Merge 3ade599515374d4cb20683ac85822189199583b2 into fd271dedfde6e192a1f1a025521070876e89e04a

* feat(api-nodes): add support for seedance-1-0-pro-fast model

* feat(api-nodes): add support for seedream-4.5 model

comfy/ldm/wan/model.py

@@ -62,6 +62,8 @@ class WanSelfAttention(nn.Module):
             x(Tensor): Shape [B, L, num_heads, C / num_heads]
             freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
         """
+        patches = transformer_options.get("patches", {})
+
         b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
 
         def qkv_fn_q(x):
@@ -86,6 +88,10 @@ class WanSelfAttention(nn.Module):
             transformer_options=transformer_options,
         )
 
+        if "attn1_patch" in patches:
+            for p in patches["attn1_patch"]:
+                x = p({"x": x, "q": q, "k": k, "transformer_options": transformer_options})
+
         x = self.o(x)
         return x
 
@@ -225,6 +231,8 @@ class WanAttentionBlock(nn.Module):
         """
         # assert e.dtype == torch.float32
 
+        patches = transformer_options.get("patches", {})
+
         if e.ndim < 4:
             e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
         else:
@@ -242,6 +250,11 @@ class WanAttentionBlock(nn.Module):
 
         # cross-attention & ffn
         x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len, transformer_options=transformer_options)
+
+        if "attn2_patch" in patches:
+            for p in patches["attn2_patch"]:
+                x = p({"x": x, "transformer_options": transformer_options})
+
         y = self.ffn(torch.addcmul(repeat_e(e[3], x), self.norm2(x), 1 + repeat_e(e[4], x)))
         x = torch.addcmul(x, y, repeat_e(e[5], x))
         return x
@@ -488,7 +501,7 @@ class WanModel(torch.nn.Module):
         self.blocks = nn.ModuleList([
             wan_attn_block_class(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)
+            for i in range(num_layers)
         ])
 
         # head
@@ -541,6 +554,7 @@ class WanModel(torch.nn.Module):
         # embeddings
         x = self.patch_embedding(x.float()).to(x.dtype)
         grid_sizes = x.shape[2:]
+        transformer_options["grid_sizes"] = grid_sizes
         x = x.flatten(2).transpose(1, 2)
 
         # time embeddings
@@ -569,6 +583,7 @@ class WanModel(torch.nn.Module):
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
         for i, block in enumerate(self.blocks):
+            transformer_options["block_idx"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -735,6 +750,7 @@ class VaceWanModel(WanModel):
         # embeddings
         x = self.patch_embedding(x.float()).to(x.dtype)
         grid_sizes = x.shape[2:]
+        transformer_options["grid_sizes"] = grid_sizes
         x = x.flatten(2).transpose(1, 2)
 
         # time embeddings
@@ -764,6 +780,7 @@ class VaceWanModel(WanModel):
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
         for i, block in enumerate(self.blocks):
+            transformer_options["block_idx"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}

