wan2.2

wan/modules/animate/animate_utils.py

+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+import torch
+import numbers
+from peft import LoraConfig
+
+
+def get_loraconfig(transformer, rank=128, alpha=128, init_lora_weights="gaussian"):
+    target_modules = []
+    for name, module in transformer.named_modules():
+        if "blocks" in name and "face" not in name and "modulation" not in name and isinstance(module, torch.nn.Linear):
+            target_modules.append(name)
+
+    transformer_lora_config = LoraConfig(
+        r=rank,
+        lora_alpha=alpha,
+        init_lora_weights=init_lora_weights,
+        target_modules=target_modules,
+    )
+    return transformer_lora_config
+
+
+
+class TensorList(object):
+
+    def __init__(self, tensors):
+        """
+        tensors: a list of torch.Tensor objects. No need to have uniform shape.
+        """
+        assert isinstance(tensors, (list, tuple))
+        assert all(isinstance(u, torch.Tensor) for u in tensors)
+        assert len(set([u.ndim for u in tensors])) == 1
+        assert len(set([u.dtype for u in tensors])) == 1
+        assert len(set([u.device for u in tensors])) == 1
+        self.tensors = tensors
+    
+    def to(self, *args, **kwargs):
+        return TensorList([u.to(*args, **kwargs) for u in self.tensors])
+    
+    def size(self, dim):
+        assert dim == 0, 'only support get the 0th size'
+        return len(self.tensors)
+    
+    def pow(self, *args, **kwargs):
+        return TensorList([u.pow(*args, **kwargs) for u in self.tensors])
+    
+    def squeeze(self, dim):
+        assert dim != 0
+        if dim > 0:
+            dim -= 1
+        return TensorList([u.squeeze(dim) for u in self.tensors])
+    
+    def type(self, *args, **kwargs):
+        return TensorList([u.type(*args, **kwargs) for u in self.tensors])
+    
+    def type_as(self, other):
+        assert isinstance(other, (torch.Tensor, TensorList))
+        if isinstance(other, torch.Tensor):
+            return TensorList([u.type_as(other) for u in self.tensors])
+        else:
+            return TensorList([u.type(other.dtype) for u in self.tensors])
+    
+    @property
+    def dtype(self):
+        return self.tensors[0].dtype
+    
+    @property
+    def device(self):
+        return self.tensors[0].device
+    
+    @property
+    def ndim(self):
+        return 1 + self.tensors[0].ndim
+    
+    def __getitem__(self, index):
+        return self.tensors[index]
+    
+    def __len__(self):
+        return len(self.tensors)
+    
+    def __add__(self, other):
+        return self._apply(other, lambda u, v: u + v)
+    
+    def __radd__(self, other):
+        return self._apply(other, lambda u, v: v + u)
+    
+    def __sub__(self, other):
+        return self._apply(other, lambda u, v: u - v)
+    
+    def __rsub__(self, other):
+        return self._apply(other, lambda u, v: v - u)
+    
+    def __mul__(self, other):
+        return self._apply(other, lambda u, v: u * v)
+    
+    def __rmul__(self, other):
+        return self._apply(other, lambda u, v: v * u)
+    
+    def __floordiv__(self, other):
+        return self._apply(other, lambda u, v: u // v)
+    
+    def __truediv__(self, other):
+        return self._apply(other, lambda u, v: u / v)
+    
+    def __rfloordiv__(self, other):
+        return self._apply(other, lambda u, v: v // u)
+    
+    def __rtruediv__(self, other):
+        return self._apply(other, lambda u, v: v / u)
+    
+    def __pow__(self, other):
+        return self._apply(other, lambda u, v: u ** v)
+    
+    def __rpow__(self, other):
+        return self._apply(other, lambda u, v: v ** u)
+    
+    def __neg__(self):
+        return TensorList([-u for u in self.tensors])
+    
+    def __iter__(self):
+        for tensor in self.tensors:
+            yield tensor
+    
+    def __repr__(self):
+        return 'TensorList: \n' + repr(self.tensors)
+
+    def _apply(self, other, op):
+        if isinstance(other, (list, tuple, TensorList)) or (
+            isinstance(other, torch.Tensor) and (
+                other.numel() > 1 or other.ndim > 1
+            )
+        ):
+            assert len(other) == len(self.tensors)
+            return TensorList([op(u, v) for u, v in zip(self.tensors, other)])
+        elif isinstance(other, numbers.Number) or (
+            isinstance(other, torch.Tensor) and (
+                other.numel() == 1 and other.ndim <= 1
+            )
+        ):
+            return TensorList([op(u, other) for u in self.tensors])
+        else:
+            raise TypeError(
+                f'unsupported operand for *: "TensorList" and "{type(other)}"'
+            )
\ No newline at end of file

wan/modules/animate/clip.py

+# Modified from ``https://github.com/openai/CLIP'' and ``https://github.com/mlfoundations/open_clip''
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+import logging
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision.transforms as T
+
+from ..attention import flash_attention
+from ..tokenizers import HuggingfaceTokenizer
+from .xlm_roberta import XLMRoberta
+
+__all__ = [
+    'XLMRobertaCLIP',
+    'clip_xlm_roberta_vit_h_14',
+    'CLIPModel',
+]
+
+
+def pos_interpolate(pos, seq_len):
+    if pos.size(1) == seq_len:
+        return pos
+    else:
+        src_grid = int(math.sqrt(pos.size(1)))
+        tar_grid = int(math.sqrt(seq_len))
+        n = pos.size(1) - src_grid * src_grid
+        return torch.cat([
+            pos[:, :n],
+            F.interpolate(
+                pos[:, n:].float().reshape(1, src_grid, src_grid, -1).permute(
+                    0, 3, 1, 2),
+                size=(tar_grid, tar_grid),
+                mode='bicubic',
+                align_corners=False).flatten(2).transpose(1, 2)
+        ],
+                         dim=1)
+
+
+class QuickGELU(nn.Module):
+
+    def forward(self, x):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class LayerNorm(nn.LayerNorm):
+
+    def forward(self, x):
+        return super().forward(x.float()).type_as(x)
+
+
+class SelfAttention(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 causal=False,
+                 attn_dropout=0.0,
+                 proj_dropout=0.0):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.causal = causal
+        self.attn_dropout = attn_dropout
+        self.proj_dropout = proj_dropout
+
+        # layers
+        self.to_qkv = nn.Linear(dim, dim * 3)
+        self.proj = nn.Linear(dim, dim)
+
+    def forward(self, x):
+        """
+        x:   [B, L, C].
