ComfyUI 0.3.53

* Update template to 0.1.67

* Update template to 0.1.68

README.md

@@ -65,7 +65,6 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Flux](https://comfyanonymous.github.io/ComfyUI_examples/flux/)
    - [Lumina Image 2.0](https://comfyanonymous.github.io/ComfyUI_examples/lumina2/)
    - [HiDream](https://comfyanonymous.github.io/ComfyUI_examples/hidream/)
-   - [Cosmos Predict2](https://comfyanonymous.github.io/ComfyUI_examples/cosmos_predict2/)
    - [Qwen Image](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/)
 - Image Editing Models
    - [Omnigen 2](https://comfyanonymous.github.io/ComfyUI_examples/omnigen/)
@@ -77,7 +76,6 @@ See what ComfyUI can do with the [example workflows](https://comfyanonymous.gith
    - [Mochi](https://comfyanonymous.github.io/ComfyUI_examples/mochi/)
    - [LTX-Video](https://comfyanonymous.github.io/ComfyUI_examples/ltxv/)
    - [Hunyuan Video](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/)
-   - [Nvidia Cosmos](https://comfyanonymous.github.io/ComfyUI_examples/cosmos/) and [Cosmos Predict2](https://comfyanonymous.github.io/ComfyUI_examples/cosmos_predict2/)
    - [Wan 2.1](https://comfyanonymous.github.io/ComfyUI_examples/wan/)
    - [Wan 2.2](https://comfyanonymous.github.io/ComfyUI_examples/wan22/)
 - Audio Models

comfy/audio_encoders/audio_encoders.py

@@ -0,0 +1,42 @@
+from .wav2vec2 import Wav2Vec2Model
+import comfy.model_management
+import comfy.ops
+import comfy.utils
+import logging
+import torchaudio
+
+
+class AudioEncoderModel():
+    def __init__(self, config):
+        self.load_device = comfy.model_management.text_encoder_device()
+        offload_device = comfy.model_management.text_encoder_offload_device()
+        self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
+        self.model = Wav2Vec2Model(dtype=self.dtype, device=offload_device, operations=comfy.ops.manual_cast)
+        self.model.eval()
+        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.model_sample_rate = 16000
+
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=False)
+
+    def get_sd(self):
+        return self.model.state_dict()
+
+    def encode_audio(self, audio, sample_rate):
+        comfy.model_management.load_model_gpu(self.patcher)
+        audio = torchaudio.functional.resample(audio, sample_rate, self.model_sample_rate)
+        out, all_layers = self.model(audio.to(self.load_device))
+        outputs = {}
+        outputs["encoded_audio"] = out
+        outputs["encoded_audio_all_layers"] = all_layers
+        return outputs
+
+
+def load_audio_encoder_from_sd(sd, prefix=""):
+    audio_encoder = AudioEncoderModel(None)
+    sd = comfy.utils.state_dict_prefix_replace(sd, {"wav2vec2.": ""})
+    m, u = audio_encoder.load_sd(sd)
+    if len(m) > 0:
+        logging.warning("missing audio encoder: {}".format(m))
+
+    return audio_encoder

comfy/audio_encoders/wav2vec2.py

@@ -0,0 +1,207 @@
+import torch
+import torch.nn as nn
+from comfy.ldm.modules.attention import optimized_attention_masked
+
+
+class LayerNormConv(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride, bias=False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.conv = operations.Conv1d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, bias=bias, device=device, dtype=dtype)
+        self.layer_norm = operations.LayerNorm(out_channels, elementwise_affine=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return torch.nn.functional.gelu(self.layer_norm(x.transpose(-2, -1)).transpose(-2, -1))
+
+
+class ConvFeatureEncoder(nn.Module):
+    def __init__(self, conv_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.conv_layers = nn.ModuleList([
+            LayerNormConv(1, conv_dim, kernel_size=10, stride=5, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=3, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=2, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+            LayerNormConv(conv_dim, conv_dim, kernel_size=2, stride=2, bias=True, device=device, dtype=dtype, operations=operations),
+        ])
+
+    def forward(self, x):
+        x = x.unsqueeze(1)
+
+        for conv in self.conv_layers:
+            x = conv(x)
+
+        return x.transpose(1, 2)
+
+
+class FeatureProjection(nn.Module):
+    def __init__(self, conv_dim, embed_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.layer_norm = operations.LayerNorm(conv_dim, eps=1e-05, device=device, dtype=dtype)
+        self.projection = operations.Linear(conv_dim, embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.layer_norm(x)
+        x = self.projection(x)
+        return x
+
+
+class PositionalConvEmbedding(nn.Module):
+    def __init__(self, embed_dim=768, kernel_size=128, groups=16):
+        super().__init__()
+        self.conv = nn.Conv1d(
+            embed_dim,
+            embed_dim,
+            kernel_size=kernel_size,
+            padding=kernel_size // 2,
+            groups=groups,
+        )
+        self.conv = torch.nn.utils.parametrizations.weight_norm(self.conv, name="weight", dim=2)
+        self.activation = nn.GELU()
+
+    def forward(self, x):
+        x = x.transpose(1, 2)
+        x = self.conv(x)[:, :, :-1]
+        x = self.activation(x)
+        x = x.transpose(1, 2)
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self,
+        embed_dim=768,
+        num_heads=12,
+        num_layers=12,
+        mlp_ratio=4.0,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        self.pos_conv_embed = PositionalConvEmbedding(embed_dim=embed_dim)
+        self.layers = nn.ModuleList([
+            TransformerEncoderLayer(
+                embed_dim=embed_dim,
+                num_heads=num_heads,
+                mlp_ratio=mlp_ratio,
+                device=device, dtype=dtype, operations=operations
+            )
+            for _ in range(num_layers)
+        ])
+
+        self.layer_norm = operations.LayerNorm(embed_dim, eps=1e-05, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None):
+        x = x + self.pos_conv_embed(x)
+        all_x = ()
+        for layer in self.layers:
+            all_x += (x,)
+            x = layer(x, mask)
+        x = self.layer_norm(x)
+        all_x += (x,)
+        return x, all_x
+
+
+class Attention(nn.Module):
+    def __init__(self, embed_dim, num_heads, bias=True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.head_dim = embed_dim // num_heads
+
+        self.k_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+        self.v_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+        self.q_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+        self.out_proj = operations.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None):
+        assert (mask is None)  # TODO?
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        out = optimized_attention_masked(q, k, v, self.num_heads)
+        return self.out_proj(out)
+
+
+class FeedForward(nn.Module):
+    def __init__(self, embed_dim, mlp_ratio, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.intermediate_dense = operations.Linear(embed_dim, int(embed_dim * mlp_ratio), device=device, dtype=dtype)
+        self.output_dense = operations.Linear(int(embed_dim * mlp_ratio), embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.intermediate_dense(x)
+        x = torch.nn.functional.gelu(x)
+        x = self.output_dense(x)
+        return x
+
+
+class TransformerEncoderLayer(nn.Module):
+    def __init__(
+        self,
+        embed_dim=768,
+        num_heads=12,
+        mlp_ratio=4.0,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        self.attention = Attention(embed_dim, num_heads, device=device, dtype=dtype, operations=operations)
+
+        self.layer_norm = operations.LayerNorm(embed_dim, device=device, dtype=dtype)
+        self.feed_forward = FeedForward(embed_dim, mlp_ratio, device=device, dtype=dtype, operations=operations)
+        self.final_layer_norm = operations.LayerNorm(embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None):
+        residual = x
+        x = self.layer_norm(x)
+        x = self.attention(x, mask=mask)
+        x = residual + x
+
+        x = x + self.feed_forward(self.final_layer_norm(x))
+        return x
+
+
+class Wav2Vec2Model(nn.Module):
+    """Complete Wav2Vec 2.0 model."""
+
+    def __init__(
+        self,
+        embed_dim=1024,
+        final_dim=256,
+        num_heads=16,
+        num_layers=24,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+
+        conv_dim = 512
+        self.feature_extractor = ConvFeatureEncoder(conv_dim, device=device, dtype=dtype, operations=operations)
+        self.feature_projection = FeatureProjection(conv_dim, embed_dim, device=device, dtype=dtype, operations=operations)
+
+        self.masked_spec_embed = nn.Parameter(torch.empty(embed_dim, device=device, dtype=dtype))
+
+        self.encoder = TransformerEncoder(
+            embed_dim=embed_dim,
+            num_heads=num_heads,
+            num_layers=num_layers,
+            device=device, dtype=dtype, operations=operations
+        )
+
+    def forward(self, x, mask_time_indices=None, return_dict=False):
+
+        x = torch.mean(x, dim=1)
+
+        x = (x - x.mean()) / torch.sqrt(x.var() + 1e-7)
+
+        features = self.feature_extractor(x)
+        features = self.feature_projection(features)
+
+        batch_size, seq_len, _ = features.shape
+
+        x, all_x = self.encoder(features)
+
+        return x, all_x

comfy/k_diffusion/sampling.py

@@ -853,6 +853,11 @@ def sample_dpmpp_2m_sde(model, x, sigmas, extra_args=None, callback=None, disabl
     return x
 
