
-        self.options[info.device_index] = info
-
-    def clone(self):
-        c = GPUOptionsGroup()
-        for opt in self.options.values():
-            c.add(opt)
-        return c
-
-    def register(self, model: ModelPatcher):
-        opts_dict = {}
-        # get devices that are valid for this model
-        devices: list[torch.device] = [model.load_device]
-        for extra_model in model.get_additional_models_with_key("multigpu"):
-            extra_model: ModelPatcher
-            devices.append(extra_model.load_device)
-        # create dictionary with actual device mapped to its GPUOptions
-        device_opts_list: list[GPUOptions] = []
-        for device in devices:
-            device_opts = self.options.get(device.index, GPUOptions(device_index=device.index, relative_speed=1.0))
-            opts_dict[device] = device_opts.create_dict()
-            device_opts_list.append(device_opts)
-        # make relative_speed relative to 1.0
-        min_speed = min([x.relative_speed for x in device_opts_list])
-        for value in opts_dict.values():
-            value['relative_speed'] /= min_speed
-        model.model_options['multigpu_options'] = opts_dict
-
-def get_torch_device_list():
-    devices = ["default"]
-    for device in comfy.model_management.get_all_torch_devices():
-        device: torch.device
-        devices.append(str(device.index))
-    return devices
-
-def get_device_from_str(device_str: str, throw_error_if_not_found=False):
-    if device_str == "default":
-        return comfy.model_management.get_torch_device()
-    for device in comfy.model_management.get_all_torch_devices():
-        device: torch.device
-        if str(device.index) == device_str:
-            return device
-    if throw_error_if_not_found:
-        raise Exception(f"Device with index '{device_str}' not found.")
-    logging.warning(f"Device with index '{device_str}' not found, using default device ({comfy.model_management.get_torch_device()}) instead.")
-
-def create_multigpu_deepclones(model: ModelPatcher, max_gpus: int, gpu_options: GPUOptionsGroup=None, reuse_loaded=False):
-    'Prepare ModelPatcher to contain deepclones of its BaseModel and related properties.'
-    model = model.clone()
-    # check if multigpu is already prepared - get the load devices from them if possible to exclude
-    skip_devices = set()
-    multigpu_models = model.get_additional_models_with_key("multigpu")
-    if len(multigpu_models) > 0:
-        for mm in multigpu_models:
-            skip_devices.add(mm.load_device)
-    skip_devices = list(skip_devices)
-
-    extra_devices = comfy.model_management.get_all_torch_devices(exclude_current=True)
-    extra_devices = extra_devices[:max_gpus-1]
-    # exclude skipped devices
-    for skip in skip_devices:
-        if skip in extra_devices:
-            extra_devices.remove(skip)
-    # create new deepclones
-    if len(extra_devices) > 0:
-        for device in extra_devices:
-            device_patcher = None
-            if reuse_loaded:
-                # check if there are any ModelPatchers currently loaded that could be referenced here after a clone
-                loaded_models: list[ModelPatcher] = comfy.model_management.loaded_models()
-                for lm in loaded_models:
-                    if lm.model is not None and lm.clone_base_uuid == model.clone_base_uuid and lm.load_device == device:
-                        device_patcher = lm.clone()
-                        logging.info(f"Reusing loaded deepclone of {device_patcher.model.__class__.__name__} for {device}")
-                        break
-            if device_patcher is None:
-                device_patcher = model.deepclone_multigpu(new_load_device=device)
-                device_patcher.is_multigpu_base_clone = True
-            multigpu_models = model.get_additional_models_with_key("multigpu")
-            multigpu_models.append(device_patcher)
-            model.set_additional_models("multigpu", multigpu_models)
-        model.match_multigpu_clones()
-        if gpu_options is None:
-            gpu_options = GPUOptionsGroup()
-        gpu_options.register(model)
-    else:
-        logging.info("No extra torch devices need initialization, skipping initializing MultiGPU Work Units.")
-    # persist skip_devices for use in sampling code
-    # if len(skip_devices) > 0 or "multigpu_skip_devices" in model.model_options:
-    #     model.model_options["multigpu_skip_devices"] = skip_devices
-    return model
-
-
-LoadBalance = namedtuple('LoadBalance', ['work_per_device', 'idle_time'])
-def load_balance_devices(model_options: dict[str], total_work: int, return_idle_time=False, work_normalized: int=None):
-    'Optimize work assigned to different devices, accounting for their relative speeds and splittable work.'
-    opts_dict = model_options['multigpu_options']
-    devices = list(model_options['multigpu_clones'].keys())
-    speed_per_device = []
-    work_per_device = []
-    # get sum of each device's relative_speed
-    total_speed = 0.0
-    for opts in opts_dict.values():
-        total_speed += opts['relative_speed']
-    # get relative work for each device;
-    # obtained by w = (W*r)/R
-    for device in devices:
-        relative_speed = opts_dict[device]['relative_speed']
-        relative_work = (total_work*relative_speed) / total_speed
-        speed_per_device.append(relative_speed)
-        work_per_device.append(relative_work)
-    # relative work must be expressed in whole numbers, but likely is a decimal;
-    # perform rounding while maintaining total sum equal to total work (sum of relative works)
-    work_per_device = round_preserved(work_per_device)
-    dict_work_per_device = {}
-    for device, relative_work in zip(devices, work_per_device):
-        dict_work_per_device[device] = relative_work
-    if not return_idle_time:
-        return LoadBalance(dict_work_per_device, None)
-    # divide relative work by relative speed to get estimated completion time of said work by each device;
-    # time here is relative and does not correspond to real-world units
-    completion_time = [w/r for w,r in zip(work_per_device, speed_per_device)]
-    # calculate relative time spent by the devices waiting on each other after their work is completed
-    idle_time = abs(min(completion_time) - max(completion_time))
-    # if need to compare work idle time, need to normalize to a common total work
-    if work_normalized:
-        idle_time *= (work_normalized/total_work)
-
-    return LoadBalance(dict_work_per_device, idle_time)
-
-def round_preserved(values: list[float]):
-    'Round all values in a list, preserving the combined sum of values.'
-    # get floor of values; casting to int does it too
-    floored = [int(x) for x in values]
-    total_floored = sum(floored)
-    # get remainder to distribute
-    remainder = round(sum(values)) - total_floored
-    # pair values with fractional portions
-    fractional = [(i, x-floored[i]) for i, x in enumerate(values)]
-    # sort by fractional part in descending order
-    fractional.sort(key=lambda x: x[1], reverse=True)
-    # distribute the remainder
-    for i in range(remainder):
-        index = fractional[i][0]
-        floored[index] += 1
-    return floored

comfy/patcher_extension.py

@@ -3,8 +3,6 @@ from typing import Callable
 
 class CallbacksMP:
     ON_CLONE = "on_clone"
-    ON_DEEPCLONE_MULTIGPU = "on_deepclone_multigpu"
-    ON_MATCH_MULTIGPU_CLONES = "on_match_multigpu_clones"
     ON_LOAD = "on_load_after"
     ON_DETACH = "on_detach_after"
     ON_CLEANUP = "on_cleanup"

comfy/sampler_helpers.py

@@ -1,9 +1,7 @@
 from __future__ import annotations
-import torch
 import uuid
 import comfy.model_management
 import comfy.conds
-import comfy.model_patcher
 import comfy.utils
 import comfy.hooks
 import comfy.patcher_extension
@@ -106,57 +104,16 @@ def cleanup_additional_models(models):
         if hasattr(m, 'cleanup'):
             m.cleanup()
 