+        """
+        b, s, c, n, d = *x.size(), self.num_heads, self.head_dim
+
+        # compute query, key, value
+        q, k, v = self.to_qkv(x).view(b, s, 3, n, d).unbind(2)
+
+        # compute attention
+        p = self.attn_dropout if self.training else 0.0
+        x = flash_attention(q, k, v, dropout_p=p, causal=self.causal, version=2)
+        x = x.reshape(b, s, c)
+
+        # output
+        x = self.proj(x)
+        x = F.dropout(x, self.proj_dropout, self.training)
+        return x
+
+
+class SwiGLU(nn.Module):
+
+    def __init__(self, dim, mid_dim):
+        super().__init__()
+        self.dim = dim
+        self.mid_dim = mid_dim
+
+        # layers
+        self.fc1 = nn.Linear(dim, mid_dim)
+        self.fc2 = nn.Linear(dim, mid_dim)
+        self.fc3 = nn.Linear(mid_dim, dim)
+
+    def forward(self, x):
+        x = F.silu(self.fc1(x)) * self.fc2(x)
+        x = self.fc3(x)
+        return x
+
+
+class AttentionBlock(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 mlp_ratio,
+                 num_heads,
+                 post_norm=False,
+                 causal=False,
+                 activation='quick_gelu',
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 norm_eps=1e-5):
+        assert activation in ['quick_gelu', 'gelu', 'swi_glu']
+        super().__init__()
+        self.dim = dim
+        self.mlp_ratio = mlp_ratio
+        self.num_heads = num_heads
+        self.post_norm = post_norm
+        self.causal = causal
+        self.norm_eps = norm_eps
+
+        # layers
+        self.norm1 = LayerNorm(dim, eps=norm_eps)
+        self.attn = SelfAttention(dim, num_heads, causal, attn_dropout,
+                                  proj_dropout)
+        self.norm2 = LayerNorm(dim, eps=norm_eps)
+        if activation == 'swi_glu':
+            self.mlp = SwiGLU(dim, int(dim * mlp_ratio))
+        else:
+            self.mlp = nn.Sequential(
+                nn.Linear(dim, int(dim * mlp_ratio)),
+                QuickGELU() if activation == 'quick_gelu' else nn.GELU(),
+                nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(proj_dropout))
+
+    def forward(self, x):
+        if self.post_norm:
+            x = x + self.norm1(self.attn(x))
+            x = x + self.norm2(self.mlp(x))
+        else:
+            x = x + self.attn(self.norm1(x))
+            x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class AttentionPool(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 mlp_ratio,
+                 num_heads,
+                 activation='gelu',
+                 proj_dropout=0.0,
+                 norm_eps=1e-5):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.mlp_ratio = mlp_ratio
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.proj_dropout = proj_dropout
+        self.norm_eps = norm_eps
+
+        # layers
+        gain = 1.0 / math.sqrt(dim)
+        self.cls_embedding = nn.Parameter(gain * torch.randn(1, 1, dim))
+        self.to_q = nn.Linear(dim, dim)
+        self.to_kv = nn.Linear(dim, dim * 2)
+        self.proj = nn.Linear(dim, dim)
+        self.norm = LayerNorm(dim, eps=norm_eps)
+        self.mlp = nn.Sequential(
+            nn.Linear(dim, int(dim * mlp_ratio)),
+            QuickGELU() if activation == 'quick_gelu' else nn.GELU(),
+            nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(proj_dropout))
+
+    def forward(self, x):
+        """
+        x:  [B, L, C].
+        """
+        b, s, c, n, d = *x.size(), self.num_heads, self.head_dim
+
+        # compute query, key, value
+        q = self.to_q(self.cls_embedding).view(1, 1, n, d).expand(b, -1, -1, -1)
+        k, v = self.to_kv(x).view(b, s, 2, n, d).unbind(2)
+
+        # compute attention
+        x = flash_attention(q, k, v, version=2)
+        x = x.reshape(b, 1, c)
+
+        # output
+        x = self.proj(x)
+        x = F.dropout(x, self.proj_dropout, self.training)
+
+        # mlp
+        x = x + self.mlp(self.norm(x))
+        return x[:, 0]
+
+
+class VisionTransformer(nn.Module):
+
+    def __init__(self,
+                 image_size=224,
+                 patch_size=16,
+                 dim=768,
+                 mlp_ratio=4,
+                 out_dim=512,
+                 num_heads=12,
+                 num_layers=12,
+                 pool_type='token',
+                 pre_norm=True,
+                 post_norm=False,
+                 activation='quick_gelu',
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 embedding_dropout=0.0,
+                 norm_eps=1e-5):
+        if image_size % patch_size != 0:
+            print(
+                '[WARNING] image_size is not divisible by patch_size',
+                flush=True)
+        assert pool_type in ('token', 'token_fc', 'attn_pool')
+        out_dim = out_dim or dim
+        super().__init__()
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_patches = (image_size // patch_size)**2
+        self.dim = dim
+        self.mlp_ratio = mlp_ratio
+        self.out_dim = out_dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.pool_type = pool_type
+        self.post_norm = post_norm
+        self.norm_eps = norm_eps
+
+        # embeddings
+        gain = 1.0 / math.sqrt(dim)
+        self.patch_embedding = nn.Conv2d(
+            3,
+            dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=not pre_norm)
+        if pool_type in ('token', 'token_fc'):
+            self.cls_embedding = nn.Parameter(gain * torch.randn(1, 1, dim))
+        self.pos_embedding = nn.Parameter(gain * torch.randn(
+            1, self.num_patches +
+            (1 if pool_type in ('token', 'token_fc') else 0), dim))
+        self.dropout = nn.Dropout(embedding_dropout)
+
+        # transformer
+        self.pre_norm = LayerNorm(dim, eps=norm_eps) if pre_norm else None
+        self.transformer = nn.Sequential(*[
+            AttentionBlock(dim, mlp_ratio, num_heads, post_norm, False,
+                           activation, attn_dropout, proj_dropout, norm_eps)
+            for _ in range(num_layers)
+        ])
+        self.post_norm = LayerNorm(dim, eps=norm_eps)
+
+        # head
+        if pool_type == 'token':
+            self.head = nn.Parameter(gain * torch.randn(dim, out_dim))
+        elif pool_type == 'token_fc':
+            self.head = nn.Linear(dim, out_dim)
+        elif pool_type == 'attn_pool':
+            self.head = AttentionPool(dim, mlp_ratio, num_heads, activation,
+                                      proj_dropout, norm_eps)
+
+    def forward(self, x, interpolation=False, use_31_block=False):
