Bump ComfyUI version to v0.3.17

Fix bug when changing the compute dtype on wan.

README.md

@@ -1,7 +1,7 @@
 <div align="center">
 
 # ComfyUI
-**The most powerful and modular diffusion model GUI and backend.**
+**The most powerful and modular visual AI engine and application.**
 
 
 [![Website][website-shield]][website-url]
@@ -31,10 +31,24 @@
 ![ComfyUI Screenshot](https://github.com/user-attachments/assets/7ccaf2c1-9b72-41ae-9a89-5688c94b7abe)
 </div>
 
-This ui will let you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. For some workflow examples and see what ComfyUI can do you can check out:
-### [ComfyUI Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
+ComfyUI lets you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. Available on Windows, Linux, and macOS.
+
+## Get Started
+
+#### [Desktop Application](https://www.comfy.org/download)
+- The easiest way to get started. 
+- Available on Windows & macOS.
+
+#### [Windows Portable Package](#installing)
+- Get the latest commits and completely portable.
+- Available on Windows.
+
+#### [Manual Install](#manual-install-windows-linux)
+Supports all operating systems and GPU types (NVIDIA, AMD, Intel, Apple Silicon, Ascend).
+
+## Examples
+See what ComfyUI can do with the [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/).
 
-### [Installing ComfyUI](#installing)
 
 ## Features
 - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.
@@ -121,7 +135,7 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 
 # Installing
 
-## Windows
+## Windows Portable
 
 There is a portable standalone build for Windows that should work for running on Nvidia GPUs or for running on your CPU only on the [releases page](https://github.com/comfyanonymous/ComfyUI/releases).
 
@@ -141,6 +155,15 @@ See the [Config file](extra_model_paths.yaml.example) to set the search paths fo
 
 To run it on services like paperspace, kaggle or colab you can use my [Jupyter Notebook](notebooks/comfyui_colab.ipynb)
 
+
+## [comfy-cli](https://docs.comfy.org/comfy-cli/getting-started)
+
+You can install and start ComfyUI using comfy-cli:
+```bash
+pip install comfy-cli
+comfy install
+```
+
 ## Manual Install (Windows, Linux)
 
 python 3.13 is supported but using 3.12 is recommended because some custom nodes and their dependencies might not support it yet.

comfy/latent_formats.py

@@ -407,3 +407,52 @@ class Cosmos1CV8x8x8(LatentFormat):
     ]
 
     latent_rgb_factors_bias = [-0.1223, -0.1889, -0.1976]
+
+class Wan21(LatentFormat):
+    latent_channels = 16
+    latent_dimensions = 3
+
+    latent_rgb_factors = [
+            [-0.1299, -0.1692,  0.2932],
+            [ 0.0671,  0.0406,  0.0442],
+            [ 0.3568,  0.2548,  0.1747],
+            [ 0.0372,  0.2344,  0.1420],
+            [ 0.0313,  0.0189, -0.0328],
+            [ 0.0296, -0.0956, -0.0665],
+            [-0.3477, -0.4059, -0.2925],
+            [ 0.0166,  0.1902,  0.1975],
+            [-0.0412,  0.0267, -0.1364],
+            [-0.1293,  0.0740,  0.1636],
+            [ 0.0680,  0.3019,  0.1128],
+            [ 0.0032,  0.0581,  0.0639],
+            [-0.1251,  0.0927,  0.1699],
+            [ 0.0060, -0.0633,  0.0005],
+            [ 0.3477,  0.2275,  0.2950],
+            [ 0.1984,  0.0913,  0.1861]
+        ]
+
+    latent_rgb_factors_bias = [-0.1835, -0.0868, -0.3360]
+
+    def __init__(self):
+        self.scale_factor = 1.0
+        self.latents_mean = torch.tensor([
+            -0.7571, -0.7089, -0.9113, 0.1075, -0.1745, 0.9653, -0.1517, 1.5508,
+            0.4134, -0.0715, 0.5517, -0.3632, -0.1922, -0.9497, 0.2503, -0.2921
+        ]).view(1, self.latent_channels, 1, 1, 1)
+        self.latents_std = torch.tensor([
+            2.8184, 1.4541, 2.3275, 2.6558, 1.2196, 1.7708, 2.6052, 2.0743,
+            3.2687, 2.1526, 2.8652, 1.5579, 1.6382, 1.1253, 2.8251, 1.9160
+        ]).view(1, self.latent_channels, 1, 1, 1)
+
+
+        self.taesd_decoder_name = None #TODO
+
+    def process_in(self, latent):
+        latents_mean = self.latents_mean.to(latent.device, latent.dtype)
+        latents_std = self.latents_std.to(latent.device, latent.dtype)
+        return (latent - latents_mean) * self.scale_factor / latents_std
+
+    def process_out(self, latent):
+        latents_mean = self.latents_mean.to(latent.device, latent.dtype)
+        latents_std = self.latents_std.to(latent.device, latent.dtype)
+        return latent * latents_std / self.scale_factor + latents_mean

