Merge remote-tracking branch 'origin/master' into worksplit-multigpu

README.md

@@ -65,18 +65,17 @@ 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/)
    - [Flux Kontext](https://comfyanonymous.github.io/ComfyUI_examples/flux/#flux-kontext-image-editing-model)
    - [HiDream E1.1](https://comfyanonymous.github.io/ComfyUI_examples/hidream/#hidream-e11)
+   - [Qwen Image Edit](https://comfyanonymous.github.io/ComfyUI_examples/qwen_image/#edit-model)
 - Video Models
    - [Stable Video Diffusion](https://comfyanonymous.github.io/ComfyUI_examples/video/)
    - [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
@@ -191,7 +190,7 @@ comfy install
 
 ## Manual Install (Windows, Linux)
 
-python 3.13 is supported but using 3.12 is recommended because some custom nodes and their dependencies might not support it yet.
+Python 3.13 is very well supported. If you have trouble with some custom node dependencies you can try 3.12
 
 Git clone this repo.
 

app/user_manager.py

@@ -363,10 +363,17 @@ class UserManager():
             if not overwrite and os.path.exists(path):
                 return web.Response(status=409, text="File already exists")
 
-            body = await request.read()
+            try:
+                body = await request.read()
 
-            with open(path, "wb") as f:
-                f.write(body)
+                with open(path, "wb") as f:
+                    f.write(body)
+            except OSError as e:
+                logging.warning(f"Error saving file '{path}': {e}")
+                return web.Response(
+                    status=400,
+                    reason="Invalid filename. Please avoid special characters like :\\/*?\"<>|"
+                )
 
             user_path = self.get_request_user_filepath(request, None)
             if full_info:

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/controlnet.py

@@ -37,6 +37,7 @@ import comfy.ldm.cascade.controlnet
 import comfy.cldm.mmdit
 import comfy.ldm.hydit.controlnet
 import comfy.ldm.flux.controlnet
+import comfy.ldm.qwen_image.controlnet
 import comfy.cldm.dit_embedder
 from typing import TYPE_CHECKING, Union
 if TYPE_CHECKING:
@@ -271,11 +272,11 @@ class ControlNet(ControlBase):
             self.cond_hint = None
             compression_ratio = self.compression_ratio
             if self.vae is not None:
-                compression_ratio *= self.vae.downscale_ratio
+                compression_ratio *= self.vae.spacial_compression_encode()
             else:
                 if self.latent_format is not None:
                     raise ValueError("This Controlnet needs a VAE but none was provided, please use a ControlNetApply node with a VAE input and connect it.")
-            self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * compression_ratio, x_noisy.shape[2] * compression_ratio, self.upscale_algorithm, "center")
+            self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[-1] * compression_ratio, x_noisy.shape[-2] * compression_ratio, self.upscale_algorithm, "center")
             self.cond_hint = self.preprocess_image(self.cond_hint)
             if self.vae is not None:
                 loaded_models = comfy.model_management.loaded_models(only_currently_used=True)
@@ -625,6 +626,15 @@ def load_controlnet_flux_instantx(sd, model_options={}):
     control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, concat_mask=concat_mask, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
     return control
 
+def load_controlnet_qwen_instantx(sd, model_options={}):
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd, model_options=model_options)
+    control_model = comfy.ldm.qwen_image.controlnet.QwenImageControlNetModel(operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
+    control_model = controlnet_load_state_dict(control_model, sd)
+    latent_format = comfy.latent_formats.Wan21()
+    extra_conds = []
+    control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
+    return control
+
 def convert_mistoline(sd):
     return comfy.utils.state_dict_prefix_replace(sd, {"single_controlnet_blocks.": "controlnet_single_blocks."})
 
@@ -698,8 +708,11 @@ def load_controlnet_state_dict(state_dict, model=None, model_options={}):
                 return load_controlnet_sd35(controlnet_data, model_options=model_options) #Stability sd3.5 format
             else:
                 return load_controlnet_mmdit(controlnet_data, model_options=model_options) #SD3 diffusers controlnet
+        elif "transformer_blocks.0.img_mlp.net.0.proj.weight" in controlnet_data:
+            return load_controlnet_qwen_instantx(controlnet_data, model_options=model_options)
         elif "controlnet_x_embedder.weight" in controlnet_data:
             return load_controlnet_flux_instantx(controlnet_data, model_options=model_options)
+
     elif "controlnet_blocks.0.linear.weight" in controlnet_data: #mistoline flux
         return load_controlnet_flux_xlabs_mistoline(convert_mistoline(controlnet_data), mistoline=True, model_options=model_options)
 

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/ace/model.py

@@ -19,6 +19,7 @@ import torch
 from torch import nn
 
 import comfy.model_management
+import comfy.patcher_extension
 
 from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
 from .attention import LinearTransformerBlock, t2i_modulate
@@ -343,7 +344,28 @@ class ACEStepTransformer2DModel(nn.Module):
         output = self.final_layer(hidden_states, embedded_timestep, output_length)
         return output
 
-    def forward(
+    def forward(self,
+        x,
+        timestep,
+        attention_mask=None,
+        context: Optional[torch.Tensor] = None,
+        text_attention_mask: Optional[torch.LongTensor] = None,
+        speaker_embeds: Optional[torch.FloatTensor] = None,
+        lyric_token_idx: Optional[torch.LongTensor] = None,
+        lyric_mask: Optional[torch.LongTensor] = None,
+        block_controlnet_hidden_states: Optional[Union[List[torch.Tensor], torch.Tensor]] = None,
+        controlnet_scale: Union[float, torch.Tensor] = 1.0,
+        lyrics_strength=1.0,
+        **kwargs
+    ):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, kwargs.get("transformer_options", {}))
+        ).execute(x, timestep, attention_mask, context, text_attention_mask, speaker_embeds, lyric_token_idx, lyric_mask, block_controlnet_hidden_states,
+                  controlnet_scale, lyrics_strength, **kwargs)
+
+    def _forward(
         self,
         x,
         timestep,

comfy/ldm/aura/mmdit.py

@@ -9,6 +9,7 @@ import torch.nn.functional as F
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.ops
+import comfy.patcher_extension
 import comfy.ldm.common_dit
 
 def modulate(x, shift, scale):
@@ -436,6 +437,13 @@ class MMDiT(nn.Module):
         return x + pos_encoding.reshape(1, -1, self.positional_encoding.shape[-1])
 
     def forward(self, x, timestep, context, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, transformer_options={}, **kwargs):
         patches_replace = transformer_options.get("patches_replace", {})
         # patchify x, add PE
         b, c, h, w = x.shape

comfy/ldm/chroma/model.py

@@ -5,6 +5,7 @@ from dataclasses import dataclass
 import torch
 from torch import Tensor, nn
 from einops import rearrange, repeat
+import comfy.patcher_extension
 import comfy.ldm.common_dit
 
 from comfy.ldm.flux.layers import (
@@ -253,6 +254,13 @@ class Chroma(nn.Module):
         return img
 
     def forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, guidance, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, guidance, control=None, transformer_options={}, **kwargs):
         bs, c, h, w = x.shape
         x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
 

comfy/ldm/cosmos/model.py

@@ -27,6 +27,8 @@ from torchvision import transforms
 from enum import Enum
 import logging
 
+import comfy.patcher_extension
+
 from .blocks import (
     FinalLayer,
     GeneralDITTransformerBlock,
@@ -435,6 +437,42 @@ class GeneralDIT(nn.Module):
         latent_condition_sigma: Optional[torch.Tensor] = None,
         condition_video_augment_sigma: Optional[torch.Tensor] = None,
         **kwargs,
+    ):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, kwargs.get("transformer_options", {}))
+        ).execute(x,
+                timesteps,
+                context,
+                attention_mask,
+                fps,
+                image_size,
+                padding_mask,
+                scalar_feature,
+                data_type,
+                latent_condition,
+                latent_condition_sigma,
+                condition_video_augment_sigma,
+                **kwargs)
+
+    def _forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        context: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        # crossattn_emb: torch.Tensor,
+        # crossattn_mask: Optional[torch.Tensor] = None,
+        fps: Optional[torch.Tensor] = None,
+        image_size: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        scalar_feature: Optional[torch.Tensor] = None,
+        data_type: Optional[DataType] = DataType.VIDEO,
+        latent_condition: Optional[torch.Tensor] = None,
+        latent_condition_sigma: Optional[torch.Tensor] = None,
+        condition_video_augment_sigma: Optional[torch.Tensor] = None,
+        **kwargs,
     ):
         """
         Args:

comfy/ldm/cosmos/predict2.py

@@ -11,6 +11,7 @@ import math
 from .position_embedding import VideoRopePosition3DEmb, LearnablePosEmbAxis
 from torchvision import transforms
 
+import comfy.patcher_extension
 from comfy.ldm.modules.attention import optimized_attention
 
 def apply_rotary_pos_emb(
@@ -805,7 +806,21 @@ class MiniTrainDIT(nn.Module):
         )
         return x_B_C_Tt_Hp_Wp
 
-    def forward(
+    def forward(self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        context: torch.Tensor,
+        fps: Optional[torch.Tensor] = None,
+        padding_mask: Optional[torch.Tensor] = None,
+        **kwargs,
+    ):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, kwargs.get("transformer_options", {}))
+        ).execute(x, timesteps, context, fps, padding_mask, **kwargs)
+
+    def _forward(
         self,
         x: torch.Tensor,
         timesteps: torch.Tensor,

comfy/ldm/flux/model.py

@@ -6,6 +6,7 @@ import torch
 from torch import Tensor, nn
 from einops import rearrange, repeat
 import comfy.ldm.common_dit
+import comfy.patcher_extension
 
 from .layers import (
     DoubleStreamBlock,
@@ -157,7 +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)
@@ -188,7 +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] :, ...]
 
@@ -214,6 +215,13 @@ class Flux(nn.Module):
         return img, repeat(img_ids, "h w c -> b (h w) c", b=bs)
 
     def forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, y, guidance, ref_latents, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, y=None, guidance=None, ref_latents=None, control=None, transformer_options={}, **kwargs):
         bs, c, h_orig, w_orig = x.shape
         patch_size = self.patch_size
 

comfy/ldm/hidream/model.py

@@ -13,6 +13,7 @@ from comfy.ldm.flux.layers import LastLayer
 
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.model_management
+import comfy.patcher_extension
 import comfy.ldm.common_dit
 
 
@@ -692,7 +693,23 @@ class HiDreamImageTransformer2DModel(nn.Module):
             raise NotImplementedError
         return x, x_masks, img_sizes
 
-    def forward(
+    def forward(self,
+        x: torch.Tensor,
+        t: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        encoder_hidden_states_llama3=None,
+        image_cond=None,
+        control = None,
+        transformer_options = {},
+    ):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, t, y, context, encoder_hidden_states_llama3, image_cond, control, transformer_options)
+
+    def _forward(
         self,
         x: torch.Tensor,
         t: torch.Tensor,

comfy/ldm/hunyuan3d/model.py

@@ -7,6 +7,7 @@ from comfy.ldm.flux.layers import (
     SingleStreamBlock,
     timestep_embedding,
 )
+import comfy.patcher_extension
 
 
 class Hunyuan3Dv2(nn.Module):
@@ -67,6 +68,13 @@ class Hunyuan3Dv2(nn.Module):
         self.final_layer = LastLayer(hidden_size, 1, in_channels, dtype=dtype, device=device, operations=operations)
 
     def forward(self, x, timestep, context, guidance=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, guidance, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, guidance=None, transformer_options={}, **kwargs):
         x = x.movedim(-1, -2)
         timestep = 1.0 - timestep
         txt = context

comfy/ldm/hunyuan_video/model.py

@@ -1,6 +1,7 @@
 #Based on Flux code because of weird hunyuan video code license.
 
 import torch
+import comfy.patcher_extension
 import comfy.ldm.flux.layers
 import comfy.ldm.modules.diffusionmodules.mmdit
 from comfy.ldm.modules.attention import optimized_attention
@@ -348,6 +349,13 @@ class HunyuanVideo(nn.Module):
         return repeat(img_ids, "t h w c -> b (t h w) c", b=bs)
 
     def forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, control=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, y, guidance, attention_mask, guiding_frame_index, ref_latent, control, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, y, guidance=None, attention_mask=None, guiding_frame_index=None, ref_latent=None, control=None, transformer_options={}, **kwargs):
         bs, c, t, h, w = x.shape
         img_ids = self.img_ids(x)
         txt_ids = torch.zeros((bs, context.shape[1], 3), device=x.device, dtype=x.dtype)

comfy/ldm/lightricks/model.py

@@ -1,5 +1,6 @@
 import torch
 from torch import nn
+import comfy.patcher_extension
 import comfy.ldm.modules.attention
 import comfy.ldm.common_dit
 from einops import rearrange
@@ -420,6 +421,13 @@ class LTXVModel(torch.nn.Module):
         self.patchifier = SymmetricPatchifier(1)
 
     def forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, attention_mask, frame_rate, transformer_options, keyframe_idxs, **kwargs)
+
+    def _forward(self, x, timestep, context, attention_mask, frame_rate=25, transformer_options={}, keyframe_idxs=None, **kwargs):
         patches_replace = transformer_options.get("patches_replace", {})
 
         orig_shape = list(x.shape)

comfy/ldm/lumina/model.py

@@ -11,6 +11,7 @@ import comfy.ldm.common_dit
 from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder
 from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
+import comfy.patcher_extension
 
 
 def modulate(x, scale):
@@ -590,8 +591,15 @@ class NextDiT(nn.Module):
 
         return padded_full_embed, mask, img_sizes, l_effective_cap_len, freqs_cis
 
-    # def forward(self, x, t, cap_feats, cap_mask):
     def forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, kwargs.get("transformer_options", {}))
+        ).execute(x, timesteps, context, num_tokens, attention_mask, **kwargs)
+
+    # def forward(self, x, t, cap_feats, cap_mask):
+    def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, **kwargs):
         t = 1.0 - timesteps
         cap_feats = context
         cap_mask = attention_mask

comfy/ldm/modules/diffusionmodules/mmdit.py

@@ -109,7 +109,7 @@ class PatchEmbed(nn.Module):
 def modulate(x, shift, scale):
     if shift is None:
         shift = torch.zeros_like(scale)
-    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+    return torch.addcmul(shift.unsqueeze(1), x, 1+ scale.unsqueeze(1))
 
 
 #################################################################################
@@ -564,10 +564,7 @@ class DismantledBlock(nn.Module):
         assert not self.pre_only
         attn1 = self.attn.post_attention(attn)
         attn2 = self.attn2.post_attention(attn2)
-        out1 = gate_msa.unsqueeze(1) * attn1
-        out2 = gate_msa2.unsqueeze(1) * attn2
-        x = x + out1
-        x = x + out2
+        x = gate_cat(x, gate_msa, gate_msa2, attn1, attn2)
         x = x + gate_mlp.unsqueeze(1) * self.mlp(
             modulate(self.norm2(x), shift_mlp, scale_mlp)
         )
@@ -594,6 +591,11 @@ class DismantledBlock(nn.Module):
             )
             return self.post_attention(attn, *intermediates)
 