comfy/ldm/wan/model_multitalk.py

@@ -0,0 +1,500 @@
+import torch
+from einops import rearrange, repeat
+import comfy
+from comfy.ldm.modules.attention import optimized_attention
+
+
+def calculate_x_ref_attn_map(visual_q, ref_k, ref_target_masks, split_num=8):
+    scale = 1.0 / visual_q.shape[-1] ** 0.5
+    visual_q = visual_q.transpose(1, 2) * scale
+
+    B, H, x_seqlens, K = visual_q.shape
+
+    x_ref_attn_maps = []
+    for class_idx, ref_target_mask in enumerate(ref_target_masks):
+        ref_target_mask = ref_target_mask.view(1, 1, 1, -1)
+
+        x_ref_attnmap = torch.zeros(B, H, x_seqlens, device=visual_q.device, dtype=visual_q.dtype)
+        chunk_size = min(max(x_seqlens // split_num, 1), x_seqlens)
+
+        for i in range(0, x_seqlens, chunk_size):
+            end_i = min(i + chunk_size, x_seqlens)
+
+            attn_chunk = visual_q[:, :, i:end_i] @ ref_k.permute(0, 2, 3, 1)  # B, H, chunk, ref_seqlens
+
+            # Apply softmax
+            attn_max = attn_chunk.max(dim=-1, keepdim=True).values
+            attn_chunk = (attn_chunk - attn_max).exp()
+            attn_sum = attn_chunk.sum(dim=-1, keepdim=True)
+            attn_chunk = attn_chunk / (attn_sum + 1e-8)
+
+            # Apply mask and sum
+            masked_attn = attn_chunk * ref_target_mask
+            x_ref_attnmap[:, :, i:end_i] = masked_attn.sum(-1) / (ref_target_mask.sum() + 1e-8)
+
+            del attn_chunk, masked_attn
+
+        # Average across heads
+        x_ref_attnmap = x_ref_attnmap.mean(dim=1)  # B, x_seqlens
+        x_ref_attn_maps.append(x_ref_attnmap)
+
+    del visual_q, ref_k
+
+    return torch.cat(x_ref_attn_maps, dim=0)
+
+def get_attn_map_with_target(visual_q, ref_k, shape, ref_target_masks=None, split_num=2):
+    """Args:
+        query (torch.tensor): B M H K
+        key (torch.tensor): B M H K
+        shape (tuple): (N_t, N_h, N_w)
+        ref_target_masks: [B, N_h * N_w]
+    """
+
+    N_t, N_h, N_w = shape
+
+    x_seqlens = N_h * N_w
+    ref_k     = ref_k[:, :x_seqlens]
+    _, seq_lens, heads, _ = visual_q.shape
+    class_num, _ = ref_target_masks.shape
+    x_ref_attn_maps = torch.zeros(class_num, seq_lens).to(visual_q)
+
+    split_chunk = heads // split_num
+
+    for i in range(split_num):
+        x_ref_attn_maps_perhead = calculate_x_ref_attn_map(
+            visual_q[:, :, i*split_chunk:(i+1)*split_chunk, :],
+            ref_k[:, :, i*split_chunk:(i+1)*split_chunk, :],
+            ref_target_masks
+            )
+        x_ref_attn_maps += x_ref_attn_maps_perhead
+
+    return x_ref_attn_maps / split_num
+
+
+def normalize_and_scale(column, source_range, target_range, epsilon=1e-8):
+    source_min, source_max = source_range
+    new_min, new_max = target_range
+    normalized = (column - source_min) / (source_max - source_min + epsilon)
+    scaled = normalized * (new_max - new_min) + new_min
+    return scaled
+
+
+def rotate_half(x):
+    x = rearrange(x, "... (d r) -> ... d r", r=2)
+    x1, x2 = x.unbind(dim=-1)
+    x = torch.stack((-x2, x1), dim=-1)
+    return rearrange(x, "... d r -> ... (d r)")
+
+
+def get_audio_embeds(encoded_audio, audio_start, audio_end):
+    audio_embs = []
+    human_num = len(encoded_audio)
+    audio_frames = encoded_audio[0].shape[0]
+
+    indices = (torch.arange(4 + 1) - 2) * 1
+
+    for human_idx in range(human_num):
+        if audio_end > audio_frames: # in case of not enough audio for current window, pad with first audio frame as that's most likely silence
+            pad_len = audio_end - audio_frames
+            pad_shape = list(encoded_audio[human_idx].shape)
+            pad_shape[0] = pad_len
+            pad_tensor = encoded_audio[human_idx][:1].repeat(pad_len, *([1] * (encoded_audio[human_idx].dim() - 1)))
+            encoded_audio_in = torch.cat([encoded_audio[human_idx], pad_tensor], dim=0)
+        else:
+            encoded_audio_in = encoded_audio[human_idx]
+        center_indices = torch.arange(audio_start, audio_end, 1).unsqueeze(1) + indices.unsqueeze(0)
+        center_indices = torch.clamp(center_indices, min=0, max=encoded_audio_in.shape[0] - 1)
+        audio_emb = encoded_audio_in[center_indices].unsqueeze(0)
+        audio_embs.append(audio_emb)
+
+    return torch.cat(audio_embs, dim=0)
+
+
+def project_audio_features(audio_proj, encoded_audio, audio_start, audio_end):
+    audio_embs = get_audio_embeds(encoded_audio, audio_start, audio_end)
+
+    first_frame_audio_emb_s = audio_embs[:, :1, ...]
+    latter_frame_audio_emb = audio_embs[:, 1:, ...]
+    latter_frame_audio_emb = rearrange(latter_frame_audio_emb, "b (n_t n) w s c -> b n_t n w s c", n=4)
+
+    middle_index = audio_proj.seq_len // 2
+
+    latter_first_frame_audio_emb = latter_frame_audio_emb[:, :, :1, :middle_index+1, ...]
+    latter_first_frame_audio_emb = rearrange(latter_first_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c")
+    latter_last_frame_audio_emb = latter_frame_audio_emb[:, :, -1:, middle_index:, ...]
+    latter_last_frame_audio_emb = rearrange(latter_last_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c")
+    latter_middle_frame_audio_emb = latter_frame_audio_emb[:, :, 1:-1, middle_index:middle_index+1, ...]
+    latter_middle_frame_audio_emb = rearrange(latter_middle_frame_audio_emb, "b n_t n w s c -> b n_t (n w) s c")
+    latter_frame_audio_emb_s = torch.cat([latter_first_frame_audio_emb, latter_middle_frame_audio_emb, latter_last_frame_audio_emb], dim=2)
+
+    audio_emb = audio_proj(first_frame_audio_emb_s, latter_frame_audio_emb_s)
+    audio_emb = torch.cat(audio_emb.split(1), dim=2)
+
+    return audio_emb
+
+
+class RotaryPositionalEmbedding1D(torch.nn.Module):
+    def __init__(self,
+                 head_dim,
+                 ):
+        super().__init__()
+        self.head_dim = head_dim
+        self.base = 10000
+
+    def precompute_freqs_cis_1d(self, pos_indices):
+        freqs = 1.0 / (self.base ** (torch.arange(0, self.head_dim, 2)[: (self.head_dim // 2)].float() / self.head_dim))
+        freqs = freqs.to(pos_indices.device)
+        freqs = torch.einsum("..., f -> ... f", pos_indices.float(), freqs)
+        freqs = repeat(freqs, "... n -> ... (n r)", r=2)
+        return freqs
+
+    def forward(self, x, pos_indices):
+        freqs_cis = self.precompute_freqs_cis_1d(pos_indices)
+
+        x_ = x.float()
+
+        freqs_cis = freqs_cis.float().to(x.device)
+        cos, sin = freqs_cis.cos(), freqs_cis.sin()
+        cos, sin = rearrange(cos, 'n d -> 1 1 n d'), rearrange(sin, 'n d -> 1 1 n d')
+        x_ = (x_ * cos) + (rotate_half(x_) * sin)
+
+        return x_.type_as(x)
+
+class SingleStreamAttention(torch.nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        encoder_hidden_states_dim: int,
+        num_heads: int,
+        qkv_bias: bool,
+        device=None, dtype=None, operations=None
+    ) -> None:
+        super().__init__()
+        self.dim = dim
+        self.encoder_hidden_states_dim = encoder_hidden_states_dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q_linear = operations.Linear(dim, dim, bias=qkv_bias, device=device, dtype=dtype)
+        self.proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+        self.kv_linear = operations.Linear(encoder_hidden_states_dim, dim * 2, bias=qkv_bias, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, encoder_hidden_states: torch.Tensor, shape=None) -> torch.Tensor:
+        N_t, N_h, N_w = shape
+
+        expected_tokens = N_t * N_h * N_w
+        actual_tokens = x.shape[1]
+        x_extra = None
+
+        if actual_tokens != expected_tokens:
+            x_extra = x[:, -N_h * N_w:, :]
+            x = x[:, :-N_h * N_w, :]
+            N_t = N_t - 1
+
+        B = x.shape[0]
+        S = N_h * N_w
+        x = x.view(B * N_t, S, self.dim)
+
+        # get q for hidden_state
+        q = self.q_linear(x).view(B * N_t, S, self.num_heads, self.head_dim)
+
+        # get kv from encoder_hidden_states # shape: (B, N, num_heads, head_dim)
+        kv = self.kv_linear(encoder_hidden_states)
+        encoder_k, encoder_v = kv.view(B * N_t, encoder_hidden_states.shape[1], 2, self.num_heads, self.head_dim).unbind(2)
+
+        #print("q.shape", q.shape) #torch.Size([21, 1024, 40, 128])
+        x = optimized_attention(
+            q.transpose(1, 2),
+            encoder_k.transpose(1, 2),
+            encoder_v.transpose(1, 2),
+            heads=self.num_heads, skip_reshape=True, skip_output_reshape=True).transpose(1, 2)
+
+        # linear transform
+        x = self.proj(x.reshape(B * N_t, S, self.dim))
+        x = x.view(B, N_t * S, self.dim)
+
+        if x_extra is not None:
+            x = torch.cat([x, torch.zeros_like(x_extra)], dim=1)
+
+        return x
+
+class SingleStreamMultiAttention(SingleStreamAttention):
+    def __init__(
+        self,
+        dim: int,
+        encoder_hidden_states_dim: int,
+        num_heads: int,
+        qkv_bias: bool,
+        class_range: int = 24,
+        class_interval: int = 4,
+        device=None, dtype=None, operations=None
+    ) -> None:
+        super().__init__(
+            dim=dim,
+            encoder_hidden_states_dim=encoder_hidden_states_dim,
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            device=device,
+            dtype=dtype,
+            operations=operations
+        )
+
+        # Rotary-embedding layout parameters
+        self.class_interval = class_interval
+        self.class_range = class_range
+        self.max_humans = self.class_range // self.class_interval
+
+        # Constant bucket used for background tokens
+        self.rope_bak = int(self.class_range // 2)
+
+        self.rope_1d = RotaryPositionalEmbedding1D(self.head_dim)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        shape=None,
+        x_ref_attn_map=None
+    ) -> torch.Tensor:
+        encoder_hidden_states = encoder_hidden_states.squeeze(0).to(x.device)
+        human_num = x_ref_attn_map.shape[0] if x_ref_attn_map is not None else 1
+        # Single-speaker fall-through
+        if human_num <= 1:
+            return super().forward(x, encoder_hidden_states, shape)
+
+        N_t, N_h, N_w = shape
+
+        x_extra = None
+        if x.shape[0] * N_t != encoder_hidden_states.shape[0]:
+            x_extra = x[:, -N_h * N_w:, :]
+            x = x[:, :-N_h * N_w, :]
+            N_t = N_t - 1
+        x = rearrange(x, "B (N_t S) C -> (B N_t) S C", N_t=N_t)
+
+        # Query projection
+        B, N, C = x.shape
+        q = self.q_linear(x)
+        q = q.view(B, N, self.num_heads, self.head_dim).permute(0, 2, 1, 3)
+
+        # Use `class_range` logic for 2 speakers
+        rope_h1 = (0, self.class_interval)
+        rope_h2 = (self.class_range - self.class_interval, self.class_range)
+        rope_bak = int(self.class_range // 2)
+
+        # Normalize and scale attention maps for each speaker
+        max_values = x_ref_attn_map.max(1).values[:, None, None]
+        min_values = x_ref_attn_map.min(1).values[:, None, None]
+        max_min_values = torch.cat([max_values, min_values], dim=2)
+
+        human1_max_value, human1_min_value = max_min_values[0, :, 0].max(), max_min_values[0, :, 1].min()
+        human2_max_value, human2_min_value = max_min_values[1, :, 0].max(), max_min_values[1, :, 1].min()
+
+        human1 = normalize_and_scale(x_ref_attn_map[0], (human1_min_value, human1_max_value), rope_h1)
+        human2 = normalize_and_scale(x_ref_attn_map[1], (human2_min_value, human2_max_value), rope_h2)
+        back = torch.full((x_ref_attn_map.size(1),), rope_bak, dtype=human1.dtype, device=human1.device)
+
+        # Token-wise speaker dominance
+        max_indices = x_ref_attn_map.argmax(dim=0)
+        normalized_map = torch.stack([human1, human2, back], dim=1)
+        normalized_pos = normalized_map[torch.arange(x_ref_attn_map.size(1)), max_indices]
+
+        # Apply rotary to Q
+        q = rearrange(q, "(B N_t) H S C -> B H (N_t S) C", N_t=N_t)
+        q = self.rope_1d(q, normalized_pos)
+        q = rearrange(q, "B H (N_t S) C -> (B N_t) H S C", N_t=N_t)
+
+        # Keys / Values
+        _, N_a, _ = encoder_hidden_states.shape
+        encoder_kv = self.kv_linear(encoder_hidden_states)
+        encoder_kv = encoder_kv.view(B, N_a, 2, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
+        encoder_k, encoder_v = encoder_kv.unbind(0)
+
+        # Rotary for keys – assign centre of each speaker bucket to its context tokens
+        per_frame = torch.zeros(N_a, dtype=encoder_k.dtype, device=encoder_k.device)
+        per_frame[: per_frame.size(0) // 2] = (rope_h1[0] + rope_h1[1]) / 2
+        per_frame[per_frame.size(0) // 2 :] = (rope_h2[0] + rope_h2[1]) / 2
+        encoder_pos = torch.cat([per_frame] * N_t, dim=0)
+
+        encoder_k = rearrange(encoder_k, "(B N_t) H S C -> B H (N_t S) C", N_t=N_t)
+        encoder_k = self.rope_1d(encoder_k, encoder_pos)
+        encoder_k = rearrange(encoder_k, "B H (N_t S) C -> (B N_t) H S C", N_t=N_t)
+
+        # Final attention
+        q = rearrange(q, "B H M K -> B M H K")
+        encoder_k = rearrange(encoder_k, "B H M K -> B M H K")
+        encoder_v = rearrange(encoder_v, "B H M K -> B M H K")
+
+        x = optimized_attention(
+            q.transpose(1, 2),
+            encoder_k.transpose(1, 2),
+            encoder_v.transpose(1, 2),
+            heads=self.num_heads, skip_reshape=True, skip_output_reshape=True).transpose(1, 2)
+
+        # Linear projection
+        x = x.reshape(B, N, C)
+        x = self.proj(x)
+
+        # Restore original layout
+        x = rearrange(x, "(B N_t) S C -> B (N_t S) C", N_t=N_t)
+        if x_extra is not None:
+            x = torch.cat([x, torch.zeros_like(x_extra)], dim=1)
+
+        return x
+
+
+class MultiTalkAudioProjModel(torch.nn.Module):
+    def __init__(
+        self,
+        seq_len: int = 5,
+        seq_len_vf: int = 12,
+        blocks: int = 12,
+        channels: int = 768,
+        intermediate_dim: int = 512,
+        out_dim: int = 768,
+        context_tokens: int = 32,
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+
+        self.seq_len = seq_len
+        self.blocks = blocks
+        self.channels = channels
+        self.input_dim = seq_len * blocks * channels
+        self.input_dim_vf = seq_len_vf * blocks * channels
+        self.intermediate_dim = intermediate_dim
+        self.context_tokens = context_tokens
+        self.out_dim = out_dim
+
+        # define multiple linear layers
+        self.proj1 = operations.Linear(self.input_dim, intermediate_dim, device=device, dtype=dtype)
+        self.proj1_vf = operations.Linear(self.input_dim_vf, intermediate_dim, device=device, dtype=dtype)
+        self.proj2 = operations.Linear(intermediate_dim, intermediate_dim, device=device, dtype=dtype)
+        self.proj3 = operations.Linear(intermediate_dim, context_tokens * out_dim, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(out_dim, device=device, dtype=dtype)
+
+    def forward(self, audio_embeds, audio_embeds_vf):
+        video_length = audio_embeds.shape[1] + audio_embeds_vf.shape[1]
+        B, _, _, S, C = audio_embeds.shape
+
+        # process audio of first frame
+        audio_embeds = rearrange(audio_embeds, "bz f w b c -> (bz f) w b c")
+        batch_size, window_size, blocks, channels = audio_embeds.shape
+        audio_embeds = audio_embeds.view(batch_size, window_size * blocks * channels)
+
+        # process audio of latter frame
+        audio_embeds_vf = rearrange(audio_embeds_vf, "bz f w b c -> (bz f) w b c")
+        batch_size_vf, window_size_vf, blocks_vf, channels_vf = audio_embeds_vf.shape
+        audio_embeds_vf = audio_embeds_vf.view(batch_size_vf, window_size_vf * blocks_vf * channels_vf)
+
+        # first projection
+        audio_embeds = torch.relu(self.proj1(audio_embeds))
+        audio_embeds_vf = torch.relu(self.proj1_vf(audio_embeds_vf))
+        audio_embeds = rearrange(audio_embeds, "(bz f) c -> bz f c", bz=B)
+        audio_embeds_vf = rearrange(audio_embeds_vf, "(bz f) c -> bz f c", bz=B)
+        audio_embeds_c = torch.concat([audio_embeds, audio_embeds_vf], dim=1)
+        batch_size_c, N_t, C_a = audio_embeds_c.shape
+        audio_embeds_c = audio_embeds_c.view(batch_size_c*N_t, C_a)
+
+        # second projection
+        audio_embeds_c = torch.relu(self.proj2(audio_embeds_c))
+
+        context_tokens = self.proj3(audio_embeds_c).reshape(batch_size_c*N_t, self.context_tokens, self.out_dim)
+
+        # normalization and reshape
+        context_tokens = self.norm(context_tokens)
+        context_tokens = rearrange(context_tokens, "(bz f) m c -> bz f m c", f=video_length)
+
+        return context_tokens
+
+
+class WanMultiTalkAttentionBlock(torch.nn.Module):
+    def __init__(self, in_dim=5120, out_dim=768, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.audio_cross_attn = SingleStreamMultiAttention(in_dim, out_dim, num_heads=40, qkv_bias=True, device=device, dtype=dtype, operations=operations)
+        self.norm_x = operations.LayerNorm(in_dim, device=device, dtype=dtype, elementwise_affine=True)
+
+
+class MultiTalkGetAttnMapPatch:
+    def __init__(self, ref_target_masks=None):
+        self.ref_target_masks = ref_target_masks
+
+    def __call__(self, kwargs):
+        transformer_options = kwargs.get("transformer_options", {})
+        x = kwargs["x"]
+
+        if self.ref_target_masks is not None:
+            x_ref_attn_map = get_attn_map_with_target(kwargs["q"], kwargs["k"], transformer_options["grid_sizes"], ref_target_masks=self.ref_target_masks.to(x.device))
+            transformer_options["x_ref_attn_map"] = x_ref_attn_map
+        return x
+
+
+class MultiTalkCrossAttnPatch:
+    def __init__(self, model_patch, audio_scale=1.0, ref_target_masks=None):
+        self.model_patch = model_patch
+        self.audio_scale = audio_scale
+        self.ref_target_masks = ref_target_masks
+
+    def __call__(self, kwargs):
+        transformer_options = kwargs.get("transformer_options", {})
+        block_idx = transformer_options.get("block_idx", None)
+        x = kwargs["x"]
+        if block_idx is None:
+            return torch.zeros_like(x)
+
+        audio_embeds = transformer_options.get("audio_embeds")
+        x_ref_attn_map = transformer_options.pop("x_ref_attn_map", None)
+
+        norm_x = self.model_patch.model.blocks[block_idx].norm_x(x)
+        x_audio = self.model_patch.model.blocks[block_idx].audio_cross_attn(
+            norm_x, audio_embeds.to(x.dtype),
+            shape=transformer_options["grid_sizes"],
+            x_ref_attn_map=x_ref_attn_map
+        )
+        x = x + x_audio * self.audio_scale
+        return x
+
+    def models(self):
+        return [self.model_patch]
+
+class MultiTalkApplyModelWrapper:
+    def __init__(self, init_latents):
+        self.init_latents = init_latents
+
+    def __call__(self, executor, x, *args, **kwargs):
+        x[:, :, :self.init_latents.shape[2]] = self.init_latents.to(x)
+        samples = executor(x, *args, **kwargs)
+        return samples
+
+
+class InfiniteTalkOuterSampleWrapper:
+    def __init__(self, motion_frames_latent, model_patch, is_extend=False):
+        self.motion_frames_latent = motion_frames_latent
+        self.model_patch = model_patch
+        self.is_extend = is_extend
+
+    def __call__(self, executor, *args, **kwargs):
+        model_patcher = executor.class_obj.model_patcher
+        model_options = executor.class_obj.model_options
+        process_latent_in = model_patcher.model.process_latent_in
+
+        # for InfiniteTalk, model input first latent(s) need to always be replaced on every step
+        if self.motion_frames_latent is not None:
+            wrappers = model_options["transformer_options"]["wrappers"]
+            w = wrappers.setdefault(comfy.patcher_extension.WrappersMP.APPLY_MODEL, {})
+            w["MultiTalk_apply_model"] = [MultiTalkApplyModelWrapper(process_latent_in(self.motion_frames_latent))]
+
+        # run the sampling process
+        result = executor(*args, **kwargs)
+
+        # insert motion frames before decoding
+        if self.is_extend:
+            overlap = self.motion_frames_latent.shape[2]
+            result = torch.cat([self.motion_frames_latent.to(result), result[:, :, overlap:]], dim=2)
+
+        return result
+
+    def to(self, device_or_dtype):
+        if isinstance(device_or_dtype, torch.device):
+            if self.motion_frames_latent is not None:
+                self.motion_frames_latent = self.motion_frames_latent.to(device_or_dtype)
+        return self