 
+@torch.no_grad()
+def sample_dpmpp_2m_sde_heun(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='heun'):
+    return sample_dpmpp_2m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
+
+
 @torch.no_grad()
 def sample_dpmpp_3m_sde(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None):
     """DPM-Solver++(3M) SDE."""
@@ -925,6 +930,16 @@ def sample_dpmpp_3m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, di
     return sample_dpmpp_3m_sde(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler)
 
 
+@torch.no_grad()
+def sample_dpmpp_2m_sde_heun_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='heun'):
+    if len(sigmas) <= 1:
+        return x
+    extra_args = {} if extra_args is None else extra_args
+    sigma_min, sigma_max = sigmas[sigmas > 0].min(), sigmas.max()
+    noise_sampler = BrownianTreeNoiseSampler(x, sigma_min, sigma_max, seed=extra_args.get("seed", None), cpu=False) if noise_sampler is None else noise_sampler
+    return sample_dpmpp_2m_sde_heun(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, eta=eta, s_noise=s_noise, noise_sampler=noise_sampler, solver_type=solver_type)
+
+
 @torch.no_grad()
 def sample_dpmpp_2m_sde_gpu(model, x, sigmas, extra_args=None, callback=None, disable=None, eta=1., s_noise=1., noise_sampler=None, solver_type='midpoint'):
     if len(sigmas) <= 1:

comfy/ldm/flux/model.py

@@ -158,7 +158,7 @@ class Flux(nn.Module):
                 if i < len(control_i):
                     add = control_i[i]
                     if add is not None:
-                        img += add
+                        img[:, :add.shape[1]] += add
 
         if img.dtype == torch.float16:
             img = torch.nan_to_num(img, nan=0.0, posinf=65504, neginf=-65504)
@@ -189,7 +189,7 @@ class Flux(nn.Module):
                 if i < len(control_o):
                     add = control_o[i]
                     if add is not None:
-                        img[:, txt.shape[1] :, ...] += add
+                        img[:, txt.shape[1] : txt.shape[1] + add.shape[1], ...] += add
 
         img = img[:, txt.shape[1] :, ...]
 

comfy/ldm/qwen_image/model.py

@@ -459,7 +459,7 @@ class QwenImageTransformer2DModel(nn.Module):
                 if i < len(control_i):
                     add = control_i[i]
                     if add is not None:
-                        hidden_states += add
+                        hidden_states[:, :add.shape[1]] += add
 
         hidden_states = self.norm_out(hidden_states, temb)
         hidden_states = self.proj_out(hidden_states)

comfy/ldm/wan/model.py

@@ -4,7 +4,7 @@ import math
 
 import torch
 import torch.nn as nn
-from einops import repeat
+from einops import rearrange
 
 from comfy.ldm.modules.attention import optimized_attention
 from comfy.ldm.flux.layers import EmbedND
@@ -153,7 +153,10 @@ def repeat_e(e, x):
         repeats = x.size(1) // e.size(1)
     if repeats == 1:
         return e
-    return torch.repeat_interleave(e, repeats, dim=1)
+    if repeats * e.size(1) == x.size(1):
+        return torch.repeat_interleave(e, repeats, dim=1)
+    else:
+        return torch.repeat_interleave(e, repeats + 1, dim=1)[:, :x.size(1)]
 
 
 class WanAttentionBlock(nn.Module):
@@ -573,6 +576,28 @@ class WanModel(torch.nn.Module):
         x = self.unpatchify(x, grid_sizes)
         return x
 
+    def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None):
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+
+        if steps_t is None:
+            steps_t = t_len
+        if steps_h is None:
+            steps_h = h_len
+        if steps_w is None:
+            steps_w = w_len
+
+        img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
+        img_ids = img_ids.reshape(1, -1, img_ids.shape[-1])
+
+        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        return freqs
+
     def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
             self._forward,
@@ -584,26 +609,16 @@ class WanModel(torch.nn.Module):
         bs, c, t, h, w = x.shape
         x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
 
-        patch_size = self.patch_size
-        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
-        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
-        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
-
+        t_len = t
         if time_dim_concat is not None:
             time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size)
             x = torch.cat([x, time_dim_concat], dim=2)
-            t_len = ((x.shape[2] + (patch_size[0] // 2)) // patch_size[0])
+            t_len = x.shape[2]
 
         if self.ref_conv is not None and "reference_latent" in kwargs:
             t_len += 1
 
-        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device, dtype=x.dtype)
-        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).reshape(-1, 1, 1)
-        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=h_len, device=x.device, dtype=x.dtype).reshape(1, -1, 1)
-        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=w_len, device=x.device, dtype=x.dtype).reshape(1, 1, -1)
-        img_ids = repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
-
-        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype)
         return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w]
 