+def gate_cat(x, gate_msa, gate_msa2, attn1, attn2):
+    out1 = gate_msa.unsqueeze(1) * attn1
+    out2 = gate_msa2.unsqueeze(1) * attn2
+    x = torch.stack([x, out1, out2], dim=0).sum(dim=0)
+    return x
 
 def block_mixing(*args, use_checkpoint=True, **kwargs):
     if use_checkpoint:

comfy/ldm/qwen_image/controlnet.py

@@ -0,0 +1,77 @@
+import torch
+import math
+
+from .model import QwenImageTransformer2DModel
+
+
+class QwenImageControlNetModel(QwenImageTransformer2DModel):
+    def __init__(
+        self,
+        extra_condition_channels=0,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs
+    ):
+        super().__init__(final_layer=False, dtype=dtype, device=device, operations=operations, **kwargs)
+        self.main_model_double = 60
+
+        # controlnet_blocks
+        self.controlnet_blocks = torch.nn.ModuleList([])
+        for _ in range(len(self.transformer_blocks)):
+            self.controlnet_blocks.append(operations.Linear(self.inner_dim, self.inner_dim, device=device, dtype=dtype))
+        self.controlnet_x_embedder = operations.Linear(self.in_channels + extra_condition_channels, self.inner_dim, device=device, dtype=dtype)
+
+    def forward(
+        self,
+        x,
+        timesteps,
+        context,
+        attention_mask=None,
+        guidance: torch.Tensor = None,
+        ref_latents=None,
+        hint=None,
+        transformer_options={},
+        **kwargs
+    ):
+        timestep = timesteps
+        encoder_hidden_states = context
+        encoder_hidden_states_mask = attention_mask
+
+        hidden_states, img_ids, orig_shape = self.process_img(x)
+        hint, _, _ = self.process_img(hint)
+
+        txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
+        txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        del ids, txt_ids, img_ids
+
+        hidden_states = self.img_in(hidden_states) + self.controlnet_x_embedder(hint)
+        encoder_hidden_states = self.txt_norm(encoder_hidden_states)
+        encoder_hidden_states = self.txt_in(encoder_hidden_states)
+
+        if guidance is not None:
+            guidance = guidance * 1000
+
+        temb = (
+            self.time_text_embed(timestep, hidden_states)
+            if guidance is None
+            else self.time_text_embed(timestep, guidance, hidden_states)
+        )
+
+        repeat = math.ceil(self.main_model_double / len(self.controlnet_blocks))
+
+        controlnet_block_samples = ()
+        for i, block in enumerate(self.transformer_blocks):
+            encoder_hidden_states, hidden_states = block(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_hidden_states_mask=encoder_hidden_states_mask,
+                temb=temb,
+                image_rotary_emb=image_rotary_emb,
+            )
+
+            controlnet_block_samples = controlnet_block_samples + (self.controlnet_blocks[i](hidden_states),) * repeat
+
+        return {"input": controlnet_block_samples[:self.main_model_double]}

comfy/ldm/qwen_image/model.py

@@ -9,6 +9,7 @@ from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
 from comfy.ldm.modules.attention import optimized_attention_masked
 from comfy.ldm.flux.layers import EmbedND
 import comfy.ldm.common_dit
+import comfy.patcher_extension
 
 class GELU(nn.Module):
     def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True, dtype=None, device=None, operations=None):
@@ -214,9 +215,9 @@ class QwenImageTransformerBlock(nn.Module):
             operations=operations,
         )
 
-    def _modulate(self, x, mod_params):
-        shift, scale, gate = mod_params.chunk(3, dim=-1)
-        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1), gate.unsqueeze(1)
+    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        shift, scale, gate = torch.chunk(mod_params, 3, dim=-1)
+        return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
 
     def forward(
         self,
@@ -248,11 +249,11 @@ class QwenImageTransformerBlock(nn.Module):
 
         img_normed2 = self.img_norm2(hidden_states)
         img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2)
-        hidden_states = hidden_states + img_gate2 * self.img_mlp(img_modulated2)
+        hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2))
 
         txt_normed2 = self.txt_norm2(encoder_hidden_states)
         txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2)
-        encoder_hidden_states = encoder_hidden_states + txt_gate2 * self.txt_mlp(txt_modulated2)
+        encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2))
 
         return encoder_hidden_states, hidden_states
 
@@ -275,7 +276,7 @@ class LastLayer(nn.Module):
     def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
         emb = self.linear(self.silu(conditioning_embedding))
         scale, shift = torch.chunk(emb, 2, dim=1)
-        x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
+        x = torch.addcmul(shift[:, None, :], self.norm(x), (1 + scale)[:, None, :])
         return x
 
 
@@ -293,6 +294,7 @@ class QwenImageTransformer2DModel(nn.Module):
         guidance_embeds: bool = False,
         axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
         image_model=None,
+        final_layer=True,
         dtype=None,
         device=None,
         operations=None,
@@ -300,6 +302,7 @@ class QwenImageTransformer2DModel(nn.Module):
         super().__init__()
         self.dtype = dtype
         self.patch_size = patch_size
+        self.in_channels = in_channels
         self.out_channels = out_channels or in_channels
         self.inner_dim = num_attention_heads * attention_head_dim
 
@@ -329,9 +332,9 @@ class QwenImageTransformer2DModel(nn.Module):
             for _ in range(num_layers)
         ])
 
-        self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
-        self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device)
-        self.gradient_checkpointing = False
+        if final_layer:
+            self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
+            self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device)
 
     def process_img(self, x, index=0, h_offset=0, w_offset=0):
         bs, c, t, h, w = x.shape
@@ -347,13 +350,20 @@ class QwenImageTransformer2DModel(nn.Module):
         h_offset = ((h_offset + (patch_size // 2)) // patch_size)
         w_offset = ((w_offset + (patch_size // 2)) // patch_size)
 
-        img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device)
         img_ids[:, :, 0] = img_ids[:, :, 1] + index
-        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)
-        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2)
         return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape
 
-    def forward(
+    def forward(self, x, timestep, context, attention_mask=None, guidance=None, ref_latents=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, attention_mask, guidance, ref_latents, transformer_options, **kwargs)
+
+    def _forward(
         self,
         x,
         timesteps,
@@ -362,6 +372,7 @@ class QwenImageTransformer2DModel(nn.Module):
         guidance: torch.Tensor = None,
         ref_latents=None,
         transformer_options={},
+        control=None,
         **kwargs
     ):
         timestep = timesteps
@@ -396,10 +407,11 @@ class QwenImageTransformer2DModel(nn.Module):
                 hidden_states = torch.cat([hidden_states, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
 
-        txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size), ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size)))
-        txt_ids = torch.linspace(txt_start, txt_start + context.shape[1], steps=context.shape[1], device=x.device, dtype=x.dtype).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
+        txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
+        txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
         ids = torch.cat((txt_ids, img_ids), dim=1)
         image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        del ids, txt_ids, img_ids
 
         hidden_states = self.img_in(hidden_states)
         encoder_hidden_states = self.txt_norm(encoder_hidden_states)
@@ -415,6 +427,7 @@ class QwenImageTransformer2DModel(nn.Module):
         )
 
         patches_replace = transformer_options.get("patches_replace", {})
+        patches = transformer_options.get("patches", {})
         blocks_replace = patches_replace.get("dit", {})
 
         for i, block in enumerate(self.transformer_blocks):
@@ -435,6 +448,19 @@ class QwenImageTransformer2DModel(nn.Module):
                     image_rotary_emb=image_rotary_emb,
                 )
 
+            if "double_block" in patches:
+                for p in patches["double_block"]:
+                    out = p({"img": hidden_states, "txt": encoder_hidden_states, "x": x, "block_index": i})
+                    hidden_states = out["img"]
+                    encoder_hidden_states = out["txt"]
+
+            if control is not None: # Controlnet
+                control_i = control.get("input")
+                if i < len(control_i):
+                    add = control_i[i]
+                    if add is not None:
+                        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

@@ -4,13 +4,14 @@ 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
 from comfy.ldm.flux.math import apply_rope
 import comfy.ldm.common_dit
 import comfy.model_management
+import comfy.patcher_extension
 
 
 def sinusoidal_embedding_1d(dim, position):
@@ -148,11 +149,14 @@ WAN_CROSSATTENTION_CLASSES = {
 
 def repeat_e(e, x):
     repeats = 1
-    if e.shape[1] > 1:
-        repeats = x.shape[1] // e.shape[1]
+    if e.size(1) > 1:
+        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):
@@ -219,15 +223,15 @@ class WanAttentionBlock(nn.Module):
 
         # self-attention
         y = self.self_attn(
-            self.norm1(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x),
+            torch.addcmul(repeat_e(e[0], x), self.norm1(x), 1 + repeat_e(e[1], x)),
             freqs)
 
-        x = x + y * repeat_e(e[2], x)
+        x = torch.addcmul(x, y, repeat_e(e[2], x))
 
         # cross-attention & ffn
         x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len)
-        y = self.ffn(self.norm2(x) * (1 + repeat_e(e[4], x)) + repeat_e(e[3], x))
-        x = x + y * repeat_e(e[5], x)
+        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
 
 
@@ -342,7 +346,7 @@ class Head(nn.Module):
         else:
             e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e.unsqueeze(2)).unbind(2)
 
-        x = (self.head(self.norm(x) * (1 + repeat_e(e[1], x)) + repeat_e(e[0], x)))
+        x = (self.head(torch.addcmul(repeat_e(e[0], x), self.norm(x), 1 + repeat_e(e[1], x))))
         return x
 
 
@@ -572,30 +576,49 @@ class WanModel(torch.nn.Module):
         x = self.unpatchify(x, grid_sizes)
         return x
 
-    def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
-        bs, c, t, h, w = x.shape
-        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
-
+    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,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, clip_fea, time_dim_concat, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
+        bs, c, t, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
+        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):
@@ -831,3 +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()
@@ -1102,9 +1110,10 @@ class WAN21(BaseModel):
             shape_image[1] = extra_channels
             image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device)
         else:
+            latent_dim = self.latent_format.latent_channels
             image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
-            for i in range(0, image.shape[1], 16):
-                image[:, i: i + 16] = self.process_latent_in(image[:, i: i + 16])
+            for i in range(0, image.shape[1], latent_dim):
+                image[:, i: i + latent_dim] = self.process_latent_in(image[:, i: i + latent_dim])
             image = utils.resize_to_batch_size(image, noise.shape[0])
 
         if extra_channels != image.shape[1] + 4:
@@ -1201,18 +1210,50 @@ class WAN21_Camera(WAN21):
             out['camera_conditions'] = comfy.conds.CONDRegular(camera_conditions)
         return out
 
-class WAN22(BaseModel):
+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(WAN21):
     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)
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
         self.image_to_video = image_to_video
 
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
-        cross_attn = kwargs.get("cross_attn", None)
-        if cross_attn is not None:
-            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
-
-        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        denoise_mask = kwargs.get("denoise_mask", None)
         if denoise_mask is not None:
             out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
         return out

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"
@@ -492,6 +494,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
     if '{}txt_norm.weight'.format(key_prefix) in state_dict_keys:  # Qwen Image
         dit_config = {}
         dit_config["image_model"] = "qwen_image"
+        dit_config["in_channels"] = state_dict['{}img_in.weight'.format(key_prefix)].shape[1]
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
         return dit_config
 
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:

comfy/model_management.py

@@ -621,7 +621,13 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
     else:
         minimum_memory_required = max(inference_memory, minimum_memory_required + extra_reserved_memory())
 
-    models = set(models)
+    models_temp = set()
+    for m in models:
+        models_temp.add(m)
+        for mm in m.model_patches_models():
+            models_temp.add(mm)
+
+    models = models_temp
 
     models_to_load = []
 

comfy/model_patcher.py

@@ -510,6 +510,9 @@ class ModelPatcher:
     def set_model_forward_timestep_embed_patch(self, patch):
         self.set_model_patch(patch, "forward_timestep_embed_patch")
 
+    def set_model_double_block_patch(self, patch):
+        self.set_model_patch(patch, "double_block")
+
     def add_object_patch(self, name, obj):
         self.object_patches[name] = obj
 
@@ -566,6 +569,30 @@ class ModelPatcher:
             if hasattr(wrap_func, "to"):
                 self.model_options["model_function_wrapper"] = wrap_func.to(device)
 
+    def model_patches_models(self):
+        to = self.model_options["transformer_options"]
+        models = []
+        if "patches" in to:
+            patches = to["patches"]
+            for name in patches:
+                patch_list = patches[name]
+                for i in range(len(patch_list)):
+                    if hasattr(patch_list[i], "models"):
+                        models += patch_list[i].models()
+        if "patches_replace" in to:
+            patches = to["patches_replace"]
+            for name in patches:
+                patch_list = patches[name]
+                for k in patch_list:
+                    if hasattr(patch_list[k], "models"):
+                        models += patch_list[k].models()
+        if "model_function_wrapper" in self.model_options:
+            wrap_func = self.model_options["model_function_wrapper"]
+            if hasattr(wrap_func, "models"):
+                models += wrap_func.models()
+
+        return models
+
     def model_dtype(self):
         if hasattr(self.model, "get_dtype"):
             return self.model.get_dtype()

comfy/patcher_extension.py

@@ -52,6 +52,7 @@ class WrappersMP:
     OUTER_SAMPLE = "outer_sample"
     PREPARE_SAMPLING = "prepare_sampling"
     SAMPLER_SAMPLE = "sampler_sample"
+    PREDICT_NOISE = "predict_noise"
     CALC_COND_BATCH = "calc_cond_batch"
     APPLY_MODEL = "apply_model"
     DIFFUSION_MODEL = "diffusion_model"

comfy/samplers.py

@@ -19,6 +19,7 @@ import comfy.model_patcher
 import comfy.patcher_extension
 import comfy.hooks
 import comfy.context_windows
+import comfy.utils
 import scipy.stats
 import numpy
 import threading
@@ -64,7 +65,7 @@ def get_area_and_mult(conds, x_in, timestep_in):
         if "mask_strength" in conds:
             mask_strength = conds["mask_strength"]
         mask = conds['mask']
-        assert (mask.shape[1:] == x_in.shape[2:])
+        # assert (mask.shape[1:] == x_in.shape[2:])
 
         mask = mask[:input_x.shape[0]]
         if area is not None:
@@ -72,7 +73,7 @@ def get_area_and_mult(conds, x_in, timestep_in):
                 mask = mask.narrow(i + 1, area[len(dims) + i], area[i])
 
         mask = mask * mask_strength
-        mask = mask.unsqueeze(1).repeat(input_x.shape[0] // mask.shape[0], input_x.shape[1], 1, 1)
+        mask = mask.unsqueeze(1).repeat((input_x.shape[0] // mask.shape[0], input_x.shape[1]) + (1, ) * (mask.ndim - 1))
     else:
         mask = torch.ones_like(input_x)
     mult = mask * strength
@@ -757,7 +758,10 @@ def resolve_areas_and_cond_masks_multidim(conditions, dims, device):
             if len(mask.shape) == len(dims):
                 mask = mask.unsqueeze(0)
             if mask.shape[1:] != dims:
-                mask = torch.nn.functional.interpolate(mask.unsqueeze(1), size=dims, mode='bilinear', align_corners=False).squeeze(1)
+                if mask.ndim < 4:
+                    mask = comfy.utils.common_upscale(mask.unsqueeze(1), dims[-1], dims[-2], 'bilinear', 'none').squeeze(1)
+                else:
+                    mask = comfy.utils.common_upscale(mask, dims[-1], dims[-2], 'bilinear', 'none')
 
             if modified.get("set_area_to_bounds", False): #TODO: handle dim != 2
                 bounds = torch.max(torch.abs(mask),dim=0).values.unsqueeze(0)
@@ -931,7 +935,7 @@ 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"]
 