comfy/ldm/modules/attention.py

@@ -30,38 +30,24 @@ ops = comfy.ops.disable_weight_init
 
 FORCE_UPCAST_ATTENTION_DTYPE = model_management.force_upcast_attention_dtype()
 
-def get_attn_precision(attn_precision):
+def get_attn_precision(attn_precision, current_dtype):
     if args.dont_upcast_attention:
         return None
-    if FORCE_UPCAST_ATTENTION_DTYPE is not None:
-        return FORCE_UPCAST_ATTENTION_DTYPE
+
+    if FORCE_UPCAST_ATTENTION_DTYPE is not None and current_dtype in FORCE_UPCAST_ATTENTION_DTYPE:
+        return FORCE_UPCAST_ATTENTION_DTYPE[current_dtype]
     return attn_precision
 
 def exists(val):
     return val is not None
 
 
-def uniq(arr):
-    return{el: True for el in arr}.keys()
-
-
 def default(val, d):
     if exists(val):
         return val
     return d
 
 
-def max_neg_value(t):
-    return -torch.finfo(t.dtype).max
-
-
-def init_(tensor):
-    dim = tensor.shape[-1]
-    std = 1 / math.sqrt(dim)
-    tensor.uniform_(-std, std)
-    return tensor
-
-
 # feedforward
 class GEGLU(nn.Module):
     def __init__(self, dim_in, dim_out, dtype=None, device=None, operations=ops):
@@ -96,7 +82,7 @@ def Normalize(in_channels, dtype=None, device=None):
     return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
 
 def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
-    attn_precision = get_attn_precision(attn_precision)
+    attn_precision = get_attn_precision(attn_precision, q.dtype)
 
     if skip_reshape:
         b, _, _, dim_head = q.shape
@@ -165,7 +151,7 @@ def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
 
 
 def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
-    attn_precision = get_attn_precision(attn_precision)
+    attn_precision = get_attn_precision(attn_precision, query.dtype)
 
     if skip_reshape:
         b, _, _, dim_head = query.shape
@@ -235,7 +221,7 @@ def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None,
     return hidden_states
 
 def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False):
-    attn_precision = get_attn_precision(attn_precision)
+    attn_precision = get_attn_precision(attn_precision, q.dtype)
 