-def preprocess_multigpu_conds(conds: dict[str, list[dict[str]]], model: ModelPatcher, model_options: dict[str]):
-    '''If multigpu acceleration required, creates deepclones of ControlNets and GLIGEN per device.'''
-    multigpu_models: list[ModelPatcher] = model.get_additional_models_with_key("multigpu")
-    if len(multigpu_models) == 0:
-        return
-    extra_devices = [x.load_device for x in multigpu_models]
-    # handle controlnets
-    controlnets: set[ControlBase] = set()
-    for k in conds:
-        for kk in conds[k]:
-            if 'control' in kk:
-                controlnets.add(kk['control'])
-    if len(controlnets) > 0:
-        # first, unload all controlnet clones
-        for cnet in list(controlnets):
-            cnet_models = cnet.get_models()
-            for cm in cnet_models:
-                comfy.model_management.unload_model_and_clones(cm, unload_additional_models=True)
-
-        # next, make sure each controlnet has a deepclone for all relevant devices
-        for cnet in controlnets:
-            curr_cnet = cnet
-            while curr_cnet is not None:
-                for device in extra_devices:
-                    if device not in curr_cnet.multigpu_clones:
-                        curr_cnet.deepclone_multigpu(device, autoregister=True)
-                curr_cnet = curr_cnet.previous_controlnet
-        # since all device clones are now present, recreate the linked list for cloned cnets per device
-        for cnet in controlnets:
-            curr_cnet = cnet
-            while curr_cnet is not None:
-                prev_cnet = curr_cnet.previous_controlnet
-                for device in extra_devices:
-                    device_cnet = curr_cnet.get_instance_for_device(device)
-                    prev_device_cnet = None
-                    if prev_cnet is not None:
-                        prev_device_cnet = prev_cnet.get_instance_for_device(device)
-                    device_cnet.set_previous_controlnet(prev_device_cnet)
-                curr_cnet = prev_cnet
-    # potentially handle gligen - since not widely used, ignored for now
 
 def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None):
-    model.match_multigpu_clones()
+    real_model: BaseModel = None
-    preprocess_multigpu_conds(conds, model, model_options)
     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)
-    real_model: BaseModel = model.model
+    real_model = model.model
 
     return real_model, conds, models
 
@@ -169,7 +126,7 @@ def cleanup_models(conds, models):
 
     cleanup_additional_models(set(control_cleanup))
 
-def prepare_model_patcher(model: ModelPatcher, conds, model_options: dict):
+def prepare_model_patcher(model: 'ModelPatcher', conds, model_options: dict):
     '''
     Registers hooks from conds.
     '''
@@ -202,18 +159,3 @@ def prepare_model_patcher(model: ModelPatcher, conds, model_options: dict):
         comfy.patcher_extension.merge_nested_dicts(to_load_options.setdefault(wc_name, {}), model_options["transformer_options"][wc_name],
                                                     copy_dict1=False)
     return to_load_options
-
-def prepare_model_patcher_multigpu_clones(model_patcher: ModelPatcher, loaded_models: list[ModelPatcher], model_options: dict):
-    '''
-    In case multigpu acceleration is enabled, prep ModelPatchers for each device.
-    '''
-    multigpu_patchers: list[ModelPatcher] = [x for x in loaded_models if x.is_multigpu_base_clone]
-    if len(multigpu_patchers) > 0:
-        multigpu_dict: dict[torch.device, ModelPatcher] = {}
-        multigpu_dict[model_patcher.load_device] = model_patcher
-        for x in multigpu_patchers:
-            x.hook_patches = comfy.model_patcher.create_hook_patches_clone(model_patcher.hook_patches, copy_tuples=True)
-            x.hook_mode = model_patcher.hook_mode # match main model's hook_mode
-            multigpu_dict[x.load_device] = x
-        model_options["multigpu_clones"] = multigpu_dict
-    return multigpu_patchers

comfy/samplers.py

@@ -1,6 +1,4 @@
 from __future__ import annotations
-
-import comfy.model_management
 from .k_diffusion import sampling as k_diffusion_sampling
 from .extra_samplers import uni_pc
 from typing import TYPE_CHECKING, Callable, NamedTuple
@@ -20,7 +18,6 @@ import comfy.patcher_extension
 import comfy.hooks
 import scipy.stats
 import numpy
-import threading
 
 
 def add_area_dims(area, num_dims):
@@ -143,7 +140,7 @@ def can_concat_cond(c1, c2):
 
     return cond_equal_size(c1.conditioning, c2.conditioning)
 
-def cond_cat(c_list, device=None):
+def cond_cat(c_list):
     temp = {}
     for x in c_list:
         for k in x:
@@ -155,8 +152,6 @@ def cond_cat(c_list, device=None):
     for k in temp:
         conds = temp[k]
         out[k] = conds[0].concat(conds[1:])
-        if device is not None and hasattr(out[k], 'to'):
-            out[k] = out[k].to(device)
 
     return out
 
@@ -210,9 +205,7 @@ def calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Ten
     )
     return executor.execute(model, conds, x_in, timestep, model_options)
 
-def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
+def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Tensor, timestep, model_options):
-    if 'multigpu_clones' in model_options:
-        return _calc_cond_batch_multigpu(model, conds, x_in, timestep, model_options)
     out_conds = []
     out_counts = []
     # separate conds by matching hooks
@@ -244,7 +237,7 @@ def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Te
     if has_default_conds:
         finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options)
 
-    model.current_patcher.prepare_state(timestep, model_options)
+    model.current_patcher.prepare_state(timestep)
 
     # run every hooked_to_run separately
     for hooks, to_run in hooked_to_run.items():
@@ -346,190 +339,6 @@ def _calc_cond_batch(model: 'BaseModel', conds: list[list[dict]], x_in: torch.Te
 