     def unpatchify(self, x, grid_sizes):
@@ -839,3 +854,468 @@ class CameraWanModel(WanModel):
         # unpatchify
         x = self.unpatchify(x, grid_sizes)
         return x
+
+
+class CausalConv1d(nn.Module):
+
+    def __init__(self,
+                 chan_in,
+                 chan_out,
+                 kernel_size=3,
+                 stride=1,
+                 dilation=1,
+                 pad_mode='replicate',
+                 operations=None,
+                 **kwargs):
+        super().__init__()
+
+        self.pad_mode = pad_mode
+        padding = (kernel_size - 1, 0)  # T
+        self.time_causal_padding = padding
+
+        self.conv = operations.Conv1d(
+            chan_in,
+            chan_out,
+            kernel_size,
+            stride=stride,
+            dilation=dilation,
+            **kwargs)
+
+    def forward(self, x):
+        x = torch.nn.functional.pad(x, self.time_causal_padding, mode=self.pad_mode)
+        return self.conv(x)
+
+
+class MotionEncoder_tc(nn.Module):
+
+    def __init__(self,
+                 in_dim: int,
+                 hidden_dim: int,
+                 num_heads=int,
+                 need_global=True,
+                 dtype=None,
+                 device=None,
+                 operations=None,):
+        factory_kwargs = {"dtype": dtype, "device": device}
+        super().__init__()
+
+        self.num_heads = num_heads
+        self.need_global = need_global
+        self.conv1_local = CausalConv1d(in_dim, hidden_dim // 4 * num_heads, 3, stride=1, operations=operations, **factory_kwargs)
+        if need_global:
+            self.conv1_global = CausalConv1d(
+                in_dim, hidden_dim // 4, 3, stride=1, operations=operations, **factory_kwargs)
+        self.norm1 = operations.LayerNorm(
+            hidden_dim // 4,
+            elementwise_affine=False,
+            eps=1e-6,
+            **factory_kwargs)
+        self.act = nn.SiLU()
+        self.conv2 = CausalConv1d(hidden_dim // 4, hidden_dim // 2, 3, stride=2, operations=operations, **factory_kwargs)
+        self.conv3 = CausalConv1d(hidden_dim // 2, hidden_dim, 3, stride=2, operations=operations, **factory_kwargs)
+
+        if need_global:
+            self.final_linear = operations.Linear(hidden_dim, hidden_dim, **factory_kwargs)
+
+        self.norm1 = operations.LayerNorm(
+            hidden_dim // 4,
+            elementwise_affine=False,
+            eps=1e-6,
+            **factory_kwargs)
+
+        self.norm2 = operations.LayerNorm(
+            hidden_dim // 2,
+            elementwise_affine=False,
+            eps=1e-6,
+            **factory_kwargs)
+
+        self.norm3 = operations.LayerNorm(
+            hidden_dim, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.padding_tokens = nn.Parameter(torch.empty(1, 1, 1, hidden_dim, **factory_kwargs))
+
+    def forward(self, x):
+        x = rearrange(x, 'b t c -> b c t')
+        x_ori = x.clone()
+        b, c, t = x.shape
+        x = self.conv1_local(x)
+        x = rearrange(x, 'b (n c) t -> (b n) t c', n=self.num_heads)
+        x = self.norm1(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv2(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm2(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv3(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm3(x)
+        x = self.act(x)
+        x = rearrange(x, '(b n) t c -> b t n c', b=b)
+        padding = comfy.model_management.cast_to(self.padding_tokens, dtype=x.dtype, device=x.device).repeat(b, x.shape[1], 1, 1)
+        x = torch.cat([x, padding], dim=-2)
+        x_local = x.clone()
+
+        if not self.need_global:
+            return x_local
+
+        x = self.conv1_global(x_ori)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm1(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv2(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm2(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv3(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm3(x)
+        x = self.act(x)
+        x = self.final_linear(x)
+        x = rearrange(x, '(b n) t c -> b t n c', b=b)
+
+        return x, x_local
+
+
+class CausalAudioEncoder(nn.Module):
+
+    def __init__(self,
+                 dim=5120,
+                 num_layers=25,
+                 out_dim=2048,
+                 video_rate=8,
+                 num_token=4,
+                 need_global=False,
+                 dtype=None,
+                 device=None,
+                 operations=None):
+        super().__init__()
+        self.encoder = MotionEncoder_tc(
+            in_dim=dim,
+            hidden_dim=out_dim,
+            num_heads=num_token,
+            need_global=need_global, dtype=dtype, device=device, operations=operations)
+        weight = torch.empty((1, num_layers, 1, 1), dtype=dtype, device=device)
+
+        self.weights = torch.nn.Parameter(weight)
+        self.act = torch.nn.SiLU()
+
+    def forward(self, features):
+        # features B * num_layers * dim * video_length
+        weights = self.act(comfy.model_management.cast_to(self.weights, dtype=features.dtype, device=features.device))
+        weights_sum = weights.sum(dim=1, keepdims=True)
+        weighted_feat = ((features * weights) / weights_sum).sum(
+            dim=1)  # b dim f
+        weighted_feat = weighted_feat.permute(0, 2, 1)  # b f dim
+        res = self.encoder(weighted_feat)  # b f n dim
+        return res  # b f n dim
+
+
+class AdaLayerNorm(nn.Module):
+    def __init__(self, embedding_dim, output_dim=None, norm_elementwise_affine=False, norm_eps=1e-5, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        output_dim = output_dim or embedding_dim * 2
+
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(embedding_dim, output_dim, dtype=dtype, device=device)
+        self.norm = operations.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine, dtype=dtype, device=device)
+
+    def forward(self, x, temb):
+        temb = self.linear(self.silu(temb))
+        shift, scale = temb.chunk(2, dim=1)
+        shift = shift[:, None, :]
+        scale = scale[:, None, :]
+        x = self.norm(x) * (1 + scale) + shift
+        return x
+
+
+class AudioInjector_WAN(nn.Module):
+
+    def __init__(self,
+                 dim=2048,
+                 num_heads=32,
+                 inject_layer=[0, 27],
+                 root_net=None,
+                 enable_adain=False,
+                 adain_dim=2048,
+                 adain_mode=None,
+                 dtype=None,
+                 device=None,
+                 operations=None):
+        super().__init__()
+        self.enable_adain = enable_adain
+        self.adain_mode = adain_mode
+        self.injected_block_id = {}
+        audio_injector_id = 0
+        for inject_id in inject_layer:
+            self.injected_block_id[inject_id] = audio_injector_id
+            audio_injector_id += 1
+
+        self.injector = nn.ModuleList([
+            WanT2VCrossAttention(
+                dim=dim,
+                num_heads=num_heads,
+                qk_norm=True, operation_settings={"operations": operations, "device": device, "dtype": dtype}
+            ) for _ in range(audio_injector_id)
+        ])
+        self.injector_pre_norm_feat = nn.ModuleList([
+            operations.LayerNorm(
+                dim,
+                elementwise_affine=False,
+                eps=1e-6, dtype=dtype, device=device
+            ) for _ in range(audio_injector_id)
+        ])
+        self.injector_pre_norm_vec = nn.ModuleList([
+            operations.LayerNorm(
+                dim,
+                elementwise_affine=False,
+                eps=1e-6, dtype=dtype, device=device
+            ) for _ in range(audio_injector_id)
+        ])
+        if enable_adain:
+            self.injector_adain_layers = nn.ModuleList([
+                AdaLayerNorm(
+                    output_dim=dim * 2, embedding_dim=adain_dim, dtype=dtype, device=device, operations=operations)
+                for _ in range(audio_injector_id)
+            ])
+            if adain_mode != "attn_norm":
+                self.injector_adain_output_layers = nn.ModuleList(
+                    [operations.Linear(dim, dim, dtype=dtype, device=device) for _ in range(audio_injector_id)])
+
+    def forward(self, x, block_id, audio_emb, audio_emb_global, seq_len):
+        audio_attn_id = self.injected_block_id.get(block_id, None)
+        if audio_attn_id is None:
+            return x
+
+        num_frames = audio_emb.shape[1]
+        input_hidden_states = rearrange(x[:, :seq_len], "b (t n) c -> (b t) n c", t=num_frames)
+        if self.enable_adain and self.adain_mode == "attn_norm":
+            audio_emb_global = rearrange(audio_emb_global, "b t n c -> (b t) n c")
+            adain_hidden_states = self.injector_adain_layers[audio_attn_id](input_hidden_states, temb=audio_emb_global[:, 0])
+            attn_hidden_states = adain_hidden_states
+        else:
+            attn_hidden_states = self.injector_pre_norm_feat[audio_attn_id](input_hidden_states)
+        audio_emb = rearrange(audio_emb, "b t n c -> (b t) n c", t=num_frames)
+        attn_audio_emb = audio_emb
+        residual_out = self.injector[audio_attn_id](x=attn_hidden_states, context=attn_audio_emb)
+        residual_out = rearrange(
+            residual_out, "(b t) n c -> b (t n) c", t=num_frames)
+        x[:, :seq_len] = x[:, :seq_len] + residual_out
+        return x
+
+
+class FramePackMotioner(nn.Module):
+    def __init__(
+            self,
+            inner_dim=1024,
+            num_heads=16,  # Used to indicate the number of heads in the backbone network; unrelated to this module's design
+            zip_frame_buckets=[
+                1, 2, 16
+            ],  # Three numbers representing the number of frames sampled for patch operations from the nearest to the farthest frames
+            drop_mode="drop",  # If not "drop", it will use "padd", meaning padding instead of deletion
+            dtype=None,
+            device=None,
+            operations=None):
+        super().__init__()
+        self.proj = operations.Conv3d(16, inner_dim, kernel_size=(1, 2, 2), stride=(1, 2, 2), dtype=dtype, device=device)
+        self.proj_2x = operations.Conv3d(16, inner_dim, kernel_size=(2, 4, 4), stride=(2, 4, 4), dtype=dtype, device=device)
+        self.proj_4x = operations.Conv3d(16, inner_dim, kernel_size=(4, 8, 8), stride=(4, 8, 8), dtype=dtype, device=device)
+        self.zip_frame_buckets = zip_frame_buckets
+
+        self.inner_dim = inner_dim
+        self.num_heads = num_heads
+
+        self.drop_mode = drop_mode
+
+    def forward(self, motion_latents, rope_embedder, add_last_motion=2):
+        lat_height, lat_width = motion_latents.shape[3], motion_latents.shape[4]
+        padd_lat = torch.zeros(motion_latents.shape[0], 16, sum(self.zip_frame_buckets), lat_height, lat_width).to(device=motion_latents.device, dtype=motion_latents.dtype)
+        overlap_frame = min(padd_lat.shape[2], motion_latents.shape[2])
+        if overlap_frame > 0:
+            padd_lat[:, :, -overlap_frame:] = motion_latents[:, :, -overlap_frame:]
+
+        if add_last_motion < 2 and self.drop_mode != "drop":
+            zero_end_frame = sum(self.zip_frame_buckets[:len(self.zip_frame_buckets) - add_last_motion - 1])
+            padd_lat[:, :, -zero_end_frame:] = 0
+
+        clean_latents_4x, clean_latents_2x, clean_latents_post = padd_lat[:, :, -sum(self.zip_frame_buckets):, :, :].split(self.zip_frame_buckets[::-1], dim=2)  # 16, 2 ,1
+
+        # patchfy
+        clean_latents_post = self.proj(clean_latents_post).flatten(2).transpose(1, 2)
+        clean_latents_2x = self.proj_2x(clean_latents_2x)
+        l_2x_shape = clean_latents_2x.shape
+        clean_latents_2x = clean_latents_2x.flatten(2).transpose(1, 2)
+        clean_latents_4x = self.proj_4x(clean_latents_4x)
+        l_4x_shape = clean_latents_4x.shape
+        clean_latents_4x = clean_latents_4x.flatten(2).transpose(1, 2)
+
+        if add_last_motion < 2 and self.drop_mode == "drop":
+            clean_latents_post = clean_latents_post[:, :
+                                                    0] if add_last_motion < 2 else clean_latents_post
+            clean_latents_2x = clean_latents_2x[:, :
+                                                0] if add_last_motion < 1 else clean_latents_2x
+
+        motion_lat = torch.cat([clean_latents_post, clean_latents_2x, clean_latents_4x], dim=1)
+
+        rope_post = rope_embedder.rope_encode(1, lat_height, lat_width, t_start=-1, device=motion_latents.device, dtype=motion_latents.dtype)
+        rope_2x = rope_embedder.rope_encode(1, lat_height, lat_width, t_start=-3, steps_h=l_2x_shape[-2], steps_w=l_2x_shape[-1], device=motion_latents.device, dtype=motion_latents.dtype)
+        rope_4x = rope_embedder.rope_encode(4, lat_height, lat_width, t_start=-19, steps_h=l_4x_shape[-2], steps_w=l_4x_shape[-1], device=motion_latents.device, dtype=motion_latents.dtype)
+
+        rope = torch.cat([rope_post, rope_2x, rope_4x], dim=1)
+        return motion_lat, rope
+
+
+class WanModel_S2V(WanModel):
+    def __init__(self,
+                 model_type='s2v',
+                 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,
+                 audio_dim=1024,
+                 num_audio_token=4,
+                 enable_adain=True,
+                 cond_dim=16,
+                 audio_inject_layers=[0, 4, 8, 12, 16, 20, 24, 27, 30, 33, 36, 39],
+                 adain_mode="attn_norm",
+                 framepack_drop_mode="padd",
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+
+        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, image_model=image_model, device=device, dtype=dtype, operations=operations)
+
+        self.trainable_cond_mask = operations.Embedding(3, self.dim, device=device, dtype=dtype)
+
+        self.casual_audio_encoder = CausalAudioEncoder(
+            dim=audio_dim,
+            out_dim=self.dim,
+            num_token=num_audio_token,
+            need_global=enable_adain, dtype=dtype, device=device, operations=operations)
+
+        if cond_dim > 0:
+            self.cond_encoder = operations.Conv3d(
+                cond_dim,
+                self.dim,
+                kernel_size=self.patch_size,
+                stride=self.patch_size, device=device, dtype=dtype)
+
+        self.audio_injector = AudioInjector_WAN(
+            dim=self.dim,
+            num_heads=self.num_heads,
+            inject_layer=audio_inject_layers,
+            root_net=self,
+            enable_adain=enable_adain,
+            adain_dim=self.dim,
+            adain_mode=adain_mode,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        self.frame_packer = FramePackMotioner(
+            inner_dim=self.dim,
+            num_heads=self.num_heads,
+            zip_frame_buckets=[1, 2, 16],
+            drop_mode=framepack_drop_mode,
+            dtype=dtype, device=device, operations=operations)
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        audio_embed=None,
+        reference_latent=None,
+        control_video=None,
+        reference_motion=None,
+        clip_fea=None,
+        freqs=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        if audio_embed is not None:
+            num_embeds = x.shape[-3] * 4
+            audio_emb_global, audio_emb = self.casual_audio_encoder(audio_embed[:, :, :, :num_embeds])
+        else:
+            audio_emb = None
+
+        # embeddings
+        bs, _, time, height, width = x.shape
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        if control_video is not None:
+            x = x + self.cond_encoder(control_video)
+
+        if t.ndim == 1:
+            t = t.unsqueeze(1).repeat(1, x.shape[2])
+
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+        seq_len = x.size(1)
+
+        cond_mask_weight = comfy.model_management.cast_to(self.trainable_cond_mask.weight, dtype=x.dtype, device=x.device).unsqueeze(1).unsqueeze(1)
+        x = x + cond_mask_weight[0]
+
+        if reference_latent is not None:
+            ref = self.patch_embedding(reference_latent.float()).to(x.dtype)
+            ref = ref.flatten(2).transpose(1, 2)
+            freqs_ref = self.rope_encode(reference_latent.shape[-3], reference_latent.shape[-2], reference_latent.shape[-1], t_start=max(30, time + 9), device=x.device, dtype=x.dtype)
+            ref = ref + cond_mask_weight[1]
+            x = torch.cat([x, ref], dim=1)
+            freqs = torch.cat([freqs, freqs_ref], dim=1)
+            t = torch.cat([t, torch.zeros((t.shape[0], reference_latent.shape[-3]), device=t.device, dtype=t.dtype)], dim=1)
+            del ref, freqs_ref
+
+        if reference_motion is not None:
+            motion_encoded, freqs_motion = self.frame_packer(reference_motion, self)
+            motion_encoded = motion_encoded + cond_mask_weight[2]
+            x = torch.cat([x, motion_encoded], dim=1)
+            freqs = torch.cat([freqs, freqs_motion], dim=1)
+
+            t = torch.repeat_interleave(t, 2, dim=1)
+            t = torch.cat([t, torch.zeros((t.shape[0], 3), device=t.device, dtype=t.dtype)], dim=1)
+            del motion_encoded, freqs_motion
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype))
+        e = e.reshape(t.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        # context
+        context = self.text_embedding(context)
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context)
+            if audio_emb is not None:
+                x = self.audio_injector(x, i, audio_emb, audio_emb_global, seq_len)
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x