@@ -1159,7 +1163,14 @@ class CFGGuider:
             self.original_conds[k] = comfy.sampler_helpers.convert_cond(conds[k])
 
     def __call__(self, *args, **kwargs):
-        return self.predict_noise(*args, **kwargs)
+        return self.outer_predict_noise(*args, **kwargs)
+
+    def outer_predict_noise(self, x, timestep, model_options={}, seed=None):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self.predict_noise,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREDICT_NOISE, self.model_options, is_model_options=True)
+        ).execute(x, timestep, model_options, seed)
 
     def predict_noise(self, x, timestep, model_options={}, seed=None):
         return sampling_function(self.inner_model, x, timestep, self.conds.get("negative", None), self.conds.get("positive", None), self.cfg, model_options=model_options, seed=seed)

comfy/sd1_clip.py

@@ -534,7 +534,10 @@ class SDTokenizer:
         min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding)
 
         text = escape_important(text)
-        parsed_weights = token_weights(text, 1.0)
+        if kwargs.get("disable_weights", False):
+            parsed_weights = [(text, 1.0)]
+        else:
+            parsed_weights = token_weights(text, 1.0)
 
         # tokenize words
         tokens = []

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/text_encoders/llama.py

@@ -27,6 +27,7 @@ class Llama2Config:
     rms_norm_add = False
     mlp_activation = "silu"
     qkv_bias = False
+    rope_dims = None
 
 @dataclass
 class Qwen25_3BConfig:
@@ -44,6 +45,7 @@ class Qwen25_3BConfig:
     rms_norm_add = False
     mlp_activation = "silu"
     qkv_bias = True
+    rope_dims = None
 
 @dataclass
 class Qwen25_7BVLI_Config:
@@ -61,6 +63,7 @@ class Qwen25_7BVLI_Config:
     rms_norm_add = False
     mlp_activation = "silu"
     qkv_bias = True
+    rope_dims = [16, 24, 24]
 
 @dataclass
 class Gemma2_2B_Config:
@@ -78,6 +81,7 @@ class Gemma2_2B_Config:
     rms_norm_add = True
     mlp_activation = "gelu_pytorch_tanh"
     qkv_bias = False
+    rope_dims = None
 
 class RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
@@ -102,7 +106,7 @@ def rotate_half(x):
     return torch.cat((-x2, x1), dim=-1)
 
 
-def precompute_freqs_cis(head_dim, position_ids, theta, device=None):
+def precompute_freqs_cis(head_dim, position_ids, theta, rope_dims=None, device=None):
     theta_numerator = torch.arange(0, head_dim, 2, device=device).float()
     inv_freq = 1.0 / (theta ** (theta_numerator / head_dim))
 
@@ -112,12 +116,20 @@ def precompute_freqs_cis(head_dim, position_ids, theta, device=None):
     emb = torch.cat((freqs, freqs), dim=-1)
     cos = emb.cos()
     sin = emb.sin()
+    if rope_dims is not None and position_ids.shape[0] > 1:
+        mrope_section = rope_dims * 2
+        cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0)
+        sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(0)
+    else:
+        cos = cos.unsqueeze(1)
+        sin = sin.unsqueeze(1)
+
     return (cos, sin)
 
 
 def apply_rope(xq, xk, freqs_cis):
-    cos = freqs_cis[0].unsqueeze(1)
-    sin = freqs_cis[1].unsqueeze(1)
+    cos = freqs_cis[0]
+    sin = freqs_cis[1]
     q_embed = (xq * cos) + (rotate_half(xq) * sin)
     k_embed = (xk * cos) + (rotate_half(xk) * sin)
     return q_embed, k_embed
@@ -292,6 +304,7 @@ class Llama2_(nn.Module):
         freqs_cis = precompute_freqs_cis(self.config.head_dim,
                                          position_ids,
                                          self.config.rope_theta,
+                                         self.config.rope_dims,
                                          device=x.device)
 
         mask = None

comfy/text_encoders/qwen_image.py

@@ -15,7 +15,7 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen25_7b", tokenizer=Qwen25_7BVLITokenizer)
         self.llama_template = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
-        self.llama_template_images = "<|im_start|>system\nDescribe the key features of the input image \\(color, shape, size, texture, objects, background\\), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n"
+        self.llama_template_images = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n<|vision_start|><|image_pad|><|vision_end|>{}<|im_end|>\n<|im_start|>assistant\n"
 
     def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, images=[], **kwargs):
         if llama_template is None:
@@ -25,7 +25,7 @@ class QwenImageTokenizer(sd1_clip.SD1Tokenizer):
                 llama_text = self.llama_template.format(text)
         else:
             llama_text = llama_template.format(text)
-        tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, **kwargs)
+        tokens = super().tokenize_with_weights(llama_text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
         key_name = next(iter(tokens))
         embed_count = 0
         qwen_tokens = tokens[key_name]

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/_input_impl/video_types.py

@@ -8,6 +8,7 @@ import av
 import io
 import json
 import numpy as np
+import math
 import torch
 from comfy_api.latest._util import VideoContainer, VideoCodec, VideoComponents
 
@@ -282,8 +283,6 @@ class VideoFromComponents(VideoInput):
             if self.__components.audio:
                 audio_sample_rate = int(self.__components.audio['sample_rate'])
                 audio_stream = output.add_stream('aac', rate=audio_sample_rate)
-                audio_stream.sample_rate = audio_sample_rate
-                audio_stream.format = 'fltp'
 
             # Encode video
             for i, frame in enumerate(self.__components.images):
@@ -298,27 +297,12 @@ class VideoFromComponents(VideoInput):
             output.mux(packet)
 
             if audio_stream and self.__components.audio:
-                # Encode audio
-                samples_per_frame = int(audio_sample_rate / frame_rate)
-                num_frames = self.__components.audio['waveform'].shape[2] // samples_per_frame
-                for i in range(num_frames):
-                    start = i * samples_per_frame
-                    end = start + samples_per_frame
-                    # TODO(Feature) - Add support for stereo audio
-                    chunk = (
-                        self.__components.audio["waveform"][0, 0, start:end]
-                        .unsqueeze(0)
-                        .contiguous()
-                        .numpy()
-                    )
-                    audio_frame = av.AudioFrame.from_ndarray(chunk, format='fltp', layout='mono')
-                    audio_frame.sample_rate = audio_sample_rate
-                    audio_frame.pts = i * samples_per_frame
-                    for packet in audio_stream.encode(audio_frame):
-                        output.mux(packet)
-
-                # Flush audio
-                for packet in audio_stream.encode(None):
-                    output.mux(packet)
-
+                waveform = self.__components.audio['waveform']
+                waveform = waveform[:, :, :math.ceil((audio_sample_rate / frame_rate) * self.__components.images.shape[0])]
+                frame = av.AudioFrame.from_ndarray(waveform.movedim(2, 1).reshape(1, -1).float().numpy(), format='flt', layout='mono' if waveform.shape[1] == 1 else 'stereo')
+                frame.sample_rate = audio_sample_rate
+                frame.pts = 0
+                output.mux(audio_stream.encode(frame))
 
+                # Flush encoder
+                output.mux(audio_stream.encode(None))

comfy_api/latest/_io.py

@@ -726,6 +726,18 @@ class SEGS(ComfyTypeIO):
 class AnyType(ComfyTypeIO):
     Type = Any
 
+@comfytype(io_type="MODEL_PATCH")
+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
@@ -1580,6 +1592,7 @@ class _IO:
     Model = Model
     ClipVision = ClipVision
     ClipVisionOutput = ClipVisionOutput
+    AudioEncoderOutput = AudioEncoderOutput
     StyleModel = StyleModel
     Gligen = Gligen
     UpscaleModel = UpscaleModel

comfy_api_nodes/apis/__init__.py

@@ -1315,6 +1315,7 @@ class KlingTaskStatus(str, Enum):
 class KlingTextToVideoModelName(str, Enum):
     kling_v1 = 'kling-v1'
     kling_v1_6 = 'kling-v1-6'
+    kling_v2_1_master = 'kling-v2-1-master'
 
 
 class KlingVideoGenAspectRatio(str, Enum):
@@ -1347,6 +1348,8 @@ class KlingVideoGenModelName(str, Enum):
     kling_v1_5 = 'kling-v1-5'
     kling_v1_6 = 'kling-v1-6'
     kling_v2_master = 'kling-v2-master'
+    kling_v2_1 = 'kling-v2-1'
+    kling_v2_1_master = 'kling-v2-1-master'
 
 
 class KlingVideoResult(BaseModel):
@@ -1620,13 +1623,14 @@ class MinimaxTaskResultResponse(BaseModel):
     task_id: str = Field(..., description='The task ID being queried.')
 
 
-class Model(str, Enum):
+class MiniMaxModel(str, Enum):
     T2V_01_Director = 'T2V-01-Director'
     I2V_01_Director = 'I2V-01-Director'
     S2V_01 = 'S2V-01'
     I2V_01 = 'I2V-01'
     I2V_01_live = 'I2V-01-live'
     T2V_01 = 'T2V-01'
+    Hailuo_02 = 'MiniMax-Hailuo-02'
 
 
 class SubjectReferenceItem(BaseModel):
@@ -1648,7 +1652,7 @@ class MinimaxVideoGenerationRequest(BaseModel):
         None,
         description='URL or base64 encoding of the first frame image. Required when model is I2V-01, I2V-01-Director, or I2V-01-live.',
     )
-    model: Model = Field(
+    model: MiniMaxModel = Field(
         ...,
         description='Required. ID of model. Options: T2V-01-Director, I2V-01-Director, S2V-01, I2V-01, I2V-01-live, T2V-01',
     )
@@ -1665,6 +1669,14 @@ class MinimaxVideoGenerationRequest(BaseModel):
         None,
         description='Only available when model is S2V-01. The model will generate a video based on the subject uploaded through this parameter.',
     )
+    duration: Optional[int] = Field(
+        None,
+        description="The length of the output video in seconds."
+    )
+    resolution: Optional[str] = Field(
+        None,
+        description="The dimensions of the video display. 1080p corresponds to 1920 x 1080 pixels, 768p corresponds to 1366 x 768 pixels."
+    )
 
 
 class MinimaxVideoGenerationResponse(BaseModel):

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

@@ -4,8 +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
 
@@ -22,6 +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,
@@ -32,6 +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,
 )
 
 
@@ -46,6 +52,16 @@ class GeminiModel(str, Enum):
 
     gemini_2_5_pro_preview_05_06 = "gemini-2.5-pro-preview-05-06"
     gemini_2_5_flash_preview_04_17 = "gemini-2.5-flash-preview-04-17"
+    gemini_2_5_pro = "gemini-2.5-pro"
+    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(
@@ -70,6 +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.
@@ -97,7 +242,7 @@ class GeminiNode(ComfyNodeABC):
                     {
                         "tooltip": "The Gemini model to use for generating responses.",
                         "options": [model.value for model in GeminiModel],
-                        "default": GeminiModel.gemini_2_5_pro_preview_05_06.value,
+                        "default": GeminiModel.gemini_2_5_pro.value,
                     },
                 ),
                 "seed": (
@@ -154,59 +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.
@@ -266,43 +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,
@@ -318,11 +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))
@@ -346,9 +401,29 @@ 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:
-            PromptServer.instance.send_progress_text(output_text, node_id=unique_id)
+            # 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,
+            )
 
         return (output_text or "Empty response from Gemini model...",)
 
@@ -437,12 +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_api_nodes/nodes_ideogram.py

@@ -1,8 +1,8 @@
-from comfy.comfy_types.node_typing import IO, ComfyNodeABC, InputTypeDict
-from inspect import cleandoc
+from io import BytesIO
+from typing_extensions import override
+from comfy_api.latest import ComfyExtension, io as comfy_io
 from PIL import Image
 import numpy as np
-import io
 import torch
 from comfy_api_nodes.apis import (
     IdeogramGenerateRequest,
@@ -246,90 +246,81 @@ def display_image_urls_on_node(image_urls, node_id):
             PromptServer.instance.send_progress_text(urls_text, node_id)
 
 
-class IdeogramV1(ComfyNodeABC):
-    """
-    Generates images using the Ideogram V1 model.
-    """
-
-    def __init__(self):
-        pass
+class IdeogramV1(comfy_io.ComfyNode):
 
     @classmethod
-    def INPUT_TYPES(cls) -> InputTypeDict:
-        return {
-            "required": {
-                "prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Prompt for the image generation",
-                    },
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="IdeogramV1",
+            display_name="Ideogram V1",
+            category="api node/image/Ideogram",
+            description="Generates images using the Ideogram V1 model.",
+            inputs=[
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Prompt for the image generation",
                 ),
-                "turbo": (
-                    IO.BOOLEAN,
-                    {
-                        "default": False,
-                        "tooltip": "Whether to use turbo mode (faster generation, potentially lower quality)",
-                    }
+                comfy_io.Boolean.Input(
+                    "turbo",
+                    default=False,
+                    tooltip="Whether to use turbo mode (faster generation, potentially lower quality)",
                 ),
-            },
-            "optional": {
-                "aspect_ratio": (
-                    IO.COMBO,
-                    {
-                        "options": list(V1_V2_RATIO_MAP.keys()),
-                        "default": "1:1",
-                        "tooltip": "The aspect ratio for image generation.",
-                    },
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=list(V1_V2_RATIO_MAP.keys()),
+                    default="1:1",
+                    tooltip="The aspect ratio for image generation.",
+                    optional=True,
                 ),
-                "magic_prompt_option": (
-                    IO.COMBO,
-                    {
-                        "options": ["AUTO", "ON", "OFF"],
-                        "default": "AUTO",
-                        "tooltip": "Determine if MagicPrompt should be used in generation",
-                    },
+                comfy_io.Combo.Input(
+                    "magic_prompt_option",
+                    options=["AUTO", "ON", "OFF"],
+                    default="AUTO",
+                    tooltip="Determine if MagicPrompt should be used in generation",
+                    optional=True,
                 ),
-                "seed": (
-                    IO.INT,
-                    {
-                        "default": 0,
-                        "min": 0,
-                        "max": 2147483647,
-                        "step": 1,
-                        "control_after_generate": True,
-                        "display": "number",
-                    },
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    control_after_generate=True,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
-                "negative_prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Description of what to exclude from the image",
-                    },
+                comfy_io.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Description of what to exclude from the image",
+                    optional=True,
                 ),
-                "num_images": (
-                    IO.INT,
-                    {"default": 1, "min": 1, "max": 8, "step": 1, "display": "number"},
+                comfy_io.Int.Input(
+                    "num_images",
+                    default=1,
+                    min=1,
+                    max=8,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
-            },
-            "hidden": {
-                "auth_token": "AUTH_TOKEN_COMFY_ORG",
-                "comfy_api_key": "API_KEY_COMFY_ORG",
-                "unique_id": "UNIQUE_ID",
-            },
-        }
+            ],
+            outputs=[
+                comfy_io.Image.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+        )
 