     return out_conds
 
-def _calc_cond_batch_multigpu(model: BaseModel, conds: list[list[dict]], x_in: torch.Tensor, timestep: torch.Tensor, model_options: dict[str]):
-    out_conds = []
-    out_counts = []
-    # separate conds by matching hooks
-    hooked_to_run: dict[comfy.hooks.HookGroup,list[tuple[tuple,int]]] = {}
-    default_conds = []
-    has_default_conds = False
-
-    output_device = x_in.device
-
-    for i in range(len(conds)):
-        out_conds.append(torch.zeros_like(x_in))
-        out_counts.append(torch.ones_like(x_in) * 1e-37)
-
-        cond = conds[i]
-        default_c = []
-        if cond is not None:
-            for x in cond:
-                if 'default' in x:
-                    default_c.append(x)
-                    has_default_conds = True
-                    continue
-                p = get_area_and_mult(x, x_in, timestep)
-                if p is None:
-                    continue
-                if p.hooks is not None:
-                    model.current_patcher.prepare_hook_patches_current_keyframe(timestep, p.hooks, model_options)
-                hooked_to_run.setdefault(p.hooks, list())
-                hooked_to_run[p.hooks] += [(p, i)]
-        default_conds.append(default_c)
-
-    if has_default_conds:
-        finalize_default_conds(model, hooked_to_run, default_conds, x_in, timestep, model_options)
-
-    model.current_patcher.prepare_state(timestep, model_options)
-
-    devices = [dev_m for dev_m in model_options['multigpu_clones'].keys()]
-    device_batched_hooked_to_run: dict[torch.device, list[tuple[comfy.hooks.HookGroup, tuple]]] = {}
-
-    total_conds = 0
-    for to_run in hooked_to_run.values():
-        total_conds += len(to_run)
-    conds_per_device = max(1, math.ceil(total_conds//len(devices)))
-    index_device = 0
-    current_device = devices[index_device]
-    # run every hooked_to_run separately
-    for hooks, to_run in hooked_to_run.items():
-        while len(to_run) > 0:
-            current_device = devices[index_device % len(devices)]
-            batched_to_run = device_batched_hooked_to_run.setdefault(current_device, [])
-            # keep track of conds currently scheduled onto this device
-            batched_to_run_length = 0
-            for btr in batched_to_run:
-                batched_to_run_length += len(btr[1])
-
-            first = to_run[0]
-            first_shape = first[0][0].shape
-            to_batch_temp = []
-            # make sure not over conds_per_device limit when creating temp batch
-            for x in range(len(to_run)):
-                if can_concat_cond(to_run[x][0], first[0]) and len(to_batch_temp) < (conds_per_device - batched_to_run_length):
-                    to_batch_temp += [x]
-
-            to_batch_temp.reverse()
-            to_batch = to_batch_temp[:1]
-
-            free_memory = model_management.get_free_memory(current_device)
-            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:
-                    to_batch = batch_amount
-                    break
-            conds_to_batch = []
-            for x in to_batch:
-                conds_to_batch.append(to_run.pop(x))
-            batched_to_run_length += len(conds_to_batch)
-
-            batched_to_run.append((hooks, conds_to_batch))
-            if batched_to_run_length >= conds_per_device:
-                index_device += 1
-
-    thread_result = collections.namedtuple('thread_result', ['output', 'mult', 'area', 'batch_chunks', 'cond_or_uncond'])
-    def _handle_batch(device: torch.device, batch_tuple: tuple[comfy.hooks.HookGroup, tuple], results: list[thread_result]):
-        model_current: BaseModel = model_options["multigpu_clones"][device].model
-        # run every hooked_to_run separately
-        with torch.no_grad():
-            for hooks, to_batch in batch_tuple:
-                input_x = []
-                mult = []
-                c = []
-                cond_or_uncond = []
-                uuids = []
-                area = []
-                control: ControlBase = None
-                patches = None
-                for x in to_batch:
-                    o = x
-                    p = o[0]
-                    input_x.append(p.input_x)
-                    mult.append(p.mult)
-                    c.append(p.conditioning)
-                    area.append(p.area)
-                    cond_or_uncond.append(o[1])
-                    uuids.append(p.uuid)
-                    control = p.control
-                    patches = p.patches
-
-                batch_chunks = len(cond_or_uncond)
-                input_x = torch.cat(input_x).to(device)
-                c = cond_cat(c, device=device)
-                timestep_ = torch.cat([timestep.to(device)] * batch_chunks)
-
-                transformer_options = model_current.current_patcher.apply_hooks(hooks=hooks)
-                if 'transformer_options' in model_options:
-                    transformer_options = comfy.patcher_extension.merge_nested_dicts(transformer_options,
-                                                                                    model_options['transformer_options'],
-                                                                                    copy_dict1=False)
-
-                if patches is not None:
-                    # TODO: replace with merge_nested_dicts function
-                    if "patches" in transformer_options:
-                        cur_patches = transformer_options["patches"].copy()
-                        for p in patches:
-                            if p in cur_patches:
-                                cur_patches[p] = cur_patches[p] + patches[p]
-                            else:
-                                cur_patches[p] = patches[p]
-                        transformer_options["patches"] = cur_patches
-                    else:
-                        transformer_options["patches"] = patches
-
-                transformer_options["cond_or_uncond"] = cond_or_uncond[:]
-                transformer_options["uuids"] = uuids[:]
-                transformer_options["sigmas"] = timestep
-                transformer_options["sample_sigmas"] = transformer_options["sample_sigmas"].to(device)
-                transformer_options["multigpu_thread_device"] = device
-
-                cast_transformer_options(transformer_options, device=device)
-                c['transformer_options'] = transformer_options
-
-                if control is not None:
-                    device_control = control.get_instance_for_device(device)
-                    c['control'] = device_control.get_control(input_x, timestep_, c, len(cond_or_uncond), transformer_options)
-
-                if 'model_function_wrapper' in model_options:
-                    output = model_options['model_function_wrapper'](model_current.apply_model, {"input": input_x, "timestep": timestep_, "c": c, "cond_or_uncond": cond_or_uncond}).to(output_device).chunk(batch_chunks)
-                else:
-                    output = model_current.apply_model(input_x, timestep_, **c).to(output_device).chunk(batch_chunks)
-                results.append(thread_result(output, mult, area, batch_chunks, cond_or_uncond))
-
-
-    results: list[thread_result] = []
-    threads: list[threading.Thread] = []
-    for device, batch_tuple in device_batched_hooked_to_run.items():
-        new_thread = threading.Thread(target=_handle_batch, args=(device, batch_tuple, results))
-        threads.append(new_thread)
-        new_thread.start()
-
-    for thread in threads:
-        thread.join()
-
-    for output, mult, area, batch_chunks, cond_or_uncond in results:
-        for o in range(batch_chunks):
-            cond_index = cond_or_uncond[o]
-            a = area[o]
-            if a is None:
-                out_conds[cond_index] += output[o] * mult[o]
-                out_counts[cond_index] += mult[o]
-            else:
-                out_c = out_conds[cond_index]
-                out_cts = out_counts[cond_index]
-                dims = len(a) // 2
-                for i in range(dims):
-                    out_c = out_c.narrow(i + 2, a[i + dims], a[i])
-                    out_cts = out_cts.narrow(i + 2, a[i + dims], a[i])
-                out_c += output[o] * mult[o]
-                out_cts += mult[o]
-
-    for i in range(len(out_conds)):
-        out_conds[i] /= out_counts[i]
-
-    return out_conds
-
 def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options): #TODO: remove
     logging.warning("WARNING: The comfy.samplers.calc_cond_uncond_batch function is deprecated please use the calc_cond_batch one instead.")
     return tuple(calc_cond_batch(model, [cond, uncond], x_in, timestep, model_options))
@@ -827,8 +636,6 @@ def pre_run_control(model, conds):
         percent_to_timestep_function = lambda a: s.percent_to_sigma(a)
         if 'control' in x:
             x['control'].pre_run(model, percent_to_timestep_function)
-            for device_cnet in x['control'].multigpu_clones.values():
-                device_cnet.pre_run(model, percent_to_timestep_function)
 
 def apply_empty_x_to_equal_area(conds, uncond, name, uncond_fill_func):
     cond_cnets = []
@@ -1071,9 +878,7 @@ def cast_to_load_options(model_options: dict[str], device=None, dtype=None):
     to_load_options = model_options.get("to_load_options", None)
     if to_load_options is None:
         return
-    cast_transformer_options(to_load_options, device, dtype)
 