comfy/model_base.py

@@ -150,6 +150,7 @@ class BaseModel(torch.nn.Module):
         logging.debug("adm {}".format(self.adm_channels))
         self.memory_usage_factor = model_config.memory_usage_factor
         self.memory_usage_factor_conds = ()
+        self.memory_usage_shape_process = {}
 
     def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
@@ -350,8 +351,15 @@ class BaseModel(torch.nn.Module):
         input_shapes = [input_shape]
         for c in self.memory_usage_factor_conds:
             shape = cond_shapes.get(c, None)
-            if shape is not None and len(shape) > 0:
-                input_shapes += shape
+            if shape is not None:
+                if c in self.memory_usage_shape_process:
+                    out = []
+                    for s in shape:
+                        out.append(self.memory_usage_shape_process[c](s))
+                    shape = out
+
+                if len(shape) > 0:
+                    input_shapes += shape
 
         if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
             dtype = self.get_dtype()
@@ -1201,6 +1209,42 @@ class WAN21_Camera(WAN21):
             out['camera_conditions'] = comfy.conds.CONDRegular(camera_conditions)
         return out
 
+class WAN22_S2V(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel_S2V)
+        self.memory_usage_factor_conds = ("reference_latent", "reference_motion")
+        self.memory_usage_shape_process = {"reference_motion": lambda shape: [shape[0], shape[1], 1.5, shape[-2], shape[-1]]}
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        audio_embed = kwargs.get("audio_embed", None)
+        if audio_embed is not None:
+            out['audio_embed'] = comfy.conds.CONDRegular(audio_embed)
+
+        reference_latents = kwargs.get("reference_latents", None)
+        if reference_latents is not None:
+            out['reference_latent'] = comfy.conds.CONDRegular(self.process_latent_in(reference_latents[-1]))
+
+        reference_motion = kwargs.get("reference_motion", None)
+        if reference_motion is not None:
+            out['reference_motion'] = comfy.conds.CONDRegular(self.process_latent_in(reference_motion))
+
+        control_video = kwargs.get("control_video", None)
+        if control_video is not None:
+            out['control_video'] = comfy.conds.CONDRegular(self.process_latent_in(control_video))
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['reference_latent'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+
+        reference_motion = kwargs.get("reference_motion", None)
+        if reference_motion is not None:
+            out['reference_motion'] = reference_motion.shape
+        return out
+
 class WAN22(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)

comfy/model_detection.py

@@ -368,6 +368,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["model_type"] = "camera"
             else:
                 dit_config["model_type"] = "camera_2.2"
+        elif '{}casual_audio_encoder.encoder.final_linear.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "s2v"
         else:
             if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
                 dit_config["model_type"] = "i2v"

comfy/samplers.py

@@ -729,7 +729,7 @@ class Sampler:
 
 KSAMPLER_NAMES = ["euler", "euler_cfg_pp", "euler_ancestral", "euler_ancestral_cfg_pp", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
                   "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_2s_ancestral_cfg_pp", "dpmpp_sde", "dpmpp_sde_gpu",
-                  "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
+                  "dpmpp_2m", "dpmpp_2m_cfg_pp", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_2m_sde_heun", "dpmpp_2m_sde_heun_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm",
                   "ipndm", "ipndm_v", "deis", "res_multistep", "res_multistep_cfg_pp", "res_multistep_ancestral", "res_multistep_ancestral_cfg_pp",
                   "gradient_estimation", "gradient_estimation_cfg_pp", "er_sde", "seeds_2", "seeds_3", "sa_solver", "sa_solver_pece"]
 

comfy/supported_models.py

@@ -700,7 +700,7 @@ class Flux(supported_models_base.BASE):
     unet_extra_config = {}
     latent_format = latent_formats.Flux
 
-    memory_usage_factor = 2.8
+    memory_usage_factor = 3.1 # TODO: debug why flux mem usage is so weird on windows.
 
     supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
@@ -1072,6 +1072,19 @@ class WAN21_Vace(WAN21_T2V):
         out = model_base.WAN21_Vace(self, image_to_video=False, device=device)
         return out
 
+class WAN22_S2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "s2v",
+    }
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN22_S2V(self, device=device)
+        return out
+
 class WAN22_T2V(WAN21_T2V):
     unet_config = {
         "image_model": "wan2.1",
@@ -1272,6 +1285,6 @@ class QwenImage(supported_models_base.BASE):
         return supported_models_base.ClipTarget(comfy.text_encoders.qwen_image.QwenImageTokenizer, comfy.text_encoders.qwen_image.te(**hunyuan_detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2, QwenImage]
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, Hunyuan3Dv2mini, Hunyuan3Dv2, HiDream, Chroma, ACEStep, Omnigen2, QwenImage]
 
 models += [SVD_img2vid]

comfy/weight_adapter/lokr.py

@@ -97,6 +97,9 @@ class LoKrAdapter(WeightAdapterBase):
             (mat1, mat2, alpha, None, None, None, None, None, None)
         )
 
+    def to_train(self):
+        return LokrDiff(self.weights)
+
     @classmethod
     def load(
         cls,

comfy_api/latest/_io.py

@@ -730,6 +730,14 @@ class AnyType(ComfyTypeIO):
 class MODEL_PATCH(ComfyTypeIO):
     Type = Any
 
+@comfytype(io_type="AUDIO_ENCODER")
+class AudioEncoder(ComfyTypeIO):
+    Type = Any
+
+@comfytype(io_type="AUDIO_ENCODER_OUTPUT")
+class AudioEncoderOutput(ComfyTypeIO):
+    Type = Any
+
 @comfytype(io_type="COMFY_MULTITYPED_V3")
 class MultiType:
     Type = Any
@@ -1584,6 +1592,7 @@ class _IO:
     Model = Model
     ClipVision = ClipVision
     ClipVisionOutput = ClipVisionOutput
+    AudioEncoderOutput = AudioEncoderOutput
     StyleModel = StyleModel
     Gligen = Gligen
     UpscaleModel = UpscaleModel

comfy_api_nodes/apis/gemini_api.py

@@ -0,0 +1,19 @@
+from __future__ import annotations
+
+from typing import List, Optional
+
+from comfy_api_nodes.apis import GeminiGenerationConfig, GeminiContent, GeminiSafetySetting, GeminiSystemInstructionContent, GeminiTool, GeminiVideoMetadata
+from pydantic import BaseModel
+
+
+class GeminiImageGenerationConfig(GeminiGenerationConfig):
+    responseModalities: Optional[List[str]] = None
+
+
+class GeminiImageGenerateContentRequest(BaseModel):
+    contents: List[GeminiContent]
+    generationConfig: Optional[GeminiImageGenerationConfig] = None
+    safetySettings: Optional[List[GeminiSafetySetting]] = None
+    systemInstruction: Optional[GeminiSystemInstructionContent] = None
+    tools: Optional[List[GeminiTool]] = None
+    videoMetadata: Optional[GeminiVideoMetadata] = None

comfy_api_nodes/nodes_gemini.py

@@ -4,11 +4,12 @@ See: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/infer
 """
 from __future__ import annotations
 
-
 import json
 import time
 import os
 import uuid
+import base64
+from io import BytesIO
 from enum import Enum
 from typing import Optional, Literal
 
@@ -25,6 +26,7 @@ from comfy_api_nodes.apis import (
     GeminiPart,
     GeminiMimeType,
 )
+from comfy_api_nodes.apis.gemini_api import GeminiImageGenerationConfig, GeminiImageGenerateContentRequest
 from comfy_api_nodes.apis.client import (
     ApiEndpoint,
     HttpMethod,
@@ -35,6 +37,7 @@ from comfy_api_nodes.apinode_utils import (
     audio_to_base64_string,
     video_to_base64_string,
     tensor_to_base64_string,
+    bytesio_to_image_tensor,
 )
 
 
@@ -53,6 +56,14 @@ class GeminiModel(str, Enum):
     gemini_2_5_flash = "gemini-2.5-flash"
 
 
+class GeminiImageModel(str, Enum):
+    """
+    Gemini Image Model Names allowed by comfy-api
+    """
+
+    gemini_2_5_flash_image_preview = "gemini-2.5-flash-image-preview"
+
+
 def get_gemini_endpoint(
     model: GeminiModel,
 ) -> ApiEndpoint[GeminiGenerateContentRequest, GeminiGenerateContentResponse]:
@@ -75,6 +86,135 @@ def get_gemini_endpoint(
     )
 
 
+def get_gemini_image_endpoint(
+    model: GeminiImageModel,
+) -> ApiEndpoint[GeminiGenerateContentRequest, GeminiGenerateContentResponse]:
+    """
+    Get the API endpoint for a given Gemini model.
+
+    Args:
+        model: The Gemini model to use, either as enum or string value.
+
+    Returns:
+        ApiEndpoint configured for the specific Gemini model.
+    """
+    if isinstance(model, str):
+        model = GeminiImageModel(model)
+    return ApiEndpoint(
+        path=f"{GEMINI_BASE_ENDPOINT}/{model.value}",
+        method=HttpMethod.POST,
+        request_model=GeminiImageGenerateContentRequest,
+        response_model=GeminiGenerateContentResponse,
+    )
+
+
+def create_image_parts(image_input: torch.Tensor) -> list[GeminiPart]:
+    """
+    Convert image tensor input to Gemini API compatible parts.
+
+    Args:
+        image_input: Batch of image tensors from ComfyUI.
+
+    Returns:
+        List of GeminiPart objects containing the encoded images.
+    """
+    image_parts: list[GeminiPart] = []
+    for image_index in range(image_input.shape[0]):
+        image_as_b64 = tensor_to_base64_string(
+            image_input[image_index].unsqueeze(0)
+        )
+        image_parts.append(
+            GeminiPart(
+                inlineData=GeminiInlineData(
+                    mimeType=GeminiMimeType.image_png,
+                    data=image_as_b64,
+                )
+            )
+        )
+    return image_parts
+
+
+def create_text_part(text: str) -> GeminiPart:
+    """
+    Create a text part for the Gemini API request.
+
+    Args:
+        text: The text content to include in the request.
+
+    Returns:
+        A GeminiPart object with the text content.
+    """
+    return GeminiPart(text=text)
+
+
+def get_parts_from_response(
+    response: GeminiGenerateContentResponse
+) -> list[GeminiPart]:
+    """
+    Extract all parts from the Gemini API response.
+
+    Args:
+        response: The API response from Gemini.
+
+    Returns:
+        List of response parts from the first candidate.
+    """
+    return response.candidates[0].content.parts
+
+
+def get_parts_by_type(
+    response: GeminiGenerateContentResponse, part_type: Literal["text"] | str
+) -> list[GeminiPart]:
+    """
+    Filter response parts by their type.
+
+    Args:
+        response: The API response from Gemini.
+        part_type: Type of parts to extract ("text" or a MIME type).
+
+    Returns:
+        List of response parts matching the requested type.
+    """
+    parts = []
+    for part in get_parts_from_response(response):
+        if part_type == "text" and hasattr(part, "text") and part.text:
+            parts.append(part)
+        elif (
+            hasattr(part, "inlineData")
+            and part.inlineData
+            and part.inlineData.mimeType == part_type
+        ):
+            parts.append(part)
+        # Skip parts that don't match the requested type
+    return parts
+
+
+def get_text_from_response(response: GeminiGenerateContentResponse) -> str:
+    """
+    Extract and concatenate all text parts from the response.
+
+    Args:
+        response: The API response from Gemini.
+
+    Returns:
+        Combined text from all text parts in the response.
+    """
+    parts = get_parts_by_type(response, "text")
+    return "\n".join([part.text for part in parts])
+
+
+def get_image_from_response(response: GeminiGenerateContentResponse) -> torch.Tensor:
+    image_tensors: list[torch.Tensor] = []
+    parts = get_parts_by_type(response, "image/png")
+    for part in parts:
+        image_data = base64.b64decode(part.inlineData.data)
+        returned_image = bytesio_to_image_tensor(BytesIO(image_data))
+        image_tensors.append(returned_image)
+    if len(image_tensors) == 0:
+        return torch.zeros((1,1024,1024,4))
+    return torch.cat(image_tensors, dim=0)
+
+
 class GeminiNode(ComfyNodeABC):
     """
     Node to generate text responses from a Gemini model.
@@ -159,59 +299,6 @@ class GeminiNode(ComfyNodeABC):
     CATEGORY = "api node/text/Gemini"
     API_NODE = True
 