+        b = x.size(0)
+
+        # embeddings
+        x = self.patch_embedding(x).flatten(2).permute(0, 2, 1)
+        if self.pool_type in ('token', 'token_fc'):
+            x = torch.cat([self.cls_embedding.expand(b, -1, -1), x], dim=1)
+        if interpolation:
+            e = pos_interpolate(self.pos_embedding, x.size(1))
+        else:
+            e = self.pos_embedding
+        x = self.dropout(x + e)
+        if self.pre_norm is not None:
+            x = self.pre_norm(x)
+
+        # transformer
+        if use_31_block:
+            x = self.transformer[:-1](x)
+            return x
+        else:
+            x = self.transformer(x)
+            return x
+
+
+class XLMRobertaWithHead(XLMRoberta):
+
+    def __init__(self, **kwargs):
+        self.out_dim = kwargs.pop('out_dim')
+        super().__init__(**kwargs)
+
+        # head
+        mid_dim = (self.dim + self.out_dim) // 2
+        self.head = nn.Sequential(
+            nn.Linear(self.dim, mid_dim, bias=False), nn.GELU(),
+            nn.Linear(mid_dim, self.out_dim, bias=False))
+
+    def forward(self, ids):
+        # xlm-roberta
+        x = super().forward(ids)
+
+        # average pooling
+        mask = ids.ne(self.pad_id).unsqueeze(-1).to(x)
+        x = (x * mask).sum(dim=1) / mask.sum(dim=1)
+
+        # head
+        x = self.head(x)
+        return x
+
+
+class XLMRobertaCLIP(nn.Module):
+
+    def __init__(self,
+                 embed_dim=1024,
+                 image_size=224,
+                 patch_size=14,
+                 vision_dim=1280,
+                 vision_mlp_ratio=4,
+                 vision_heads=16,
+                 vision_layers=32,
+                 vision_pool='token',
+                 vision_pre_norm=True,
+                 vision_post_norm=False,
+                 activation='gelu',
+                 vocab_size=250002,
+                 max_text_len=514,
+                 type_size=1,
+                 pad_id=1,
+                 text_dim=1024,
+                 text_heads=16,
+                 text_layers=24,
+                 text_post_norm=True,
+                 text_dropout=0.1,
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 embedding_dropout=0.0,
+                 norm_eps=1e-5):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.vision_dim = vision_dim
+        self.vision_mlp_ratio = vision_mlp_ratio
+        self.vision_heads = vision_heads
+        self.vision_layers = vision_layers
+        self.vision_pre_norm = vision_pre_norm
+        self.vision_post_norm = vision_post_norm
+        self.activation = activation
+        self.vocab_size = vocab_size
+        self.max_text_len = max_text_len
+        self.type_size = type_size
+        self.pad_id = pad_id
+        self.text_dim = text_dim
+        self.text_heads = text_heads
+        self.text_layers = text_layers
+        self.text_post_norm = text_post_norm
+        self.norm_eps = norm_eps
+
+        # models
+        self.visual = VisionTransformer(
+            image_size=image_size,
+            patch_size=patch_size,
+            dim=vision_dim,
+            mlp_ratio=vision_mlp_ratio,
+            out_dim=embed_dim,
+            num_heads=vision_heads,
+            num_layers=vision_layers,
+            pool_type=vision_pool,
+            pre_norm=vision_pre_norm,
+            post_norm=vision_post_norm,
+            activation=activation,
+            attn_dropout=attn_dropout,
+            proj_dropout=proj_dropout,
+            embedding_dropout=embedding_dropout,
+            norm_eps=norm_eps)
+        self.textual = XLMRobertaWithHead(
+            vocab_size=vocab_size,
+            max_seq_len=max_text_len,
+            type_size=type_size,
+            pad_id=pad_id,
+            dim=text_dim,
+            out_dim=embed_dim,
+            num_heads=text_heads,
+            num_layers=text_layers,
+            post_norm=text_post_norm,
+            dropout=text_dropout)
+        self.log_scale = nn.Parameter(math.log(1 / 0.07) * torch.ones([]))
+
+    def forward(self, imgs, txt_ids):
+        """
+        imgs:       [B, 3, H, W] of torch.float32.
+        - mean:     [0.48145466, 0.4578275, 0.40821073]
+        - std:      [0.26862954, 0.26130258, 0.27577711]
+        txt_ids:    [B, L] of torch.long.
+                    Encoded by data.CLIPTokenizer.
+        """
+        xi = self.visual(imgs)
+        xt = self.textual(txt_ids)
+        return xi, xt
+
+    def param_groups(self):
+        groups = [{
+            'params': [
+                p for n, p in self.named_parameters()
+                if 'norm' in n or n.endswith('bias')
+            ],
+            'weight_decay': 0.0
+        }, {
+            'params': [
+                p for n, p in self.named_parameters()
+                if not ('norm' in n or n.endswith('bias'))
+            ]
+        }]
+        return groups
+
+
+def _clip(pretrained=False,
+          pretrained_name=None,
+          model_cls=XLMRobertaCLIP,
+          return_transforms=False,
+          return_tokenizer=False,
+          tokenizer_padding='eos',
+          dtype=torch.float32,
+          device='cpu',
+          **kwargs):
+    # init a model on device
+    with torch.device(device):
+        model = model_cls(**kwargs)
+
+    # set device
+    model = model.to(dtype=dtype, device=device)
+    output = (model,)
+
+    # init transforms
+    if return_transforms:
+        # mean and std
+        if 'siglip' in pretrained_name.lower():
+            mean, std = [0.5, 0.5, 0.5], [0.5, 0.5, 0.5]
+        else:
+            mean = [0.48145466, 0.4578275, 0.40821073]
+            std = [0.26862954, 0.26130258, 0.27577711]
+
+        # transforms
+        transforms = T.Compose([
+            T.Resize((model.image_size, model.image_size),
+                     interpolation=T.InterpolationMode.BICUBIC),
+            T.ToTensor(),
+            T.Normalize(mean=mean, std=std)
+        ])
+        output += (transforms,)
+    return output[0] if len(output) == 1 else output
+
+
+def clip_xlm_roberta_vit_h_14(
+        pretrained=False,
+        pretrained_name='open-clip-xlm-roberta-large-vit-huge-14',
+        **kwargs):
+    cfg = dict(
+        embed_dim=1024,
+        image_size=224,
+        patch_size=14,
+        vision_dim=1280,
+        vision_mlp_ratio=4,
+        vision_heads=16,
+        vision_layers=32,
+        vision_pool='token',
+        activation='gelu',
+        vocab_size=250002,
+        max_text_len=514,
+        type_size=1,
+        pad_id=1,