comfy_api/latest/__init__.py

@@ -5,9 +5,9 @@ from typing import Type, TYPE_CHECKING
 from comfy_api.internal import ComfyAPIBase
 from comfy_api.internal.singleton import ProxiedSingleton
 from comfy_api.internal.async_to_sync import create_sync_class
-from comfy_api.latest._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput
-from comfy_api.latest._input_impl import VideoFromFile, VideoFromComponents
-from comfy_api.latest._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL
+from ._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput
+from ._input_impl import VideoFromFile, VideoFromComponents
+from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL
 from . import _io_public as io
 from . import _ui_public as ui
 # from comfy_api.latest._resources import _RESOURCES as resources  #noqa: F401
@@ -80,7 +80,7 @@ class ComfyExtension(ABC):
     async def on_load(self) -> None:
         """
         Called when an extension is loaded.
-        This should be used to initialize any global resources neeeded by the extension.
+        This should be used to initialize any global resources needed by the extension.
         """
 
     @abstractmethod

comfy_api/latest/_input/video_types.py

@@ -4,7 +4,7 @@ from fractions import Fraction
 from typing import Optional, Union, IO
 import io
 import av
-from comfy_api.util import VideoContainer, VideoCodec, VideoComponents
+from .._util import VideoContainer, VideoCodec, VideoComponents
 
 class VideoInput(ABC):
     """

comfy_api/latest/_input_impl/video_types.py

@@ -3,14 +3,14 @@ from av.container import InputContainer
 from av.subtitles.stream import SubtitleStream
 from fractions import Fraction
 from typing import Optional
-from comfy_api.latest._input import AudioInput, VideoInput
+from .._input import AudioInput, VideoInput
 import av
 import io
 import json
 import numpy as np
 import math
 import torch
-from comfy_api.latest._util import VideoContainer, VideoCodec, VideoComponents
+from .._util import VideoContainer, VideoCodec, VideoComponents
 
 
 def container_to_output_format(container_format: str | None) -> str | None:

comfy_api/latest/_io.py

@@ -26,7 +26,7 @@ if TYPE_CHECKING:
     from comfy_api.input import VideoInput
 from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classproperty, copy_class, first_real_override, is_class,
     prune_dict, shallow_clone_class)
-from comfy_api.latest._resources import Resources, ResourcesLocal
+from ._resources import Resources, ResourcesLocal
 from comfy_execution.graph_utils import ExecutionBlocker
 from ._util import MESH, VOXEL
 
@@ -763,7 +763,7 @@ class AnyType(ComfyTypeIO):
     Type = Any
 
 @comfytype(io_type="MODEL_PATCH")
-class MODEL_PATCH(ComfyTypeIO):
+class ModelPatch(ComfyTypeIO):
     Type = Any
 
 @comfytype(io_type="AUDIO_ENCODER")
@@ -1863,6 +1863,7 @@ __all__ = [
     "ControlNet",
     "Vae",
     "Model",
+    "ModelPatch",
     "ClipVision",
     "ClipVisionOutput",
     "AudioEncoder",

comfy_api/latest/_ui.py

@@ -22,7 +22,7 @@ import folder_paths
 
 # used for image preview
 from comfy.cli_args import args
-from comfy_api.latest._io import ComfyNode, FolderType, Image, _UIOutput
+from ._io import ComfyNode, FolderType, Image, _UIOutput
 
 
 class SavedResult(dict):

comfy_api/latest/_util/video_types.py

@@ -3,7 +3,7 @@ from dataclasses import dataclass
 from enum import Enum
 from fractions import Fraction
 from typing import Optional
-from comfy_api.latest._input import ImageInput, AudioInput
+from .._input import ImageInput, AudioInput
 
 class VideoCodec(str, Enum):
     AUTO = "auto"