-def cast_transformer_options(transformer_options: dict[str], device=None, dtype=None):
     casts = []
     if device is not None:
         casts.append(device)
@@ -1082,17 +887,18 @@ def cast_transformer_options(transformer_options: dict[str], device=None, dtype=
     # if nothing to apply, do nothing
     if len(casts) == 0:
         return
+
     # try to call .to on patches
-    if "patches" in transformer_options:
+    if "patches" in to_load_options:
-        patches = transformer_options["patches"]
+        patches = to_load_options["patches"]
         for name in patches:
             patch_list = patches[name]
             for i in range(len(patch_list)):
                 if hasattr(patch_list[i], "to"):
                     for cast in casts:
                         patch_list[i] = patch_list[i].to(cast)
-    if "patches_replace" in transformer_options:
+    if "patches_replace" in to_load_options:
-        patches = transformer_options["patches_replace"]
+        patches = to_load_options["patches_replace"]
         for name in patches:
             patch_list = patches[name]
             for k in patch_list:
@@ -1102,8 +908,8 @@ def cast_transformer_options(transformer_options: dict[str], device=None, dtype=
     # try to call .to on any wrappers/callbacks
     wrappers_and_callbacks = ["wrappers", "callbacks"]
     for wc_name in wrappers_and_callbacks:
-        if wc_name in transformer_options:
+        if wc_name in to_load_options:
-            wc: dict[str, list] = transformer_options[wc_name]
+            wc: dict[str, list] = to_load_options[wc_name]
             for wc_dict in wc.values():
                 for wc_list in wc_dict.values():
                     for i in range(len(wc_list)):
@@ -1111,6 +917,7 @@ def cast_transformer_options(transformer_options: dict[str], device=None, dtype=
                             for cast in casts:
                                 wc_list[i] = wc_list[i].to(cast)
 
+
 class CFGGuider:
     def __init__(self, model_patcher: ModelPatcher):
         self.model_patcher = model_patcher
@@ -1156,8 +963,6 @@ class CFGGuider:
         self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options)
         device = self.model_patcher.load_device
 
-        multigpu_patchers = comfy.sampler_helpers.prepare_model_patcher_multigpu_clones(self.model_patcher, self.loaded_models, self.model_options)
-
         if denoise_mask is not None:
             denoise_mask = comfy.sampler_helpers.prepare_mask(denoise_mask, noise.shape, device)
 
@@ -1168,13 +973,9 @@ class CFGGuider:
 
         try:
             self.model_patcher.pre_run()
-            for multigpu_patcher in multigpu_patchers:
-                multigpu_patcher.pre_run()
             output = self.inner_sample(noise, latent_image, device, sampler, sigmas, denoise_mask, callback, disable_pbar, seed)
         finally:
             self.model_patcher.cleanup()
-            for multigpu_patcher in multigpu_patchers:
-                multigpu_patcher.cleanup()
 
         comfy.sampler_helpers.cleanup_models(self.conds, self.loaded_models)
         del self.inner_model

comfy/sd.py

@@ -14,6 +14,7 @@ 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 yaml
 import math
 
@@ -412,6 +413,17 @@ class VAE:
                 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]
             else:
                 logging.warning("WARNING: No VAE weights detected, VAE not initalized.")
                 self.first_stage_model = None
@@ -498,7 +510,7 @@ class VAE:
         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:
@@ -510,7 +522,7 @@ class VAE:
 
             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

comfy/supported_models.py

@@ -506,6 +506,22 @@ 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,
@@ -959,6 +975,44 @@ class WAN21_I2V(WAN21_T2V):
         out = model_base.WAN21(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, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V]
+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
+
+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, Hunyuan3Dv2mini, Hunyuan3Dv2]
 
 models += [SVD_img2vid]

comfy_execution/graph.py

@@ -1,6 +1,9 @@
-import nodes
+from __future__ import annotations
+from typing import Type, Literal
 
+import nodes
 from comfy_execution.graph_utils import is_link
+from comfy.comfy_types.node_typing import ComfyNodeABC, InputTypeDict, InputTypeOptions
 
 class DependencyCycleError(Exception):
     pass
@@ -54,7 +57,22 @@ class DynamicPrompt:
     def get_original_prompt(self):
         return self.original_prompt
 
-def get_input_info(class_def, input_name, valid_inputs=None):
+def get_input_info(
+    class_def: Type[ComfyNodeABC],
+    input_name: str,
+    valid_inputs: InputTypeDict | None = None
+) -> tuple[str, Literal["required", "optional", "hidden"], InputTypeOptions] | tuple[None, None, None]:
+    """Get the input type, category, and extra info for a given input name.
+
+    Arguments:
+        class_def: The class definition of the node.
+        input_name: The name of the input to get info for.
+        valid_inputs: The valid inputs for the node, or None to use the class_def.INPUT_TYPES().
+
+    Returns:
+        tuple[str, str, dict] | tuple[None, None, None]: The input type, category, and extra info for the input name.
+    """
+
     valid_inputs = valid_inputs or class_def.INPUT_TYPES()
     input_info = None
     input_category = None
@@ -126,7 +144,7 @@ class TopologicalSort:
                     from_node_id, from_socket = value
                     if subgraph_nodes is not None and from_node_id not in subgraph_nodes:
                         continue
-                    input_type, input_category, input_info = self.get_input_info(unique_id, input_name)
+                    _, _, input_info = self.get_input_info(unique_id, input_name)
                     is_lazy = input_info is not None and "lazy" in input_info and input_info["lazy"]
                     if (include_lazy or not is_lazy) and not self.is_cached(from_node_id):
                         node_ids.append(from_node_id)