+        text_dim=1024,
+        text_heads=16,
+        text_layers=24,
+        text_post_norm=True,
+        text_dropout=0.1,
+        attn_dropout=0.0,
+        proj_dropout=0.0,
+        embedding_dropout=0.0)
+    cfg.update(**kwargs)
+    return _clip(pretrained, pretrained_name, XLMRobertaCLIP, **cfg)
+
+
+class CLIPModel:
+
+    def __init__(self, dtype, device, checkpoint_path, tokenizer_path):
+        self.dtype = dtype
+        self.device = device
+        self.checkpoint_path = checkpoint_path
+        self.tokenizer_path = tokenizer_path
+
+        # init model
+        self.model, self.transforms = clip_xlm_roberta_vit_h_14(
+            pretrained=False,
+            return_transforms=True,
+            return_tokenizer=False,
+            dtype=dtype,
+            device=device)
+        self.model = self.model.eval().requires_grad_(False)
+        logging.info(f'loading {checkpoint_path}')
+        self.model.load_state_dict(
+            torch.load(checkpoint_path, map_location='cpu'))
+
+        # init tokenizer
+        self.tokenizer = HuggingfaceTokenizer(
+            name=tokenizer_path,
+            seq_len=self.model.max_text_len - 2,
+            clean='whitespace')
+
+    def visual(self, videos):
+        # preprocess
+        size = (self.model.image_size,) * 2
+        videos = torch.cat([
+            F.interpolate(
+                u.transpose(0, 1),
+                size=size,
+                mode='bicubic',
+                align_corners=False) for u in videos
+        ])
+        videos = self.transforms.transforms[-1](videos.mul_(0.5).add_(0.5))
+
+        # forward
+        with torch.cuda.amp.autocast(dtype=self.dtype):
+            out = self.model.visual(videos, use_31_block=True)
+            return out
\ No newline at end of file

wan/modules/animate/face_blocks.py

+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+from torch import nn
+import torch
+from typing import Tuple, Optional
+from einops import rearrange
+import torch.nn.functional as F
+import math
+from ...distributed.util import gather_forward, get_rank, get_world_size
+
+
+try:
+    from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
+except ImportError:
+    flash_attn_func = None
+
+MEMORY_LAYOUT = {
+    "flash": (
+        lambda x: x.view(x.shape[0] * x.shape[1], *x.shape[2:]),
+        lambda x: x,
+    ),
+    "torch": (
+        lambda x: x.transpose(1, 2),
+        lambda x: x.transpose(1, 2),
+    ),
+    "vanilla": (
+        lambda x: x.transpose(1, 2),
+        lambda x: x.transpose(1, 2),
+    ),
+}
+
+
+def attention(
+    q,
+    k,
+    v,
+    mode="flash",
+    drop_rate=0,
+    attn_mask=None,
+    causal=False,
+    max_seqlen_q=None,
+    batch_size=1,
+):
+    """
+    Perform QKV self attention.
+
+    Args:
+        q (torch.Tensor): Query tensor with shape [b, s, a, d], where a is the number of heads.
+        k (torch.Tensor): Key tensor with shape [b, s1, a, d]
+        v (torch.Tensor): Value tensor with shape [b, s1, a, d]
+        mode (str): Attention mode. Choose from 'self_flash', 'cross_flash', 'torch', and 'vanilla'.
+        drop_rate (float): Dropout rate in attention map. (default: 0)
+        attn_mask (torch.Tensor): Attention mask with shape [b, s1] (cross_attn), or [b, a, s, s1] (torch or vanilla).
+            (default: None)
+        causal (bool): Whether to use causal attention. (default: False)
+        cu_seqlens_q (torch.Tensor): dtype torch.int32. The cumulative sequence lengths of the sequences in the batch,
+            used to index into q.
+        cu_seqlens_kv (torch.Tensor): dtype torch.int32. The cumulative sequence lengths of the sequences in the batch,
+            used to index into kv.
+        max_seqlen_q (int): The maximum sequence length in the batch of q.
+        max_seqlen_kv (int): The maximum sequence length in the batch of k and v.
+
+    Returns:
+        torch.Tensor: Output tensor after self attention with shape [b, s, ad]
+    """
+    pre_attn_layout, post_attn_layout = MEMORY_LAYOUT[mode]
+
+    if mode == "torch":
+        if attn_mask is not None and attn_mask.dtype != torch.bool:
+            attn_mask = attn_mask.to(q.dtype)
+        x = F.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask, dropout_p=drop_rate, is_causal=causal)
+
+    elif mode == "flash":
+        x = flash_attn_func(
+            q,
+            k,
+            v,
+        )
+        x = x.view(batch_size, max_seqlen_q, x.shape[-2], x.shape[-1])  # reshape x to [b, s, a, d]
+    elif mode == "vanilla":
+        scale_factor = 1 / math.sqrt(q.size(-1))
+
+        b, a, s, _ = q.shape
+        s1 = k.size(2)
+        attn_bias = torch.zeros(b, a, s, s1, dtype=q.dtype, device=q.device)
+        if causal:
+            # Only applied to self attention
+            assert attn_mask is None, "Causal mask and attn_mask cannot be used together"
+            temp_mask = torch.ones(b, a, s, s, dtype=torch.bool, device=q.device).tril(diagonal=0)
+            attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
+            attn_bias.to(q.dtype)
+
+        if attn_mask is not None:
+            if attn_mask.dtype == torch.bool:
+                attn_bias.masked_fill_(attn_mask.logical_not(), float("-inf"))
+            else:
+                attn_bias += attn_mask
+
+        attn = (q @ k.transpose(-2, -1)) * scale_factor
+        attn += attn_bias
+        attn = attn.softmax(dim=-1)
+        attn = torch.dropout(attn, p=drop_rate, train=True)
+        x = attn @ v
+    else:
+        raise NotImplementedError(f"Unsupported attention mode: {mode}")
+
+    x = post_attn_layout(x)
+    b, s, a, d = x.shape
+    out = x.reshape(b, s, -1)
+    return out
+
+
+class CausalConv1d(nn.Module):
+
+    def __init__(self, chan_in, chan_out, kernel_size=3, stride=1, dilation=1, pad_mode="replicate", **kwargs):
+        super().__init__()
+
+        self.pad_mode = pad_mode
+        padding = (kernel_size - 1, 0)  # T
+        self.time_causal_padding = padding
+
+        self.conv = nn.Conv1d(chan_in, chan_out, kernel_size, stride=stride, dilation=dilation, **kwargs)
+
+    def forward(self, x):
+        x = F.pad(x, self.time_causal_padding, mode=self.pad_mode)
+        return self.conv(x)
+
+
+
+class FaceEncoder(nn.Module):
+    def __init__(self, in_dim: int, hidden_dim: int, num_heads=int, dtype=None, device=None):
+        factory_kwargs = {"dtype": dtype, "device": device}
+        super().__init__()
+
+        self.num_heads = num_heads
+        self.conv1_local = CausalConv1d(in_dim, 1024 * num_heads, 3, stride=1)