     if skip_reshape:
         b, _, _, dim_head = q.shape

comfy/ldm/wan/model.py

@@ -0,0 +1,485 @@
+# original version: https://github.com/Wan-Video/Wan2.1/blob/main/wan/modules/model.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+import math
+
+import torch
+import torch.nn as nn
+from einops import repeat
+
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.flux.math import apply_rope
+from comfy.ldm.modules.diffusionmodules.mmdit import RMSNorm
+import comfy.ldm.common_dit
+import comfy.model_management
+
+
+def sinusoidal_embedding_1d(dim, position):
+    # preprocess
+    assert dim % 2 == 0
+    half = dim // 2
+    position = position.type(torch.float32)
+
+    # calculation
+    sinusoid = torch.outer(
+        position, torch.pow(10000, -torch.arange(half).to(position).div(half)))
+    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+    return x
+
+
+class WanSelfAttention(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 eps=1e-6, operation_settings={}):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.eps = eps
+
+        # layers
+        self.q = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.k = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.v = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.o = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.norm_q = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+        self.norm_k = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+
+    def forward(self, x, freqs):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L, num_heads, C / num_heads]
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
+
+        # query, key, value function
+        def qkv_fn(x):
+            q = self.norm_q(self.q(x)).view(b, s, n, d)
+            k = self.norm_k(self.k(x)).view(b, s, n, d)
+            v = self.v(x).view(b, s, n * d)
+            return q, k, v
+
+        q, k, v = qkv_fn(x)
+        q, k = apply_rope(q, k, freqs)
+
+        x = optimized_attention(
+            q.view(b, s, n * d),
+            k.view(b, s, n * d),
+            v,
+            heads=self.num_heads,
+        )
+
+        x = self.o(x)
+        return x
+
+
+class WanT2VCrossAttention(WanSelfAttention):
+
+    def forward(self, x, context):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+        """
+        # compute query, key, value
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(context))
+        v = self.v(context)
+
+        # compute attention
+        x = optimized_attention(q, k, v, heads=self.num_heads)
+
+        x = self.o(x)
+        return x
+
+
+class WanI2VCrossAttention(WanSelfAttention):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 eps=1e-6, operation_settings={}):
+        super().__init__(dim, num_heads, window_size, qk_norm, eps, operation_settings=operation_settings)
+
+        self.k_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.v_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        # self.alpha = nn.Parameter(torch.zeros((1, )))
+        self.norm_k_img = RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+
+    def forward(self, x, context):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+        """
+        context_img = context[:, :257]
+        context = context[:, 257:]
+
+        # compute query, key, value
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(context))
+        v = self.v(context)
+        k_img = self.norm_k_img(self.k_img(context_img))
+        v_img = self.v_img(context_img)
+        img_x = optimized_attention(q, k_img, v_img, heads=self.num_heads)
+        # compute attention
+        x = optimized_attention(q, k, v, heads=self.num_heads)
+
+        # output
+        x = x + img_x
+        x = self.o(x)
+        return x
+
+
+WAN_CROSSATTENTION_CLASSES = {
+    't2v_cross_attn': WanT2VCrossAttention,
+    'i2v_cross_attn': WanI2VCrossAttention,
+}
+
+
+class WanAttentionBlock(nn.Module):
+
+    def __init__(self,
+                 cross_attn_type,
+                 dim,
+                 ffn_dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=False,
+                 eps=1e-6, operation_settings={}):
+        super().__init__()
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # layers
+        self.norm1 = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.self_attn = WanSelfAttention(dim, num_heads, window_size, qk_norm,
+                                          eps, operation_settings=operation_settings)
+        self.norm3 = operation_settings.get("operations").LayerNorm(
+            dim, eps,
+            elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if cross_attn_norm else nn.Identity()
+        self.cross_attn = WAN_CROSSATTENTION_CLASSES[cross_attn_type](dim,
+                                                                      num_heads,
+                                                                      (-1, -1),
+                                                                      qk_norm,
+                                                                      eps, operation_settings=operation_settings)
+        self.norm2 = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.ffn = nn.Sequential(
+            operation_settings.get("operations").Linear(dim, ffn_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.GELU(approximate='tanh'),
+            operation_settings.get("operations").Linear(ffn_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        # modulation
+        self.modulation = nn.Parameter(torch.empty(1, 6, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+    def forward(
+        self,
+        x,
+        e,
+        freqs,
+        context,
+    ):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L, C]
+            e(Tensor): Shape [B, 6, C]
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        # assert e.dtype == torch.float32
+
+        e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        # assert e[0].dtype == torch.float32
+
+        # self-attention
+        y = self.self_attn(
+            self.norm1(x) * (1 + e[1]) + e[0],
+            freqs)
+
+        x = x + y * e[2]
+
+        # cross-attention & ffn function
+        def cross_attn_ffn(x, context, e):
+            x = x + self.cross_attn(self.norm3(x), context)
+            y = self.ffn(self.norm2(x) * (1 + e[4]) + e[3])
+            x = x + y * e[5]
+            return x
+
+        x = cross_attn_ffn(x, context, e)
+        return x
+
+
+class Head(nn.Module):
+
+    def __init__(self, dim, out_dim, patch_size, eps=1e-6, operation_settings={}):
+        super().__init__()
+        self.dim = dim
+        self.out_dim = out_dim
+        self.patch_size = patch_size
+        self.eps = eps
+
+        # layers
+        out_dim = math.prod(patch_size) * out_dim
+        self.norm = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.head = operation_settings.get("operations").Linear(dim, out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        # modulation
+        self.modulation = nn.Parameter(torch.empty(1, 2, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+    def forward(self, x, e):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            e(Tensor): Shape [B, C]
+        """
+        # assert e.dtype == torch.float32
+        e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e.unsqueeze(1)).chunk(2, dim=1)
+        x = (self.head(self.norm(x) * (1 + e[1]) + e[0]))
+        return x
+
+
+class MLPProj(torch.nn.Module):
+
+    def __init__(self, in_dim, out_dim, operation_settings={}):
+        super().__init__()
+