comfy_extras/nodes_hunyuan3d.py

@@ -0,0 +1,415 @@
+import torch
+import os
+import json
+import struct
+import numpy as np
+from comfy.ldm.modules.diffusionmodules.mmdit import get_1d_sincos_pos_embed_from_grid_torch
+import folder_paths
+import comfy.model_management
+from comfy.cli_args import args
+
+
+class EmptyLatentHunyuan3Dv2:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"resolution": ("INT", {"default": 3072, "min": 1, "max": 8192}),
+                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096, "tooltip": "The number of latent images in the batch."}),
+                             }}
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "generate"
+
+    CATEGORY = "latent/3d"
+
+    def generate(self, resolution, batch_size):
+        latent = torch.zeros([batch_size, 64, resolution], device=comfy.model_management.intermediate_device())
+        return ({"samples": latent, "type": "hunyuan3dv2"}, )
+
+
+class Hunyuan3Dv2Conditioning:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"clip_vision_output": ("CLIP_VISION_OUTPUT",),
+                             }}
+
+    RETURN_TYPES = ("CONDITIONING", "CONDITIONING")
+    RETURN_NAMES = ("positive", "negative")
+
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/video_models"
+
+    def encode(self, clip_vision_output):
+        embeds = clip_vision_output.last_hidden_state
+        positive = [[embeds, {}]]
+        negative = [[torch.zeros_like(embeds), {}]]
+        return (positive, negative)
+
+
+class Hunyuan3Dv2ConditioningMultiView:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {},
+                "optional": {"front": ("CLIP_VISION_OUTPUT",),
+                             "left": ("CLIP_VISION_OUTPUT",),
+                             "back": ("CLIP_VISION_OUTPUT",),
+                             "right": ("CLIP_VISION_OUTPUT",), }}
+
+    RETURN_TYPES = ("CONDITIONING", "CONDITIONING")
+    RETURN_NAMES = ("positive", "negative")
+
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/video_models"
+
+    def encode(self, front=None, left=None, back=None, right=None):
+        all_embeds = [front, left, back, right]
+        out = []
+        pos_embeds = None
+        for i, e in enumerate(all_embeds):
+            if e is not None:
+                if pos_embeds is None:
+                    pos_embeds = get_1d_sincos_pos_embed_from_grid_torch(e.last_hidden_state.shape[-1], torch.arange(4))
+                out.append(e.last_hidden_state + pos_embeds[i].reshape(1, 1, -1))
+
+        embeds = torch.cat(out, dim=1)
+        positive = [[embeds, {}]]
+        negative = [[torch.zeros_like(embeds), {}]]
+        return (positive, negative)
+
+
+class VOXEL:
+    def __init__(self, data):
+        self.data = data
+
+
+class VAEDecodeHunyuan3D:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"samples": ("LATENT", ),
+                             "vae": ("VAE", ),
+                             "num_chunks": ("INT", {"default": 8000, "min": 1000, "max": 500000}),
+                             "octree_resolution": ("INT", {"default": 256, "min": 16, "max": 512}),
+                             }}
+    RETURN_TYPES = ("VOXEL",)
+    FUNCTION = "decode"
+
+    CATEGORY = "latent/3d"
+
+    def decode(self, vae, samples, num_chunks, octree_resolution):
+        voxels = VOXEL(vae.decode(samples["samples"], vae_options={"num_chunks": num_chunks, "octree_resolution": octree_resolution}))
+        return (voxels, )
+
+
+def voxel_to_mesh(voxels, threshold=0.5, device=None):
+    if device is None:
+        device = torch.device("cpu")
+    voxels = voxels.to(device)
+
+    binary = (voxels > threshold).float()
+    padded = torch.nn.functional.pad(binary, (1, 1, 1, 1, 1, 1), 'constant', 0)
+
+    D, H, W = binary.shape
+
+    neighbors = torch.tensor([
+        [0, 0, 1],
+        [0, 0, -1],
+        [0, 1, 0],
+        [0, -1, 0],
+        [1, 0, 0],
+        [-1, 0, 0]
+    ], device=device)
+
+    z, y, x = torch.meshgrid(
+        torch.arange(D, device=device),
+        torch.arange(H, device=device),
+        torch.arange(W, device=device),
+        indexing='ij'
+    )
+    voxel_indices = torch.stack([z.flatten(), y.flatten(), x.flatten()], dim=1)
+
+    solid_mask = binary.flatten() > 0
+    solid_indices = voxel_indices[solid_mask]
+
+    corner_offsets = [
+        torch.tensor([
+            [0, 0, 1], [0, 1, 1], [1, 1, 1], [1, 0, 1]
+        ], device=device),
+        torch.tensor([
+            [0, 0, 0], [1, 0, 0], [1, 1, 0], [0, 1, 0]
+        ], device=device),
+        torch.tensor([
+            [0, 1, 0], [1, 1, 0], [1, 1, 1], [0, 1, 1]
+        ], device=device),
+        torch.tensor([
+            [0, 0, 0], [0, 0, 1], [1, 0, 1], [1, 0, 0]
+        ], device=device),
+        torch.tensor([
+            [1, 0, 1], [1, 1, 1], [1, 1, 0], [1, 0, 0]
+        ], device=device),
+        torch.tensor([
+            [0, 1, 0], [0, 1, 1], [0, 0, 1], [0, 0, 0]
+        ], device=device)
+    ]
+
+    all_vertices = []
+    all_indices = []
+
+    vertex_count = 0
+
+    for face_idx, offset in enumerate(neighbors):
+        neighbor_indices = solid_indices + offset
+
+        padded_indices = neighbor_indices + 1
+
+        is_exposed = padded[
+            padded_indices[:, 0],
+            padded_indices[:, 1],
+            padded_indices[:, 2]
+        ] == 0
+
+        if not is_exposed.any():
+            continue
+
+        exposed_indices = solid_indices[is_exposed]
+
+        corners = corner_offsets[face_idx].unsqueeze(0)
+
+        face_vertices = exposed_indices.unsqueeze(1) + corners
+
+        all_vertices.append(face_vertices.reshape(-1, 3))
+
+        num_faces = exposed_indices.shape[0]
+        face_indices = torch.arange(
+            vertex_count,
+            vertex_count + 4 * num_faces,
+            device=device
+        ).reshape(-1, 4)
+
+        all_indices.append(torch.stack([face_indices[:, 0], face_indices[:, 1], face_indices[:, 2]], dim=1))
+        all_indices.append(torch.stack([face_indices[:, 0], face_indices[:, 2], face_indices[:, 3]], dim=1))
+
+        vertex_count += 4 * num_faces
+
+    if len(all_vertices) > 0:
+        vertices = torch.cat(all_vertices, dim=0)
+        faces = torch.cat(all_indices, dim=0)
+    else:
+        vertices = torch.zeros((1, 3))
+        faces = torch.zeros((1, 3))
+
+    v_min = 0
+    v_max = max(voxels.shape)
+
+    vertices = vertices - (v_min + v_max) / 2
+
+    scale = (v_max - v_min) / 2
+    if scale > 0:
+        vertices = vertices / scale
+
+    vertices = torch.fliplr(vertices)
+    return vertices, faces
+
+
+class MESH:
+    def __init__(self, vertices, faces):
+        self.vertices = vertices
+        self.faces = faces
+
+
+class VoxelToMeshBasic:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"voxel": ("VOXEL", ),
+                             "threshold": ("FLOAT", {"default": 0.6, "min": -1.0, "max": 1.0, "step": 0.01}),
+                             }}
+    RETURN_TYPES = ("MESH",)
+    FUNCTION = "decode"
+
+    CATEGORY = "3d"
+
+    def decode(self, voxel, threshold):
+        vertices = []
+        faces = []
+        for x in voxel.data:
+            v, f = voxel_to_mesh(x, threshold=threshold, device=None)
+            vertices.append(v)
+            faces.append(f)
+
+        return (MESH(torch.stack(vertices), torch.stack(faces)), )
+
+
+def save_glb(vertices, faces, filepath, metadata=None):
+    """
+    Save PyTorch tensor vertices and faces as a GLB file without external dependencies.
+
+    Parameters:
+    vertices: torch.Tensor of shape (N, 3) - The vertex coordinates
+    faces: torch.Tensor of shape (M, 4) or (M, 3) - The face indices (quad or triangle faces)
+    filepath: str - Output filepath (should end with .glb)
+    """
+
+    # Convert tensors to numpy arrays
+    vertices_np = vertices.cpu().numpy().astype(np.float32)
+    faces_np = faces.cpu().numpy().astype(np.uint32)
+
+    vertices_buffer = vertices_np.tobytes()
+    indices_buffer = faces_np.tobytes()
+
+    def pad_to_4_bytes(buffer):
+        padding_length = (4 - (len(buffer) % 4)) % 4
+        return buffer + b'\x00' * padding_length
+
+    vertices_buffer_padded = pad_to_4_bytes(vertices_buffer)
+    indices_buffer_padded = pad_to_4_bytes(indices_buffer)
+
+    buffer_data = vertices_buffer_padded + indices_buffer_padded
+
+    vertices_byte_length = len(vertices_buffer)
+    vertices_byte_offset = 0
+    indices_byte_length = len(indices_buffer)
+    indices_byte_offset = len(vertices_buffer_padded)
+
+    gltf = {
+        "asset": {"version": "2.0", "generator": "ComfyUI"},
+        "buffers": [
+            {
+                "byteLength": len(buffer_data)
+            }
+        ],
+        "bufferViews": [
+            {
+                "buffer": 0,
+                "byteOffset": vertices_byte_offset,
+                "byteLength": vertices_byte_length,
+                "target": 34962  # ARRAY_BUFFER
+            },
+            {
+                "buffer": 0,
+                "byteOffset": indices_byte_offset,
+                "byteLength": indices_byte_length,
+                "target": 34963  # ELEMENT_ARRAY_BUFFER
+            }
+        ],
+        "accessors": [
+            {
+                "bufferView": 0,
+                "byteOffset": 0,
+                "componentType": 5126,  # FLOAT
+                "count": len(vertices_np),
+                "type": "VEC3",
+                "max": vertices_np.max(axis=0).tolist(),
+                "min": vertices_np.min(axis=0).tolist()
+            },
+            {
+                "bufferView": 1,
+                "byteOffset": 0,
+                "componentType": 5125,  # UNSIGNED_INT
+                "count": faces_np.size,
+                "type": "SCALAR"
+            }
+        ],
+        "meshes": [
+            {
+                "primitives": [
+                    {
+                        "attributes": {
+                            "POSITION": 0
+                        },
+                        "indices": 1,
+                        "mode": 4  # TRIANGLES
+                    }
+                ]
+            }
+        ],
+        "nodes": [
+            {
+                "mesh": 0
+            }
+        ],
+        "scenes": [
+            {
+                "nodes": [0]
+            }
+        ],
+        "scene": 0
+    }
+
+    if metadata is not None:
+        gltf["asset"]["extras"] = metadata
+
+    # Convert the JSON to bytes
+    gltf_json = json.dumps(gltf).encode('utf8')
+
+    def pad_json_to_4_bytes(buffer):
+        padding_length = (4 - (len(buffer) % 4)) % 4
+        return buffer + b' ' * padding_length
+
+    gltf_json_padded = pad_json_to_4_bytes(gltf_json)
+
+    # Create the GLB header
+    # Magic glTF
+    glb_header = struct.pack('<4sII', b'glTF', 2, 12 + 8 + len(gltf_json_padded) + 8 + len(buffer_data))
+
+    # Create JSON chunk header (chunk type 0)
+    json_chunk_header = struct.pack('<II', len(gltf_json_padded), 0x4E4F534A)  # "JSON" in little endian
+
+    # Create BIN chunk header (chunk type 1)
+    bin_chunk_header = struct.pack('<II', len(buffer_data), 0x004E4942)  # "BIN\0" in little endian
+
+    # Write the GLB file
+    with open(filepath, 'wb') as f:
+        f.write(glb_header)
+        f.write(json_chunk_header)
+        f.write(gltf_json_padded)
+        f.write(bin_chunk_header)
+        f.write(buffer_data)
+
+    return filepath
+
+
+class SaveGLB:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"mesh": ("MESH", ),
+                             "filename_prefix": ("STRING", {"default": "mesh/ComfyUI"}), },
+                "hidden": {"prompt": "PROMPT", "extra_pnginfo": "EXTRA_PNGINFO"}, }
+
+    RETURN_TYPES = ()
+    FUNCTION = "save"
+
+    OUTPUT_NODE = True
+
+    CATEGORY = "3d"
+
+    def save(self, mesh, filename_prefix, prompt=None, extra_pnginfo=None):
+        full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, folder_paths.get_output_directory())
+        results = []
+
+        metadata = {}
+        if not args.disable_metadata:
+            if prompt is not None:
+                metadata["prompt"] = json.dumps(prompt)
+            if extra_pnginfo is not None:
+                for x in extra_pnginfo:
+                    metadata[x] = json.dumps(extra_pnginfo[x])
+
+        for i in range(mesh.vertices.shape[0]):
+            f = f"{filename}_{counter:05}_.glb"
+            save_glb(mesh.vertices[i], mesh.faces[i], os.path.join(full_output_folder, f), metadata)
+            results.append({
+                "filename": f,
+                "subfolder": subfolder,
+                "type": "output"
+            })
+            counter += 1
+        return {"ui": {"3d": results}}
+
+
+NODE_CLASS_MAPPINGS = {
+    "EmptyLatentHunyuan3Dv2": EmptyLatentHunyuan3Dv2,
+    "Hunyuan3Dv2Conditioning": Hunyuan3Dv2Conditioning,
+    "Hunyuan3Dv2ConditioningMultiView": Hunyuan3Dv2ConditioningMultiView,
+    "VAEDecodeHunyuan3D": VAEDecodeHunyuan3D,
+    "VoxelToMeshBasic": VoxelToMeshBasic,
+    "SaveGLB": SaveGLB,
+}

