# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved. import math import random 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 .attention import flash_attention, attention from torch.utils.checkpoint import checkpoint from einops import rearrange from .audio_proj import AudioProjModel import warnings warnings.filterwarnings('ignore') __all__ = ['WanModel'] def sinusoidal_embedding_1d(dim, position): # preprocess assert dim % 2 == 0 half = dim // 2 position = position.type(torch.float32) # Changed float64 to float32 here # 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 @amp.autocast(enabled=False) def rope_params(max_seq_len, dim, theta=10000): assert dim % 2 == 0 freqs = torch.outer( torch.arange(max_seq_len), 1.0 / torch.pow(theta, torch.arange(0, dim, 2).to(torch.float32).div(dim))) # Changed float64 to float32 here freqs = torch.polar(torch.ones_like(freqs), freqs) return freqs @amp.autocast(enabled=False) def rope_apply(x, grid_sizes, freqs): n, c = x.size(2), x.size(3) // 2 # split freqs freqs = freqs.split([c - 2 * (c // 3), c // 3, c // 3], dim=1) # loop over samples output = [] for i, (f, h, w) in enumerate(grid_sizes.tolist()): seq_len = f * h * w # precompute multipliers x_i = torch.view_as_complex(x[i, :seq_len].to(torch.float32).reshape( seq_len, n, -1, 2)) # Changed float64 to float32 here freqs_i = torch.cat([ freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1), freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1), freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1) ], dim=-1).reshape(seq_len, 1, -1) # apply rotary embedding x_i = torch.view_as_real(x_i * freqs_i).flatten(2) x_i = torch.cat([x_i, x[i, seq_len:]]) # append to collection output.append(x_i) return torch.stack(output).float() class WanRMSNorm(nn.Module): def __init__(self, dim, eps=1e-5): super().__init__() self.dim = dim self.eps = eps self.weight = nn.Parameter(torch.ones(dim)) def forward(self, x): r""" Args: x(Tensor): Shape [B, L, C] """ return self._norm(x.float()).type_as(x) * self.weight def _norm(self, x): return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps) class WanLayerNorm(nn.LayerNorm): def __init__(self, dim, eps=1e-6, elementwise_affine=False): super().__init__(dim, elementwise_affine=elementwise_affine, eps=eps) def forward(self, x): r""" Args: x(Tensor): Shape [B, L, C] """ return super().forward(x.float()).type_as(x) class WanSelfAttention(nn.Module): def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, eps=1e-6): 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 = nn.Linear(dim, dim) self.k = nn.Linear(dim, dim) self.v = nn.Linear(dim, dim) self.o = nn.Linear(dim, dim) self.norm_q = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity() self.norm_k = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity() def forward(self, x, seq_lens, grid_sizes, freqs): r""" 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 WanT2VCrossAttention(WanSelfAttention): def forward(self, x, context, context_lens): r""" Args: x(Tensor): Shape [B, L1, C] context(Tensor): Shape [B, L2, C] context_lens(Tensor): Shape [B] """ 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) # compute attention x = flash_attention(q, k, v, k_lens=context_lens) # output x = x.flatten(2) 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): super().__init__(dim, num_heads, window_size, qk_norm, eps) self.k_img = nn.Linear(dim, dim) self.v_img = nn.Linear(dim, dim) # self.alpha = nn.Parameter(torch.zeros((1, ))) self.norm_k_img = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity() def forward(self, x, context, context_lens): r""" Args: x(Tensor): Shape [B, L1, C] context(Tensor): Shape [B, L2, C] context_lens(Tensor): Shape [B] """ context_img = context[:, :257] context = context[:, 257:] 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) 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) img_x = img_x.flatten(2) x = x + img_x x = self.o(x) return x class WanA2VCrossAttention(WanSelfAttention): def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, eps=1e-6): super().