+        self.proj = torch.nn.Sequential(
+            operation_settings.get("operations").LayerNorm(in_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), operation_settings.get("operations").Linear(in_dim, in_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+            torch.nn.GELU(), operation_settings.get("operations").Linear(in_dim, out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+            operation_settings.get("operations").LayerNorm(out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+    def forward(self, image_embeds):
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class WanModel(torch.nn.Module):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='t2v',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+        r"""
+        Initialize the diffusion model backbone.
+
+        Args:
+            model_type (`str`, *optional*, defaults to 't2v'):
+                Model variant - 't2v' (text-to-video) or 'i2v' (image-to-video)
+            patch_size (`tuple`, *optional*, defaults to (1, 2, 2)):
+                3D patch dimensions for video embedding (t_patch, h_patch, w_patch)
+            text_len (`int`, *optional*, defaults to 512):
+                Fixed length for text embeddings
+            in_dim (`int`, *optional*, defaults to 16):
+                Input video channels (C_in)
+            dim (`int`, *optional*, defaults to 2048):
+                Hidden dimension of the transformer
+            ffn_dim (`int`, *optional*, defaults to 8192):
+                Intermediate dimension in feed-forward network
+            freq_dim (`int`, *optional*, defaults to 256):
+                Dimension for sinusoidal time embeddings
+            text_dim (`int`, *optional*, defaults to 4096):
+                Input dimension for text embeddings
+            out_dim (`int`, *optional*, defaults to 16):
+                Output video channels (C_out)
+            num_heads (`int`, *optional*, defaults to 16):
+                Number of attention heads
+            num_layers (`int`, *optional*, defaults to 32):
+                Number of transformer blocks
+            window_size (`tuple`, *optional*, defaults to (-1, -1)):
+                Window size for local attention (-1 indicates global attention)
+            qk_norm (`bool`, *optional*, defaults to True):
+                Enable query/key normalization
+            cross_attn_norm (`bool`, *optional*, defaults to False):
+                Enable cross-attention normalization
+            eps (`float`, *optional*, defaults to 1e-6):
+                Epsilon value for normalization layers
+        """
+
+        super().__init__()
+        self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        assert model_type in ['t2v', 'i2v']
+        self.model_type = model_type
+
+        self.patch_size = patch_size
+        self.text_len = text_len
+        self.in_dim = in_dim
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.freq_dim = freq_dim
+        self.text_dim = text_dim
+        self.out_dim = out_dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # embeddings
+        self.patch_embedding = operations.Conv3d(
+            in_dim, dim, kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), dtype=torch.float32)
+        self.text_embedding = nn.Sequential(
+            operations.Linear(text_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.GELU(approximate='tanh'),
+            operations.Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        self.time_embedding = nn.Sequential(
+            operations.Linear(freq_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.SiLU(), operations.Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+        self.time_projection = nn.Sequential(nn.SiLU(), operations.Linear(dim, dim * 6, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        # blocks
+        cross_attn_type = 't2v_cross_attn' if model_type == 't2v' else 'i2v_cross_attn'
+        self.blocks = nn.ModuleList([
+            WanAttentionBlock(cross_attn_type, dim, ffn_dim, num_heads,
+                              window_size, qk_norm, cross_attn_norm, eps, operation_settings=operation_settings)
+            for _ in range(num_layers)
+        ])
+
+        # head
+        self.head = Head(dim, out_dim, patch_size, eps, operation_settings=operation_settings)
+
+        d = dim // num_heads
+        self.rope_embedder = EmbedND(dim=d, theta=10000.0, axes_dim=[d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)])
+
+        if model_type == 'i2v':
+            self.img_emb = MLPProj(1280, dim, operation_settings=operation_settings)
+        else:
+            self.img_emb = None
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        clip_fea=None,
+        freqs=None,
+    ):
+        r"""
+        Forward pass through the diffusion model
+
+        Args:
+            x (Tensor):
+                List of input video tensors with shape [B, C_in, F, H, W]
+            t (Tensor):
+                Diffusion timesteps tensor of shape [B]
+            context (List[Tensor]):
+                List of text embeddings each with shape [B, L, C]
+            seq_len (`int`):
+                Maximum sequence length for positional encoding
+            clip_fea (Tensor, *optional*):
+                CLIP image features for image-to-video mode
+            y (List[Tensor], *optional*):
+                Conditional video inputs for image-to-video mode, same shape as x
+
+        Returns:
+            List[Tensor]:
+                List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8]
+        """
+        # embeddings
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
+        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
+
+        # context
+        context = self.text_embedding(torch.cat([context, context.new_zeros(context.size(0), self.text_len - context.size(1), context.size(2))], dim=1))
+
+        if clip_fea is not None and self.img_emb is not None:
+            context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+            context = torch.concat([context_clip, context], dim=1)
+
+        # arguments
+        kwargs = dict(
+            e=e0,
+            freqs=freqs,
+            context=context)
+
+        for block in self.blocks:
+            x = block(x, **kwargs)
+
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+        # return [u.float() for u in x]
+
+    def forward(self, x, timestep, context, clip_fea=None, **kwargs):
+        bs, c, t, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
+        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
+
+        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs)[:, :, :t, :h, :w]
+
+    def unpatchify(self, x, grid_sizes):
+        r"""
+        Reconstruct video tensors from patch embeddings.
+
+        Args:
+            x (List[Tensor]):
+                List of patchified features, each with shape [L, C_out * prod(patch_size)]
+            grid_sizes (Tensor):
+                Original spatial-temporal grid dimensions before patching,
+                    shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches)
+
+        Returns:
+            List[Tensor]:
+                Reconstructed video tensors with shape [L, C_out, F, H / 8, W / 8]
+        """
+
+        c = self.out_dim
+        u = x
+        b = u.shape[0]
+        u = u[:, :math.prod(grid_sizes)].view(b, *grid_sizes, *self.patch_size, c)
+        u = torch.einsum('bfhwpqrc->bcfphqwr', u)
+        u = u.reshape(b, c, *[i * j for i, j in zip(grid_sizes, self.patch_size)])
+        return u