-    RETURN_TYPES = (IO.IMAGE,)
-    FUNCTION = "api_call"
-    CATEGORY = "api node/image/Ideogram"
-    DESCRIPTION = cleandoc(__doc__ or "")
-    API_NODE = True
-
-    async def api_call(
-        self,
+    @classmethod
+    async def execute(
+        cls,
         prompt,
         turbo=False,
         aspect_ratio="1:1",
@@ -337,13 +328,15 @@ class IdeogramV1(ComfyNodeABC):
         seed=0,
         negative_prompt="",
         num_images=1,
-        unique_id=None,
-        **kwargs,
     ):
         # Determine the model based on turbo setting
         aspect_ratio = V1_V2_RATIO_MAP.get(aspect_ratio, None)
         model = "V_1_TURBO" if turbo else "V_1"
 
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
         operation = SynchronousOperation(
             endpoint=ApiEndpoint(
                 path="/proxy/ideogram/generate",
@@ -364,7 +357,7 @@ class IdeogramV1(ComfyNodeABC):
                     negative_prompt=negative_prompt if negative_prompt else None,
                 )
             ),
-            auth_kwargs=kwargs,
+            auth_kwargs=auth,
         )
 
         response = await operation.execute()
@@ -377,93 +370,85 @@ class IdeogramV1(ComfyNodeABC):
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
 
-        display_image_urls_on_node(image_urls, unique_id)
-        return (await download_and_process_images(image_urls),)
+        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_and_process_images(image_urls))
 
 
-class IdeogramV2(ComfyNodeABC):
-    """
-    Generates images using the Ideogram V2 model.
-    """
-
-    def __init__(self):
-        pass
+class IdeogramV2(comfy_io.ComfyNode):
 
     @classmethod
-    def INPUT_TYPES(cls) -> InputTypeDict:
-        return {
-            "required": {
-                "prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Prompt for the image generation",
-                    },
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="IdeogramV2",
+            display_name="Ideogram V2",
+            category="api node/image/Ideogram",
+            description="Generates images using the Ideogram V2 model.",
+            inputs=[
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Prompt for the image generation",
                 ),
-                "turbo": (
-                    IO.BOOLEAN,
-                    {
-                        "default": False,
-                        "tooltip": "Whether to use turbo mode (faster generation, potentially lower quality)",
-                    }
+                comfy_io.Boolean.Input(
+                    "turbo",
+                    default=False,
+                    tooltip="Whether to use turbo mode (faster generation, potentially lower quality)",
                 ),
-            },
-            "optional": {
-                "aspect_ratio": (
-                    IO.COMBO,
-                    {
-                        "options": list(V1_V2_RATIO_MAP.keys()),
-                        "default": "1:1",
-                        "tooltip": "The aspect ratio for image generation. Ignored if resolution is not set to AUTO.",
-                    },
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=list(V1_V2_RATIO_MAP.keys()),
+                    default="1:1",
+                    tooltip="The aspect ratio for image generation. Ignored if resolution is not set to AUTO.",
+                    optional=True,
                 ),
-                "resolution": (
-                    IO.COMBO,
-                    {
-                        "options": list(V1_V1_RES_MAP.keys()),
-                        "default": "Auto",
-                        "tooltip": "The resolution for image generation. If not set to AUTO, this overrides the aspect_ratio setting.",
-                    },
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=list(V1_V1_RES_MAP.keys()),
+                    default="Auto",
+                    tooltip="The resolution for image generation. "
+                            "If not set to AUTO, this overrides the aspect_ratio setting.",
+                    optional=True,
                 ),
-                "magic_prompt_option": (
-                    IO.COMBO,
-                    {
-                        "options": ["AUTO", "ON", "OFF"],
-                        "default": "AUTO",
-                        "tooltip": "Determine if MagicPrompt should be used in generation",
-                    },
+                comfy_io.Combo.Input(
+                    "magic_prompt_option",
+                    options=["AUTO", "ON", "OFF"],
+                    default="AUTO",
+                    tooltip="Determine if MagicPrompt should be used in generation",
+                    optional=True,
                 ),
-                "seed": (
-                    IO.INT,
-                    {
-                        "default": 0,
-                        "min": 0,
-                        "max": 2147483647,
-                        "step": 1,
-                        "control_after_generate": True,
-                        "display": "number",
-                    },
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    control_after_generate=True,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
-                "style_type": (
-                    IO.COMBO,
-                    {
-                        "options": ["AUTO", "GENERAL", "REALISTIC", "DESIGN", "RENDER_3D", "ANIME"],
-                        "default": "NONE",
-                        "tooltip": "Style type for generation (V2 only)",
-                    },
+                comfy_io.Combo.Input(
+                    "style_type",
+                    options=["AUTO", "GENERAL", "REALISTIC", "DESIGN", "RENDER_3D", "ANIME"],
+                    default="NONE",
+                    tooltip="Style type for generation (V2 only)",
+                    optional=True,
                 ),
-                "negative_prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Description of what to exclude from the image",
-                    },
+                comfy_io.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Description of what to exclude from the image",
+                    optional=True,
                 ),
-                "num_images": (
-                    IO.INT,
-                    {"default": 1, "min": 1, "max": 8, "step": 1, "display": "number"},
+                comfy_io.Int.Input(
+                    "num_images",
+                    default=1,
+                    min=1,
+                    max=8,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
                 #"color_palette": (
                 #    IO.STRING,
@@ -473,22 +458,20 @@ class IdeogramV2(ComfyNodeABC):
                 #        "tooltip": "Color palette preset name or hex colors with weights",
                 #    },
                 #),
-            },
-            "hidden": {
-                "auth_token": "AUTH_TOKEN_COMFY_ORG",
-                "comfy_api_key": "API_KEY_COMFY_ORG",
-                "unique_id": "UNIQUE_ID",
-            },
-        }
+            ],
+            outputs=[
+                comfy_io.Image.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+        )
 
-    RETURN_TYPES = (IO.IMAGE,)
-    FUNCTION = "api_call"
-    CATEGORY = "api node/image/Ideogram"
-    DESCRIPTION = cleandoc(__doc__ or "")
-    API_NODE = True
-
-    async def api_call(
-        self,
+    @classmethod
+    async def execute(
+        cls,
         prompt,
         turbo=False,
         aspect_ratio="1:1",
@@ -499,8 +482,6 @@ class IdeogramV2(ComfyNodeABC):
         negative_prompt="",
         num_images=1,
         color_palette="",
-        unique_id=None,
-        **kwargs,
     ):
         aspect_ratio = V1_V2_RATIO_MAP.get(aspect_ratio, None)
         resolution = V1_V1_RES_MAP.get(resolution, None)
@@ -517,6 +498,10 @@ class IdeogramV2(ComfyNodeABC):
         else:
             final_aspect_ratio = aspect_ratio if aspect_ratio != "ASPECT_1_1" else None
 
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
         operation = SynchronousOperation(
             endpoint=ApiEndpoint(
                 path="/proxy/ideogram/generate",
@@ -540,7 +525,7 @@ class IdeogramV2(ComfyNodeABC):
                     color_palette=color_palette if color_palette else None,
                 )
             ),
-            auth_kwargs=kwargs,
+            auth_kwargs=auth,
         )
 
         response = await operation.execute()
@@ -553,108 +538,99 @@ class IdeogramV2(ComfyNodeABC):
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
 
-        display_image_urls_on_node(image_urls, unique_id)
-        return (await download_and_process_images(image_urls),)
+        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_and_process_images(image_urls))
 
-class IdeogramV3(ComfyNodeABC):
-    """
-    Generates images using the Ideogram V3 model. Supports both regular image generation from text prompts and image editing with mask.
-    """
 
-    def __init__(self):
-        pass
+class IdeogramV3(comfy_io.ComfyNode):
 
     @classmethod
-    def INPUT_TYPES(cls) -> InputTypeDict:
-        return {
-            "required": {
-                "prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Prompt for the image generation or editing",
-                    },
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="IdeogramV3",
+            display_name="Ideogram V3",
+            category="api node/image/Ideogram",
+            description="Generates images using the Ideogram V3 model. "
+                        "Supports both regular image generation from text prompts and image editing with mask.",
+            inputs=[
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Prompt for the image generation or editing",
                 ),
-            },
-            "optional": {
-                "image": (
-                    IO.IMAGE,
-                    {
-                        "default": None,
-                        "tooltip": "Optional reference image for image editing.",
-                    },
+                comfy_io.Image.Input(
+                    "image",
+                    tooltip="Optional reference image for image editing.",
+                    optional=True,
                 ),
-                "mask": (
-                    IO.MASK,
-                    {
-                        "default": None,
-                        "tooltip": "Optional mask for inpainting (white areas will be replaced)",
-                    },
+                comfy_io.Mask.Input(
+                    "mask",
+                    tooltip="Optional mask for inpainting (white areas will be replaced)",
+                    optional=True,
                 ),
-                "aspect_ratio": (
-                    IO.COMBO,
-                    {
-                        "options": list(V3_RATIO_MAP.keys()),
-                        "default": "1:1",
-                        "tooltip": "The aspect ratio for image generation. Ignored if resolution is not set to Auto.",
-                    },
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=list(V3_RATIO_MAP.keys()),
+                    default="1:1",
+                    tooltip="The aspect ratio for image generation. Ignored if resolution is not set to Auto.",
+                    optional=True,
                 ),
-                "resolution": (
-                    IO.COMBO,
-                    {
-                        "options": V3_RESOLUTIONS,
-                        "default": "Auto",
-                        "tooltip": "The resolution for image generation. If not set to Auto, this overrides the aspect_ratio setting.",
-                    },
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=V3_RESOLUTIONS,
+                    default="Auto",
+                    tooltip="The resolution for image generation. "
+                            "If not set to Auto, this overrides the aspect_ratio setting.",
+                    optional=True,
                 ),
-                "magic_prompt_option": (
-                    IO.COMBO,
-                    {
-                        "options": ["AUTO", "ON", "OFF"],
-                        "default": "AUTO",
-                        "tooltip": "Determine if MagicPrompt should be used in generation",
-                    },
+                comfy_io.Combo.Input(
+                    "magic_prompt_option",
+                    options=["AUTO", "ON", "OFF"],
+                    default="AUTO",
+                    tooltip="Determine if MagicPrompt should be used in generation",
+                    optional=True,
                 ),
-                "seed": (
-                    IO.INT,
-                    {
-                        "default": 0,
-                        "min": 0,
-                        "max": 2147483647,
-                        "step": 1,
-                        "control_after_generate": True,
-                        "display": "number",
-                    },
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    control_after_generate=True,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
-                "num_images": (
-                    IO.INT,
-                    {"default": 1, "min": 1, "max": 8, "step": 1, "display": "number"},
+                comfy_io.Int.Input(
+                    "num_images",
+                    default=1,
+                    min=1,
+                    max=8,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    optional=True,
                 ),
-                "rendering_speed": (
-                    IO.COMBO,
-                    {
-                        "options": ["BALANCED", "TURBO", "QUALITY"],
-                        "default": "BALANCED",
-                        "tooltip": "Controls the trade-off between generation speed and quality",
-                    },
+                comfy_io.Combo.Input(
+                    "rendering_speed",
+                    options=["BALANCED", "TURBO", "QUALITY"],
+                    default="BALANCED",
+                    tooltip="Controls the trade-off between generation speed and quality",
+                    optional=True,
                 ),
-            },
-            "hidden": {
-                "auth_token": "AUTH_TOKEN_COMFY_ORG",
-                "comfy_api_key": "API_KEY_COMFY_ORG",
-                "unique_id": "UNIQUE_ID",
-            },
-        }
+            ],
+            outputs=[
+                comfy_io.Image.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+        )
 
-    RETURN_TYPES = (IO.IMAGE,)
-    FUNCTION = "api_call"
-    CATEGORY = "api node/image/Ideogram"
-    DESCRIPTION = cleandoc(__doc__ or "")
-    API_NODE = True
-
-    async def api_call(
-        self,
+    @classmethod
+    async def execute(
+        cls,
         prompt,
         image=None,
         mask=None,
@@ -664,9 +640,11 @@ class IdeogramV3(ComfyNodeABC):
         seed=0,
         num_images=1,
         rendering_speed="BALANCED",
-        unique_id=None,
-        **kwargs,
     ):
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
         # Check if both image and mask are provided for editing mode
         if image is not None and mask is not None:
             # Edit mode
@@ -686,7 +664,7 @@ class IdeogramV3(ComfyNodeABC):
             # Process image
             img_np = (input_tensor.numpy() * 255).astype(np.uint8)
             img = Image.fromarray(img_np)
-            img_byte_arr = io.BytesIO()
+            img_byte_arr = BytesIO()
             img.save(img_byte_arr, format="PNG")
             img_byte_arr.seek(0)
             img_binary = img_byte_arr
@@ -695,7 +673,7 @@ class IdeogramV3(ComfyNodeABC):
             # Process mask - white areas will be replaced
             mask_np = (mask.squeeze().cpu().numpy() * 255).astype(np.uint8)
             mask_img = Image.fromarray(mask_np)
-            mask_byte_arr = io.BytesIO()
+            mask_byte_arr = BytesIO()
             mask_img.save(mask_byte_arr, format="PNG")
             mask_byte_arr.seek(0)
             mask_binary = mask_byte_arr
@@ -729,7 +707,7 @@ class IdeogramV3(ComfyNodeABC):
                     "mask": mask_binary,
                 },
                 content_type="multipart/form-data",
-                auth_kwargs=kwargs,
+                auth_kwargs=auth,
             )
 
         elif image is not None or mask is not None:
@@ -770,7 +748,7 @@ class IdeogramV3(ComfyNodeABC):
                     response_model=IdeogramGenerateResponse,
                 ),
                 request=gen_request,
-                auth_kwargs=kwargs,
+                auth_kwargs=auth,
             )
 
         # Execute the operation and process response
@@ -784,18 +762,18 @@ class IdeogramV3(ComfyNodeABC):
         if not image_urls:
             raise Exception("No image URLs were generated in the response")
 
-        display_image_urls_on_node(image_urls, unique_id)
-        return (await download_and_process_images(image_urls),)
+        display_image_urls_on_node(image_urls, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_and_process_images(image_urls))
 