+        self.norm1 = nn.LayerNorm(hidden_dim // 8, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+        self.act = nn.SiLU()
+        self.conv2 = CausalConv1d(1024, 1024, 3, stride=2)
+        self.conv3 = CausalConv1d(1024, 1024, 3, stride=2)
+
+        self.out_proj = nn.Linear(1024, hidden_dim)
+        self.norm1 = nn.LayerNorm(1024, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.norm2 = nn.LayerNorm(1024, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.norm3 = nn.LayerNorm(1024, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.padding_tokens = nn.Parameter(torch.zeros(1, 1, 1, hidden_dim))
+
+    def forward(self, x):
+        
+        x = rearrange(x, "b t c -> b c t")
+        b, c, t = x.shape
+
+        x = self.conv1_local(x)
+        x = rearrange(x, "b (n c) t -> (b n) t c", n=self.num_heads)
+        
+        x = self.norm1(x)
+        x = self.act(x)
+        x = rearrange(x, "b t c -> b c t")
+        x = self.conv2(x)
+        x = rearrange(x, "b c t -> b t c")
+        x = self.norm2(x)
+        x = self.act(x)
+        x = rearrange(x, "b t c -> b c t")
+        x = self.conv3(x)
+        x = rearrange(x, "b c t -> b t c")
+        x = self.norm3(x)
+        x = self.act(x)
+        x = self.out_proj(x)
+        x = rearrange(x, "(b n) t c -> b t n c", b=b)
+        padding = self.padding_tokens.repeat(b, x.shape[1], 1, 1)
+        x = torch.cat([x, padding], dim=-2)
+        x_local = x.clone()
+
+        return x_local
+
+
+
+class RMSNorm(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        elementwise_affine=True,
+        eps: float = 1e-6,
+        device=None,
+        dtype=None,
+    ):
+        """
+        Initialize the RMSNorm normalization layer.
+
+        Args:
+            dim (int): The dimension of the input tensor.
+            eps (float, optional): A small value added to the denominator for numerical stability. Default is 1e-6.
+
+        Attributes:
+            eps (float): A small value added to the denominator for numerical stability.
+            weight (nn.Parameter): Learnable scaling parameter.
+
+        """
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        self.eps = eps
+        if elementwise_affine:
+            self.weight = nn.Parameter(torch.ones(dim, **factory_kwargs))
+
+    def _norm(self, x):
+        """
+        Apply the RMSNorm normalization to the input tensor.
+
+        Args:
+            x (torch.Tensor): The input tensor.
+
+        Returns:
+            torch.Tensor: The normalized tensor.
+
+        """
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        """
+        Forward pass through the RMSNorm layer.
+
+        Args:
+            x (torch.Tensor): The input tensor.
+
+        Returns:
+            torch.Tensor: The output tensor after applying RMSNorm.
+
+        """
+        output = self._norm(x.float()).type_as(x)
+        if hasattr(self, "weight"):
+            output = output * self.weight
+        return output
+
+
+def get_norm_layer(norm_layer):
+    """
+    Get the normalization layer.
+
+    Args:
+        norm_layer (str): The type of normalization layer.
+
+    Returns:
+        norm_layer (nn.Module): The normalization layer.
+    """
+    if norm_layer == "layer":
+        return nn.LayerNorm
+    elif norm_layer == "rms":
+        return RMSNorm
+    else:
+        raise NotImplementedError(f"Norm layer {norm_layer} is not implemented")
+
+
+class FaceAdapter(nn.Module):
+    def __init__(
+        self,
+        hidden_dim: int,
+        heads_num: int,
+        qk_norm: bool = True,
+        qk_norm_type: str = "rms",
+        num_adapter_layers: int = 1,
+        dtype=None,
+        device=None,
+    ):
+
+        factory_kwargs = {"dtype": dtype, "device": device}
+        super().__init__()
+        self.hidden_size = hidden_dim
+        self.heads_num = heads_num
+        self.fuser_blocks = nn.ModuleList(
+            [
+                FaceBlock(
+                    self.hidden_size,
+                    self.heads_num,
+                    qk_norm=qk_norm,
+                    qk_norm_type=qk_norm_type,
+                    **factory_kwargs,
+                )
+                for _ in range(num_adapter_layers)
+            ]
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        motion_embed: torch.Tensor,
+        idx: int,
+        freqs_cis_q: Tuple[torch.Tensor, torch.Tensor] = None,
+        freqs_cis_k: Tuple[torch.Tensor, torch.Tensor] = None,
+    ) -> torch.Tensor:
+
+        return self.fuser_blocks[idx](x, motion_embed, freqs_cis_q, freqs_cis_k)
+
+
+
+class FaceBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        heads_num: int,
+        qk_norm: bool = True,
+        qk_norm_type: str = "rms",
+        qk_scale: float = None,
+        dtype: Optional[torch.dtype] = None,
+        device: Optional[torch.device] = None,
+    ):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+
+        self.deterministic = False
+        self.hidden_size = hidden_size
+        self.heads_num = heads_num
+        head_dim = hidden_size // heads_num
+        self.scale = qk_scale or head_dim**-0.5
+       
+        self.linear1_kv = nn.Linear(hidden_size, hidden_size * 2, **factory_kwargs)
+        self.linear1_q = nn.Linear(hidden_size, hidden_size, **factory_kwargs)
+
+        self.linear2 = nn.Linear(hidden_size, hidden_size, **factory_kwargs)
+
+        qk_norm_layer = get_norm_layer(qk_norm_type)
+        self.q_norm = (
+            qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, **factory_kwargs) if qk_norm else nn.Identity()
+        )
+        self.k_norm = (
+            qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, **factory_kwargs) if qk_norm else nn.Identity()
+        )
+
+        self.pre_norm_feat = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.pre_norm_motion = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        motion_vec: torch.Tensor,
+        motion_mask: Optional[torch.Tensor] = None,
+        use_context_parallel=False,
+    ) -> torch.Tensor:
+        
+        B, T, N, C = motion_vec.shape
+        T_comp = T
+
+        x_motion = self.pre_norm_motion(motion_vec)
+        x_feat = self.pre_norm_feat(x)
+
+        kv = self.linear1_kv(x_motion)
+        q = self.linear1_q(x_feat)
+
+        k, v = rearrange(kv, "B L N (K H D) -> K B L N H D", K=2, H=self.heads_num)
+        q = rearrange(q, "B S (H D) -> B S H D", H=self.heads_num)
+
+        # Apply QK-Norm if needed.