comfy_api_nodes/apis/bytedance_api.py

@@ -0,0 +1,144 @@
+from typing import Literal
+
+from pydantic import BaseModel, Field
+
+
+class Text2ImageTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    prompt: str = Field(...)
+    response_format: str | None = Field("url")
+    size: str | None = Field(None)
+    seed: int | None = Field(0, ge=0, le=2147483647)
+    guidance_scale: float | None = Field(..., ge=1.0, le=10.0)
+    watermark: bool | None = Field(True)
+
+
+class Image2ImageTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    prompt: str = Field(...)
+    response_format: str | None = Field("url")
+    image: str = Field(..., description="Base64 encoded string or image URL")
+    size: str | None = Field("adaptive")
+    seed: int | None = Field(..., ge=0, le=2147483647)
+    guidance_scale: float | None = Field(..., ge=1.0, le=10.0)
+    watermark: bool | None = Field(True)
+
+
+class Seedream4Options(BaseModel):
+    max_images: int = Field(15)
+
+
+class Seedream4TaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    prompt: str = Field(...)
+    response_format: str = Field("url")
+    image: list[str] | None = Field(None, description="Image URLs")
+    size: str = Field(...)
+    seed: int = Field(..., ge=0, le=2147483647)
+    sequential_image_generation: str = Field("disabled")
+    sequential_image_generation_options: Seedream4Options = Field(Seedream4Options(max_images=15))
+    watermark: bool = Field(True)
+
+
+class ImageTaskCreationResponse(BaseModel):
+    model: str = Field(...)
+    created: int = Field(..., description="Unix timestamp (in seconds) indicating time when the request was created.")
+    data: list = Field([], description="Contains information about the generated image(s).")
+    error: dict = Field({}, description="Contains `code` and `message` fields in case of error.")
+
+
+class TaskTextContent(BaseModel):
+    type: str = Field("text")
+    text: str = Field(...)
+
+
+class TaskImageContentUrl(BaseModel):
+    url: str = Field(...)
+
+
+class TaskImageContent(BaseModel):
+    type: str = Field("image_url")
+    image_url: TaskImageContentUrl = Field(...)
+    role: Literal["first_frame", "last_frame", "reference_image"] | None = Field(None)
+
+
+class Text2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    content: list[TaskTextContent] = Field(..., min_length=1)
+
+
+class Image2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    content: list[TaskTextContent | TaskImageContent] = Field(..., min_length=2)
+
+
+class TaskCreationResponse(BaseModel):
+    id: str = Field(...)
+
+
+class TaskStatusError(BaseModel):
+    code: str = Field(...)
+    message: str = Field(...)
+
+
+class TaskStatusResult(BaseModel):
+    video_url: str = Field(...)
+
+
+class TaskStatusResponse(BaseModel):
+    id: str = Field(...)
+    model: str = Field(...)
+    status: Literal["queued", "running", "cancelled", "succeeded", "failed"] = Field(...)
+    error: TaskStatusError | None = Field(None)
+    content: TaskStatusResult | None = Field(None)
+
+
+RECOMMENDED_PRESETS = [
+    ("1024x1024 (1:1)", 1024, 1024),
+    ("864x1152 (3:4)", 864, 1152),
+    ("1152x864 (4:3)", 1152, 864),
+    ("1280x720 (16:9)", 1280, 720),
+    ("720x1280 (9:16)", 720, 1280),
+    ("832x1248 (2:3)", 832, 1248),
+    ("1248x832 (3:2)", 1248, 832),
+    ("1512x648 (21:9)", 1512, 648),
+    ("2048x2048 (1:1)", 2048, 2048),
+    ("Custom", None, None),
+]
+
+RECOMMENDED_PRESETS_SEEDREAM_4 = [
+    ("2048x2048 (1:1)", 2048, 2048),
+    ("2304x1728 (4:3)", 2304, 1728),
+    ("1728x2304 (3:4)", 1728, 2304),
+    ("2560x1440 (16:9)", 2560, 1440),
+    ("1440x2560 (9:16)", 1440, 2560),
+    ("2496x1664 (3:2)", 2496, 1664),
+    ("1664x2496 (2:3)", 1664, 2496),
+    ("3024x1296 (21:9)", 3024, 1296),
+    ("4096x4096 (1:1)", 4096, 4096),
+    ("Custom", None, None),
+]
+
+# The time in this dictionary are given for 10 seconds duration.
+VIDEO_TASKS_EXECUTION_TIME = {
+    "seedance-1-0-lite-t2v-250428": {
+        "480p": 40,
+        "720p": 60,
+        "1080p": 90,
+    },
+    "seedance-1-0-lite-i2v-250428": {
+        "480p": 40,
+        "720p": 60,
+        "1080p": 90,
+    },
+    "seedance-1-0-pro-250528": {
+        "480p": 70,
+        "720p": 85,
+        "1080p": 115,
+    },
+    "seedance-1-0-pro-fast-251015": {
+        "480p": 50,
+        "720p": 65,
+        "1080p": 100,
+    },
+}

comfy_api_nodes/apis/gemini_api.py

@@ -84,15 +84,7 @@ class GeminiSystemInstructionContent(BaseModel):
         description="A list of ordered parts that make up a single message. "
         "Different parts may have different IANA MIME types.",
     )
-    role: GeminiRole = Field(
-        ...,
-        description="The identity of the entity that creates the message. "
-        "The following values are supported: "
-        "user: This indicates that the message is sent by a real person, typically a user-generated message. "
-        "model: This indicates that the message is generated by the model. "
-        "The model value is used to insert messages from model into the conversation during multi-turn conversations. "
-        "For non-multi-turn conversations, this field can be left blank or unset.",
-    )
+    role: GeminiRole | None = Field(..., description="The role field of systemInstruction may be ignored.")
 
 
 class GeminiFunctionDeclaration(BaseModel):

comfy_api_nodes/apis/veo_api.py

@@ -85,7 +85,7 @@ class Response1(BaseModel):
     raiMediaFilteredReasons: Optional[list[str]] = Field(
         None, description='Reasons why media was filtered by responsible AI policies'
     )
-    videos: Optional[list[Video]] = None
+    videos: Optional[list[Video]] = Field(None)
 
 
 class VeoGenVidPollResponse(BaseModel):

comfy_api_nodes/nodes_bytedance.py

@@ -1,13 +1,27 @@
 import logging
 import math
-from enum import Enum
-from typing import Literal, Optional, Union
 
 import torch
-from pydantic import BaseModel, Field
 from typing_extensions import override
 
-from comfy_api.latest import IO, ComfyExtension
+from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api_nodes.apis.bytedance_api import (
+    RECOMMENDED_PRESETS,
+    RECOMMENDED_PRESETS_SEEDREAM_4,
+    VIDEO_TASKS_EXECUTION_TIME,
+    Image2ImageTaskCreationRequest,
+    Image2VideoTaskCreationRequest,
+    ImageTaskCreationResponse,
+    Seedream4Options,
+    Seedream4TaskCreationRequest,
+    TaskCreationResponse,
+    TaskImageContent,
+    TaskImageContentUrl,
+    TaskStatusResponse,
+    TaskTextContent,
+    Text2ImageTaskCreationRequest,
+    Text2VideoTaskCreationRequest,
+)
 from comfy_api_nodes.util import (
     ApiEndpoint,
     download_url_to_image_tensor,
@@ -29,162 +43,6 @@ BYTEPLUS_TASK_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks"
 BYTEPLUS_TASK_STATUS_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks"  # + /{task_id}
 
 
-class Text2ImageModelName(str, Enum):
-    seedream_3 = "seedream-3-0-t2i-250415"
-
-
-class Image2ImageModelName(str, Enum):
-    seededit_3 = "seededit-3-0-i2i-250628"
-
-
-class Text2VideoModelName(str, Enum):
-    seedance_1_pro = "seedance-1-0-pro-250528"
-    seedance_1_lite = "seedance-1-0-lite-t2v-250428"
-
-
-class Image2VideoModelName(str, Enum):
-    """note(August 31): Pro model only supports FirstFrame: https://docs.byteplus.com/en/docs/ModelArk/1520757"""
-
-    seedance_1_pro = "seedance-1-0-pro-250528"
-    seedance_1_lite = "seedance-1-0-lite-i2v-250428"
-
-
-class Text2ImageTaskCreationRequest(BaseModel):
-    model: Text2ImageModelName = Text2ImageModelName.seedream_3
-    prompt: str = Field(...)
-    response_format: Optional[str] = Field("url")
-    size: Optional[str] = Field(None)
-    seed: Optional[int] = Field(0, ge=0, le=2147483647)
-    guidance_scale: Optional[float] = Field(..., ge=1.0, le=10.0)
-    watermark: Optional[bool] = Field(True)
-
-
-class Image2ImageTaskCreationRequest(BaseModel):
-    model: Image2ImageModelName = Image2ImageModelName.seededit_3
-    prompt: str = Field(...)
-    response_format: Optional[str] = Field("url")
-    image: str = Field(..., description="Base64 encoded string or image URL")
-    size: Optional[str] = Field("adaptive")
-    seed: Optional[int] = Field(..., ge=0, le=2147483647)
-    guidance_scale: Optional[float] = Field(..., ge=1.0, le=10.0)
-    watermark: Optional[bool] = Field(True)
-
-
-class Seedream4Options(BaseModel):
-    max_images: int = Field(15)
-
-
-class Seedream4TaskCreationRequest(BaseModel):
-    model: str = Field("seedream-4-0-250828")
-    prompt: str = Field(...)
-    response_format: str = Field("url")
-    image: Optional[list[str]] = Field(None, description="Image URLs")
-    size: str = Field(...)
-    seed: int = Field(..., ge=0, le=2147483647)
-    sequential_image_generation: str = Field("disabled")
-    sequential_image_generation_options: Seedream4Options = Field(Seedream4Options(max_images=15))
-    watermark: bool = Field(True)
-
-
-class ImageTaskCreationResponse(BaseModel):
-    model: str = Field(...)
-    created: int = Field(..., description="Unix timestamp (in seconds) indicating time when the request was created.")
-    data: list = Field([], description="Contains information about the generated image(s).")
-    error: dict = Field({}, description="Contains `code` and `message` fields in case of error.")
-
-
-class TaskTextContent(BaseModel):
-    type: str = Field("text")
-    text: str = Field(...)
-
-
-class TaskImageContentUrl(BaseModel):
-    url: str = Field(...)
-
-
-class TaskImageContent(BaseModel):
-    type: str = Field("image_url")
-    image_url: TaskImageContentUrl = Field(...)
-    role: Optional[Literal["first_frame", "last_frame", "reference_image"]] = Field(None)
-
-
-class Text2VideoTaskCreationRequest(BaseModel):
-    model: Text2VideoModelName = Text2VideoModelName.seedance_1_pro
-    content: list[TaskTextContent] = Field(..., min_length=1)
-
-
-class Image2VideoTaskCreationRequest(BaseModel):
-    model: Image2VideoModelName = Image2VideoModelName.seedance_1_pro
-    content: list[Union[TaskTextContent, TaskImageContent]] = Field(..., min_length=2)
-
-
-class TaskCreationResponse(BaseModel):
-    id: str = Field(...)
-
-
-class TaskStatusError(BaseModel):
-    code: str = Field(...)
-    message: str = Field(...)
-
-
-class TaskStatusResult(BaseModel):
-    video_url: str = Field(...)
-
-
-class TaskStatusResponse(BaseModel):
-    id: str = Field(...)
-    model: str = Field(...)
-    status: Literal["queued", "running", "cancelled", "succeeded", "failed"] = Field(...)
-    error: Optional[TaskStatusError] = Field(None)
-    content: Optional[TaskStatusResult] = Field(None)
-
-
-RECOMMENDED_PRESETS = [
-    ("1024x1024 (1:1)", 1024, 1024),
-    ("864x1152 (3:4)", 864, 1152),
-    ("1152x864 (4:3)", 1152, 864),
-    ("1280x720 (16:9)", 1280, 720),
-    ("720x1280 (9:16)", 720, 1280),
-    ("832x1248 (2:3)", 832, 1248),
-    ("1248x832 (3:2)", 1248, 832),
-    ("1512x648 (21:9)", 1512, 648),
-    ("2048x2048 (1:1)", 2048, 2048),
-    ("Custom", None, None),
-]
-
-RECOMMENDED_PRESETS_SEEDREAM_4 = [
-    ("2048x2048 (1:1)", 2048, 2048),
-    ("2304x1728 (4:3)", 2304, 1728),
-    ("1728x2304 (3:4)", 1728, 2304),
-    ("2560x1440 (16:9)", 2560, 1440),
-    ("1440x2560 (9:16)", 1440, 2560),
-    ("2496x1664 (3:2)", 2496, 1664),
-    ("1664x2496 (2:3)", 1664, 2496),
-    ("3024x1296 (21:9)", 3024, 1296),
-    ("4096x4096 (1:1)", 4096, 4096),
-    ("Custom", None, None),
-]
-
-# The time in this dictionary are given for 10 seconds duration.
-VIDEO_TASKS_EXECUTION_TIME = {
-    "seedance-1-0-lite-t2v-250428": {
-        "480p": 40,
-        "720p": 60,
-        "1080p": 90,
-    },
-    "seedance-1-0-lite-i2v-250428": {
-        "480p": 40,
-        "720p": 60,
-        "1080p": 90,
-    },
-    "seedance-1-0-pro-250528": {
-        "480p": 70,
-        "720p": 85,
-        "1080p": 115,
-    },
-}
-
-
 def get_image_url_from_response(response: ImageTaskCreationResponse) -> str:
     if response.error:
         error_msg = f"ByteDance request failed. Code: {response.error['code']}, message: {response.error['message']}"
@@ -194,13 +52,6 @@ def get_image_url_from_response(response: ImageTaskCreationResponse) -> str:
     return response.data[0]["url"]
 