 
-NODE_CLASS_MAPPINGS = {
-    "IdeogramV1": IdeogramV1,
-    "IdeogramV2": IdeogramV2,
-    "IdeogramV3": IdeogramV3,
-}
+class IdeogramExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[comfy_io.ComfyNode]]:
+        return [
+            IdeogramV1,
+            IdeogramV2,
+            IdeogramV3,
+        ]
 
-NODE_DISPLAY_NAME_MAPPINGS = {
-    "IdeogramV1": "Ideogram V1",
-    "IdeogramV2": "Ideogram V2",
-    "IdeogramV3": "Ideogram V3",
-}
+async def comfy_entrypoint() -> IdeogramExtension:
+    return IdeogramExtension()

comfy_api_nodes/nodes_kling.py

@@ -421,6 +421,8 @@ class KlingTextToVideoNode(KlingNodeBase):
             "pro mode / 10s duration / kling-v2-master": ("pro", "10", "kling-v2-master"),
             "standard mode / 5s duration / kling-v2-master": ("std", "5", "kling-v2-master"),
             "standard mode / 10s duration / kling-v2-master": ("std", "10", "kling-v2-master"),
+            "pro mode / 5s duration / kling-v2-1-master": ("pro", "5", "kling-v2-1-master"),
+            "pro mode / 10s duration / kling-v2-1-master": ("pro", "10", "kling-v2-1-master"),
         }
 
     @classmethod

comfy_api_nodes/nodes_minimax.py

@@ -1,3 +1,4 @@
+from inspect import cleandoc
 from typing import Union
 import logging
 import torch
@@ -10,7 +11,7 @@ from comfy_api_nodes.apis import (
     MinimaxFileRetrieveResponse,
     MinimaxTaskResultResponse,
     SubjectReferenceItem,
-    Model
+    MiniMaxModel
 )
 from comfy_api_nodes.apis.client import (
     ApiEndpoint,
@@ -84,7 +85,6 @@ class MinimaxTextToVideoNode:
     FUNCTION = "generate_video"
     CATEGORY = "api node/video/MiniMax"
     API_NODE = True
-    OUTPUT_NODE = True
 
     async def generate_video(
         self,
@@ -121,7 +121,7 @@ class MinimaxTextToVideoNode:
                 response_model=MinimaxVideoGenerationResponse,
             ),
             request=MinimaxVideoGenerationRequest(
-                model=Model(model),
+                model=MiniMaxModel(model),
                 prompt=prompt_text,
                 callback_url=None,
                 first_frame_image=image_url,
@@ -251,7 +251,6 @@ class MinimaxImageToVideoNode(MinimaxTextToVideoNode):
     FUNCTION = "generate_video"
     CATEGORY = "api node/video/MiniMax"
     API_NODE = True
-    OUTPUT_NODE = True
 
 
 class MinimaxSubjectToVideoNode(MinimaxTextToVideoNode):
@@ -313,7 +312,181 @@ class MinimaxSubjectToVideoNode(MinimaxTextToVideoNode):
     FUNCTION = "generate_video"
     CATEGORY = "api node/video/MiniMax"
     API_NODE = True
-    OUTPUT_NODE = True
+
+
+class MinimaxHailuoVideoNode:
+    """Generates videos from prompt, with optional start frame using the new MiniMax Hailuo-02 model."""
+
+    @classmethod
+    def INPUT_TYPES(s):
+        return {
+            "required": {
+                "prompt_text": (
+                    "STRING",
+                    {
+                        "multiline": True,
+                        "default": "",
+                        "tooltip": "Text prompt to guide the video generation.",
+                    },
+                ),
+            },
+            "optional": {
+                "seed": (
+                    IO.INT,
+                    {
+                        "default": 0,
+                        "min": 0,
+                        "max": 0xFFFFFFFFFFFFFFFF,
+                        "control_after_generate": True,
+                        "tooltip": "The random seed used for creating the noise.",
+                    },
+                ),
+                "first_frame_image": (
+                    IO.IMAGE,
+                    {
+                        "tooltip": "Optional image to use as the first frame to generate a video."
+                    },
+                ),
+                "prompt_optimizer": (
+                    IO.BOOLEAN,
+                    {
+                        "tooltip": "Optimize prompt to improve generation quality when needed.",
+                        "default": True,
+                    },
+                ),
+                "duration": (
+                    IO.COMBO,
+                    {
+                        "tooltip": "The length of the output video in seconds.",
+                        "default": 6,
+                        "options": [6, 10],
+                    },
+                ),
+                "resolution": (
+                    IO.COMBO,
+                    {
+                        "tooltip": "The dimensions of the video display. "
+                                   "1080p corresponds to 1920 x 1080 pixels, 768p corresponds to 1366 x 768 pixels.",
+                        "default": "768P",
+                        "options": ["768P", "1080P"],
+                    },
+                ),
+            },
+            "hidden": {
+                "auth_token": "AUTH_TOKEN_COMFY_ORG",
+                "comfy_api_key": "API_KEY_COMFY_ORG",
+                "unique_id": "UNIQUE_ID",
+            },
+        }
+
+    RETURN_TYPES = ("VIDEO",)
+    DESCRIPTION = cleandoc(__doc__ or "")
+    FUNCTION = "generate_video"
+    CATEGORY = "api node/video/MiniMax"
+    API_NODE = True
+
+    async def generate_video(
+        self,
+        prompt_text,
+        seed=0,
+        first_frame_image: torch.Tensor=None, # used for ImageToVideo
+        prompt_optimizer=True,
+        duration=6,
+        resolution="768P",
+        model="MiniMax-Hailuo-02",
+        unique_id: Union[str, None]=None,
+        **kwargs,
+    ):
+        if first_frame_image is None:
+            validate_string(prompt_text, field_name="prompt_text")
+
+        if model == "MiniMax-Hailuo-02" and resolution.upper() == "1080P" and duration != 6:
+            raise Exception(
+                "When model is MiniMax-Hailuo-02 and resolution is 1080P, duration is limited to 6 seconds."
+            )
+
+        # upload image, if passed in
+        image_url = None
+        if first_frame_image is not None:
+            image_url = (await upload_images_to_comfyapi(first_frame_image, max_images=1, auth_kwargs=kwargs))[0]
+
+        video_generate_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path="/proxy/minimax/video_generation",
+                method=HttpMethod.POST,
+                request_model=MinimaxVideoGenerationRequest,
+                response_model=MinimaxVideoGenerationResponse,
+            ),
+            request=MinimaxVideoGenerationRequest(
+                model=MiniMaxModel(model),
+                prompt=prompt_text,
+                callback_url=None,
+                first_frame_image=image_url,
+                prompt_optimizer=prompt_optimizer,
+                duration=duration,
+                resolution=resolution,
+            ),
+            auth_kwargs=kwargs,
+        )
+        response = await video_generate_operation.execute()
+
+        task_id = response.task_id
+        if not task_id:
+            raise Exception(f"MiniMax generation failed: {response.base_resp}")
+
+        average_duration = 120 if resolution == "768P" else 240
+        video_generate_operation = PollingOperation(
+            poll_endpoint=ApiEndpoint(
+                path="/proxy/minimax/query/video_generation",
+                method=HttpMethod.GET,
+                request_model=EmptyRequest,
+                response_model=MinimaxTaskResultResponse,
+                query_params={"task_id": task_id},
+            ),
+            completed_statuses=["Success"],
+            failed_statuses=["Fail"],
+            status_extractor=lambda x: x.status.value,
+            estimated_duration=average_duration,
+            node_id=unique_id,
+            auth_kwargs=kwargs,
+        )
+        task_result = await video_generate_operation.execute()
+
+        file_id = task_result.file_id
+        if file_id is None:
+            raise Exception("Request was not successful. Missing file ID.")
+        file_retrieve_operation = SynchronousOperation(
+            endpoint=ApiEndpoint(
+                path="/proxy/minimax/files/retrieve",
+                method=HttpMethod.GET,
+                request_model=EmptyRequest,
+                response_model=MinimaxFileRetrieveResponse,
+                query_params={"file_id": int(file_id)},
+            ),
+            request=EmptyRequest(),
+            auth_kwargs=kwargs,
+        )
+        file_result = await file_retrieve_operation.execute()
+
+        file_url = file_result.file.download_url
+        if file_url is None:
+            raise Exception(
+                f"No video was found in the response. Full response: {file_result.model_dump()}"
+            )
+        logging.info(f"Generated video URL: {file_url}")
+        if unique_id:
+            if hasattr(file_result.file, "backup_download_url"):
+                message = f"Result URL: {file_url}\nBackup URL: {file_result.file.backup_download_url}"
+            else:
+                message = f"Result URL: {file_url}"
+            PromptServer.instance.send_progress_text(message, unique_id)
+
+        video_io = await download_url_to_bytesio(file_url)
+        if video_io is None:
+            error_msg = f"Failed to download video from {file_url}"
+            logging.error(error_msg)
+            raise Exception(error_msg)
+        return (VideoFromFile(video_io),)
 
 
 # A dictionary that contains all nodes you want to export with their names
@@ -322,6 +495,7 @@ NODE_CLASS_MAPPINGS = {
     "MinimaxTextToVideoNode": MinimaxTextToVideoNode,
     "MinimaxImageToVideoNode": MinimaxImageToVideoNode,
     # "MinimaxSubjectToVideoNode": MinimaxSubjectToVideoNode,
+    "MinimaxHailuoVideoNode": MinimaxHailuoVideoNode,
 }
 
 # A dictionary that contains the friendly/humanly readable titles for the nodes
@@ -329,4 +503,5 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "MinimaxTextToVideoNode": "MiniMax Text to Video",
     "MinimaxImageToVideoNode": "MiniMax Image to Video",
     "MinimaxSubjectToVideoNode": "MiniMax Subject to Video",
+    "MinimaxHailuoVideoNode": "MiniMax Hailuo Video",
 }

comfy_api_nodes/nodes_openai.py

@@ -80,6 +80,9 @@ class SupportedOpenAIModel(str, Enum):
     gpt_4_1 = "gpt-4.1"
     gpt_4_1_mini = "gpt-4.1-mini"
     gpt_4_1_nano = "gpt-4.1-nano"
+    gpt_5 = "gpt-5"
+    gpt_5_mini = "gpt-5-mini"
+    gpt_5_nano = "gpt-5-nano"
 
 
 class OpenAIDalle2(ComfyNodeABC):
@@ -995,7 +998,7 @@ NODE_DISPLAY_NAME_MAPPINGS = {
     "OpenAIDalle2": "OpenAI DALL·E 2",
     "OpenAIDalle3": "OpenAI DALL·E 3",
     "OpenAIGPTImage1": "OpenAI GPT Image 1",
-    "OpenAIChatNode": "OpenAI Chat",
-    "OpenAIInputFiles": "OpenAI Chat Input Files",
-    "OpenAIChatConfig": "OpenAI Chat Advanced Options",
+    "OpenAIChatNode": "OpenAI ChatGPT",
+    "OpenAIInputFiles": "OpenAI ChatGPT Input Files",
+    "OpenAIChatConfig": "OpenAI ChatGPT Advanced Options",
 }

comfy_api_nodes/nodes_veo2.py

@@ -1,17 +1,18 @@
-import io
 import logging
 import base64
 import aiohttp
 import torch
+from io import BytesIO
 from typing import Optional
+from typing_extensions import override
 
-from comfy.comfy_types.node_typing import IO, ComfyNodeABC
+from comfy_api.latest import ComfyExtension, io as comfy_io
 from comfy_api.input_impl.video_types import VideoFromFile
 from comfy_api_nodes.apis import (
     VeoGenVidRequest,
     VeoGenVidResponse,
     VeoGenVidPollRequest,
-    VeoGenVidPollResponse
+    VeoGenVidPollResponse,
 )
 from comfy_api_nodes.apis.client import (
     ApiEndpoint,
@@ -22,7 +23,7 @@ from comfy_api_nodes.apis.client import (
 
 from comfy_api_nodes.apinode_utils import (
     downscale_image_tensor,
-    tensor_to_base64_string
+    tensor_to_base64_string,
 )
 
 AVERAGE_DURATION_VIDEO_GEN = 32
@@ -50,7 +51,7 @@ def get_video_url_from_response(poll_response: VeoGenVidPollResponse) -> Optiona
     return None
 
 
-class VeoVideoGenerationNode(ComfyNodeABC):
+class VeoVideoGenerationNode(comfy_io.ComfyNode):
     """
     Generates videos from text prompts using Google's Veo API.
 
@@ -59,101 +60,93 @@ class VeoVideoGenerationNode(ComfyNodeABC):
     """
 
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Text description of the video",
-                    },
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="VeoVideoGenerationNode",
+            display_name="Google Veo 2 Video Generation",
+            category="api node/video/Veo",
+            description="Generates videos from text prompts using Google's Veo 2 API",
+            inputs=[
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Text description of the video",
                 ),
-                "aspect_ratio": (
-                    IO.COMBO,
-                    {
-                        "options": ["16:9", "9:16"],
-                        "default": "16:9",
-                        "tooltip": "Aspect ratio of the output video",
-                    },
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=["16:9", "9:16"],
+                    default="16:9",
+                    tooltip="Aspect ratio of the output video",
                 ),
-            },
-            "optional": {
-                "negative_prompt": (
-                    IO.STRING,
-                    {
-                        "multiline": True,
-                        "default": "",
-                        "tooltip": "Negative text prompt to guide what to avoid in the video",
-                    },
+                comfy_io.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Negative text prompt to guide what to avoid in the video",
+                    optional=True,
                 ),
-                "duration_seconds": (
-                    IO.INT,
-                    {
-                        "default": 5,
-                        "min": 5,
-                        "max": 8,
-                        "step": 1,
-                        "display": "number",
-                        "tooltip": "Duration of the output video in seconds",
-                    },
+                comfy_io.Int.Input(
+                    "duration_seconds",
+                    default=5,
+                    min=5,
+                    max=8,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
                 ),
-                "enhance_prompt": (
-                    IO.BOOLEAN,
-                    {
-                        "default": True,
-                        "tooltip": "Whether to enhance the prompt with AI assistance",
-                    }
+                comfy_io.Boolean.Input(
+                    "enhance_prompt",
+                    default=True,
+                    tooltip="Whether to enhance the prompt with AI assistance",
+                    optional=True,
                 ),
-                "person_generation": (
-                    IO.COMBO,
-                    {
-                        "options": ["ALLOW", "BLOCK"],
-                        "default": "ALLOW",
-                        "tooltip": "Whether to allow generating people in the video",
-                    },
+                comfy_io.Combo.Input(
+                    "person_generation",
+                    options=["ALLOW", "BLOCK"],
+                    default="ALLOW",
+                    tooltip="Whether to allow generating people in the video",
+                    optional=True,
                 ),
-                "seed": (
-                    IO.INT,
-                    {
-                        "default": 0,
-                        "min": 0,
-                        "max": 0xFFFFFFFF,
-                        "step": 1,
-                        "display": "number",
-                        "control_after_generate": True,
-                        "tooltip": "Seed for video generation (0 for random)",
-                    },
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=0xFFFFFFFF,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
                 ),
-                "image": (IO.IMAGE, {
-                    "default": None,
-                    "tooltip": "Optional reference image to guide video generation",
-                }),
-                "model": (
-                    IO.COMBO,
-                    {
-                        "options": ["veo-2.0-generate-001"],
-                        "default": "veo-2.0-generate-001",
-                        "tooltip": "Veo 2 model to use for video generation",
-                    },
+                comfy_io.Image.Input(
+                    "image",
+                    tooltip="Optional reference image to guide video generation",
+                    optional=True,
                 ),
-            },
-            "hidden": {
-                "auth_token": "AUTH_TOKEN_COMFY_ORG",
-                "comfy_api_key": "API_KEY_COMFY_ORG",
-                "unique_id": "UNIQUE_ID",
-            },
-        }
+                comfy_io.Combo.Input(
+                    "model",
+                    options=["veo-2.0-generate-001"],
+                    default="veo-2.0-generate-001",
+                    tooltip="Veo 2 model to use for video generation",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
 