+        q = self.q_norm(q).to(v)
+        k = self.k_norm(k).to(v)
+
+        k = rearrange(k, "B L N H D -> (B L) N H D")  
+        v = rearrange(v, "B L N H D -> (B L) N H D") 
+
+        if use_context_parallel:
+            q = gather_forward(q, dim=1)
+
+        q = rearrange(q, "B (L S) H D -> (B L) S H D", L=T_comp)  
+        # Compute attention.
+        attn = attention(
+            q,
+            k,
+            v,
+            max_seqlen_q=q.shape[1],
+            batch_size=q.shape[0],
+        )
+
+        attn = rearrange(attn, "(B L) S C -> B (L S) C", L=T_comp)
+        if use_context_parallel:
+            attn = torch.chunk(attn, get_world_size(), dim=1)[get_rank()]
+
+        output = self.linear2(attn)
+
+        if motion_mask is not None:
+            output = output * rearrange(motion_mask, "B T H W -> B (T H W)").unsqueeze(-1)
+
+        return output
\ No newline at end of file

wan/modules/animate/model_animate.py

+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+import math
+import types
+from copy import deepcopy
+from einops import  rearrange
+from typing import List
+import numpy as np
+import torch
+import torch.cuda.amp as amp
+import torch.nn as nn
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.loaders import PeftAdapterMixin
+
+from ...distributed.sequence_parallel import (
+    distributed_attention,
+    gather_forward,
+    get_rank,
+    get_world_size,
+)
+
+
+from ..model import (
+    Head,
+    WanAttentionBlock,
+    WanLayerNorm,
+    WanRMSNorm,
+    WanModel,
+    WanSelfAttention,
+    flash_attention,
+    rope_params,
+    sinusoidal_embedding_1d,
+    rope_apply
+)
+
+from .face_blocks import FaceEncoder, FaceAdapter
+from .motion_encoder import Generator
+
+class HeadAnimate(Head):
+
+    def forward(self, x, e):
+        """
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            e(Tensor): Shape [B, L1, C]
+        """
+        assert e.dtype == torch.float32
+        with amp.autocast(dtype=torch.float32):
+            e = (self.modulation + e.unsqueeze(1)).chunk(2, dim=1)
+            x = (self.head(self.norm(x) * (1 + e[1]) + e[0]))
+        return x
+
+
+class WanAnimateSelfAttention(WanSelfAttention):
+
+    def forward(self, x, seq_lens, grid_sizes, freqs):
+        """
+        Args:
+            x(Tensor): Shape [B, L, num_heads, C / num_heads]
+            seq_lens(Tensor): Shape [B]
+            grid_sizes(Tensor): Shape [B, 3], the second dimension contains (F, H, W)
+            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)
+
+        x = flash_attention(
+            q=rope_apply(q, grid_sizes, freqs),
+            k=rope_apply(k, grid_sizes, freqs),
+            v=v,
+            k_lens=seq_lens,
+            window_size=self.window_size)
+
+        # output
+        x = x.flatten(2)
+        x = self.o(x)
+        return x
+
+
+class WanAnimateCrossAttention(WanSelfAttention):
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        window_size=(-1, -1),
+        qk_norm=True,
+        eps=1e-6,
+        use_img_emb=True
+    ):
+        super().__init__(
+            dim,
+            num_heads,
+            window_size,
+            qk_norm,
+            eps
+        )
+        self.use_img_emb = use_img_emb
+
+        if use_img_emb:
+            self.k_img = nn.Linear(dim, dim)
+            self.v_img = nn.Linear(dim, dim)
+            self.norm_k_img = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
+
+    def forward(self, x, context, context_lens):
+        """
+        x:              [B, L1, C].
+        context:        [B, L2, C].
+        context_lens:   [B].
+        """
+        if self.use_img_emb:
+            context_img = context[:, :257]
+            context = context[:, 257:]
+        else:
+            context = context
+
+        b, n, d = x.size(0), self.num_heads, self.head_dim
+
+        # compute query, key, value
+        q = self.norm_q(self.q(x)).view(b, -1, n, d)
+        k = self.norm_k(self.k(context)).view(b, -1, n, d)
+        v = self.v(context).view(b, -1, n, d)
+
+        if self.use_img_emb:
+            k_img = self.norm_k_img(self.k_img(context_img)).view(b, -1, n, d)
+            v_img = self.v_img(context_img).view(b, -1, n, d)
+            img_x = flash_attention(q, k_img, v_img, k_lens=None)
+        # compute attention
+        x = flash_attention(q, k, v, k_lens=context_lens)
+
+        # output
+        x = x.flatten(2)
+
+        if self.use_img_emb:
+            img_x = img_x.flatten(2)
+            x = x + img_x
+
+        x = self.o(x)
+        return x
+
+
+class WanAnimateAttentionBlock(nn.Module):
+    def __init__(self,
+                 dim,
+                 ffn_dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 use_img_emb=True):
+
+        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 = WanLayerNorm(dim, eps)
+        self.self_attn = WanAnimateSelfAttention(dim, num_heads, window_size, qk_norm, eps)
+        
+        self.norm3 = WanLayerNorm(
+            dim, eps, elementwise_affine=True
+        ) if cross_attn_norm else nn.Identity()
+            
+        self.cross_attn = WanAnimateCrossAttention(dim, num_heads, (-1, -1), qk_norm, eps, use_img_emb=use_img_emb)
+        self.norm2 = WanLayerNorm(dim, eps)
+        self.ffn = nn.Sequential(
+            nn.Linear(dim, ffn_dim),
+            nn.GELU(approximate='tanh'),
+            nn.Linear(ffn_dim, dim)
+        )
+
+        # modulation
+        self.modulation = nn.Parameter(torch.randn(1, 6, dim) / dim ** 0.5)
+
+    def forward(
+        self,
+        x,
+        e,
+        seq_lens,
+        grid_sizes,
+        freqs,
+        context,
+        context_lens,
+    ):
+        """
+        Args:
+            x(Tensor): Shape [B, L, C]
+            e(Tensor): Shape [B, L1, 6, C]
+            seq_lens(Tensor): Shape [B], length of each sequence in batch
+            grid_sizes(Tensor): Shape [B, 3], the second dimension contains (F, H, W)
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        assert e.dtype == torch.float32
+        with amp.autocast(dtype=torch.float32):
+            e = (self.modulation + e).chunk(6, dim=1)
+        assert e[0].dtype == torch.float32
+
+        # self-attention
+        y = self.self_attn(
+            self.norm1(x).float() * (1 + e[1]) + e[0], seq_lens, grid_sizes, freqs
+        )
+        with amp.autocast(dtype=torch.float32):
+            x = x + y * e[2]
+
+        # cross-attention & ffn function
+        def cross_attn_ffn(x, context, context_lens, e):
+            x = x + self.cross_attn(self.norm3(x), context, context_lens)
+            y = self.ffn(self.norm2(x).float() * (1 + e[4]) + e[3])
+            with amp.autocast(dtype=torch.float32):
+                x = x + y * e[5]
+            return x
+
+        x = cross_attn_ffn(x, context, context_lens, e)
+        return x
+
+
+class MLPProj(torch.nn.Module):
+    def __init__(self, in_dim, out_dim):
+        super().__init__()
+
+        self.proj = torch.nn.Sequential(
+            torch.nn.LayerNorm(in_dim),
+            torch.nn.Linear(in_dim, in_dim),
+            torch.nn.GELU(),
+            torch.nn.Linear(in_dim, out_dim),