comfy/ldm/wan/vae.py

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

comfy/model_base.py

@@ -35,6 +35,7 @@ import comfy.ldm.lightricks.model
 import comfy.ldm.hunyuan_video.model
 import comfy.ldm.cosmos.model
 import comfy.ldm.lumina.model
+import comfy.ldm.wan.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -927,3 +928,47 @@ class Lumina2(BaseModel):
         if cross_attn is not None:
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
+
+class WAN21(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
+        self.image_to_video = image_to_video
+
+    def concat_cond(self, **kwargs):
+        if not self.image_to_video:
+            return None
+
+        image = kwargs.get("concat_latent_image", None)
+        noise = kwargs.get("noise", None)
+        device = kwargs["device"]
+
+        if image is None:
+            image = torch.zeros_like(noise)
+
+        image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+        image = self.process_latent_in(image)
+        image = utils.resize_to_batch_size(image, noise.shape[0])
+
+        mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if mask is None:
+            mask = torch.zeros_like(noise)[:, :4]
+        else:
+            mask = 1.0 - torch.mean(mask, dim=1, keepdim=True)
+            mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            if mask.shape[-3] < noise.shape[-3]:
+                mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0)
+            mask = mask.repeat(1, 4, 1, 1, 1)
+            mask = utils.resize_to_batch_size(mask, noise.shape[0])
+
+        return torch.cat((mask, image), dim=1)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        clip_vision_output = kwargs.get("clip_vision_output", None)
+        if clip_vision_output is not None:
+            out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.penultimate_hidden_states)
+        return out

comfy/model_detection.py

@@ -299,6 +299,27 @@ def detect_unet_config(state_dict, key_prefix):
         dit_config["axes_lens"] = [300, 512, 512]
         return dit_config
 
+    if '{}head.modulation'.format(key_prefix) in state_dict_keys:  # Wan 2.1
+        dit_config = {}
+        dit_config["image_model"] = "wan2.1"
+        dim = state_dict['{}head.modulation'.format(key_prefix)].shape[-1]
+        dit_config["dim"] = dim
+        dit_config["num_heads"] = dim // 128
+        dit_config["ffn_dim"] = state_dict['{}blocks.0.ffn.0.weight'.format(key_prefix)].shape[0]
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')
+        dit_config["patch_size"] = (1, 2, 2)
+        dit_config["freq_dim"] = 256
+        dit_config["window_size"] = (-1, -1)
+        dit_config["qk_norm"] = True
+        dit_config["cross_attn_norm"] = True
+        dit_config["eps"] = 1e-6
+        dit_config["in_dim"] = state_dict['{}patch_embedding.weight'.format(key_prefix)].shape[1]
+        if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "i2v"
+        else:
+            dit_config["model_type"] = "t2v"
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 

comfy/model_management.py

@@ -256,9 +256,12 @@ if ENABLE_PYTORCH_ATTENTION:
     torch.backends.cuda.enable_flash_sdp(True)
     torch.backends.cuda.enable_mem_efficient_sdp(True)
 