__init__(dim, num_heads, window_size, qk_norm, eps) self.k_img = nn.Linear(dim, dim) self.v_img = nn.Linear(dim, dim) # self.alpha = nn.Parameter(torch.zeros((1, ))) self.norm_k_img = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity() self.k_audio = nn.Linear(dim, dim) self.v_audio = nn.Linear(dim, dim) self.norm_k_audio = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity() def forward(self, x, context, context_lens, temporal_mask=None, face_mask_list=None): r""" Args: x(Tensor): Shape [B, L1, C] context(Tensor): Shape [B, L2, C] context_lens(Tensor): Shape [B] temporal_mask(Tensor): Shape [B, L2] face_mask_list(list): Shape [n, B, L1] """ context_img = context[1] context_audio = context[2] context = context[0] 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) 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) k_audio = self.norm_k_audio(self.k_audio(context_audio)).view(b, -1, n, d) v_audio = self.v_audio(context_audio).view(b, -1, n, d) img_x = flash_attention(q, k_img, v_img, k_lens=None) if temporal_mask is not None: audio_x = attention(q, k_audio, v_audio, k_lens=None, attn_mask=temporal_mask) else: audio_x = flash_attention(q, k_audio, v_audio, k_lens=None) # compute attention x = flash_attention(q, k, v, k_lens=context_lens) # output x = x.flatten(2) img_x = img_x.flatten(2) audio_x = audio_x.flatten(2) x = x + img_x + audio_x x = self.o(x) return x class WanAF2VCrossAttention(WanSelfAttention): """ For audio CA output, apply additional Ref attention Ref cond input may come from face recognition embedding / clip embedding / 3d vae token """ def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True, eps=1e-6, use_concat_attention=True): # New parameter to control whether to use concat mode super().__init__(dim, num_heads, window_size, qk_norm, eps) self.k_img = nn.Linear(dim, dim) self.v_img = nn.Linear(dim, dim) # self.alpha = nn.Parameter(torch.zeros((1, ))) self.norm_k_img = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity() self.k_audio = nn.Linear(dim, dim) self.v_audio = nn.Linear(dim, dim) self.norm_k_audio = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity() self.k_face = nn.Linear(dim, dim) self.v_face = nn.Linear(dim, dim) self.norm_k_face = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity() # New parameter to control attention mode self.use_concat_attention = use_concat_attention def forward( self, x, context, context_lens, temporal_mask=None, face_mask_list=None, use_token_mask=True, ): r""" Args: x(Tensor): Shape [B, L1, C] context(Tensor): Shape [B, L2, C] context_lens(Tensor): Shape [B] temporal_mask(Tensor): Shape [B, L2] Usage example: # Original mode (separated attention) model = WanModel(model_type='a2v_af', use_concat_attention=False) # New mode (concat attention) model = WanModel(model_type='a2v_af', use_concat_attention=True) # In new mode, face token is always visible, audio part follows temporal_mask logic """ # [text, image, audio list, audio ref list] context_img = context[1] context_audios = context[2] face_context_list = context[3] context = context[0] 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) 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) """New face kv for audio focus n people means n face attn operations """ k_face_list = [] v_face_list = [] k_audio_list = [] v_audio_list = [] # Ensure audio and face lists have consistent length min_length = min(len(context_audios), len(face_context_list)) # print(f"WanAF2VCrossAttention: Processing {min_length} audio-face pairs") for i in range(min_length): context_audio = context_audios[i] face_context = face_context_list[i] # Extract audio features k_audio = self.norm_k_audio(self.k_audio(context_audio)).view(b, -1, n, d) v_audio = self.v_audio(context_audio).view(b, -1, n, d) k_audio_list.append(k_audio) v_audio_list.append(v_audio) # Extract face features k_face = self.norm_k_face(self.k_face(face_context)).view(b, -1, n, d) v_face = self.v_face(face_context).view(b, -1, n, d) k_face_list.append(k_face) v_face_list.append(v_face) # text attn x = flash_attention(q, k, v, k_lens=context_lens) # ref image attn img_x = flash_attention(q, k_img, v_img, k_lens=None) """ For each id, execute identity-aware audio ca Method 1: Add residual connection between each person, make it causal Method 2: No residual connection Method 3: Add residual connection only at the end, preserve original driving information Method 4: Don't split into two steps, directly do three-modal CA (video, audio, face) """ af_output_list = [] # Ensure all lists have consistent length min_length = min(len(k_face_list), len(v_face_list), len(k_audio_list), len(v_audio_list), len(face_mask_list)) # print(f"Processing {min_length} audio-face pairs") for i in range(min_length): k_face = k_face_list[i] v_face = v_face_list[i] k_audio = k_audio_list[i] v_audio = v_audio_list[i] face_mask = face_mask_list[i] # concat face and audio features k_concat = torch.cat([k_face, k_audio], dim=1) # [B, L_face+L_audio, n, d] v_concat = torch.cat([v_face, v_audio], dim=1) # [B, L_face+L_audio, n, d] # Construct attention mask if temporal_mask is not None: # Get face token count face_len = k_face.shape[1] audio_len = k_audio.shape[1] # Create new mask: face part all True, audio part follows original mask # Fix dimensions: [B, 1, seq_len_q, seq_len_kv] new_mask = torch.ones((b, 1, q.shape[1], face_len + audio_len), dtype=torch.bool, device=temporal_mask.device) # face part is always visible new_mask[..., :face_len] = True # audio part follows original mask logic - need to adjust temporal_mask shape # temporal_mask shape is [B, 1, seq_len_q, audio_len] if temporal_mask.shape[-1] == audio_len: # Ensure dimension match new_mask[..., face_len:] = temporal_mask # [B, 1, seq_len_q, audio_len] audio_x = attention(q, k_concat, v_concat, k_lens=None, attn_mask=new_mask) else: # When no mask, all tokens are visible audio_x = flash_attention(q, k_concat, v_concat, k_lens=None) if use_token_mask: # Multiply output by face_mask af_output_list.append(audio_x.flatten(2) * face_mask) else: af_output_list.append(audio_x.flatten(2)) # output x = x.flatten(2) img_x = img_x.flatten(2) x = x + img_x for af_output in af_output_list: x = x + af_output x = self.o(x) return x WAN_CROSSATTENTION_CLASSES = { 't2v_cross_attn': WanT2VCrossAttention, 'i2v_cross_attn': WanI2VCrossAttention, 'a2v_cross_attn': WanA2VCrossAttention, 'a2v_cross_attn_af': WanAF2VCrossAttention } 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, use_concat_attention=False): # New parameter 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 = WanSelfAttention(dim, num_heads, window_size, qk_norm, eps) self.norm3 = WanLayerNorm( dim, eps, elementwise_affine=True) if cross_attn_norm else nn.Identity() # Create corresponding cross attention based on cross_attn_type if cross_attn_type == 'a2v_cross_attn_af': self.cross_attn = WAN_CROSSATTENTION_CLASSES[cross_attn_type](dim, num_heads, (-1, -1), qk_norm, eps, use_concat_attention) # Pass new parameter else: self.cross_attn = WAN_CROSSATTENTION_CLASSES[cross_attn_type](dim, num_heads, (-1, -1), qk_norm, eps) 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, temporal_mask=None, # For audio alignment face_mask_list=None, # Multi-person binding human_mask_list=None, # Multi-person binding (deprecated, set to None) use_token_mask=True, ): r""" Args: x(Tensor): Shape [B, L, C] e(Tensor): Shape [B, 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, temporal_mask=None): if isinstance(self.cross_attn, WanAF2VCrossAttention): # human_mask_list is now None, no longer used x = x + self.cross_attn( self.norm3(x), context, context_lens, temporal_mask, face_mask_list, use_token_mask=use_token_mask ) elif isinstance(self.cross_attn, WanA2VCrossAttention): x = x + self.cross_attn(self.norm3(x), context, context_lens, temporal_mask) else: 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, temporal_mask) return x class Head(nn.Module): def __init__(self, dim, out_dim, patch_size, eps=1e-6): 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 = WanLayerNorm(dim, eps) self.head = nn.Linear(dim, out_dim) # modulation self.modulation = nn.Parameter(torch.randn(1, 2, dim) / dim**0.5) def forward(self, x, e): r""" Args: x(Tensor): Shape [B, L1, C] e(Tensor): Shape [B, 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 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 WanModel(ModelMixin, ConfigMixin): r""" Wan diffusion backbone supporting both text-to-video and image-to-video. """ ignore_for_config = [ 'patch_size', 'cross_attn_norm', 'qk_norm', 'text_dim', 'window_size' ] _no_split_modules = ['WanAttentionBlock'] @register_to_config 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, temporal_align=True, use_concat_attention=False): # New parameter to control concat attention mode 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 temporal_align (`bool`, *optional*, defaults to True): Enable temporal alignment for audio features use_concat_attention (`bool`, *optional*, defaults to False): Use concatenated face and audio features for attention computation """ super().