-    def get_parts_from_response(
-        self, response: GeminiGenerateContentResponse
-    ) -> list[GeminiPart]:
-        """
-        Extract all parts from the Gemini API response.
-
-        Args:
-            response: The API response from Gemini.
-
-        Returns:
-            List of response parts from the first candidate.
-        """
-        return response.candidates[0].content.parts
-
-    def get_parts_by_type(
-        self, response: GeminiGenerateContentResponse, part_type: Literal["text"] | str
-    ) -> list[GeminiPart]:
-        """
-        Filter response parts by their type.
-
-        Args:
-            response: The API response from Gemini.
-            part_type: Type of parts to extract ("text" or a MIME type).
-
-        Returns:
-            List of response parts matching the requested type.
-        """
-        parts = []
-        for part in self.get_parts_from_response(response):
-            if part_type == "text" and hasattr(part, "text") and part.text:
-                parts.append(part)
-            elif (
-                hasattr(part, "inlineData")
-                and part.inlineData
-                and part.inlineData.mimeType == part_type
-            ):
-                parts.append(part)
-            # Skip parts that don't match the requested type
-        return parts
-
-    def get_text_from_response(self, response: GeminiGenerateContentResponse) -> str:
-        """
-        Extract and concatenate all text parts from the response.
-
-        Args:
-            response: The API response from Gemini.
-
-        Returns:
-            Combined text from all text parts in the response.
-        """
-        parts = self.get_parts_by_type(response, "text")
-        return "\n".join([part.text for part in parts])
-
     def create_video_parts(self, video_input: IO.VIDEO, **kwargs) -> list[GeminiPart]:
         """
         Convert video input to Gemini API compatible parts.
@@ -271,43 +358,6 @@ class GeminiNode(ComfyNodeABC):
             )
         return audio_parts
 
-    def create_image_parts(self, image_input: torch.Tensor) -> list[GeminiPart]:
-        """
-        Convert image tensor input to Gemini API compatible parts.
-
-        Args:
-            image_input: Batch of image tensors from ComfyUI.
-
-        Returns:
-            List of GeminiPart objects containing the encoded images.
-        """
-        image_parts: list[GeminiPart] = []
-        for image_index in range(image_input.shape[0]):
-            image_as_b64 = tensor_to_base64_string(
-                image_input[image_index].unsqueeze(0)
-            )
-            image_parts.append(
-                GeminiPart(
-                    inlineData=GeminiInlineData(
-                        mimeType=GeminiMimeType.image_png,
-                        data=image_as_b64,
-                    )
-                )
-            )
-        return image_parts
-
-    def create_text_part(self, text: str) -> GeminiPart:
-        """
-        Create a text part for the Gemini API request.
-
-        Args:
-            text: The text content to include in the request.
-
-        Returns:
-            A GeminiPart object with the text content.
-        """
-        return GeminiPart(text=text)
-
     async def api_call(
         self,
         prompt: str,
@@ -323,11 +373,11 @@ class GeminiNode(ComfyNodeABC):
         validate_string(prompt, strip_whitespace=False)
 
         # Create parts list with text prompt as the first part
-        parts: list[GeminiPart] = [self.create_text_part(prompt)]
+        parts: list[GeminiPart] = [create_text_part(prompt)]
 
         # Add other modal parts
         if images is not None:
-            image_parts = self.create_image_parts(images)
+            image_parts = create_image_parts(images)
             parts.extend(image_parts)
         if audio is not None:
             parts.extend(self.create_audio_parts(audio))
@@ -351,7 +401,7 @@ class GeminiNode(ComfyNodeABC):
         ).execute()
 
         # Get result output
-        output_text = self.get_text_from_response(response)
+        output_text = get_text_from_response(response)
         if unique_id and output_text:
             # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button.
             render_spec = {
@@ -462,12 +512,162 @@ class GeminiInputFiles(ComfyNodeABC):
         return (files,)
 
 
+class GeminiImage(ComfyNodeABC):
+    """
+    Node to generate text and image responses from a Gemini model.
+
+    This node allows users to interact with Google's Gemini AI models, providing
+    multimodal inputs (text, images, files) to generate coherent
+    text and image responses. The node works with the latest Gemini models, handling the
+    API communication and response parsing.
+    """
+    @classmethod
+    def INPUT_TYPES(cls) -> InputTypeDict:
+        return {
+            "required": {
+                "prompt": (
+                    IO.STRING,
+                    {
+                        "multiline": True,
+                        "default": "",
+                        "tooltip": "Text prompt for generation",
+                    },
+                ),
+                "model": (
+                    IO.COMBO,
+                    {
+                        "tooltip": "The Gemini model to use for generating responses.",
+                        "options": [model.value for model in GeminiImageModel],
+                        "default": GeminiImageModel.gemini_2_5_flash_image_preview.value,
+                    },
+                ),
+                "seed": (
+                    IO.INT,
+                    {
+                        "default": 42,
+                        "min": 0,
+                        "max": 0xFFFFFFFFFFFFFFFF,
+                        "control_after_generate": True,
+                        "tooltip": "When seed is fixed to a specific value, the model makes a best effort to provide the same response for repeated requests. Deterministic output isn't guaranteed. Also, changing the model or parameter settings, such as the temperature, can cause variations in the response even when you use the same seed value. By default, a random seed value is used.",
+                    },
+                ),
+            },
+            "optional": {
+                "images": (
+                    IO.IMAGE,
+                    {
+                        "default": None,
+                        "tooltip": "Optional image(s) to use as context for the model. To include multiple images, you can use the Batch Images node.",
+                    },
+                ),
+                "files": (
+                    "GEMINI_INPUT_FILES",
+                    {
+                        "default": None,
+                        "tooltip": "Optional file(s) to use as context for the model. Accepts inputs from the Gemini Generate Content Input Files node.",
+                    },
+                ),
+                # TODO: later we can add this parameter later
+                # "n": (
+                #     IO.INT,
+                #     {
+                #         "default": 1,
+                #         "min": 1,
+                #         "max": 8,
+                #         "step": 1,
+                #         "display": "number",
+                #         "tooltip": "How many images to generate",
+                #     },
+                # ),
+            },
+            "hidden": {
+                "auth_token": "AUTH_TOKEN_COMFY_ORG",
+                "comfy_api_key": "API_KEY_COMFY_ORG",
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+
+    RETURN_TYPES = (IO.IMAGE, IO.STRING)
+    FUNCTION = "api_call"
+    CATEGORY = "api node/image/Gemini"
+    DESCRIPTION = "Edit images synchronously via Google API."
+    API_NODE = True
+
+    async def api_call(
+        self,
+        prompt: str,
+        model: GeminiImageModel,
+        images: Optional[IO.IMAGE] = None,
+        files: Optional[list[GeminiPart]] = None,
+        n=1,
+        unique_id: Optional[str] = None,
+        **kwargs,
+    ):
+        # Validate inputs
+        validate_string(prompt, strip_whitespace=True, min_length=1)
+        # Create parts list with text prompt as the first part
+        parts: list[GeminiPart] = [create_text_part(prompt)]
+
+        # Add other modal parts
+        if images is not None:
+            image_parts = create_image_parts(images)
+            parts.extend(image_parts)
+        if files is not None:
+            parts.extend(files)
+
+        response = await SynchronousOperation(
+            endpoint=get_gemini_image_endpoint(model),
+            request=GeminiImageGenerateContentRequest(
+                contents=[
+                    GeminiContent(
+                        role="user",
+                        parts=parts,
+                    ),
+                ],
+                generationConfig=GeminiImageGenerationConfig(
+                    responseModalities=["TEXT","IMAGE"]
+                )
+            ),
+            auth_kwargs=kwargs,
+        ).execute()
+
+        output_image = get_image_from_response(response)
+        output_text = get_text_from_response(response)
+        if unique_id and output_text:
+            # Not a true chat history like the OpenAI Chat node. It is emulated so the frontend can show a copy button.
+            render_spec = {
+                "node_id": unique_id,
+                "component": "ChatHistoryWidget",
+                "props": {
+                    "history": json.dumps(
+                        [
+                            {
+                                "prompt": prompt,
+                                "response": output_text,
+                                "response_id": str(uuid.uuid4()),
+                                "timestamp": time.time(),
+                            }
+                        ]
+                    ),
+                },
+            }
+            PromptServer.instance.send_sync(
+                "display_component",
+                render_spec,
+            )
+
+        output_text = output_text or "Empty response from Gemini model..."
+        return (output_image, output_text,)
+
+
 NODE_CLASS_MAPPINGS = {
     "GeminiNode": GeminiNode,
+    "GeminiImageNode": GeminiImage,
     "GeminiInputFiles": GeminiInputFiles,
 }
 