+
+PRIORITIZE_FP16 = False  # TODO: remove and replace with something that shows exactly which dtype is faster than the other
 try:
     if is_nvidia() and args.fast:
         torch.backends.cuda.matmul.allow_fp16_accumulation = True
+        PRIORITIZE_FP16 = True  # TODO: limit to cards where it actually boosts performance
 except:
     pass
 
@@ -681,6 +684,10 @@ def unet_dtype(device=None, model_params=0, supported_dtypes=[torch.float16, tor
         if model_params * 2 > free_model_memory:
             return fp8_dtype
 
+    if PRIORITIZE_FP16:
+        if torch.float16 in supported_dtypes and should_use_fp16(device=device, model_params=model_params):
+            return torch.float16
+
     for dt in supported_dtypes:
         if dt == torch.float16 and should_use_fp16(device=device, model_params=model_params):
             if torch.float16 in supported_dtypes:
@@ -947,7 +954,7 @@ def force_upcast_attention_dtype():
         upcast = True
 
     if upcast:
-        return torch.float32
+        return {torch.float16: torch.float32}
     else:
         return None
 

comfy/model_patcher.py

@@ -639,7 +639,7 @@ class ModelPatcher:
                         mem_counter += module_mem
                         load_completely.append((module_mem, n, m, params))
 
-                if cast_weight:
+                if cast_weight and hasattr(m, "comfy_cast_weights"):
                     m.prev_comfy_cast_weights = m.comfy_cast_weights
                     m.comfy_cast_weights = True
 

comfy/sd.py

@@ -12,6 +12,7 @@ from .ldm.audio.autoencoder import AudioOobleckVAE
 import comfy.ldm.genmo.vae.model
 import comfy.ldm.lightricks.vae.causal_video_autoencoder
 import comfy.ldm.cosmos.vae
+import comfy.ldm.wan.vae
 import yaml
 import math
 
@@ -37,6 +38,7 @@ import comfy.text_encoders.lt
 import comfy.text_encoders.hunyuan_video
 import comfy.text_encoders.cosmos
 import comfy.text_encoders.lumina2
+import comfy.text_encoders.wan
 
 import comfy.model_patcher
 import comfy.lora
@@ -392,6 +394,18 @@ class VAE:
                 self.memory_used_decode = lambda shape, dtype: (50 * shape[2] * shape[3] * shape[4] * (8 * 8 * 8)) * model_management.dtype_size(dtype)
                 self.memory_used_encode = lambda shape, dtype: (50 * (round((shape[2] + 7) / 8) * 8) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
                 self.working_dtypes = [torch.bfloat16, torch.float32]
+            elif "decoder.middle.0.residual.0.gamma" in sd:
+                self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
+                self.upscale_index_formula = (4, 8, 8)
+                self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8)
+                self.downscale_index_formula = (4, 8, 8)
+                self.latent_dim = 3
+                self.latent_channels = 16
+                ddconfig = {"dim": 96, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "dropout": 0.0}
+                self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig)
+                self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+                self.memory_used_encode = lambda shape, dtype: 6000 * shape[3] * shape[4] * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: 7000 * shape[3] * shape[4] * (8 * 8) * model_management.dtype_size(dtype)
             else:
                 logging.warning("WARNING: No VAE weights detected, VAE not initalized.")
                 self.first_stage_model = None
@@ -659,6 +673,7 @@ class CLIPType(Enum):
     PIXART = 10
     COSMOS = 11
     LUMINA2 = 12
+    WAN = 13
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -763,6 +778,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             elif clip_type == CLIPType.PIXART:
                 clip_target.clip = comfy.text_encoders.pixart_t5.pixart_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.pixart_t5.PixArtTokenizer
+            elif clip_type == CLIPType.WAN:
+                clip_target.clip = comfy.text_encoders.wan.te(**t5xxl_detect(clip_data))
+                clip_target.tokenizer = comfy.text_encoders.wan.WanT5Tokenizer
+                tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
             else: #CLIPType.MOCHI
                 clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer

comfy/supported_models.py

@@ -16,6 +16,7 @@ import comfy.text_encoders.lt
 import comfy.text_encoders.hunyuan_video
 import comfy.text_encoders.cosmos
 import comfy.text_encoders.lumina2
+import comfy.text_encoders.wan
 
 from . import supported_models_base
 from . import latent_formats
@@ -895,6 +896,49 @@ class Lumina2(supported_models_base.BASE):
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}gemma2_2b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.lumina2.LuminaTokenizer, comfy.text_encoders.lumina2.te(**hunyuan_detect))
 