 
-def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, None]:
-    """Returns the video URL from the task status response if it exists."""
-    if hasattr(response, "content") and response.content:
-        return response.content.video_url
-    return None
-
-
 class ByteDanceImageNode(IO.ComfyNode):
 
     @classmethod
@@ -211,12 +62,7 @@ class ByteDanceImageNode(IO.ComfyNode):
             category="api node/image/ByteDance",
             description="Generate images using ByteDance models via api based on prompt",
             inputs=[
-                IO.Combo.Input(
-                    "model",
-                    options=Text2ImageModelName,
-                    default=Text2ImageModelName.seedream_3,
-                    tooltip="Model name",
-                ),
+                IO.Combo.Input("model", options=["seedream-3-0-t2i-250415"]),
                 IO.String.Input(
                     "prompt",
                     multiline=True,
@@ -335,12 +181,7 @@ class ByteDanceImageEditNode(IO.ComfyNode):
             category="api node/image/ByteDance",
             description="Edit images using ByteDance models via api based on prompt",
             inputs=[
-                IO.Combo.Input(
-                    "model",
-                    options=Image2ImageModelName,
-                    default=Image2ImageModelName.seededit_3,
-                    tooltip="Model name",
-                ),
+                IO.Combo.Input("model", options=["seededit-3-0-i2i-250628"]),
                 IO.Image.Input(
                     "image",
                     tooltip="The base image to edit",
@@ -394,7 +235,7 @@ class ByteDanceImageEditNode(IO.ComfyNode):
     async def execute(
         cls,
         model: str,
-        image: torch.Tensor,
+        image: Input.Image,
         prompt: str,
         seed: int,
         guidance_scale: float,
@@ -434,7 +275,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
             inputs=[
                 IO.Combo.Input(
                     "model",
-                    options=["seedream-4-0-250828"],
+                    options=["seedream-4-5-251128", "seedream-4-0-250828"],
                     tooltip="Model name",
                 ),
                 IO.String.Input(
@@ -459,7 +300,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                     default=2048,
                     min=1024,
                     max=4096,
-                    step=64,
+                    step=8,
                     tooltip="Custom width for image. Value is working only if `size_preset` is set to `Custom`",
                     optional=True,
                 ),
@@ -468,7 +309,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                     default=2048,
                     min=1024,
                     max=4096,
-                    step=64,
+                    step=8,
                     tooltip="Custom height for image. Value is working only if `size_preset` is set to `Custom`",
                     optional=True,
                 ),
@@ -532,7 +373,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
         cls,
         model: str,
         prompt: str,
-        image: torch.Tensor = None,
+        image: Input.Image | None = None,
         size_preset: str = RECOMMENDED_PRESETS_SEEDREAM_4[0][0],
         width: int = 2048,
         height: int = 2048,
@@ -555,6 +396,18 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                 raise ValueError(
                     f"Custom size out of range: {w}x{h}. " "Both width and height must be between 1024 and 4096 pixels."
                 )
+        out_num_pixels = w * h
+        mp_provided = out_num_pixels / 1_000_000.0
+        if "seedream-4-5" in model and out_num_pixels < 3686400:
+            raise ValueError(
+                f"Minimum image resolution that Seedream 4.5 can generate is 3.68MP, "
+                f"but {mp_provided:.2f}MP provided."
+            )
+        if "seedream-4-0" in model and out_num_pixels < 921600:
+            raise ValueError(
+                f"Minimum image resolution that the selected model can generate is 0.92MP, "
+                f"but {mp_provided:.2f}MP provided."
+            )
         n_input_images = get_number_of_images(image) if image is not None else 0
         if n_input_images > 10:
             raise ValueError(f"Maximum of 10 reference images are supported, but {n_input_images} received.")
@@ -607,9 +460,8 @@ class ByteDanceTextToVideoNode(IO.ComfyNode):
             inputs=[
                 IO.Combo.Input(
                     "model",
-                    options=Text2VideoModelName,
-                    default=Text2VideoModelName.seedance_1_pro,
-                    tooltip="Model name",
+                    options=["seedance-1-0-pro-250528", "seedance-1-0-lite-t2v-250428", "seedance-1-0-pro-fast-251015"],
+                    default="seedance-1-0-pro-fast-251015",
                 ),
                 IO.String.Input(
                     "prompt",
@@ -714,9 +566,8 @@ class ByteDanceImageToVideoNode(IO.ComfyNode):
             inputs=[
                 IO.Combo.Input(
                     "model",
-                    options=Image2VideoModelName,
-                    default=Image2VideoModelName.seedance_1_pro,
-                    tooltip="Model name",
+                    options=["seedance-1-0-pro-250528", "seedance-1-0-lite-t2v-250428", "seedance-1-0-pro-fast-251015"],
+                    default="seedance-1-0-pro-fast-251015",
                 ),
                 IO.String.Input(
                     "prompt",
@@ -787,7 +638,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode):
         cls,
         model: str,
         prompt: str,
-        image: torch.Tensor,
+        image: Input.Image,
         resolution: str,
         aspect_ratio: str,
         duration: int,
@@ -833,9 +684,8 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode):
             inputs=[
                 IO.Combo.Input(
                     "model",
-                    options=[model.value for model in Image2VideoModelName],
-                    default=Image2VideoModelName.seedance_1_lite.value,
-                    tooltip="Model name",
+                    options=["seedance-1-0-pro-250528", "seedance-1-0-lite-i2v-250428"],
+                    default="seedance-1-0-lite-i2v-250428",
                 ),
                 IO.String.Input(
                     "prompt",
@@ -910,8 +760,8 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode):
         cls,
         model: str,
         prompt: str,
-        first_frame: torch.Tensor,
-        last_frame: torch.Tensor,
+        first_frame: Input.Image,
+        last_frame: Input.Image,
         resolution: str,
         aspect_ratio: str,
         duration: int,
@@ -968,9 +818,8 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
             inputs=[
                 IO.Combo.Input(
                     "model",
-                    options=[Image2VideoModelName.seedance_1_lite.value],
-                    default=Image2VideoModelName.seedance_1_lite.value,
-                    tooltip="Model name",
+                    options=["seedance-1-0-pro-250528", "seedance-1-0-lite-i2v-250428"],
+                    default="seedance-1-0-lite-i2v-250428",
                 ),
                 IO.String.Input(
                     "prompt",
@@ -1034,7 +883,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
         cls,
         model: str,
         prompt: str,
-        images: torch.Tensor,
+        images: Input.Image,
         resolution: str,
         aspect_ratio: str,
         duration: int,
@@ -1069,8 +918,8 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
 
 async def process_video_task(
     cls: type[IO.ComfyNode],
-    payload: Union[Text2VideoTaskCreationRequest, Image2VideoTaskCreationRequest],
-    estimated_duration: Optional[int],
+    payload: Text2VideoTaskCreationRequest | Image2VideoTaskCreationRequest,
+    estimated_duration: int | None,
 ) -> IO.NodeOutput:
     initial_response = await sync_op(
         cls,
@@ -1085,7 +934,7 @@ async def process_video_task(
         estimated_duration=estimated_duration,
         response_model=TaskStatusResponse,
     )
-    return IO.NodeOutput(await download_url_to_video_output(get_video_url_from_task_status(response)))
+    return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
 
 
 def raise_if_text_params(prompt: str, text_params: list[str]) -> None:

comfy_api_nodes/nodes_gemini.py

@@ -13,8 +13,7 @@ import torch
 from typing_extensions import override
 
 import folder_paths
-from comfy_api.latest import IO, ComfyExtension, Input
-from comfy_api.util import VideoCodec, VideoContainer
+from comfy_api.latest import IO, ComfyExtension, Input, Types
 from comfy_api_nodes.apis.gemini_api import (
     GeminiContent,
     GeminiFileData,
@@ -27,6 +26,8 @@ from comfy_api_nodes.apis.gemini_api import (
     GeminiMimeType,
     GeminiPart,
     GeminiRole,
+    GeminiSystemInstructionContent,
+    GeminiTextPart,
     Modality,
 )
 from comfy_api_nodes.util import (
@@ -43,6 +44,14 @@ from comfy_api_nodes.util import (
 