 NODE_DISPLAY_NAME_MAPPINGS = {
     "GeminiNode": "Google Gemini",
+    "GeminiImageNode": "Google Gemini Image",
     "GeminiInputFiles": "Gemini Input Files",
 }

comfy_extras/nodes_audio_encoder.py

@@ -0,0 +1,44 @@
+import folder_paths
+import comfy.audio_encoders.audio_encoders
+import comfy.utils
+
+
+class AudioEncoderLoader:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "audio_encoder_name": (folder_paths.get_filename_list("audio_encoders"), ),
+                             }}
+    RETURN_TYPES = ("AUDIO_ENCODER",)
+    FUNCTION = "load_model"
+
+    CATEGORY = "loaders"
+
+    def load_model(self, audio_encoder_name):
+        audio_encoder_name = folder_paths.get_full_path_or_raise("audio_encoders", audio_encoder_name)
+        sd = comfy.utils.load_torch_file(audio_encoder_name, safe_load=True)
+        audio_encoder = comfy.audio_encoders.audio_encoders.load_audio_encoder_from_sd(sd)
+        if audio_encoder is None:
+            raise RuntimeError("ERROR: audio encoder file is invalid and does not contain a valid model.")
+        return (audio_encoder,)
+
+
+class AudioEncoderEncode:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "audio_encoder": ("AUDIO_ENCODER",),
+                              "audio": ("AUDIO",),
+                             }}
+    RETURN_TYPES = ("AUDIO_ENCODER_OUTPUT",)
+    FUNCTION = "encode"
+
+    CATEGORY = "conditioning"
+
+    def encode(self, audio_encoder, audio):
+        output = audio_encoder.encode_audio(audio["waveform"], audio["sample_rate"])
+        return (output,)
+
+
+NODE_CLASS_MAPPINGS = {
+    "AudioEncoderLoader": AudioEncoderLoader,
+    "AudioEncoderEncode": AudioEncoderEncode,
+}

comfy_extras/nodes_easycache.py

@@ -28,6 +28,7 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
     input_change = None
     do_easycache = easycache.should_do_easycache(sigmas)
     if do_easycache:
+        easycache.check_metadata(x)
         # if first cond marked this step for skipping, skip it and use appropriate cached values
         if easycache.skip_current_step:
             if easycache.verbose:
@@ -92,6 +93,7 @@ def lazycache_predict_noise_wrapper(executor, *args, **kwargs):
     input_change = None
     do_easycache = easycache.should_do_easycache(timestep)
     if do_easycache:
+        easycache.check_metadata(x)
         if easycache.has_x_prev_subsampled():
             if easycache.has_x_prev_subsampled():
                 input_change = (easycache.subsample(x, clone=False) - easycache.x_prev_subsampled).flatten().abs().mean()
@@ -194,6 +196,7 @@ class EasyCacheHolder:
         # how to deal with mismatched dims
         self.allow_mismatch = True
         self.cut_from_start = True
+        self.state_metadata = None
 
     def is_past_end_timestep(self, timestep: float) -> bool:
         return not (timestep[0] > self.end_t).item()
@@ -283,6 +286,17 @@ class EasyCacheHolder:
     def has_first_cond_uuid(self, uuids: list[UUID]) -> bool:
         return self.first_cond_uuid in uuids
 
+    def check_metadata(self, x: torch.Tensor) -> bool:
+        metadata = (x.device, x.dtype, x.shape[1:])
+        if self.state_metadata is None:
+            self.state_metadata = metadata
+            return True
+        if metadata == self.state_metadata:
+            return True
+        logging.warn(f"{self.name} - Tensor shape, dtype or device changed, resetting state")
+        self.reset()
+        return False
+
     def reset(self):
         self.relative_transformation_rate = 0.0
         self.cumulative_change_rate = 0.0
@@ -299,6 +313,7 @@ class EasyCacheHolder:
         del self.uuid_cache_diffs
         self.uuid_cache_diffs = {}
         self.total_steps_skipped = 0
+        self.state_metadata = None
         return self
 
     def clone(self):
@@ -360,6 +375,7 @@ class LazyCacheHolder:
         self.output_change_rates = []
         self.approx_output_change_rates = []
         self.total_steps_skipped = 0
+        self.state_metadata = None
 
     def has_cache_diff(self) -> bool:
         return self.cache_diff is not None
@@ -404,6 +420,17 @@ class LazyCacheHolder:
     def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor):
         self.cache_diff = output - x
 
+    def check_metadata(self, x: torch.Tensor) -> bool:
+        metadata = (x.device, x.dtype, x.shape)
+        if self.state_metadata is None:
+            self.state_metadata = metadata
+            return True
+        if metadata == self.state_metadata:
+            return True
+        logging.warn(f"{self.name} - Tensor shape, dtype or device changed, resetting state")
+        self.reset()
+        return False
+
     def reset(self):
         self.relative_transformation_rate = 0.0
         self.cumulative_change_rate = 0.0
@@ -412,7 +439,14 @@ class LazyCacheHolder:
         self.approx_output_change_rates = []
         del self.cache_diff
         self.cache_diff = None
+        del self.x_prev_subsampled
+        self.x_prev_subsampled = None
+        del self.output_prev_subsampled
+        self.output_prev_subsampled = None
+        del self.output_prev_norm
+        self.output_prev_norm = None
         self.total_steps_skipped = 0
+        self.state_metadata = None
         return self
 
     def clone(self):

comfy_extras/nodes_latent.py

@@ -105,6 +105,38 @@ class LatentInterpolate:
         samples_out["samples"] = st * (m1 * ratio + m2 * (1.0 - ratio))
         return (samples_out,)
 
+class LatentConcat:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "samples1": ("LATENT",), "samples2": ("LATENT",), "dim": (["x", "-x", "y", "-y", "t", "-t"], )}}
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "op"
+
+    CATEGORY = "latent/advanced"
+
+    def op(self, samples1, samples2, dim):
+        samples_out = samples1.copy()
+
+        s1 = samples1["samples"]
+        s2 = samples2["samples"]
+        s2 = comfy.utils.repeat_to_batch_size(s2, s1.shape[0])
+
+        if "-" in dim:
+            c = (s2, s1)
+        else:
+            c = (s1, s2)
+
+        if "x" in dim:
+            dim = -1
+        elif "y" in dim:
+            dim = -2
+        elif "t" in dim:
+            dim = -3
+
+        samples_out["samples"] = torch.cat(c, dim=dim)
+        return (samples_out,)
+
 class LatentBatch:
     @classmethod
     def INPUT_TYPES(s):
@@ -279,6 +311,7 @@ NODE_CLASS_MAPPINGS = {
     "LatentSubtract": LatentSubtract,
     "LatentMultiply": LatentMultiply,
     "LatentInterpolate": LatentInterpolate,
+    "LatentConcat": LatentConcat,
     "LatentBatch": LatentBatch,
     "LatentBatchSeedBehavior": LatentBatchSeedBehavior,
     "LatentApplyOperation": LatentApplyOperation,

comfy_extras/nodes_model_patch.py

@@ -89,6 +89,7 @@ class DiffSynthCnetPatch:
         self.strength = strength
         self.mask = mask
         self.encoded_image = model_patch.model.process_input_latent_image(self.encode_latent_cond(image))
+        self.encoded_image_size = (image.shape[1], image.shape[2])
 
     def encode_latent_cond(self, image):
         latent_image = self.vae.encode(image)
@@ -106,14 +107,15 @@ class DiffSynthCnetPatch:
         x = kwargs.get("x")
         img = kwargs.get("img")
         block_index = kwargs.get("block_index")
-        if self.encoded_image is None or self.encoded_image.shape[1:] != img.shape[1:]:
-            spacial_compression = self.vae.spacial_compression_encode()
+        spacial_compression = self.vae.spacial_compression_encode()
+        if self.encoded_image is None or self.encoded_image_size != (x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression):
             image_scaled = comfy.utils.common_upscale(self.image.movedim(-1, 1), x.shape[-1] * spacial_compression, x.shape[-2] * spacial_compression, "area", "center")
             loaded_models = comfy.model_management.loaded_models(only_currently_used=True)
             self.encoded_image = self.model_patch.model.process_input_latent_image(self.encode_latent_cond(image_scaled.movedim(1, -1)))
+            self.encoded_image_size = (image_scaled.shape[-2], image_scaled.shape[-1])
             comfy.model_management.load_models_gpu(loaded_models)
 
-        img = img + (self.model_patch.model.control_block(img, self.encoded_image.to(img.dtype), block_index) * self.strength)
+        img[:, :self.encoded_image.shape[1]] += (self.model_patch.model.control_block(img[:, :self.encoded_image.shape[1]], self.encoded_image.to(img.dtype), block_index) * self.strength)
         kwargs['img'] = img
         return kwargs
 

comfy_extras/nodes_wan.py

@@ -786,6 +786,180 @@ class WanTrackToVideo(io.ComfyNode):
         return io.NodeOutput(positive, negative, out_latent)
 