-models = [Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2]
+class WAN21_T2V(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "t2v",
+    }
+
+    sampling_settings = {
+        "shift": 8.0,
+    }
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Wan21
+
+    memory_usage_factor = 1.0
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.memory_usage_factor = self.unet_config.get("dim", 2000) / 2000
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}umt5xxl.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.wan.WanT5Tokenizer, comfy.text_encoders.wan.te(**t5_detect))
+
+class WAN21_I2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "i2v",
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        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, HunyuanVideo, CosmosT2V, CosmosI2V, Lumina2, WAN21_T2V, WAN21_I2V]
 
 models += [SVD_img2vid]

comfy/text_encoders/lumina2.py

@@ -19,11 +19,6 @@ class LuminaTokenizer(sd1_clip.SD1Tokenizer):
 
 class Gemma2_2BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
-        llama_scaled_fp8 = model_options.get("llama_scaled_fp8", None)
-        if llama_scaled_fp8 is not None:
-            model_options = model_options.copy()
-            model_options["scaled_fp8"] = llama_scaled_fp8
-
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma2_2B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
 
@@ -35,10 +30,10 @@ class LuminaModel(sd1_clip.SD1ClipModel):
 def te(dtype_llama=None, llama_scaled_fp8=None):
     class LuminaTEModel_(LuminaModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
-            if llama_scaled_fp8 is not None and "llama_scaled_fp8" not in model_options:
+            if llama_scaled_fp8 is not None and "scaled_fp8" not in model_options:
                 model_options = model_options.copy()
-                model_options["llama_scaled_fp8"] = llama_scaled_fp8
-                if dtype_llama is not None:
-                    dtype = dtype_llama
+                model_options["scaled_fp8"] = llama_scaled_fp8
+            if dtype_llama is not None:
+                dtype = dtype_llama
             super().__init__(device=device, dtype=dtype, model_options=model_options)
     return LuminaTEModel_

comfy/text_encoders/umt5_config_xxl.json

@@ -0,0 +1,22 @@
+{
+  "d_ff": 10240,
+  "d_kv": 64,
+  "d_model": 4096,
+  "decoder_start_token_id": 0,
+  "dropout_rate": 0.1,
+  "eos_token_id": 1,
+  "dense_act_fn": "gelu_pytorch_tanh",
+  "initializer_factor": 1.0,
+  "is_encoder_decoder": true,
+  "is_gated_act": true,
+  "layer_norm_epsilon": 1e-06,
+  "model_type": "umt5",
+  "num_decoder_layers": 24,
+  "num_heads": 64,
+  "num_layers": 24,
+  "output_past": true,
+  "pad_token_id": 0,
+  "relative_attention_num_buckets": 32,
+  "tie_word_embeddings": false,
+  "vocab_size": 256384
+}

comfy/text_encoders/wan.py

@@ -0,0 +1,37 @@
+from comfy import sd1_clip
+from .spiece_tokenizer import SPieceTokenizer
+import comfy.text_encoders.t5
+import os
+
+class UMT5XXlModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, model_options={}):
+        textmodel_json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "umt5_config_xxl.json")
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"end": 1, "pad": 0}, model_class=comfy.text_encoders.t5.T5, enable_attention_masks=True, zero_out_masked=True, model_options=model_options)
+
+class UMT5XXlTokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = tokenizer_data.get("spiece_model", None)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=4096, embedding_key='umt5xxl', tokenizer_class=SPieceTokenizer, has_start_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=0)
+
+    def state_dict(self):
+        return {"spiece_model": self.tokenizer.serialize_model()}
+
+
+class WanT5Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="umt5xxl", tokenizer=UMT5XXlTokenizer)
+
+class WanT5Model(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, name="umt5xxl", clip_model=UMT5XXlModel, **kwargs)
+
+def te(dtype_t5=None, t5xxl_scaled_fp8=None):
+    class WanTEModel(WanT5Model):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if t5xxl_scaled_fp8 is not None and "scaled_fp8" not in model_options:
+                model_options = model_options.copy()
+                model_options["scaled_fp8"] = t5xxl_scaled_fp8
+            if dtype_t5 is not None:
+                dtype = dtype_t5
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return WanTEModel