 GEMINI_BASE_ENDPOINT = "/proxy/vertexai/gemini"
 GEMINI_MAX_INPUT_FILE_SIZE = 20 * 1024 * 1024  # 20 MB
+GEMINI_IMAGE_SYS_PROMPT = (
+    "You are an expert image-generation engine. You must ALWAYS produce an image.\n"
+    "Interpret all user input—regardless of "
+    "format, intent, or abstraction—as literal visual directives for image composition.\n"
+    "If a prompt is conversational or lacks specific visual details, "
+    "you must creatively invent a concrete visual scenario that depicts the concept.\n"
+    "Prioritize generating the visual representation above any text, formatting, or conversational requests."
+)
 
 
 class GeminiModel(str, Enum):
@@ -68,7 +77,7 @@ class GeminiImageModel(str, Enum):
 
 async def create_image_parts(
     cls: type[IO.ComfyNode],
-    images: torch.Tensor,
+    images: Input.Image,
     image_limit: int = 0,
 ) -> list[GeminiPart]:
     image_parts: list[GeminiPart] = []
@@ -154,8 +163,8 @@ def get_text_from_response(response: GeminiGenerateContentResponse) -> str:
     return "\n".join([part.text for part in parts])
 
 
-def get_image_from_response(response: GeminiGenerateContentResponse) -> torch.Tensor:
-    image_tensors: list[torch.Tensor] = []
+def get_image_from_response(response: GeminiGenerateContentResponse) -> Input.Image:
+    image_tensors: list[Input.Image] = []
     parts = get_parts_by_type(response, "image/png")
     for part in parts:
         image_data = base64.b64decode(part.inlineData.data)
@@ -277,6 +286,13 @@ class GeminiNode(IO.ComfyNode):
                     tooltip="Optional file(s) to use as context for the model. "
                     "Accepts inputs from the Gemini Generate Content Input Files node.",
                 ),
+                IO.String.Input(
+                    "system_prompt",
+                    multiline=True,
+                    default="",
+                    optional=True,
+                    tooltip="Foundational instructions that dictate an AI's behavior.",
+                ),
             ],
             outputs=[
                 IO.String.Output(),
@@ -293,7 +309,9 @@ class GeminiNode(IO.ComfyNode):
     def create_video_parts(cls, video_input: Input.Video) -> list[GeminiPart]:
         """Convert video input to Gemini API compatible parts."""
 
-        base_64_string = video_to_base64_string(video_input, container_format=VideoContainer.MP4, codec=VideoCodec.H264)
+        base_64_string = video_to_base64_string(
+            video_input, container_format=Types.VideoContainer.MP4, codec=Types.VideoCodec.H264
+        )
         return [
             GeminiPart(
                 inlineData=GeminiInlineData(
@@ -343,10 +361,11 @@ class GeminiNode(IO.ComfyNode):
         prompt: str,
         model: str,
         seed: int,
-        images: torch.Tensor | None = None,
+        images: Input.Image | None = None,
         audio: Input.Audio | None = None,
         video: Input.Video | None = None,
         files: list[GeminiPart] | None = None,
+        system_prompt: str = "",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=False)
 
@@ -363,7 +382,10 @@ class GeminiNode(IO.ComfyNode):
         if files is not None:
             parts.extend(files)
 
-        # Create response
+        gemini_system_prompt = None
+        if system_prompt:
+            gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None)
+
         response = await sync_op(
             cls,
             endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"),
@@ -373,7 +395,8 @@ class GeminiNode(IO.ComfyNode):
                         role=GeminiRole.user,
                         parts=parts,
                     )
-                ]
+                ],
+                systemInstruction=gemini_system_prompt,
             ),
             response_model=GeminiGenerateContentResponse,
             price_extractor=calculate_tokens_price,
@@ -523,6 +546,13 @@ class GeminiImage(IO.ComfyNode):
                     "'IMAGE+TEXT' to return both the generated image and a text response.",
                     optional=True,
                 ),
+                IO.String.Input(
+                    "system_prompt",
+                    multiline=True,
+                    default=GEMINI_IMAGE_SYS_PROMPT,
+                    optional=True,
+                    tooltip="Foundational instructions that dictate an AI's behavior.",
+                ),
             ],
             outputs=[
                 IO.Image.Output(),
@@ -542,10 +572,11 @@ class GeminiImage(IO.ComfyNode):
         prompt: str,
         model: str,
         seed: int,
-        images: torch.Tensor | None = None,
+        images: Input.Image | None = None,
         files: list[GeminiPart] | None = None,
         aspect_ratio: str = "auto",
         response_modalities: str = "IMAGE+TEXT",
+        system_prompt: str = "",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
         parts: list[GeminiPart] = [GeminiPart(text=prompt)]
@@ -559,6 +590,10 @@ class GeminiImage(IO.ComfyNode):
         if files is not None:
             parts.extend(files)
 
+        gemini_system_prompt = None
+        if system_prompt:
+            gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None)
+
         response = await sync_op(
             cls,
             endpoint=ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"),
@@ -570,6 +605,7 @@ class GeminiImage(IO.ComfyNode):
                     responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]),
                     imageConfig=None if aspect_ratio == "auto" else image_config,
                 ),
+                systemInstruction=gemini_system_prompt,
             ),
             response_model=GeminiGenerateContentResponse,
             price_extractor=calculate_tokens_price,
@@ -640,6 +676,13 @@ class GeminiImage2(IO.ComfyNode):
                     tooltip="Optional file(s) to use as context for the model. "
                     "Accepts inputs from the Gemini Generate Content Input Files node.",
                 ),
+                IO.String.Input(
+                    "system_prompt",
+                    multiline=True,
+                    default=GEMINI_IMAGE_SYS_PROMPT,
+                    optional=True,
+                    tooltip="Foundational instructions that dictate an AI's behavior.",
+                ),
             ],
             outputs=[
                 IO.Image.Output(),
@@ -662,8 +705,9 @@ class GeminiImage2(IO.ComfyNode):
         aspect_ratio: str,
         resolution: str,
         response_modalities: str,
-        images: torch.Tensor | None = None,
+        images: Input.Image | None = None,
         files: list[GeminiPart] | None = None,
+        system_prompt: str = "",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
 
@@ -679,6 +723,10 @@ class GeminiImage2(IO.ComfyNode):
         if aspect_ratio != "auto":
             image_config.aspectRatio = aspect_ratio
 
+        gemini_system_prompt = None
+        if system_prompt:
+            gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None)
+
         response = await sync_op(
             cls,
             ApiEndpoint(path=f"{GEMINI_BASE_ENDPOINT}/{model}", method="POST"),
@@ -690,6 +738,7 @@ class GeminiImage2(IO.ComfyNode):
                     responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]),
                     imageConfig=image_config,
                 ),
+                systemInstruction=gemini_system_prompt,
             ),
             response_model=GeminiGenerateContentResponse,
             price_extractor=calculate_tokens_price,

comfy_api_nodes/nodes_ltxv.py

@@ -1,12 +1,9 @@
 from io import BytesIO
-from typing import Optional
 
-import torch
 from pydantic import BaseModel, Field
 from typing_extensions import override
 
-from comfy_api.input_impl import VideoFromFile
-from comfy_api.latest import IO, ComfyExtension
+from comfy_api.latest import IO, ComfyExtension, Input, InputImpl
 from comfy_api_nodes.util import (
     ApiEndpoint,
     get_number_of_images,
@@ -26,9 +23,9 @@ class ExecuteTaskRequest(BaseModel):
     model: str = Field(...)
     duration: int = Field(...)
     resolution: str = Field(...)
-    fps: Optional[int] = Field(25)
-    generate_audio: Optional[bool] = Field(True)
-    image_uri: Optional[str] = Field(None)
+    fps: int | None = Field(25)
+    generate_audio: bool | None = Field(True)
+    image_uri: str | None = Field(None)
 
 
 class TextToVideoNode(IO.ComfyNode):
@@ -103,7 +100,7 @@ class TextToVideoNode(IO.ComfyNode):
             as_binary=True,
             max_retries=1,
         )
-        return IO.NodeOutput(VideoFromFile(BytesIO(response)))
+        return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(response)))
 
 
 class ImageToVideoNode(IO.ComfyNode):
@@ -153,7 +150,7 @@ class ImageToVideoNode(IO.ComfyNode):
     @classmethod
     async def execute(
         cls,
-        image: torch.Tensor,
+        image: Input.Image,
         model: str,
         prompt: str,
         duration: int,
@@ -183,7 +180,7 @@ class ImageToVideoNode(IO.ComfyNode):
             as_binary=True,
             max_retries=1,
         )
-        return IO.NodeOutput(VideoFromFile(BytesIO(response)))
+        return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(response)))
 
 
 class LtxvApiExtension(ComfyExtension):

comfy_api_nodes/nodes_moonvalley.py

@@ -1,11 +1,8 @@
 import logging
-from typing import Optional
 
-import torch
 from typing_extensions import override
 
-from comfy_api.input import VideoInput
-from comfy_api.latest import IO, ComfyExtension
+from comfy_api.latest import IO, ComfyExtension, Input
 from comfy_api_nodes.apis import (
     MoonvalleyPromptResponse,
     MoonvalleyTextToVideoInferenceParams,
@@ -61,7 +58,7 @@ def validate_task_creation_response(response) -> None:
         raise RuntimeError(error_msg)
 
 
-def validate_video_to_video_input(video: VideoInput) -> VideoInput:
+def validate_video_to_video_input(video: Input.Video) -> Input.Video:
     """
     Validates and processes video input for Moonvalley Video-to-Video generation.
 
@@ -82,7 +79,7 @@ def validate_video_to_video_input(video: VideoInput) -> VideoInput:
     return _validate_and_trim_duration(video)
 
 
-def _get_video_dimensions(video: VideoInput) -> tuple[int, int]:
+def _get_video_dimensions(video: Input.Video) -> tuple[int, int]:
     """Extracts video dimensions with error handling."""
     try:
         return video.get_dimensions()
@@ -106,7 +103,7 @@ def _validate_video_dimensions(width: int, height: int) -> None:
         raise ValueError(f"Resolution {width}x{height} not supported. Supported: {supported_list}")
 
 
-def _validate_and_trim_duration(video: VideoInput) -> VideoInput:
+def _validate_and_trim_duration(video: Input.Video) -> Input.Video:
     """Validates video duration and trims to 5 seconds if needed."""
     duration = video.get_duration()
     _validate_minimum_duration(duration)
@@ -119,7 +116,7 @@ def _validate_minimum_duration(duration: float) -> None:
         raise ValueError("Input video must be at least 5 seconds long.")
 