comfy_extras/nodes_load_3d.py

@@ -21,8 +21,8 @@ class Load3D():
             "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}),
         }}
 
-    RETURN_TYPES = ("IMAGE", "MASK", "STRING")
+    RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE", "IMAGE")
-    RETURN_NAMES = ("image", "mask", "mesh_path")
+    RETURN_NAMES = ("image", "mask", "mesh_path", "normal", "lineart")
 
     FUNCTION = "process"
     EXPERIMENTAL = True
@@ -32,12 +32,16 @@ class Load3D():
     def process(self, model_file, image, **kwargs):
         image_path = folder_paths.get_annotated_filepath(image['image'])
         mask_path = folder_paths.get_annotated_filepath(image['mask'])
+        normal_path = folder_paths.get_annotated_filepath(image['normal'])
+        lineart_path = folder_paths.get_annotated_filepath(image['lineart'])
 
         load_image_node = nodes.LoadImage()
         output_image, ignore_mask = load_image_node.load_image(image=image_path)
         ignore_image, output_mask = load_image_node.load_image(image=mask_path)
+        normal_image, ignore_mask2 = load_image_node.load_image(image=normal_path)
+        lineart_image, ignore_mask3 = load_image_node.load_image(image=lineart_path)
 
-        return output_image, output_mask, model_file,
+        return output_image, output_mask, model_file, normal_image, lineart_image
 
 class Load3DAnimation():
     @classmethod
@@ -55,8 +59,8 @@ class Load3DAnimation():
             "height": ("INT", {"default": 1024, "min": 1, "max": 4096, "step": 1}),
         }}
 
-    RETURN_TYPES = ("IMAGE", "MASK", "STRING")
+    RETURN_TYPES = ("IMAGE", "MASK", "STRING", "IMAGE")
-    RETURN_NAMES = ("image", "mask", "mesh_path")
+    RETURN_NAMES = ("image", "mask", "mesh_path", "normal")
 
     FUNCTION = "process"
     EXPERIMENTAL = True
@@ -66,12 +70,14 @@ class Load3DAnimation():
     def process(self, model_file, image, **kwargs):
         image_path = folder_paths.get_annotated_filepath(image['image'])
         mask_path = folder_paths.get_annotated_filepath(image['mask'])
+        normal_path = folder_paths.get_annotated_filepath(image['normal'])
 
         load_image_node = nodes.LoadImage()
         output_image, ignore_mask = load_image_node.load_image(image=image_path)
         ignore_image, output_mask = load_image_node.load_image(image=mask_path)
+        normal_image, ignore_mask2 = load_image_node.load_image(image=normal_path)
 
-        return output_image, output_mask, model_file,
+        return output_image, output_mask, model_file, normal_image
 
 class Preview3D():
     @classmethod

comfy_extras/nodes_model_advanced.py

@@ -20,10 +20,6 @@ class LCM(comfy.model_sampling.EPS):
 
         return c_out * x0 + c_skip * model_input
 
-class X0(comfy.model_sampling.EPS):
-    def calculate_denoised(self, sigma, model_output, model_input):
-        return model_output
-
 class ModelSamplingDiscreteDistilled(comfy.model_sampling.ModelSamplingDiscrete):
     original_timesteps = 50
 
@@ -56,7 +52,7 @@ class ModelSamplingDiscrete:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "model": ("MODEL",),
-                              "sampling": (["eps", "v_prediction", "lcm", "x0"],),
+                              "sampling": (["eps", "v_prediction", "lcm", "x0", "img_to_img"],),
                               "zsnr": ("BOOLEAN", {"default": False}),
                               }}
 