comfy_extras/nodes_video.py

@@ -59,6 +59,7 @@ class SaveWEBM:
             frame = av.VideoFrame.from_ndarray(torch.clamp(frame[..., :3] * 255, min=0, max=255).to(device=torch.device("cpu"), dtype=torch.uint8).numpy(), format="rgb24")
             for packet in stream.encode(frame):
                 container.mux(packet)
+        container.mux(stream.encode())
         container.close()
 
         results = [{

comfy_extras/nodes_wan.py

@@ -0,0 +1,54 @@
+import nodes
+import node_helpers
+import torch
+import comfy.model_management
+import comfy.utils
+
+
+class WanImageToVideo:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"positive": ("CONDITIONING", ),
+                             "negative": ("CONDITIONING", ),
+                             "vae": ("VAE", ),
+                             "width": ("INT", {"default": 1280, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "height": ("INT", {"default": 720, "min": 16, "max": nodes.MAX_RESOLUTION, "step": 16}),
+                             "length": ("INT", {"default": 121, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 4}),
+                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096}),
+                },
+                "optional": {"clip_vision_output": ("CLIP_VISION_OUTPUT", ),
+                             "start_image": ("IMAGE", ),
+                }}
+
+    RETURN_TYPES = ("CONDITIONING", "CONDITIONING", "LATENT")
+    RETURN_NAMES = ("positive", "negative", "latent")
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning/video_models"
+
+    def encode(self, positive, negative, vae, width, height, length, batch_size, start_image=None, clip_vision_output=None):
+        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+        if start_image is not None:
+            start_image = comfy.utils.common_upscale(start_image[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            image = torch.ones((length, height, width, start_image.shape[-1]), device=start_image.device, dtype=start_image.dtype) * 0.5
+            image[:start_image.shape[0]] = start_image
+
+            concat_latent_image = vae.encode(image[:, :, :, :3])
+            mask = torch.ones((1, 1, latent.shape[2], concat_latent_image.shape[-2], concat_latent_image.shape[-1]), device=start_image.device, dtype=start_image.dtype)
+            mask[:, :, :((start_image.shape[0] - 1) // 4) + 1] = 0.0
+
+            positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
+            negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": mask})
+
+        if clip_vision_output is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output})
+            negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output})
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return (positive, negative, out_latent)
+
+
+NODE_CLASS_MAPPINGS = {
+    "WanImageToVideo": WanImageToVideo,
+}

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.15"
+__version__ = "0.3.17"

nodes.py

@@ -914,7 +914,7 @@ class CLIPLoader:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": { "clip_name": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2"], ),
+                              "type": (["stable_diffusion", "stable_cascade", "sd3", "stable_audio", "mochi", "ltxv", "pixart", "cosmos", "lumina2", "wan"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),
@@ -924,7 +924,7 @@ class CLIPLoader:
 
     CATEGORY = "advanced/loaders"
 
-    DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 / clip-g / clip-l\nstable_audio: t5\nmochi: t5\ncosmos: old t5 xxl\nlumina2: gemma 2 2B"
+    DESCRIPTION = "[Recipes]\n\nstable_diffusion: clip-l\nstable_cascade: clip-g\nsd3: t5 xxl/ clip-g / clip-l\nstable_audio: t5 base\nmochi: t5 xxl\ncosmos: old t5 xxl\nlumina2: gemma 2 2B\nwan: umt5 xxl"
 
     def load_clip(self, clip_name, type="stable_diffusion", device="default"):
         if type == "stable_cascade":
@@ -943,6 +943,8 @@ class CLIPLoader:
             clip_type = comfy.sd.CLIPType.COSMOS
         elif type == "lumina2":
             clip_type = comfy.sd.CLIPType.LUMINA2
+        elif type == "wan":
+            clip_type = comfy.sd.CLIPType.WAN
         else:
             clip_type = comfy.sd.CLIPType.STABLE_DIFFUSION
 
@@ -2267,6 +2269,7 @@ def init_builtin_extra_nodes():
         "nodes_cosmos.py",
         "nodes_video.py",
         "nodes_lumina2.py",
+        "nodes_wan.py",
     ]
 
     import_failed = []

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.15"
+version = "0.3.17"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"