-    RETURN_TYPES = (IO.VIDEO,)
-    FUNCTION = "generate_video"
-    CATEGORY = "api node/video/Veo"
-    DESCRIPTION = "Generates videos from text prompts using Google's Veo 2 API"
-    API_NODE = True
-
-    async def generate_video(
-        self,
+    @classmethod
+    async def execute(
+        cls,
         prompt,
         aspect_ratio="16:9",
         negative_prompt="",
@@ -164,8 +157,6 @@ class VeoVideoGenerationNode(ComfyNodeABC):
         image=None,
         model="veo-2.0-generate-001",
         generate_audio=False,
-        unique_id: Optional[str] = None,
-        **kwargs,
     ):
         # Prepare the instances for the request
         instances = []
@@ -202,6 +193,10 @@ class VeoVideoGenerationNode(ComfyNodeABC):
         if "veo-3.0" in model:
             parameters["generateAudio"] = generate_audio
 
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
         # Initial request to start video generation
         initial_operation = SynchronousOperation(
             endpoint=ApiEndpoint(
@@ -214,7 +209,7 @@ class VeoVideoGenerationNode(ComfyNodeABC):
                 instances=instances,
                 parameters=parameters
             ),
-            auth_kwargs=kwargs,
+            auth_kwargs=auth,
         )
 
         initial_response = await initial_operation.execute()
@@ -248,10 +243,10 @@ class VeoVideoGenerationNode(ComfyNodeABC):
             request=VeoGenVidPollRequest(
                 operationName=operation_name
             ),
-            auth_kwargs=kwargs,
+            auth_kwargs=auth,
             poll_interval=5.0,
             result_url_extractor=get_video_url_from_response,
-            node_id=unique_id,
+            node_id=cls.hidden.unique_id,
             estimated_duration=AVERAGE_DURATION_VIDEO_GEN,
         )
 
@@ -304,10 +299,10 @@ class VeoVideoGenerationNode(ComfyNodeABC):
         logging.info("Video generation completed successfully")
 
         # Convert video data to BytesIO object
-        video_io = io.BytesIO(video_data)
+        video_io = BytesIO(video_data)
 
         # Return VideoFromFile object
-        return (VideoFromFile(video_io),)
+        return comfy_io.NodeOutput(VideoFromFile(video_io))
 
 
 class Veo3VideoGenerationNode(VeoVideoGenerationNode):
@@ -323,51 +318,104 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode):
     """
 
     @classmethod
-    def INPUT_TYPES(s):
-        parent_input = super().INPUT_TYPES()
-
-        # Update model options for Veo 3
-        parent_input["optional"]["model"] = (
-            IO.COMBO,
-            {
-                "options": ["veo-3.0-generate-001", "veo-3.0-fast-generate-001"],
-                "default": "veo-3.0-generate-001",
-                "tooltip": "Veo 3 model to use for video generation",
-            },
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="Veo3VideoGenerationNode",
+            display_name="Google Veo 3 Video Generation",
+            category="api node/video/Veo",
+            description="Generates videos from text prompts using Google's Veo 3 API",
+            inputs=[
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Text description of the video",
+                ),
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=["16:9", "9:16"],
+                    default="16:9",
+                    tooltip="Aspect ratio of the output video",
+                ),
+                comfy_io.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Negative text prompt to guide what to avoid in the video",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "duration_seconds",
+                    default=8,
+                    min=8,
+                    max=8,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds (Veo 3 only supports 8 seconds)",
+                    optional=True,
+                ),
+                comfy_io.Boolean.Input(
+                    "enhance_prompt",
+                    default=True,
+                    tooltip="Whether to enhance the prompt with AI assistance",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "person_generation",
+                    options=["ALLOW", "BLOCK"],
+                    default="ALLOW",
+                    tooltip="Whether to allow generating people in the video",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=0xFFFFFFFF,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Image.Input(
+                    "image",
+                    tooltip="Optional reference image to guide video generation",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "model",
+                    options=["veo-3.0-generate-001", "veo-3.0-fast-generate-001"],
+                    default="veo-3.0-generate-001",
+                    tooltip="Veo 3 model to use for video generation",
+                    optional=True,
+                ),
+                comfy_io.Boolean.Input(
+                    "generate_audio",
+                    default=False,
+                    tooltip="Generate audio for the video. Supported by all Veo 3 models.",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
         )
 
-        # Add generateAudio parameter
-        parent_input["optional"]["generate_audio"] = (
-            IO.BOOLEAN,
-            {
-                "default": False,
-                "tooltip": "Generate audio for the video. Supported by all Veo 3 models.",
-            }
-        )
 
-        # Update duration constraints for Veo 3 (only 8 seconds supported)
-        parent_input["optional"]["duration_seconds"] = (
-            IO.INT,
-            {
-                "default": 8,
-                "min": 8,
-                "max": 8,
-                "step": 1,
-                "display": "number",
-                "tooltip": "Duration of the output video in seconds (Veo 3 only supports 8 seconds)",
-            },
-        )
+class VeoExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[comfy_io.ComfyNode]]:
+        return [
+            VeoVideoGenerationNode,
+            Veo3VideoGenerationNode,
+        ]
 
-        return parent_input
-
-
-# Register the nodes
-NODE_CLASS_MAPPINGS = {
-    "VeoVideoGenerationNode": VeoVideoGenerationNode,
-    "Veo3VideoGenerationNode": Veo3VideoGenerationNode,
-}
-
-NODE_DISPLAY_NAME_MAPPINGS = {
-    "VeoVideoGenerationNode": "Google Veo 2 Video Generation",
-    "Veo3VideoGenerationNode": "Google Veo 3 Video Generation",
-}
+async def comfy_entrypoint() -> VeoExtension:
+    return VeoExtension()

comfy_api_nodes/nodes_vidu.py

@@ -0,0 +1,622 @@
+import logging
+from enum import Enum
+from typing import Any, Callable, Optional, Literal, TypeVar
+from typing_extensions import override
+
+import torch
+from pydantic import BaseModel, Field
+
+from comfy_api.latest import ComfyExtension, io as comfy_io
+from comfy_api_nodes.util.validation_utils import (
+    validate_aspect_ratio_closeness,
+    validate_image_dimensions,
+    validate_image_aspect_ratio_range,
+    get_number_of_images,
+)
+from comfy_api_nodes.apis.client import (
+    ApiEndpoint,
+    HttpMethod,
+    SynchronousOperation,
+    PollingOperation,
+    EmptyRequest,
+)
+from comfy_api_nodes.apinode_utils import download_url_to_video_output, upload_images_to_comfyapi
+
+
+VIDU_TEXT_TO_VIDEO = "/proxy/vidu/text2video"
+VIDU_IMAGE_TO_VIDEO = "/proxy/vidu/img2video"
+VIDU_REFERENCE_VIDEO = "/proxy/vidu/reference2video"
+VIDU_START_END_VIDEO = "/proxy/vidu/start-end2video"
+VIDU_GET_GENERATION_STATUS = "/proxy/vidu/tasks/%s/creations"
+
+R = TypeVar("R")
+
+class VideoModelName(str, Enum):
+    vidu_q1 = 'viduq1'
+
+
+class AspectRatio(str, Enum):
+    r_16_9 = "16:9"
+    r_9_16 = "9:16"
+    r_1_1 = "1:1"
+
+
+class Resolution(str, Enum):
+    r_1080p = "1080p"
+
+
+class MovementAmplitude(str, Enum):
+    auto = "auto"
+    small = "small"
+    medium = "medium"
+    large = "large"
+
+
+class TaskCreationRequest(BaseModel):
+    model: VideoModelName = VideoModelName.vidu_q1
+    prompt: Optional[str] = Field(None, max_length=1500)
+    duration: Optional[Literal[5]] = 5
+    seed: Optional[int] = Field(0, ge=0, le=2147483647)
+    aspect_ratio: Optional[AspectRatio] = AspectRatio.r_16_9
+    resolution: Optional[Resolution] = Resolution.r_1080p
+    movement_amplitude: Optional[MovementAmplitude] = MovementAmplitude.auto
+    images: Optional[list[str]] = Field(None, description="Base64 encoded string or image URL")
+
+
+class TaskStatus(str, Enum):
+    created = "created"
+    queueing = "queueing"
+    processing = "processing"
+    success = "success"
+    failed = "failed"
+
+
+class TaskCreationResponse(BaseModel):
+    task_id: str = Field(...)
+    state: TaskStatus = Field(...)
+    created_at: str = Field(...)
+    code: Optional[int] = Field(None, description="Error code")
+
+
+class TaskResult(BaseModel):
+    id: str = Field(..., description="Creation id")
+    url: str = Field(..., description="The URL of the generated results, valid for one hour")
+    cover_url: str = Field(..., description="The cover URL of the generated results, valid for one hour")
+
+
+class TaskStatusResponse(BaseModel):
+    state: TaskStatus = Field(...)
+    err_code: Optional[str] = Field(None)
+    creations: list[TaskResult] = Field(..., description="Generated results")
+
+
+async def poll_until_finished(
+    auth_kwargs: dict[str, str],
+    api_endpoint: ApiEndpoint[Any, R],
+    result_url_extractor: Optional[Callable[[R], str]] = None,
+    estimated_duration: Optional[int] = None,
+    node_id: Optional[str] = None,
+) -> R:
+    return await PollingOperation(
+        poll_endpoint=api_endpoint,
+        completed_statuses=[TaskStatus.success.value],
+        failed_statuses=[TaskStatus.failed.value],
+        status_extractor=lambda response: response.state.value,
+        auth_kwargs=auth_kwargs,
+        result_url_extractor=result_url_extractor,
+        estimated_duration=estimated_duration,
+        node_id=node_id,
+        poll_interval=16.0,
+        max_poll_attempts=256,
+    ).execute()
+
+
+def get_video_url_from_response(response) -> Optional[str]:
+    if response.creations:
+        return response.creations[0].url
+    return None
+
+
+def get_video_from_response(response) -> TaskResult:
+    if not response.creations:
+        error_msg = f"Vidu request does not contain results. State: {response.state}, Error Code: {response.err_code}"
+        logging.info(error_msg)
+        raise RuntimeError(error_msg)
+    logging.info("Vidu task %s succeeded. Video URL: %s", response.creations[0].id, response.creations[0].url)
+    return response.creations[0]
+
+
+async def execute_task(
+    vidu_endpoint: str,
+    auth_kwargs: Optional[dict[str, str]],
+    payload: TaskCreationRequest,
+    estimated_duration: int,
+    node_id: str,
+) -> R:
+    response = await SynchronousOperation(
+        endpoint=ApiEndpoint(
+            path=vidu_endpoint,
+            method=HttpMethod.POST,
+            request_model=TaskCreationRequest,
+            response_model=TaskCreationResponse,
+        ),
+        request=payload,
+        auth_kwargs=auth_kwargs,
+    ).execute()
+    if response.state == TaskStatus.failed:
+        error_msg = f"Vidu request failed. Code: {response.code}"
+        logging.error(error_msg)
+        raise RuntimeError(error_msg)
+    return await poll_until_finished(
+        auth_kwargs,
+        ApiEndpoint(
+            path=VIDU_GET_GENERATION_STATUS % response.task_id,
+            method=HttpMethod.GET,
+            request_model=EmptyRequest,
+            response_model=TaskStatusResponse,
+        ),
+        result_url_extractor=get_video_url_from_response,
+        estimated_duration=estimated_duration,
+        node_id=node_id,
+    )
+
+
+class ViduTextToVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduTextToVideoNode",
+            display_name="Vidu Text To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate video from text prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="A textual description for video generation",
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=[model.value for model in AspectRatio],
+                    default=AspectRatio.r_16_9.value,
+                    tooltip="The aspect ratio of the output video",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        prompt: str,
+        duration: int,
+        seed: int,
+        aspect_ratio: str,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        if not prompt:
+            raise ValueError("The prompt field is required and cannot be empty.")
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            aspect_ratio=aspect_ratio,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        results = await execute_task(VIDU_TEXT_TO_VIDEO, auth, payload, 320, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduImageToVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduImageToVideoNode",
+            display_name="Vidu Image To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate video from image and optional prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.Image.Input(
+                    "image",
+                    tooltip="An image to be used as the start frame of the generated video",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="A textual description for video generation",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        image: torch.Tensor,
+        prompt: str,
+        duration: int,
+        seed: int,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        if get_number_of_images(image) > 1:
+            raise ValueError("Only one input image is allowed.")
+        validate_image_aspect_ratio_range(image, (1, 4), (4, 1))
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        payload.images = await upload_images_to_comfyapi(
+            image,
+            max_images=1,
+            mime_type="image/png",
+            auth_kwargs=auth,
+        )
+        results = await execute_task(VIDU_IMAGE_TO_VIDEO, auth, payload, 120, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduReferenceVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduReferenceVideoNode",
+            display_name="Vidu Reference To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate video from multiple images and prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.Image.Input(
+                    "images",
+                    tooltip="Images to use as references to generate a video with consistent subjects (max 7 images).",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="A textual description for video generation",
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "aspect_ratio",
+                    options=[model.value for model in AspectRatio],
+                    default=AspectRatio.r_16_9.value,
+                    tooltip="The aspect ratio of the output video",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        images: torch.Tensor,
+        prompt: str,
+        duration: int,
+        seed: int,
+        aspect_ratio: str,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        if not prompt:
+            raise ValueError("The prompt field is required and cannot be empty.")
+        a = get_number_of_images(images)
+        if a > 7:
+            raise ValueError("Too many images, maximum allowed is 7.")
+        for image in images:
+            validate_image_aspect_ratio_range(image, (1, 4), (4, 1))
+            validate_image_dimensions(image, min_width=128, min_height=128)
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            aspect_ratio=aspect_ratio,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        payload.images = await upload_images_to_comfyapi(
+            images,
+            max_images=7,
+            mime_type="image/png",
+            auth_kwargs=auth,
+        )
+        results = await execute_task(VIDU_REFERENCE_VIDEO, auth, payload, 120, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduStartEndToVideoNode(comfy_io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return comfy_io.Schema(
+            node_id="ViduStartEndToVideoNode",
+            display_name="Vidu Start End To Video Generation",
+            category="api node/video/Vidu",
+            description="Generate a video from start and end frames and a prompt",
+            inputs=[
+                comfy_io.Combo.Input(
+                    "model",
+                    options=[model.value for model in VideoModelName],
+                    default=VideoModelName.vidu_q1.value,
+                    tooltip="Model name",
+                ),
+                comfy_io.Image.Input(
+                    "first_frame",
+                    tooltip="Start frame",
+                ),
+                comfy_io.Image.Input(
+                    "end_frame",
+                    tooltip="End frame",
+                ),
+                comfy_io.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="A textual description for video generation",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "duration",
+                    default=5,
+                    min=5,
+                    max=5,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    tooltip="Duration of the output video in seconds",
+                    optional=True,
+                ),
+                comfy_io.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=comfy_io.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed for video generation (0 for random)",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "resolution",
+                    options=[model.value for model in Resolution],
+                    default=Resolution.r_1080p.value,
+                    tooltip="Supported values may vary by model & duration",
+                    optional=True,
+                ),
+                comfy_io.Combo.Input(
+                    "movement_amplitude",
+                    options=[model.value for model in MovementAmplitude],
+                    default=MovementAmplitude.auto.value,
+                    tooltip="The movement amplitude of objects in the frame",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                comfy_io.Video.Output(),
+            ],
+            hidden=[
+                comfy_io.Hidden.auth_token_comfy_org,
+                comfy_io.Hidden.api_key_comfy_org,
+                comfy_io.Hidden.unique_id,
+            ],
+            is_api_node=True,
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        first_frame: torch.Tensor,
+        end_frame: torch.Tensor,
+        prompt: str,
+        duration: int,
+        seed: int,
+        resolution: str,
+        movement_amplitude: str,
+    ) -> comfy_io.NodeOutput:
+        validate_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False)
+        payload = TaskCreationRequest(
+            model_name=model,
+            prompt=prompt,
+            duration=duration,
+            seed=seed,
+            resolution=resolution,
+            movement_amplitude=movement_amplitude,
+        )
+        auth = {
+            "auth_token": cls.hidden.auth_token_comfy_org,
+            "comfy_api_key": cls.hidden.api_key_comfy_org,
+        }
+        payload.images = [
+            (await upload_images_to_comfyapi(frame, max_images=1, mime_type="image/png", auth_kwargs=auth))[0]
+            for frame in (first_frame, end_frame)
+        ]
+        results = await execute_task(VIDU_START_END_VIDEO, auth, payload, 96, cls.hidden.unique_id)
+        return comfy_io.NodeOutput(await download_url_to_video_output(get_video_from_response(results).url))
+
+
+class ViduExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[comfy_io.ComfyNode]]:
+        return [
+            ViduTextToVideoNode,
+            ViduImageToVideoNode,
+            ViduReferenceVideoNode,
+            ViduStartEndToVideoNode,
+        ]
+
+async def comfy_entrypoint() -> ViduExtension:
+    return ViduExtension()