+            torch.nn.LayerNorm(out_dim),
+        )
+
+    def forward(self, image_embeds):
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+class WanAnimateModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
+    _no_split_modules = ['WanAttentionBlock']
+
+    @register_to_config
+    def __init__(self,
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=36,
+                 dim=5120,
+                 ffn_dim=13824,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=40,
+                 num_layers=40,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 motion_encoder_dim=512,
+                 use_context_parallel=False,
+                 use_img_emb=True):
+
+        super().__init__()
+        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
+        self.motion_encoder_dim = motion_encoder_dim
+        self.use_context_parallel = use_context_parallel
+        self.use_img_emb = use_img_emb
+
+        # embeddings
+        self.patch_embedding = nn.Conv3d(
+            in_dim, dim, kernel_size=patch_size, stride=patch_size)
+
+        self.pose_patch_embedding = nn.Conv3d(
+            16, dim, kernel_size=patch_size, stride=patch_size
+        )
+
+        self.text_embedding = nn.Sequential(
+            nn.Linear(text_dim, dim), nn.GELU(approximate='tanh'),
+            nn.Linear(dim, dim))
+
+        self.time_embedding = nn.Sequential(
+            nn.Linear(freq_dim, dim), nn.SiLU(), nn.Linear(dim, dim))
+        self.time_projection = nn.Sequential(nn.SiLU(), nn.Linear(dim, dim * 6))
+
+        # blocks
+        self.blocks = nn.ModuleList([
+            WanAnimateAttentionBlock(dim, ffn_dim, num_heads, window_size, qk_norm,
+                              cross_attn_norm, eps, use_img_emb) for _ in range(num_layers)
+        ])
+
+        # head
+        self.head = HeadAnimate(dim, out_dim, patch_size, eps)
+
+        # buffers (don't use register_buffer otherwise dtype will be changed in to())
+        assert (dim % num_heads) == 0 and (dim // num_heads) % 2 == 0
+        d = dim // num_heads
+        self.freqs = torch.cat([
+            rope_params(1024, d - 4 * (d // 6)),
+            rope_params(1024, 2 * (d // 6)),
+            rope_params(1024, 2 * (d // 6))
+        ], dim=1)
+
+        self.img_emb = MLPProj(1280, dim)
+        
+        # initialize weights
+        self.init_weights()
+
+        self.motion_encoder = Generator(size=512, style_dim=512, motion_dim=20)
+        self.face_adapter = FaceAdapter(
+            heads_num=self.num_heads,
+            hidden_dim=self.dim,
+            num_adapter_layers=self.num_layers // 5,
+        )
+
+        self.face_encoder = FaceEncoder(
+            in_dim=motion_encoder_dim,
+            hidden_dim=self.dim,
+            num_heads=4,
+        )
+
+    def after_patch_embedding(self, x: List[torch.Tensor], pose_latents, face_pixel_values):
+        pose_latents = [self.pose_patch_embedding(u.unsqueeze(0)) for u in pose_latents]
+        for x_, pose_latents_ in zip(x, pose_latents):
+            x_[:, :, 1:] += pose_latents_
+        
+        b,c,T,h,w = face_pixel_values.shape
+        face_pixel_values = rearrange(face_pixel_values, "b c t h w -> (b t) c h w")
+
+        encode_bs = 8
+        face_pixel_values_tmp = []
+        for i in range(math.ceil(face_pixel_values.shape[0]/encode_bs)):
+            face_pixel_values_tmp.append(self.motion_encoder.get_motion(face_pixel_values[i*encode_bs:(i+1)*encode_bs]))
+
+        motion_vec = torch.cat(face_pixel_values_tmp)
+        
+        motion_vec = rearrange(motion_vec, "(b t) c -> b t c", t=T)
+        motion_vec = self.face_encoder(motion_vec)
+
+        B, L, H, C = motion_vec.shape
+        pad_face = torch.zeros(B, 1, H, C).type_as(motion_vec)
+        motion_vec = torch.cat([pad_face, motion_vec], dim=1)
+        return x, motion_vec
+
+
+    def after_transformer_block(self, block_idx, x, motion_vec, motion_masks=None):
+        if block_idx % 5 == 0:
+            adapter_args = [x, motion_vec, motion_masks, self.use_context_parallel]
+            residual_out = self.face_adapter.fuser_blocks[block_idx // 5](*adapter_args)
+            x = residual_out + x
+        return x
+
+
+    def forward(
+        self,
+        x,
+        t,
+        clip_fea,
+        context,
+        seq_len,
+        y=None,
+        pose_latents=None, 
+        face_pixel_values=None
+    ):
+        # params
+        device = self.patch_embedding.weight.device
+        if self.freqs.device != device:
+            self.freqs = self.freqs.to(device)
+
+        if y is not None:
+            x = [torch.cat([u, v], dim=0) for u, v in zip(x, y)]
+
+        # embeddings
+        x = [self.patch_embedding(u.unsqueeze(0)) for u in x]
+        x, motion_vec = self.after_patch_embedding(x, pose_latents, face_pixel_values)
+
+        grid_sizes = torch.stack(
+            [torch.tensor(u.shape[2:], dtype=torch.long) for u in x])
+        x = [u.flatten(2).transpose(1, 2) for u in x]
+        seq_lens = torch.tensor([u.size(1) for u in x], dtype=torch.long)
+        assert seq_lens.max() <= seq_len
+        x = torch.cat([
+            torch.cat([u, u.new_zeros(1, seq_len - u.size(1), u.size(2))],
+                      dim=1) for u in x
+        ])
+
+        # time embeddings
+        with amp.autocast(dtype=torch.float32):
+            e = self.time_embedding(
+                sinusoidal_embedding_1d(self.freq_dim, t).float()
+            )
+            e0 = self.time_projection(e).unflatten(1, (6, self.dim))
+            assert e.dtype == torch.float32 and e0.dtype == torch.float32
+
+        # context
+        context_lens = None
+        context = self.text_embedding(
+            torch.stack([
+                torch.cat(
+                    [u, u.new_zeros(self.text_len - u.size(0), u.size(1))])
+                for u in context
+            ]))
+
+        if self.use_img_emb:
+            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,
+            seq_lens=seq_lens,
+            grid_sizes=grid_sizes,
+            freqs=self.freqs,
+            context=context,
+            context_lens=context_lens)
+
+        if self.use_context_parallel:
+            x = torch.chunk(x, get_world_size(), dim=1)[get_rank()]
+
+        for idx, block in enumerate(self.blocks):
+            x = block(x, **kwargs)
+            x = self.after_transformer_block(idx, x, motion_vec)
+
+        # head
+        x = self.head(x, e)
+
+        if self.use_context_parallel:
+            x = gather_forward(x, dim=1)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return [u.float() for u in x]
+
+
+    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 [C_out, F, H / 8, W / 8]
+        """
+
+        c = self.out_dim
+        out = []
+        for u, v in zip(x, grid_sizes.tolist()):
+            u = u[:math.prod(v)].view(*v, *self.patch_size, c)
+            u = torch.einsum('fhwpqrc->cfphqwr', u)
+            u = u.reshape(c, *[i * j for i, j in zip(v, self.patch_size)])
+            out.append(u)
+        return out
+
+    def init_weights(self):
+        r"""
+        Initialize model parameters using Xavier initialization.