 
-def _trim_if_too_long(video: VideoInput, duration: float) -> VideoInput:
+def _trim_if_too_long(video: Input.Video, duration: float) -> Input.Video:
     """Trims video to 5 seconds if longer."""
     if duration > 5:
         return trim_video(video, 5)
@@ -241,7 +238,7 @@ class MoonvalleyImg2VideoNode(IO.ComfyNode):
     @classmethod
     async def execute(
         cls,
-        image: torch.Tensor,
+        image: Input.Image,
         prompt: str,
         negative_prompt: str,
         resolution: str,
@@ -362,9 +359,9 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode):
         prompt: str,
         negative_prompt: str,
         seed: int,
-        video: Optional[VideoInput] = None,
+        video: Input.Video | None = None,
         control_type: str = "Motion Transfer",
-        motion_intensity: Optional[int] = 100,
+        motion_intensity: int | None = 100,
         steps=33,
         prompt_adherence=4.5,
     ) -> IO.NodeOutput:

comfy_api_nodes/nodes_runway.py

@@ -11,12 +11,11 @@ User Guides:
 
 """
 
-from typing import Union, Optional
-from typing_extensions import override
 from enum import Enum
 
-import torch
+from typing_extensions import override
 
+from comfy_api.latest import IO, ComfyExtension, Input, InputImpl
 from comfy_api_nodes.apis import (
     RunwayImageToVideoRequest,
     RunwayImageToVideoResponse,
@@ -44,8 +43,6 @@ from comfy_api_nodes.util import (
     sync_op,
     poll_op,
 )
-from comfy_api.input_impl import VideoFromFile
-from comfy_api.latest import ComfyExtension, IO
 
 PATH_IMAGE_TO_VIDEO = "/proxy/runway/image_to_video"
 PATH_TEXT_TO_IMAGE = "/proxy/runway/text_to_image"
@@ -80,7 +77,7 @@ class RunwayGen3aAspectRatio(str, Enum):
     field_1280_768 = "1280:768"
 
 
-def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, None]:
+def get_video_url_from_task_status(response: TaskStatusResponse) -> str | None:
     """Returns the video URL from the task status response if it exists."""
     if hasattr(response, "output") and len(response.output) > 0:
         return response.output[0]
@@ -89,13 +86,13 @@ def get_video_url_from_task_status(response: TaskStatusResponse) -> Union[str, N
 
 def extract_progress_from_task_status(
     response: TaskStatusResponse,
-) -> Union[float, None]:
+) -> float | None:
     if hasattr(response, "progress") and response.progress is not None:
         return response.progress * 100
     return None
 
 
-def get_image_url_from_task_status(response: TaskStatusResponse) -> Union[str, None]:
+def get_image_url_from_task_status(response: TaskStatusResponse) -> str | None:
     """Returns the image URL from the task status response if it exists."""
     if hasattr(response, "output") and len(response.output) > 0:
         return response.output[0]
@@ -103,7 +100,7 @@ def get_image_url_from_task_status(response: TaskStatusResponse) -> Union[str, N
 
 
 async def get_response(
-    cls: type[IO.ComfyNode], task_id: str, estimated_duration: Optional[int] = None
+    cls: type[IO.ComfyNode], task_id: str, estimated_duration: int | None = None
 ) -> TaskStatusResponse:
     """Poll the task status until it is finished then get the response."""
     return await poll_op(
@@ -119,8 +116,8 @@ async def get_response(
 async def generate_video(
     cls: type[IO.ComfyNode],
     request: RunwayImageToVideoRequest,
-    estimated_duration: Optional[int] = None,
-) -> VideoFromFile:
+    estimated_duration: int | None = None,
+) -> InputImpl.VideoFromFile:
     initial_response = await sync_op(
         cls,
         endpoint=ApiEndpoint(path=PATH_IMAGE_TO_VIDEO, method="POST"),
@@ -193,7 +190,7 @@ class RunwayImageToVideoNodeGen3a(IO.ComfyNode):
     async def execute(
         cls,
         prompt: str,
-        start_frame: torch.Tensor,
+        start_frame: Input.Image,
         duration: str,
         ratio: str,
         seed: int,
@@ -283,7 +280,7 @@ class RunwayImageToVideoNodeGen4(IO.ComfyNode):
     async def execute(
         cls,
         prompt: str,
-        start_frame: torch.Tensor,
+        start_frame: Input.Image,
         duration: str,
         ratio: str,
         seed: int,
@@ -381,8 +378,8 @@ class RunwayFirstLastFrameNode(IO.ComfyNode):
     async def execute(
         cls,
         prompt: str,
-        start_frame: torch.Tensor,
-        end_frame: torch.Tensor,
+        start_frame: Input.Image,
+        end_frame: Input.Image,
         duration: str,
         ratio: str,
         seed: int,
@@ -467,7 +464,7 @@ class RunwayTextToImageNode(IO.ComfyNode):
         cls,
         prompt: str,
         ratio: str,
-        reference_image: Optional[torch.Tensor] = None,
+        reference_image: Input.Image | None = None,
     ) -> IO.NodeOutput:
         validate_string(prompt, min_length=1)
 

comfy_api_nodes/nodes_veo2.py

@@ -1,11 +1,9 @@
 import base64
 from io import BytesIO
 
-import torch
 from typing_extensions import override
 
-from comfy_api.input_impl.video_types import VideoFromFile
-from comfy_api.latest import IO, ComfyExtension
+from comfy_api.latest import IO, ComfyExtension, Input, InputImpl
 from comfy_api_nodes.apis.veo_api import (
     VeoGenVidPollRequest,
     VeoGenVidPollResponse,
@@ -232,7 +230,7 @@ class VeoVideoGenerationNode(IO.ComfyNode):
 
             # Check if video is provided as base64 or URL
             if hasattr(video, "bytesBase64Encoded") and video.bytesBase64Encoded:
-                return IO.NodeOutput(VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded))))
+                return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded))))
 
             if hasattr(video, "gcsUri") and video.gcsUri:
                 return IO.NodeOutput(await download_url_to_video_output(video.gcsUri))
@@ -431,8 +429,8 @@ class Veo3FirstLastFrameNode(IO.ComfyNode):
         aspect_ratio: str,
         duration: int,
         seed: int,
-        first_frame: torch.Tensor,
-        last_frame: torch.Tensor,
+        first_frame: Input.Image,
+        last_frame: Input.Image,
         model: str,
         generate_audio: bool,
     ):
@@ -493,7 +491,7 @@ class Veo3FirstLastFrameNode(IO.ComfyNode):
         if response.videos:
             video = response.videos[0]
             if video.bytesBase64Encoded:
-                return IO.NodeOutput(VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded))))
+                return IO.NodeOutput(InputImpl.VideoFromFile(BytesIO(base64.b64decode(video.bytesBase64Encoded))))
             if video.gcsUri:
                 return IO.NodeOutput(await download_url_to_video_output(video.gcsUri))
             raise Exception("Video returned but no data or URL was provided")

comfy_extras/nodes_model_patch.py

@@ -244,6 +244,14 @@ class ModelPatchLoader:
         elif 'control_all_x_embedder.2-1.weight' in sd: # alipai z image fun controlnet
             sd = z_image_convert(sd)
             model = comfy.ldm.lumina.controlnet.ZImage_Control(device=comfy.model_management.unet_offload_device(), dtype=dtype, operations=comfy.ops.manual_cast)
+        elif "audio_proj.proj1.weight" in sd:
+            model = MultiTalkModelPatch(
+                    audio_window=5, context_tokens=32, vae_scale=4,
+                    in_dim=sd["blocks.0.audio_cross_attn.proj.weight"].shape[0],
+                    intermediate_dim=sd["audio_proj.proj1.weight"].shape[0],
+                    out_dim=sd["audio_proj.norm.weight"].shape[0],
+                    device=comfy.model_management.unet_offload_device(),
+                    operations=comfy.ops.manual_cast)
 
         model.load_state_dict(sd)
         model = comfy.model_patcher.ModelPatcher(model, load_device=comfy.model_management.get_torch_device(), offload_device=comfy.model_management.unet_offload_device())
@@ -435,6 +443,40 @@ class USOStyleReference:
         return (model_patched,)
 
 
+from comfy.ldm.wan.model_multitalk import WanMultiTalkAttentionBlock, MultiTalkAudioProjModel
+
+class MultiTalkModelPatch(torch.nn.Module):
+    def __init__(
+        self,
+        audio_window: int = 5,
+        intermediate_dim: int = 512,
+        in_dim: int = 5120,
+        out_dim: int = 768,
+        context_tokens: int = 32,
+        vae_scale: int = 4,
+        num_layers: int = 40,
+
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+        self.audio_proj = MultiTalkAudioProjModel(
+                seq_len=audio_window,
+                seq_len_vf=audio_window+vae_scale-1,
+                intermediate_dim=intermediate_dim,
+                out_dim=out_dim,
+                context_tokens=context_tokens,
+                device=device,
+                dtype=dtype,
+                operations=operations
+        )
+        self.blocks = torch.nn.ModuleList(
+            [
+                WanMultiTalkAttentionBlock(in_dim, out_dim, device=device, dtype=dtype, operations=operations)
+                for _ in range(num_layers)
+            ]
+        )
+
+
 NODE_CLASS_MAPPINGS = {
     "ModelPatchLoader": ModelPatchLoader,
     "QwenImageDiffsynthControlnet": QwenImageDiffsynthControlnet,

comfy_extras/nodes_video.py

@@ -8,10 +8,7 @@ import json
 from typing import Optional
 from typing_extensions import override
 from fractions import Fraction
-from comfy_api.input import AudioInput, ImageInput, VideoInput
-from comfy_api.input_impl import VideoFromComponents, VideoFromFile
-from comfy_api.util import VideoCodec, VideoComponents, VideoContainer
-from comfy_api.latest import ComfyExtension, io, ui
+from comfy_api.latest import ComfyExtension, io, ui, Input, InputImpl, Types
 from comfy.cli_args import args
 
 class SaveWEBM(io.ComfyNode):
@@ -28,7 +25,6 @@ class SaveWEBM(io.ComfyNode):
                 io.Float.Input("fps", default=24.0, min=0.01, max=1000.0, step=0.01),
                 io.Float.Input("crf", default=32.0, min=0, max=63.0, step=1, tooltip="Higher crf means lower quality with a smaller file size, lower crf means higher quality higher filesize."),
             ],
-            outputs=[],
             hidden=[io.Hidden.prompt, io.Hidden.extra_pnginfo],
             is_output_node=True,
         )
@@ -79,16 +75,15 @@ class SaveVideo(io.ComfyNode):
             inputs=[
                 io.Video.Input("video", tooltip="The video to save."),
                 io.String.Input("filename_prefix", default="video/ComfyUI", tooltip="The prefix for the file to save. This may include formatting information such as %date:yyyy-MM-dd% or %Empty Latent Image.width% to include values from nodes."),
-                io.Combo.Input("format", options=VideoContainer.as_input(), default="auto", tooltip="The format to save the video as."),
-                io.Combo.Input("codec", options=VideoCodec.as_input(), default="auto", tooltip="The codec to use for the video."),
+                io.Combo.Input("format", options=Types.VideoContainer.as_input(), default="auto", tooltip="The format to save the video as."),
+                io.Combo.Input("codec", options=Types.VideoCodec.as_input(), default="auto", tooltip="The codec to use for the video."),
             ],
-            outputs=[],
             hidden=[io.Hidden.prompt, io.Hidden.extra_pnginfo],
             is_output_node=True,
         )
 