 
+def linear_interpolation(features, input_fps, output_fps, output_len=None):
+    """
+    features: shape=[1, T, 512]
+    input_fps: fps for audio, f_a
+    output_fps: fps for video, f_m
+    output_len: video length
+    """
+    features = features.transpose(1, 2)  # [1, 512, T]
+    seq_len = features.shape[2] / float(input_fps)  # T/f_a
+    if output_len is None:
+        output_len = int(seq_len * output_fps)  # f_m*T/f_a
+    output_features = torch.nn.functional.interpolate(
+        features, size=output_len, align_corners=True,
+        mode='linear')  # [1, 512, output_len]
+    return output_features.transpose(1, 2)  # [1, output_len, 512]
+
+
+def get_sample_indices(original_fps,
+                       total_frames,
+                       target_fps,
+                       num_sample,
+                       fixed_start=None):
+    required_duration = num_sample / target_fps
+    required_origin_frames = int(np.ceil(required_duration * original_fps))
+    if required_duration > total_frames / original_fps:
+        raise ValueError("required_duration must be less than video length")
+
+    if not fixed_start is None and fixed_start >= 0:
+        start_frame = fixed_start
+    else:
+        max_start = total_frames - required_origin_frames
+        if max_start < 0:
+            raise ValueError("video length is too short")
+        start_frame = np.random.randint(0, max_start + 1)
+    start_time = start_frame / original_fps
+
+    end_time = start_time + required_duration
+    time_points = np.linspace(start_time, end_time, num_sample, endpoint=False)
+
+    frame_indices = np.round(np.array(time_points) * original_fps).astype(int)
+    frame_indices = np.clip(frame_indices, 0, total_frames - 1)
+    return frame_indices
+
+
+def get_audio_embed_bucket_fps(audio_embed, fps=16, batch_frames=81, m=0, video_rate=30):
+    num_layers, audio_frame_num, audio_dim = audio_embed.shape
+
+    if num_layers > 1:
+        return_all_layers = True
+    else:
+        return_all_layers = False
+
+    scale = video_rate / fps
+
+    min_batch_num = int(audio_frame_num / (batch_frames * scale)) + 1
+
+    bucket_num = min_batch_num * batch_frames
+    padd_audio_num = math.ceil(min_batch_num * batch_frames / fps * video_rate) - audio_frame_num
+    batch_idx = get_sample_indices(
+        original_fps=video_rate,
+        total_frames=audio_frame_num + padd_audio_num,
+        target_fps=fps,
+        num_sample=bucket_num,
+        fixed_start=0)
+    batch_audio_eb = []
+    audio_sample_stride = int(video_rate / fps)
+    for bi in batch_idx:
+        if bi < audio_frame_num:
+
+            chosen_idx = list(
+                range(bi - m * audio_sample_stride, bi + (m + 1) * audio_sample_stride, audio_sample_stride))
+            chosen_idx = [0 if c < 0 else c for c in chosen_idx]
+            chosen_idx = [
+                audio_frame_num - 1 if c >= audio_frame_num else c
+                for c in chosen_idx
+            ]
+
+            if return_all_layers:
+                frame_audio_embed = audio_embed[:, chosen_idx].flatten(
+                    start_dim=-2, end_dim=-1)
+            else:
+                frame_audio_embed = audio_embed[0][chosen_idx].flatten()
+        else:
+            frame_audio_embed = torch.zeros([audio_dim * (2 * m + 1)], device=audio_embed.device) if not return_all_layers \
+                else torch.zeros([num_layers, audio_dim * (2 * m + 1)], device=audio_embed.device)
+        batch_audio_eb.append(frame_audio_embed)
+    batch_audio_eb = torch.cat([c.unsqueeze(0) for c in batch_audio_eb], dim=0)
+
+    return batch_audio_eb, min_batch_num
+
+
+class WanSoundImageToVideo(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanSoundImageToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                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=77, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.AudioEncoderOutput.Input("audio_encoder_output", optional=True),
+                io.Image.Input("ref_image", optional=True),
+                io.Image.Input("control_video", optional=True),
+                io.Image.Input("ref_motion", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+            is_experimental=True,
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, ref_image=None, audio_encoder_output=None, control_video=None, ref_motion=None) -> io.NodeOutput:
+        latent_t = ((length - 1) // 4) + 1
+        if audio_encoder_output is not None:
+            feat = torch.cat(audio_encoder_output["encoded_audio_all_layers"])
+            video_rate = 30
+            fps = 16
+            feat = linear_interpolation(feat, input_fps=50, output_fps=video_rate)
+            audio_embed_bucket, num_repeat = get_audio_embed_bucket_fps(feat, fps=fps, batch_frames=latent_t * 4, m=0, video_rate=video_rate)
+            audio_embed_bucket = audio_embed_bucket.unsqueeze(0)
+            if len(audio_embed_bucket.shape) == 3:
+                audio_embed_bucket = audio_embed_bucket.permute(0, 2, 1)
+            elif len(audio_embed_bucket.shape) == 4:
+                audio_embed_bucket = audio_embed_bucket.permute(0, 2, 3, 1)
+
+            positive = node_helpers.conditioning_set_values(positive, {"audio_embed": audio_embed_bucket})
+            negative = node_helpers.conditioning_set_values(negative, {"audio_embed": audio_embed_bucket * 0.0})
+
+        if ref_image is not None:
+            ref_image = comfy.utils.common_upscale(ref_image[:1].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            ref_latent = vae.encode(ref_image[:, :, :, :3])
+            positive = node_helpers.conditioning_set_values(positive, {"reference_latents": [ref_latent]}, append=True)
+            negative = node_helpers.conditioning_set_values(negative, {"reference_latents": [ref_latent]}, append=True)
+
+        if ref_motion is not None:
+            if ref_motion.shape[0] > 73:
+                ref_motion = ref_motion[-73:]
+
+            ref_motion = comfy.utils.common_upscale(ref_motion.movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+
+            if ref_motion.shape[0] < 73:
+                r = torch.ones([73, height, width, 3]) * 0.5
+                r[-ref_motion.shape[0]:] = ref_motion
+                ref_motion = r
+
+            ref_motion = vae.encode(ref_motion[:, :, :, :3])
+            positive = node_helpers.conditioning_set_values(positive, {"reference_motion": ref_motion})
+            negative = node_helpers.conditioning_set_values(negative, {"reference_motion": ref_motion})
+
+        latent = torch.zeros([batch_size, 16, latent_t, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+
+        control_video_out = comfy.latent_formats.Wan21().process_out(torch.zeros_like(latent))
+        if control_video is not None:
+            control_video = comfy.utils.common_upscale(control_video[:length].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            control_video = vae.encode(control_video[:, :, :, :3])
+            control_video_out[:, :, :control_video.shape[2]] = control_video
+
+        # TODO: check if zero is better than none if none provided
+        positive = node_helpers.conditioning_set_values(positive, {"control_video": control_video_out})
+        negative = node_helpers.conditioning_set_values(negative, {"control_video": control_video_out})
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return io.NodeOutput(positive, negative, out_latent)
+
+
 class Wan22ImageToVideoLatent(io.ComfyNode):
     @classmethod
     def define_schema(cls):
@@ -844,6 +1018,7 @@ class WanExtension(ComfyExtension):
             TrimVideoLatent,
             WanCameraImageToVideo,
             WanPhantomSubjectToVideo,
+            WanSoundImageToVideo,
             Wan22ImageToVideoLatent,
         ]
 

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.3.52"
+__version__ = "0.3.53"

folder_paths.py

@@ -48,6 +48,8 @@ folder_names_and_paths["classifiers"] = ([os.path.join(models_dir, "classifiers"
 
 folder_names_and_paths["model_patches"] = ([os.path.join(models_dir, "model_patches")], supported_pt_extensions)
 
+folder_names_and_paths["audio_encoders"] = ([os.path.join(models_dir, "audio_encoders")], supported_pt_extensions)
+
 output_directory = os.path.join(base_path, "output")
 temp_directory = os.path.join(base_path, "temp")
 input_directory = os.path.join(base_path, "input")

nodes.py

@@ -2324,6 +2324,7 @@ async def init_builtin_extra_nodes():
         "nodes_qwen.py",
         "nodes_model_patch.py",
         "nodes_easycache.py",
+        "nodes_audio_encoder.py",
     ]
 
     import_failed = []

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.52"
+version = "0.3.53"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,5 +1,5 @@
-comfyui-frontend-package==1.25.10
-comfyui-workflow-templates==0.1.65
+comfyui-frontend-package==1.25.11
+comfyui-workflow-templates==0.1.68
 comfyui-embedded-docs==0.2.6
 torch
 torchsde