comfy_api_nodes/util/validation_utils.py

@@ -53,6 +53,53 @@ def validate_image_aspect_ratio(
         )
 
 
+def validate_image_aspect_ratio_range(
+    image: torch.Tensor,
+    min_ratio: tuple[float, float],  # e.g. (1, 4)
+    max_ratio: tuple[float, float],  # e.g. (4, 1)
+    *,
+    strict: bool = True,             # True -> (min, max); False -> [min, max]
+) -> float:
+    a1, b1 = min_ratio
+    a2, b2 = max_ratio
+    if a1 <= 0 or b1 <= 0 or a2 <= 0 or b2 <= 0:
+        raise ValueError("Ratios must be positive, like (1, 4) or (4, 1).")
+    lo, hi = (a1 / b1), (a2 / b2)
+    if lo > hi:
+        lo, hi = hi, lo
+        a1, b1, a2, b2 = a2, b2, a1, b1  # swap only for error text
+    w, h = get_image_dimensions(image)
+    if w <= 0 or h <= 0:
+        raise ValueError(f"Invalid image dimensions: {w}x{h}")
+    ar = w / h
+    ok = (lo < ar < hi) if strict else (lo <= ar <= hi)
+    if not ok:
+        op = "<" if strict else "≤"
+        raise ValueError(f"Image aspect ratio {ar:.6g} is outside allowed range: {a1}:{b1} {op} ratio {op} {a2}:{b2}")
+    return ar
+
+
+def validate_aspect_ratio_closeness(
+    start_img,
+    end_img,
+    min_rel: float,
+    max_rel: float,
+    *,
+    strict: bool = False,   # True => exclusive, False => inclusive
+) -> None:
+    w1, h1 = get_image_dimensions(start_img)
+    w2, h2 = get_image_dimensions(end_img)
+    if min(w1, h1, w2, h2) <= 0:
+        raise ValueError("Invalid image dimensions")
+    ar1 = w1 / h1
+    ar2 = w2 / h2
+    # Normalize so it is symmetric (no need to check both ar1/ar2 and ar2/ar1)
+    closeness = max(ar1, ar2) / min(ar1, ar2)
+    limit = max(max_rel, 1.0 / min_rel)  # for 0.8..1.25 this is 1.25
+    if (closeness >= limit) if strict else (closeness > limit):
+        raise ValueError(f"Aspect ratios must be close: start/end={ar1/ar2:.4f}, allowed range {min_rel}–{max_rel}.")
+
+
 def validate_video_dimensions(
     video: VideoInput,
     min_width: Optional[int] = None,
@@ -98,3 +145,9 @@ def validate_video_duration(
         raise ValueError(
             f"Video duration must be at most {max_duration}s, got {duration}s"
         )
+
+
+def get_number_of_images(images):
+    if isinstance(images, torch.Tensor):
+        return images.shape[0] if images.ndim >= 4 else 1
+    return len(images)

comfy_extras/nodes_ace.py

@@ -1,49 +1,63 @@
 import torch
+from typing_extensions import override
+
 import comfy.model_management
 import node_helpers
+from comfy_api.latest import ComfyExtension, io
 
-class TextEncodeAceStepAudio:
+
+class TextEncodeAceStepAudio(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {
-            "clip": ("CLIP", ),
-            "tags": ("STRING", {"multiline": True, "dynamicPrompts": True}),
-            "lyrics": ("STRING", {"multiline": True, "dynamicPrompts": True}),
-            "lyrics_strength": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-            }}
-    RETURN_TYPES = ("CONDITIONING",)
-    FUNCTION = "encode"
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncodeAceStepAudio",
+            category="conditioning",
+            inputs=[
+                io.Clip.Input("clip"),
+                io.String.Input("tags", multiline=True, dynamic_prompts=True),
+                io.String.Input("lyrics", multiline=True, dynamic_prompts=True),
+                io.Float.Input("lyrics_strength", default=1.0, min=0.0, max=10.0, step=0.01),
+            ],
+            outputs=[io.Conditioning.Output()],
+        )
 
-    CATEGORY = "conditioning"
-
-    def encode(self, clip, tags, lyrics, lyrics_strength):
+    @classmethod
+    def execute(cls, clip, tags, lyrics, lyrics_strength) -> io.NodeOutput:
         tokens = clip.tokenize(tags, lyrics=lyrics)
         conditioning = clip.encode_from_tokens_scheduled(tokens)
         conditioning = node_helpers.conditioning_set_values(conditioning, {"lyrics_strength": lyrics_strength})
-        return (conditioning, )
+        return io.NodeOutput(conditioning)
 
 
-class EmptyAceStepLatentAudio:
-    def __init__(self):
-        self.device = comfy.model_management.intermediate_device()
+class EmptyAceStepLatentAudio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="EmptyAceStepLatentAudio",
+            category="latent/audio",
+            inputs=[
+                io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.1),
+                io.Int.Input(
+                    "batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch."
+                ),
+            ],
+            outputs=[io.Latent.Output()],
+        )
 
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": {"seconds": ("FLOAT", {"default": 120.0, "min": 1.0, "max": 1000.0, "step": 0.1}),
-                             "batch_size": ("INT", {"default": 1, "min": 1, "max": 4096, "tooltip": "The number of latent images in the batch."}),
-                             }}
-    RETURN_TYPES = ("LATENT",)
-    FUNCTION = "generate"
-
-    CATEGORY = "latent/audio"
-
-    def generate(self, seconds, batch_size):
+    def execute(cls, seconds, batch_size) -> io.NodeOutput:
         length = int(seconds * 44100 / 512 / 8)
-        latent = torch.zeros([batch_size, 8, 16, length], device=self.device)
-        return ({"samples": latent, "type": "audio"}, )
+        latent = torch.zeros([batch_size, 8, 16, length], device=comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": latent, "type": "audio"})
 
 
-NODE_CLASS_MAPPINGS = {
-    "TextEncodeAceStepAudio": TextEncodeAceStepAudio,
-    "EmptyAceStepLatentAudio": EmptyAceStepLatentAudio,
-}
+class AceExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            TextEncodeAceStepAudio,
+            EmptyAceStepLatentAudio,
+        ]
+
+async def comfy_entrypoint() -> AceExtension:
+    return AceExtension()

comfy_extras/nodes_advanced_samplers.py

@@ -1,8 +1,13 @@
+import numpy as np
+import torch
+from tqdm.auto import trange
+from typing_extensions import override
+
+import comfy.model_patcher
 import comfy.samplers
 import comfy.utils
-import torch
-import numpy as np
-from tqdm.auto import trange
+from comfy.k_diffusion.sampling import to_d
+from comfy_api.latest import ComfyExtension, io
 
 
 @torch.no_grad()
@@ -33,30 +38,29 @@ def sample_lcm_upscale(model, x, sigmas, extra_args=None, callback=None, disable
     return x
 
 
-class SamplerLCMUpscale:
-    upscale_methods = ["bislerp", "nearest-exact", "bilinear", "area", "bicubic"]
+class SamplerLCMUpscale(io.ComfyNode):
+    UPSCALE_METHODS = ["bislerp", "nearest-exact", "bilinear", "area", "bicubic"]
 
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required":
-                    {"scale_ratio": ("FLOAT", {"default": 1.0, "min": 0.1, "max": 20.0, "step": 0.01}),
-                     "scale_steps": ("INT", {"default": -1, "min": -1, "max": 1000, "step": 1}),
-                     "upscale_method": (s.upscale_methods,),
-                      }
-               }
-    RETURN_TYPES = ("SAMPLER",)
-    CATEGORY = "sampling/custom_sampling/samplers"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="SamplerLCMUpscale",
+            category="sampling/custom_sampling/samplers",
+            inputs=[
+                io.Float.Input("scale_ratio", default=1.0, min=0.1, max=20.0, step=0.01),
+                io.Int.Input("scale_steps", default=-1, min=-1, max=1000, step=1),
+                io.Combo.Input("upscale_method", options=cls.UPSCALE_METHODS),
+            ],
+            outputs=[io.Sampler.Output()],
+        )
 
-    FUNCTION = "get_sampler"
-
-    def get_sampler(self, scale_ratio, scale_steps, upscale_method):
+    @classmethod
+    def execute(cls, scale_ratio, scale_steps, upscale_method) -> io.NodeOutput:
         if scale_steps < 0:
             scale_steps = None
         sampler = comfy.samplers.KSAMPLER(sample_lcm_upscale, extra_options={"total_upscale": scale_ratio, "upscale_steps": scale_steps, "upscale_method": upscale_method})
-        return (sampler, )
+        return io.NodeOutput(sampler)
 
-from comfy.k_diffusion.sampling import to_d
-import comfy.model_patcher
 
 @torch.no_grad()
 def sample_euler_pp(model, x, sigmas, extra_args=None, callback=None, disable=None):
@@ -82,30 +86,36 @@ def sample_euler_pp(model, x, sigmas, extra_args=None, callback=None, disable=No
     return x
 
 
-class SamplerEulerCFGpp:
+class SamplerEulerCFGpp(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required":
-                    {"version": (["regular", "alternative"],),}
-               }
-    RETURN_TYPES = ("SAMPLER",)
-    # CATEGORY = "sampling/custom_sampling/samplers"
-    CATEGORY = "_for_testing"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="SamplerEulerCFGpp",
+            display_name="SamplerEulerCFG++",
+            category="_for_testing",  # "sampling/custom_sampling/samplers"
+            inputs=[
+                io.Combo.Input("version", options=["regular", "alternative"]),
+            ],
+            outputs=[io.Sampler.Output()],
+            is_experimental=True,
+        )
 
-    FUNCTION = "get_sampler"
-
-    def get_sampler(self, version):
+    @classmethod
+    def execute(cls, version) -> io.NodeOutput:
         if version == "alternative":
             sampler = comfy.samplers.KSAMPLER(sample_euler_pp)
         else:
             sampler = comfy.samplers.ksampler("euler_cfg_pp")
-        return (sampler, )
+        return io.NodeOutput(sampler)
 
-NODE_CLASS_MAPPINGS = {
-    "SamplerLCMUpscale": SamplerLCMUpscale,
-    "SamplerEulerCFGpp": SamplerEulerCFGpp,
-}
 
-NODE_DISPLAY_NAME_MAPPINGS = {
-    "SamplerEulerCFGpp": "SamplerEulerCFG++",
-}
+class AdvancedSamplersExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            SamplerLCMUpscale,
+            SamplerEulerCFGpp,
+        ]
+
+async def comfy_entrypoint() -> AdvancedSamplersExtension:
+    return AdvancedSamplersExtension()

comfy_extras/nodes_apg.py

@@ -1,4 +1,8 @@
 import torch
+from typing_extensions import override
+
+from comfy_api.latest import ComfyExtension, io
+
 
 def project(v0, v1):
     v1 = torch.nn.functional.normalize(v1, dim=[-1, -2, -3])
@@ -6,22 +10,45 @@ def project(v0, v1):
     v0_orthogonal = v0 - v0_parallel
     return v0_parallel, v0_orthogonal
 
-class APG:
+class APG(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "model": ("MODEL",),
-                "eta": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01, "tooltip": "Controls the scale of the parallel guidance vector. Default CFG behavior at a setting of 1."}),
-                "norm_threshold": ("FLOAT", {"default": 5.0, "min": 0.0, "max": 50.0, "step": 0.1, "tooltip": "Normalize guidance vector to this value, normalization disable at a setting of 0."}),
-                "momentum": ("FLOAT", {"default": 0.0, "min": -5.0, "max": 1.0, "step": 0.01, "tooltip":"Controls a running average of guidance during diffusion, disabled at a setting of 0."}),
-            }
-        }
-    RETURN_TYPES = ("MODEL",)
-    FUNCTION = "patch"
-    CATEGORY = "sampling/custom_sampling"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="APG",
+            display_name="Adaptive Projected Guidance",
+            category="sampling/custom_sampling",
+            inputs=[
+                io.Model.Input("model"),
+                io.Float.Input(
+                    "eta",
+                    default=1.0,
+                    min=-10.0,
+                    max=10.0,
+                    step=0.01,
+                    tooltip="Controls the scale of the parallel guidance vector. Default CFG behavior at a setting of 1.",
+                ),
+                io.Float.Input(
+                    "norm_threshold",
+                    default=5.0,
+                    min=0.0,
+                    max=50.0,
+                    step=0.1,
+                    tooltip="Normalize guidance vector to this value, normalization disable at a setting of 0.",
+                ),
+                io.Float.Input(
+                    "momentum",
+                    default=0.0,
+                    min=-5.0,
+                    max=1.0,
+                    step=0.01,
+                    tooltip="Controls a running average of guidance during diffusion, disabled at a setting of 0.",
+                ),
+            ],
+            outputs=[io.Model.Output()],
+        )
 