+        """
+
+        # basic init
+        for m in self.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.xavier_uniform_(m.weight)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+
+        # init embeddings
+        nn.init.xavier_uniform_(self.patch_embedding.weight.flatten(1))
+        for m in self.text_embedding.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, std=.02)
+        for m in self.time_embedding.modules():
+            if isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, std=.02)
+
+        # init output layer
+        nn.init.zeros_(self.head.head.weight)

wan/modules/animate/preprocess/UserGuider.md

+# Wan-animate Preprocessing User Guider
+
+## 1. Introductions
+
+
+Wan-animate offers two generation modes: `animation` and `replacement`. While both modes extract the skeleton from the reference video, they each have a distinct preprocessing pipeline.
+
+### 1.1 Animation Mode
+
+In this mode, it is highly recommended to enable pose retargeting, especially if the body proportions of the reference and driving characters are dissimilar.
+
+ - A simplified version of pose retargeting pipeline is provided to help developers quickly implement this functionality.
+
+ - **NOTE:** Due to the potential complexity of input data, the results from this simplified retargeting version are NOT guaranteed to be perfect. It is strongly advised to verify the preprocessing results before proceeding.
+
+ - Community contributions to improve on this feature are welcome.
+
+### 1.2 Replacement Mode
+
+ - Pose retargeting is DISABLED by default in this mode. This is a deliberate choice to account for potential spatial interactions between the character and the environment.
+
+ - **WARNING**: If there is a significant mismatch in body proportions between the reference and driving characters, artifacts or deformations may appear in the final output.
+
+ - A simplified version for extracting the character's mask is also provided.
+ - **WARNING:** This mask extraction process is designed for **single-person videos ONLY** and may produce incorrect results or fail in multi-person videos (incorrect pose tracking). For multi-person video, users are required to either develop their own solution or integrate a suitable open-source tool.
+
+---
+
+## 2. Preprocessing Instructions and Recommendations
+
+### 2.1 Basic Usage
+
+- The preprocessing process requires some additional models, including pose detection (mandatory), and mask extraction and image editing models (optional, as needed). Place them according to the following directory structure:
+```
+    /path/to/your/ckpt_path/
+    ├── det/
+    │ └── yolov10m.onnx
+    ├── pose2d/
+    │ └── vitpose_h_wholebody.onnx
+    ├── sam2/
+    │ └── sam2_hiera_large.pt
+    └── FLUX.1-Kontext-dev/
+```
+- `video_path`, `refer_path`, and `save_path` correspond to the paths for the input driving video, the character image, and the preprocessed results.
+
+- When using `animation` mode, two videos, `src_face.mp4` and `src_pose.mp4`, will be generated in `save_path`. When using `replacement` mode, two additional videos, `src_bg.mp4` and `src_mask.mp4`, will also be generated.
+
+- The `resolution_area` parameter determines the resolution for both preprocessing and the generation model. Its size is determined by pixel area.
+
+- The `fps` parameter can specify the frame rate for video processing. A lower frame rate can improve generation efficiency, but may cause stuttering or choppiness.
+
+---
+
+### 2.2 Animation Mode
+
+- We support three forms: not using pose retargeting, using basic pose retargeting, and using enhanced pose retargeting based on the `FLUX.1-Kontext-dev` image editing model. These are specified via the `retarget_flag` and `use_flux` parameters.
+
+- Specifying `retarget_flag` to use basic pose retargeting requires ensuring that both the reference character and the character in the first frame of the driving video are in a front-facing, stretched pose.
+
+- Other than that, we recommend using enhanced pose retargeting by specifying both `retarget_flag` and `use_flux`. **NOTE:** Due to the limited capabilities of `FLUX.1-Kontext-dev`, it is NOT guaranteed to produce the expected results (e.g., consistency is not maintained, the pose is incorrect, etc.). It is recommended to check the intermediate results as well as the finally generated pose video; both are stored in `save_path`. Of course, users can also use a better image editing model, or explore the prompts for Flux on their own.
+
+---
+
+### 2.3 Replacement Mode
+
+- Specifying `replace_flag` to enable data preprocessing for this mode. The preprocessing will additionally process a mask for the character in the video, and its size and shape can be adjusted by specifying some parameters.
+- `iterations` and `k` can make the mask larger, covering more area.
+- `w_len` and `h_len` can adjust the mask's shape. Smaller values will make the outline coarser, while larger values will make it finer.
+
+- A smaller, finer-contoured mask can allow for more of the original background to be preserved, but may potentially limit the character's generation area (considering potential appearance differences, this can lead to some shape leakage). A larger, coarser mask can allow the character generation to be more flexible and consistent, but because it includes more of the background, it might affect the background's consistency. We recommend users to adjust the relevant parameters based on their specific input data.
\ No newline at end of file

wan/modules/animate/preprocess/__init__.py

+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+from .process_pipepline import ProcessPipeline
+from .video_predictor import SAM2VideoPredictor
\ No newline at end of file