__init__() self.checkpoint_enabled = True assert model_type in ['t2v', 'i2v', 'a2v', 'a2v_af'] 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 self.has_temporal_align = temporal_align self.use_concat_attention = use_concat_attention # Save new parameter # embeddings self.patch_embedding = nn.Conv3d( in_dim, 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 attn_type = { 't2v':'t2v_cross_attn', 'i2v':'i2v_cross_attn', 'a2v':'a2v_cross_attn', 'a2v_af':'a2v_cross_attn_af' } # blocks cross_attn_type = attn_type[model_type] self.blocks = nn.ModuleList([ WanAttentionBlock(cross_attn_type, dim, ffn_dim, num_heads, window_size, qk_norm, cross_attn_norm, eps, self.use_concat_attention) # Pass new parameter for _ in range(num_layers) ]) # head self.head = Head(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) if model_type == 'i2v': self.img_emb = MLPProj(1280, dim) elif model_type=='a2v': self.img_emb = MLPProj(1280, dim) self.audio_emb = AudioProjModel(seq_len=5, blocks=12, channels=768, intermediate_dim=512, output_dim=dim, context_tokens=32,) elif model_type=='a2v_af': self.img_emb = MLPProj(1280, dim) self.audio_emb = AudioProjModel(seq_len=5, blocks=12, channels=768, intermediate_dim=512, output_dim=dim, context_tokens=32,) self.audio_ref_emb = MLPProj(1280, dim) # Used for audio ref attention # initialize weights self.init_weights() def enable_gradient_checkpointing(self,use_reentrant=False): self.checkpoint_enabled = True self._use_reentrant = use_reentrant def forward( self, x, t, context, seq_len, clip_fea=None, y=None, audio_feature=None, audio_ref_features=None, # For audio ref face_mask_list=None, # Multi-person binding human_mask_list=None, # Multi-person binding (deprecated, set to None) masks_flattened=False, use_token_mask=True, ): r""" Forward pass through the diffusion model Args: x (List[Tensor]): List of input video tensors, each with shape [C_in, F, H, W] t (Tensor): Diffusion timesteps tensor of shape [B] context (List[Tensor]): List of text embeddings each with shape [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 audio_ref_features (List[Tensor], *optional*): Conditional audio features for audio-to-video mode Returns: List[Tensor]: List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8] """ if self.model_type == 'i2v': assert clip_fea is not None and y is not None c, f, h, w = x[0].shape h, w = h//self.patch_size[-2], w//self.patch_size[-1] b = len(x) # 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)] # x arrangement: [noisy frames, mask, ref frame + padding frames] # embeddings, before: [[36, F, H, W], ...] x = [self.patch_embedding(u.unsqueeze(0)) for u in x] # after: [[1, 1536, F, H/2 , W/2], ...] 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] # [[1, seq_len, 1536], ...] # Also flatten mask for each id if use_token_mask: if not masks_flattened and face_mask_list is not None: for m_index in range(len(face_mask_list)): # Only take first channel face_mask_list[m_index] = [m[0].flatten(0) for m in face_mask_list[m_index]] face_mask_list[m_index] = torch.stack(face_mask_list[m_index]) # [B, seq_len] # Add a dimension at the end face_mask_list[m_index] = face_mask_list[m_index][..., None] 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 ])) # print("="*25,self.model_type) if self.model_type=="i2v": context_clip = self.img_emb(clip_fea) # bs x 257 x dim context = torch.concat([context_clip, context], dim=1) elif self.model_type == 'a2v_af': # New list mode: supports multiple audio and faces if "ref_face_list" in audio_ref_features and "audio_list" in audio_ref_features: # Use new list mode ref_face_list = audio_ref_features["ref_face_list"] audio_list = audio_ref_features["audio_list"] # Process audio feature list audio_embeding_list = [] for i, audio_feat in enumerate(audio_list): audio_embeding = self.audio_emb(audio_feat) audio_embeding_list.append(audio_embeding) # Process face feature list ref_context_list = [] for i, ref_features in enumerate(ref_face_list): audio_ref_embeding = self.audio_ref_emb(ref_features) ref_context_list.append(audio_ref_embeding) # Original a2v required features context_clip = self.img_emb(clip_fea) # bs x 257 x dim # [text, image, audio list, audio ref list] context = [context] context.append(context_clip) context.append(audio_embeding_list) context.append(ref_context_list) # Currently