     @classmethod
-    def execute(cls, video: VideoInput, filename_prefix, format: str, codec) -> io.NodeOutput:
+    def execute(cls, video: Input.Video, filename_prefix, format: str, codec) -> io.NodeOutput:
         width, height = video.get_dimensions()
         full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(
             filename_prefix,
@@ -105,10 +100,10 @@ class SaveVideo(io.ComfyNode):
                 metadata["prompt"] = cls.hidden.prompt
             if len(metadata) > 0:
                 saved_metadata = metadata
-        file = f"{filename}_{counter:05}_.{VideoContainer.get_extension(format)}"
+        file = f"{filename}_{counter:05}_.{Types.VideoContainer.get_extension(format)}"
         video.save_to(
             os.path.join(full_output_folder, file),
-            format=VideoContainer(format),
+            format=Types.VideoContainer(format),
             codec=codec,
             metadata=saved_metadata
         )
@@ -135,9 +130,9 @@ class CreateVideo(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, images: ImageInput, fps: float, audio: Optional[AudioInput] = None) -> io.NodeOutput:
+    def execute(cls, images: Input.Image, fps: float, audio: Optional[Input.Audio] = None) -> io.NodeOutput:
         return io.NodeOutput(
-            VideoFromComponents(VideoComponents(images=images, audio=audio, frame_rate=Fraction(fps)))
+            InputImpl.VideoFromComponents(Types.VideoComponents(images=images, audio=audio, frame_rate=Fraction(fps)))
         )
 
 class GetVideoComponents(io.ComfyNode):
@@ -159,11 +154,11 @@ class GetVideoComponents(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, video: VideoInput) -> io.NodeOutput:
+    def execute(cls, video: Input.Video) -> io.NodeOutput:
         components = video.get_components()
-
         return io.NodeOutput(components.images, components.audio, float(components.frame_rate))
 
+
 class LoadVideo(io.ComfyNode):
     @classmethod
     def define_schema(cls):
@@ -185,7 +180,7 @@ class LoadVideo(io.ComfyNode):
     @classmethod
     def execute(cls, file) -> io.NodeOutput:
         video_path = folder_paths.get_annotated_filepath(file)
-        return io.NodeOutput(VideoFromFile(video_path))
+        return io.NodeOutput(InputImpl.VideoFromFile(video_path))
 
     @classmethod
     def fingerprint_inputs(s, file):

comfy_extras/nodes_wan.py

@@ -11,6 +11,7 @@ import numpy as np
 from typing import Tuple
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, io
+import logging
 
 class WanImageToVideo(io.ComfyNode):
     @classmethod
@@ -1288,6 +1289,155 @@ class Wan22ImageToVideoLatent(io.ComfyNode):
         return io.NodeOutput(out_latent)
 
 
+from comfy.ldm.wan.model_multitalk import InfiniteTalkOuterSampleWrapper, MultiTalkCrossAttnPatch, MultiTalkGetAttnMapPatch, project_audio_features
+class WanInfiniteTalkToVideo(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanInfiniteTalkToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Model.Input("model"),
+                io.ModelPatch.Input("model_patch"),
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=832, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=480, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("start_image", optional=True),
+                io.AudioEncoderOutput.Input("audio_encoder_output_1"),
+                io.AudioEncoderOutput.Input("audio_encoder_output_2", optional=True),
+                io.Mask.Input("mask_1", optional=True, tooltip="Mask for the first speaker, required if using two audio inputs."),
+                io.Mask.Input("mask_2", optional=True, tooltip="Mask for the second speaker, required if using two audio inputs."),
+                io.Int.Input("motion_frame_count", default=9, min=1, max=33, step=1, tooltip="Number of previous frames to use as motion context."),
+                io.Float.Input("audio_scale", default=1.0, min=-10.0, max=10.0, step=0.01),
+                io.Image.Input("previous_frames", optional=True),
+            ],
+            outputs=[
+                io.Model.Output(display_name="model"),
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+                io.Int.Output(display_name="trim_image"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model, model_patch, positive, negative, vae, width, height, length, audio_encoder_output_1, motion_frame_count,
+                start_image=None, previous_frames=None, audio_scale=None, clip_vision_output=None, audio_encoder_output_2=None, mask_1=None, mask_2=None) -> io.NodeOutput:
+
+        if previous_frames is not None and previous_frames.shape[0] < motion_frame_count:
+            raise ValueError("Not enough previous frames provided.")
+        if audio_encoder_output_2 is not None:
+            if mask_1 is None or mask_2 is None:
+                raise ValueError("Masks must be provided if two audio encoder outputs are used.")
+
+        ref_masks = None
+        if mask_1 is not None and mask_2 is not None:
+            if audio_encoder_output_2 is None:
+                raise ValueError("Second audio encoder output must be provided if two masks are used.")
+            ref_masks = torch.cat([mask_1, mask_2])
+
+        latent = torch.zeros([1, 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])
+            concat_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)
+            concat_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": concat_mask})
+            negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent_image, "concat_mask": concat_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})
+
+        model_patched = model.clone()
+
+        encoded_audio_list = []
+        seq_lengths = []
+
+        for audio_encoder_output in [audio_encoder_output_1, audio_encoder_output_2]:
+            if audio_encoder_output is None:
+                continue
+            all_layers = audio_encoder_output["encoded_audio_all_layers"]
+            encoded_audio = torch.stack(all_layers, dim=0).squeeze(1)[1:]  # shape: [num_layers, T, 512]
+            encoded_audio = linear_interpolation(encoded_audio, input_fps=50, output_fps=25).movedim(0, 1) # shape: [T, num_layers, 512]
+            encoded_audio_list.append(encoded_audio)
+            seq_lengths.append(encoded_audio.shape[0])
+
+        # Pad / combine depending on multi_audio_type
+        multi_audio_type = "add"
+        if len(encoded_audio_list) > 1:
+            if multi_audio_type == "para":
+                max_len = max(seq_lengths)
+                padded = []
+                for emb in encoded_audio_list:
+                    if emb.shape[0] < max_len:
+                        pad = torch.zeros(max_len - emb.shape[0], *emb.shape[1:], dtype=emb.dtype)
+                        emb = torch.cat([emb, pad], dim=0)
+                    padded.append(emb)
+                encoded_audio_list = padded
+            elif multi_audio_type == "add":
+                total_len = sum(seq_lengths)
+                full_list = []
+                offset = 0
+                for emb, seq_len in zip(encoded_audio_list, seq_lengths):
+                    full = torch.zeros(total_len, *emb.shape[1:], dtype=emb.dtype)
+                    full[offset:offset+seq_len] = emb
+                    full_list.append(full)
+                    offset += seq_len
+                encoded_audio_list = full_list
+
+        token_ref_target_masks = None
+        if ref_masks is not None:
+            token_ref_target_masks = torch.nn.functional.interpolate(
+                ref_masks.unsqueeze(0), size=(latent.shape[-2] // 2, latent.shape[-1] // 2), mode='nearest')[0]
+            token_ref_target_masks = (token_ref_target_masks > 0).view(token_ref_target_masks.shape[0], -1)
+
+        # when extending from previous frames
+        if previous_frames is not None:
+            motion_frames = comfy.utils.common_upscale(previous_frames[-motion_frame_count:].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            frame_offset = previous_frames.shape[0] - motion_frame_count
+
+            audio_start = frame_offset
+            audio_end = audio_start + length
+            logging.info(f"InfiniteTalk: Processing audio frames {audio_start} - {audio_end}")
+
+            motion_frames_latent = vae.encode(motion_frames[:, :, :, :3])
+            trim_image = motion_frame_count
+        else:
+            audio_start = trim_image = 0
+            audio_end = length
+            motion_frames_latent = concat_latent_image[:, :, :1]
+
+        audio_embed = project_audio_features(model_patch.model.audio_proj, encoded_audio_list, audio_start, audio_end).to(model_patched.model_dtype())
+        model_patched.model_options["transformer_options"]["audio_embeds"] = audio_embed
+
+        # add outer sample wrapper
+        model_patched.add_wrapper_with_key(
+            comfy.patcher_extension.WrappersMP.OUTER_SAMPLE,
+            "infinite_talk_outer_sample",
+            InfiniteTalkOuterSampleWrapper(
+                motion_frames_latent,
+                model_patch,
+                is_extend=previous_frames is not None,
+            ))
+        # add cross-attention patch
+        model_patched.set_model_patch(MultiTalkCrossAttnPatch(model_patch, audio_scale), "attn2_patch")
+        if token_ref_target_masks is not None:
+            model_patched.set_model_patch(MultiTalkGetAttnMapPatch(token_ref_target_masks), "attn1_patch")
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return io.NodeOutput(model_patched, positive, negative, out_latent, trim_image)
+
+
 class WanExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
@@ -1307,6 +1457,7 @@ class WanExtension(ComfyExtension):
             WanHuMoImageToVideo,
             WanAnimateToVideo,
             Wan22ImageToVideoLatent,
+            WanInfiniteTalkToVideo,
         ]
 
 async def comfy_entrypoint() -> WanExtension:

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.33.10
-comfyui-workflow-templates==0.7.25
+comfyui-workflow-templates==0.7.51
 comfyui-embedded-docs==0.3.1
 torch
 torchsde