@@ -77,7 +73,9 @@ class ModelSamplingDiscrete:
             sampling_type = LCM
             sampling_base = ModelSamplingDiscreteDistilled
         elif sampling == "x0":
-            sampling_type = X0
+            sampling_type = comfy.model_sampling.X0
+        elif sampling == "img_to_img":
+            sampling_type = comfy.model_sampling.IMG_TO_IMG
 
         class ModelSamplingAdvanced(sampling_base, sampling_type):
             pass

comfy_extras/nodes_model_merging_model_specific.py

@@ -244,6 +244,30 @@ class ModelMergeCosmos14B(comfy_extras.nodes_model_merging.ModelMergeBlocks):
 
         return {"required": arg_dict}
 
+class ModelMergeWAN2_1(comfy_extras.nodes_model_merging.ModelMergeBlocks):
+    CATEGORY = "advanced/model_merging/model_specific"
+    DESCRIPTION = "1.3B model has 30 blocks, 14B model has 40 blocks. Image to video model has the extra img_emb."
+
+    @classmethod
+    def INPUT_TYPES(s):
+        arg_dict = { "model1": ("MODEL",),
+                              "model2": ("MODEL",)}
+
+        argument = ("FLOAT", {"default": 1.0, "min": 0.0, "max": 1.0, "step": 0.01})
+
+        arg_dict["patch_embedding."] = argument
+        arg_dict["time_embedding."] = argument
+        arg_dict["time_projection."] = argument
+        arg_dict["text_embedding."] = argument
+        arg_dict["img_emb."] = argument
+
+        for i in range(40):
+            arg_dict["blocks.{}.".format(i)] = argument
+
+        arg_dict["head."] = argument
+
+        return {"required": arg_dict}
+
 NODE_CLASS_MAPPINGS = {
     "ModelMergeSD1": ModelMergeSD1,
     "ModelMergeSD2": ModelMergeSD1, #SD1 and SD2 have the same blocks
@@ -256,4 +280,5 @@ NODE_CLASS_MAPPINGS = {
     "ModelMergeLTXV": ModelMergeLTXV,
     "ModelMergeCosmos7B": ModelMergeCosmos7B,
     "ModelMergeCosmos14B": ModelMergeCosmos14B,
+    "ModelMergeWAN2_1": ModelMergeWAN2_1,
 }

comfy_extras/nodes_morphology.py

@@ -2,6 +2,7 @@ import torch
 import comfy.model_management
 
 from kornia.morphology import dilation, erosion, opening, closing, gradient, top_hat, bottom_hat
+import kornia.color
 
 
 class Morphology:
@@ -40,8 +41,45 @@ class Morphology:
         img_out = output.to(comfy.model_management.intermediate_device()).movedim(1, -1)
         return (img_out,)
 
+
+class ImageRGBToYUV:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "image": ("IMAGE",),
+                              }}
+
+    RETURN_TYPES = ("IMAGE", "IMAGE", "IMAGE")
+    RETURN_NAMES = ("Y", "U", "V")
+    FUNCTION = "execute"
+
+    CATEGORY = "image/batch"
+
+    def execute(self, image):
+        out = kornia.color.rgb_to_ycbcr(image.movedim(-1, 1)).movedim(1, -1)
+        return (out[..., 0:1].expand_as(image), out[..., 1:2].expand_as(image), out[..., 2:3].expand_as(image))
+
+class ImageYUVToRGB:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"Y": ("IMAGE",),
+                             "U": ("IMAGE",),
+                             "V": ("IMAGE",),
+                              }}
+
+    RETURN_TYPES = ("IMAGE",)
+    FUNCTION = "execute"
+
+    CATEGORY = "image/batch"
+
+    def execute(self, Y, U, V):
+        image = torch.cat([torch.mean(Y, dim=-1, keepdim=True), torch.mean(U, dim=-1, keepdim=True), torch.mean(V, dim=-1, keepdim=True)], dim=-1)
+        out = kornia.color.ycbcr_to_rgb(image.movedim(-1, 1)).movedim(1, -1)
+        return (out,)
+
 NODE_CLASS_MAPPINGS = {
     "Morphology": Morphology,
+    "ImageRGBToYUV": ImageRGBToYUV,
+    "ImageYUVToRGB": ImageYUVToRGB,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {

comfy_extras/nodes_multigpu.py

@@ -1,108 +0,0 @@
-from __future__ import annotations
-from inspect import cleandoc
-
-from typing import TYPE_CHECKING
-if TYPE_CHECKING:
-    from comfy.model_patcher import ModelPatcher
-import comfy.multigpu
-
-from nodes import VAELoader
-
-
-class VAELoaderDevice(VAELoader):
-    NodeId = "VAELoaderDevice"
-    NodeName = "Load VAE MultiGPU"
-    @classmethod
-    def INPUT_TYPES(cls):
-        return {
-            "required": {
-                "vae_name": (cls.vae_list(), ),
-                "load_device": (comfy.multigpu.get_torch_device_list(), ),
-            }
-        }
-    
-    FUNCTION = "load_vae_device"
-    CATEGORY = "advanced/multigpu/loaders"
-
-    def load_vae_device(self, vae_name, load_device: str):
-        device = comfy.multigpu.get_device_from_str(load_device)
-        return self.load_vae(vae_name, device)
-
-class MultiGPUWorkUnitsNode:
-    """
-    Prepares model to have sampling accelerated via splitting work units.
-
-    Should be placed after nodes that modify the model object itself, such as compile or attention-switch nodes.
-
-    Other than those exceptions, this node can be placed in any order.
-    """
-
-    NodeId = "MultiGPU_WorkUnits"
-    NodeName = "MultiGPU Work Units"
-    @classmethod
-    def INPUT_TYPES(cls):
-        return {
-            "required": {
-                "model": ("MODEL",),
-                "max_gpus" : ("INT", {"default": 8, "min": 1, "step": 1}),
-            },
-            "optional": {
-                "gpu_options": ("GPU_OPTIONS",)
-            }
-        }
-
-    RETURN_TYPES = ("MODEL",)
-    FUNCTION = "init_multigpu"
-    CATEGORY = "advanced/multigpu"
-    DESCRIPTION = cleandoc(__doc__)
-
-    def init_multigpu(self, model: ModelPatcher, max_gpus: int, gpu_options: comfy.multigpu.GPUOptionsGroup=None):
-        model = comfy.multigpu.create_multigpu_deepclones(model, max_gpus, gpu_options, reuse_loaded=True)
-        return (model,)
-
-class MultiGPUOptionsNode:
-    """
-    Select the relative speed of GPUs in the special case they have significantly different performance from one another.
-    """
-
-    NodeId = "MultiGPU_Options"
-    NodeName = "MultiGPU Options"
-    @classmethod
-    def INPUT_TYPES(cls):
-        return {
-            "required": {
-                "device_index": ("INT", {"default": 0, "min": 0, "max": 64}),
-                "relative_speed": ("FLOAT", {"default": 1.0, "min": 0.0, "step": 0.01})
-            },
-            "optional": {
-                "gpu_options": ("GPU_OPTIONS",)
-            }
-        }
-
-    RETURN_TYPES = ("GPU_OPTIONS",)
-    FUNCTION = "create_gpu_options"
-    CATEGORY = "advanced/multigpu"
-    DESCRIPTION = cleandoc(__doc__)
-
-    def create_gpu_options(self, device_index: int, relative_speed: float, gpu_options: comfy.multigpu.GPUOptionsGroup=None):
-        if not gpu_options:
-            gpu_options = comfy.multigpu.GPUOptionsGroup()
-        gpu_options.clone()
-
-        opt = comfy.multigpu.GPUOptions(device_index=device_index, relative_speed=relative_speed)
-        gpu_options.add(opt)
-
-        return (gpu_options,)
-
-
-node_list = [
-    MultiGPUWorkUnitsNode,
-    MultiGPUOptionsNode,
-    VAELoaderDevice,
-]
-NODE_CLASS_MAPPINGS = {}
-NODE_DISPLAY_NAME_MAPPINGS = {}
-
-for node in node_list:
-    NODE_CLASS_MAPPINGS[node.NodeId] = node
-    NODE_DISPLAY_NAME_MAPPINGS[node.NodeId] = node.NodeName