testing does not use temporal_mask self.has_temporal_align = True if self.has_temporal_align and len(audio_embeding_list) > 0 and audio_embeding_list[0] is not None: # Use first audio's shape to build temporal_mask audio_shape = audio_embeding_list[0].shape temporal_mask = torch.zeros((f, audio_shape[-3]), dtype=torch.bool, device=x.device) temporal_mask[0] = True # First frame image and all speech compute attention # print(f"temporal_mask {temporal_mask.shape},{torch.sum(temporal_mask)}") for i in range(1, f): temporal_mask[i, (i - 1)* 4 + 1: i*4 + 1]=True # In dataloader, audio is already taken with sliding window of 5, no need to do overlap here temporal_mask = temporal_mask.reshape(f, 1, 1 , audio_shape[-3], 1).repeat(1, h, w, 1, audio_shape[-2]) # print(f"temporal_mask {temporal_mask.shape},{h},{w},{torch.sum(temporal_mask)}") temporal_mask = rearrange(temporal_mask, 'f h w c d -> (f h w) (c d)').contiguous()[None,None,...] temporal_mask = temporal_mask.expand(b, 1, temporal_mask.shape[-2], temporal_mask.shape[-1]) else: temporal_mask = None # arguments kwargs = dict( e=e0, seq_lens=seq_lens, grid_sizes=grid_sizes, freqs=self.freqs, context=context, context_lens=context_lens, temporal_mask=temporal_mask, # For audio alignment face_mask_list=face_mask_list, # Multi-person binding human_mask_list=None, # human_mask_list no longer used use_token_mask=use_token_mask ) def create_custom_forward(module): def custom_forward(x, **kwargs): # Explicitly accept x and **kwargs return module(x, **kwargs) return custom_forward for block in self.blocks: if self.training and self.checkpoint_enabled: x = checkpoint( create_custom_forward(block), x, # Positional argument **kwargs, # Keyword arguments use_reentrant=False, ) else: x = block(x, **kwargs) # head x = self.head(x, e) # 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 @classmethod def from_pretrained(cls, pretrained_model_name_or_path, config: dict = None, **kwargs): import glob import os from omegaconf import ListConfig from typing import Union if isinstance(pretrained_model_name_or_path, str) and os.path.isdir(pretrained_model_name_or_path) and (config is None) and not pretrained_model_name_or_path.endswith('.pth'): print(">>> Using diffusers from_pretrained with provided config") return super().from_pretrained(pretrained_model_name_or_path, **kwargs) else: # === Custom loading logic === print(">>> Using custom from_pretrained with provided config") from diffusers.models.model_loading_utils import load_model_dict_into_meta, load_state_dict import accelerate torch_dtype = kwargs.pop("torch_dtype", torch.bfloat16) map_location = kwargs.pop("map_location", 'cpu') # step 1. Initialize model with accelerate.init_empty_weights(): model = cls.from_config(config) # step 2. Find weight files if isinstance(pretrained_model_name_or_path, Union[list, ListConfig]): weight_files = pretrained_model_name_or_path elif os.path.isdir(pretrained_model_name_or_path): weight_files = glob.glob(f'{pretrained_model_name_or_path}/*.safetensors') else: weight_files = [pretrained_model_name_or_path] state_dict = {} for wf in weight_files: _state_dict = load_state_dict(wf, map_location=map_location) if "model" in _state_dict: state_dict.update(_state_dict["model"]) else: state_dict.update(_state_dict) del _state_dict empty_state_dict = model.state_dict() n_miss = 0 n_unexpect = 0 for param_name in model.state_dict().keys(): if param_name not in state_dict: n_miss+=1 for param_name in state_dict.keys(): if param_name not in model.state_dict(): n_unexpect+=1 # Initialize weights for missing modules for name, param in empty_state_dict.items(): if name not in state_dict: if param.dim() > 1: state_dict[name] = nn.init.xavier_uniform_(torch.zeros(param.shape)) elif '.norm_' in name: state_dict[name] = nn.init.constant_(torch.zeros(param.shape), 1) else: state_dict[name] = nn.init.zeros_(torch.zeros(param.shape)) state_dict = {k:v.to(dtype=torch.bfloat16) for k, v in state_dict.items()} # step 3. Load weights load_model_dict_into_meta(model, state_dict, dtype=torch_dtype) n_updated = len(empty_state_dict.keys()) - n_miss print(f"{n_updated} parameters are loaded from {pretrained_model_name_or_path}, {n_miss} parameters are miss, {n_unexpect} parameters are unexpected.") del state_dict return model 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)