comfy_extras/nodes_primitive.py

@@ -0,0 +1,79 @@
+# Primitive nodes that are evaluated at backend.
+from __future__ import annotations
+
+from comfy.comfy_types.node_typing import ComfyNodeABC, InputTypeDict, IO
+
+
+class String(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {"value": (IO.STRING, {})},
+        }
+
+    RETURN_TYPES = (IO.STRING,)
+    FUNCTION = "execute"
+    CATEGORY = "utils/primitive"
+
+    def execute(self, value: str) -> tuple[str]:
+        return (value,)
+
+
+class Int(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {"value": (IO.INT, {"control_after_generate": True})},
+        }
+
+    RETURN_TYPES = (IO.INT,)
+    FUNCTION = "execute"
+    CATEGORY = "utils/primitive"
+
+    def execute(self, value: int) -> tuple[int]:
+        return (value,)
+
+
+class Float(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {"value": (IO.FLOAT, {})},
+        }
+
+    RETURN_TYPES = (IO.FLOAT,)
+    FUNCTION = "execute"
+    CATEGORY = "utils/primitive"
+
+    def execute(self, value: float) -> tuple[float]:
+        return (value,)
+
+
+class Boolean(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {"value": (IO.BOOLEAN, {})},
+        }
+
+    RETURN_TYPES = (IO.BOOLEAN,)
+    FUNCTION = "execute"
+    CATEGORY = "utils/primitive"
+
+    def execute(self, value: bool) -> tuple[bool]:
+        return (value,)
+
+
+NODE_CLASS_MAPPINGS = {
+    "PrimitiveString": String,
+    "PrimitiveInt": Int,
+    "PrimitiveFloat": Float,
+    "PrimitiveBoolean": Boolean,
+}
+
+NODE_DISPLAY_NAME_MAPPINGS = {
+    "PrimitiveString": "String",
+    "PrimitiveInt": "Int",
+    "PrimitiveFloat": "Float",
+    "PrimitiveBoolean": "Boolean",
+}

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.26"
+__version__ = "0.3.27"

execution.py

@@ -93,7 +93,7 @@ def get_input_data(inputs, class_def, unique_id, outputs=None, dynprompt=None, e
     missing_keys = {}
     for x in inputs:
         input_data = inputs[x]
-        input_type, input_category, input_info = get_input_info(class_def, x, valid_inputs)
+        _, input_category, input_info = get_input_info(class_def, x, valid_inputs)
         def mark_missing():
             missing_keys[x] = True
             input_data_all[x] = (None,)
@@ -555,7 +555,7 @@ def validate_inputs(prompt, item, validated):
     received_types = {}
 
     for x in valid_inputs:
-        type_input, input_category, extra_info = get_input_info(obj_class, x, class_inputs)
+        input_type, input_category, extra_info = get_input_info(obj_class, x, class_inputs)
         assert extra_info is not None
         if x not in inputs:
             if input_category == "required":
@@ -571,7 +571,7 @@ def validate_inputs(prompt, item, validated):
             continue
 
         val = inputs[x]
-        info = (type_input, extra_info)
+        info = (input_type, extra_info)
         if isinstance(val, list):
             if len(val) != 2:
                 error = {
@@ -592,8 +592,8 @@ def validate_inputs(prompt, item, validated):
             r = nodes.NODE_CLASS_MAPPINGS[o_class_type].RETURN_TYPES
             received_type = r[val[1]]
             received_types[x] = received_type
-            if 'input_types' not in validate_function_inputs and not validate_node_input(received_type, type_input):
+            if 'input_types' not in validate_function_inputs and not validate_node_input(received_type, input_type):
-                details = f"{x}, received_type({received_type}) mismatch input_type({type_input})"
+                details = f"{x}, received_type({received_type}) mismatch input_type({input_type})"
                 error = {
                     "type": "return_type_mismatch",
                     "message": "Return type mismatch between linked nodes",
@@ -641,22 +641,22 @@ def validate_inputs(prompt, item, validated):
                     val = val["__value__"]
                     inputs[x] = val
 
-                if type_input == "INT":
+                if input_type == "INT":
                     val = int(val)
                     inputs[x] = val
-                if type_input == "FLOAT":
+                if input_type == "FLOAT":
                     val = float(val)
                     inputs[x] = val
-                if type_input == "STRING":
+                if input_type == "STRING":
                     val = str(val)
                     inputs[x] = val
-                if type_input == "BOOLEAN":
+                if input_type == "BOOLEAN":
                     val = bool(val)
                     inputs[x] = val
             except Exception as ex:
                 error = {
                     "type": "invalid_input_type",
-                    "message": f"Failed to convert an input value to a {type_input} value",
+                    "message": f"Failed to convert an input value to a {input_type} value",
                     "details": f"{x}, {val}, {ex}",
                     "extra_info": {
                         "input_name": x,
@@ -696,18 +696,19 @@ def validate_inputs(prompt, item, validated):
                     errors.append(error)
                     continue
 
-                if isinstance(type_input, list):
+                if isinstance(input_type, list):
-                    if val not in type_input:
+                    combo_options = input_type
+                    if val not in combo_options:
                         input_config = info
                         list_info = ""
 
                         # Don't send back gigantic lists like if they're lots of
                         # scanned model filepaths
-                        if len(type_input) > 20:
+                        if len(combo_options) > 20:
-                            list_info = f"(list of length {len(type_input)})"
+                            list_info = f"(list of length {len(combo_options)})"
                             input_config = None
                         else:
-                            list_info = str(type_input)
+                            list_info = str(combo_options)
 
                         error = {
                             "type": "value_not_in_list",

nodes.py

@@ -763,13 +763,13 @@ class VAELoader:
     CATEGORY = "loaders"
 
     #TODO: scale factor?
-    def load_vae(self, vae_name, device=None):
+    def load_vae(self, vae_name):
         if vae_name in ["taesd", "taesdxl", "taesd3", "taef1"]:
             sd = self.load_taesd(vae_name)
         else:
             vae_path = folder_paths.get_full_path_or_raise("vae", vae_name)
             sd = comfy.utils.load_torch_file(vae_path)
-        vae = comfy.sd.VAE(sd=sd, device=device)
+        vae = comfy.sd.VAE(sd=sd)
         vae.throw_exception_if_invalid()
         return (vae,)
 
@@ -2259,12 +2259,14 @@ def init_builtin_extra_nodes():
         "nodes_mahiro.py",
         "nodes_lt.py",
         "nodes_hooks.py",
-        "nodes_multigpu.py",
         "nodes_load_3d.py",
         "nodes_cosmos.py",
         "nodes_video.py",
         "nodes_lumina2.py",
         "nodes_wan.py",
+        "nodes_lotus.py",
+        "nodes_hunyuan3d.py",
+        "nodes_primitive.py",
     ]
 
     import_failed = []

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.26"
+version = "0.3.27"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.12.14
+comfyui-frontend-package==1.14.5
 torch
 torchsde
 torchvision