-    def patch(self, model, eta, norm_threshold, momentum):
+    @classmethod
+    def execute(cls, model, eta, norm_threshold, momentum) -> io.NodeOutput:
         running_avg = 0
         prev_sigma = None
 
@@ -65,12 +92,15 @@ class APG:
 
         m = model.clone()
         m.set_model_sampler_pre_cfg_function(pre_cfg_function)
-        return (m,)
+        return io.NodeOutput(m)
 
-NODE_CLASS_MAPPINGS = {
-    "APG": APG,
-}
 
-NODE_DISPLAY_NAME_MAPPINGS = {
-    "APG": "Adaptive Projected Guidance",
-}
+class ApgExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            APG,
+        ]
+
+async def comfy_entrypoint() -> ApgExtension:
+    return ApgExtension()

comfy_extras/nodes_attention_multiply.py

@@ -1,3 +1,7 @@
+from typing_extensions import override
+
+from comfy_api.latest import ComfyExtension, io
+
 
 def attention_multiply(attn, model, q, k, v, out):
     m = model.clone()
@@ -16,57 +20,71 @@ def attention_multiply(attn, model, q, k, v, out):
     return m
 
 
-class UNetSelfAttentionMultiply:
+class UNetSelfAttentionMultiply(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "model": ("MODEL",),
-                              "q": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "k": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "v": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "out": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              }}
-    RETURN_TYPES = ("MODEL",)
-    FUNCTION = "patch"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="UNetSelfAttentionMultiply",
+            category="_for_testing/attention_experiments",
+            inputs=[
+                io.Model.Input("model"),
+                io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("k", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("v", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("out", default=1.0, min=0.0, max=10.0, step=0.01),
+            ],
+            outputs=[io.Model.Output()],
+            is_experimental=True,
+        )
 
-    CATEGORY = "_for_testing/attention_experiments"
-
-    def patch(self, model, q, k, v, out):
+    @classmethod
+    def execute(cls, model, q, k, v, out) -> io.NodeOutput:
         m = attention_multiply("attn1", model, q, k, v, out)
-        return (m, )
+        return io.NodeOutput(m)
 
-class UNetCrossAttentionMultiply:
+
+class UNetCrossAttentionMultiply(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "model": ("MODEL",),
-                              "q": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "k": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "v": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "out": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              }}
-    RETURN_TYPES = ("MODEL",)
-    FUNCTION = "patch"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="UNetCrossAttentionMultiply",
+            category="_for_testing/attention_experiments",
+            inputs=[
+                io.Model.Input("model"),
+                io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("k", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("v", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("out", default=1.0, min=0.0, max=10.0, step=0.01),
+            ],
+            outputs=[io.Model.Output()],
+            is_experimental=True,
+        )
 
-    CATEGORY = "_for_testing/attention_experiments"
-
-    def patch(self, model, q, k, v, out):
+    @classmethod
+    def execute(cls, model, q, k, v, out) -> io.NodeOutput:
         m = attention_multiply("attn2", model, q, k, v, out)
-        return (m, )
+        return io.NodeOutput(m)
 
-class CLIPAttentionMultiply:
+
+class CLIPAttentionMultiply(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "clip": ("CLIP",),
-                              "q": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "k": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "v": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "out": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              }}
-    RETURN_TYPES = ("CLIP",)
-    FUNCTION = "patch"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="CLIPAttentionMultiply",
+            category="_for_testing/attention_experiments",
+            inputs=[
+                io.Clip.Input("clip"),
+                io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("k", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("v", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("out", default=1.0, min=0.0, max=10.0, step=0.01),
+            ],
+            outputs=[io.Clip.Output()],
+            is_experimental=True,
+        )
 
-    CATEGORY = "_for_testing/attention_experiments"
-
-    def patch(self, clip, q, k, v, out):
+    @classmethod
+    def execute(cls, clip, q, k, v, out) -> io.NodeOutput:
         m = clip.clone()
         sd = m.patcher.model_state_dict()
 
@@ -79,23 +97,28 @@ class CLIPAttentionMultiply:
                 m.add_patches({key: (None,)}, 0.0, v)
             if key.endswith("self_attn.out_proj.weight") or key.endswith("self_attn.out_proj.bias"):
                 m.add_patches({key: (None,)}, 0.0, out)
-        return (m, )
+        return io.NodeOutput(m)
 
-class UNetTemporalAttentionMultiply:
+
+class UNetTemporalAttentionMultiply(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {"required": { "model": ("MODEL",),
-                              "self_structural": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "self_temporal": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "cross_structural": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              "cross_temporal": ("FLOAT", {"default": 1.0, "min": 0.0, "max": 10.0, "step": 0.01}),
-                              }}
-    RETURN_TYPES = ("MODEL",)
-    FUNCTION = "patch"
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="UNetTemporalAttentionMultiply",
+            category="_for_testing/attention_experiments",
+            inputs=[
+                io.Model.Input("model"),
+                io.Float.Input("self_structural", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("self_temporal", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("cross_structural", default=1.0, min=0.0, max=10.0, step=0.01),
+                io.Float.Input("cross_temporal", default=1.0, min=0.0, max=10.0, step=0.01),
+            ],
+            outputs=[io.Model.Output()],
+            is_experimental=True,
+        )
 
-    CATEGORY = "_for_testing/attention_experiments"
-
-    def patch(self, model, self_structural, self_temporal, cross_structural, cross_temporal):
+    @classmethod
+    def execute(cls, model, self_structural, self_temporal, cross_structural, cross_temporal) -> io.NodeOutput:
         m = model.clone()
         sd = model.model_state_dict()
 
@@ -110,11 +133,18 @@ class UNetTemporalAttentionMultiply:
                     m.add_patches({k: (None,)}, 0.0, cross_temporal)
                 else:
                     m.add_patches({k: (None,)}, 0.0, cross_structural)
-        return (m, )
+        return io.NodeOutput(m)
 
-NODE_CLASS_MAPPINGS = {
-    "UNetSelfAttentionMultiply": UNetSelfAttentionMultiply,
-    "UNetCrossAttentionMultiply": UNetCrossAttentionMultiply,
-    "CLIPAttentionMultiply": CLIPAttentionMultiply,
-    "UNetTemporalAttentionMultiply": UNetTemporalAttentionMultiply,
-}
+
+class AttentionMultiplyExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            UNetSelfAttentionMultiply,
+            UNetCrossAttentionMultiply,
+            CLIPAttentionMultiply,
+            UNetTemporalAttentionMultiply,
+        ]
+
+async def comfy_entrypoint() -> AttentionMultiplyExtension:
+    return AttentionMultiplyExtension()

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

@@ -0,0 +1,493 @@
+from __future__ import annotations
+from typing import TYPE_CHECKING, Union
+from comfy_api.latest import io, ComfyExtension
+import comfy.patcher_extension
+import logging
+import torch
+import comfy.model_patcher
+if TYPE_CHECKING:
+    from uuid import UUID
+
+
+def easycache_forward_wrapper(executor, *args, **kwargs):
+    # get values from args
+    x: torch.Tensor = args[0]
+    transformer_options: dict[str] = args[-1]
+    if not isinstance(transformer_options, dict):
+        transformer_options = kwargs.get("transformer_options")
+        if not transformer_options:
+            transformer_options = args[-2]
+    easycache: EasyCacheHolder = transformer_options["easycache"]
+    sigmas = transformer_options["sigmas"]
+    uuids = transformer_options["uuids"]
+    if sigmas is not None and easycache.is_past_end_timestep(sigmas):
+        return executor(*args, **kwargs)
+    # prepare next x_prev
+    has_first_cond_uuid = easycache.has_first_cond_uuid(uuids)
+    next_x_prev = x
+    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:
+                logging.info(f"EasyCache [verbose] - was marked to skip this step by {easycache.first_cond_uuid}. Present uuids: {uuids}")
+            return easycache.apply_cache_diff(x, uuids)
+        if easycache.initial_step:
+            easycache.first_cond_uuid = uuids[0]
+            has_first_cond_uuid = easycache.has_first_cond_uuid(uuids)
+            easycache.initial_step = False
+        if has_first_cond_uuid:
+            if easycache.has_x_prev_subsampled():
+                input_change = (easycache.subsample(x, uuids, clone=False) - easycache.x_prev_subsampled).flatten().abs().mean()
+            if easycache.has_output_prev_norm() and easycache.has_relative_transformation_rate():
+                approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm
+                easycache.cumulative_change_rate += approx_output_change_rate
+                if easycache.cumulative_change_rate < easycache.reuse_threshold:
+                    if easycache.verbose:
+                        logging.info(f"EasyCache [verbose] - skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
+                    # other conds should also skip this step, and instead use their cached values
+                    easycache.skip_current_step = True
+                    return easycache.apply_cache_diff(x, uuids)
+                else:
+                    if easycache.verbose:
+                        logging.info(f"EasyCache [verbose] - NOT skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
+                    easycache.cumulative_change_rate = 0.0
+
+    output: torch.Tensor = executor(*args, **kwargs)
+    if has_first_cond_uuid and easycache.has_output_prev_norm():
+        output_change = (easycache.subsample(output, uuids, clone=False) - easycache.output_prev_subsampled).flatten().abs().mean()
+        if easycache.verbose:
+            output_change_rate = output_change / easycache.output_prev_norm
+            easycache.output_change_rates.append(output_change_rate.item())
+        if easycache.has_relative_transformation_rate():
+            approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm
+            easycache.approx_output_change_rates.append(approx_output_change_rate.item())
+            if easycache.verbose:
+                logging.info(f"EasyCache [verbose] - approx_output_change_rate: {approx_output_change_rate}")
+        if input_change is not None:
+            easycache.relative_transformation_rate = output_change / input_change
+        if easycache.verbose:
+            logging.info(f"EasyCache [verbose] - output_change_rate: {output_change_rate}")
+    # TODO: allow cache_diff to be offloaded
+    easycache.update_cache_diff(output, next_x_prev, uuids)
+    if has_first_cond_uuid:
+        easycache.x_prev_subsampled = easycache.subsample(next_x_prev, uuids)
+        easycache.output_prev_subsampled = easycache.subsample(output, uuids)
+        easycache.output_prev_norm = output.flatten().abs().mean()
+        if easycache.verbose:
+            logging.info(f"EasyCache [verbose] - x_prev_subsampled: {easycache.x_prev_subsampled.shape}")
+    return output
+
+def lazycache_predict_noise_wrapper(executor, *args, **kwargs):
+    # get values from args
+    x: torch.Tensor = args[0]
+    timestep: float = args[1]
+    model_options: dict[str] = args[2]
+    easycache: LazyCacheHolder = model_options["transformer_options"]["easycache"]
+    if easycache.is_past_end_timestep(timestep):
+        return executor(*args, **kwargs)
+    # prepare next x_prev
+    next_x_prev = x
+    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()
+            if easycache.has_output_prev_norm() and easycache.has_relative_transformation_rate():
+                approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm
+                easycache.cumulative_change_rate += approx_output_change_rate
+                if easycache.cumulative_change_rate < easycache.reuse_threshold:
+                    if easycache.verbose:
+                        logging.info(f"LazyCache [verbose] - skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
+                    # other conds should also skip this step, and instead use their cached values
+                    easycache.skip_current_step = True
+                    return easycache.apply_cache_diff(x)
+                else:
+                    if easycache.verbose:
+                        logging.info(f"LazyCache [verbose] - NOT skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
+                    easycache.cumulative_change_rate = 0.0
+    output: torch.Tensor = executor(*args, **kwargs)
+    if easycache.has_output_prev_norm():
+        output_change = (easycache.subsample(output, clone=False) - easycache.output_prev_subsampled).flatten().abs().mean()
+        if easycache.verbose:
+            output_change_rate = output_change / easycache.output_prev_norm
+            easycache.output_change_rates.append(output_change_rate.item())
+        if easycache.has_relative_transformation_rate():
+            approx_output_change_rate = (easycache.relative_transformation_rate * input_change) / easycache.output_prev_norm
+            easycache.approx_output_change_rates.append(approx_output_change_rate.item())
+            if easycache.verbose:
+                logging.info(f"LazyCache [verbose] - approx_output_change_rate: {approx_output_change_rate}")
+        if input_change is not None:
+            easycache.relative_transformation_rate = output_change / input_change
+        if easycache.verbose:
+            logging.info(f"LazyCache [verbose] - output_change_rate: {output_change_rate}")
+    # TODO: allow cache_diff to be offloaded
+    easycache.update_cache_diff(output, next_x_prev)
+    easycache.x_prev_subsampled = easycache.subsample(next_x_prev)
+    easycache.output_prev_subsampled = easycache.subsample(output)
+    easycache.output_prev_norm = output.flatten().abs().mean()
+    if easycache.verbose:
+        logging.info(f"LazyCache [verbose] - x_prev_subsampled: {easycache.x_prev_subsampled.shape}")
+    return output
+
+def easycache_calc_cond_batch_wrapper(executor, *args, **kwargs):
+    model_options = args[-1]
+    easycache: EasyCacheHolder = model_options["transformer_options"]["easycache"]
+    easycache.skip_current_step = False
+    # TODO: check if first_cond_uuid is active at this timestep; otherwise, EasyCache needs to be partially reset
+    return executor(*args, **kwargs)
+
+def easycache_sample_wrapper(executor, *args, **kwargs):
+    """
+    This OUTER_SAMPLE wrapper makes sure easycache is prepped for current run, and all memory usage is cleared at the end.
+    """
+    try:
+        guider = executor.class_obj
+        orig_model_options = guider.model_options
+        guider.model_options = comfy.model_patcher.create_model_options_clone(orig_model_options)
+        # clone and prepare timesteps
+        guider.model_options["transformer_options"]["easycache"] = guider.model_options["transformer_options"]["easycache"].clone().prepare_timesteps(guider.model_patcher.model.model_sampling)
+        easycache: Union[EasyCacheHolder, LazyCacheHolder] = guider.model_options['transformer_options']['easycache']
+        logging.info(f"{easycache.name} enabled - threshold: {easycache.reuse_threshold}, start_percent: {easycache.start_percent}, end_percent: {easycache.end_percent}")
+        return executor(*args, **kwargs)
+    finally:
+        easycache = guider.model_options['transformer_options']['easycache']
+        output_change_rates = easycache.output_change_rates
+        approx_output_change_rates = easycache.approx_output_change_rates
+        if easycache.verbose:
+            logging.info(f"{easycache.name} [verbose] - output_change_rates {len(output_change_rates)}: {output_change_rates}")
+            logging.info(f"{easycache.name} [verbose] - approx_output_change_rates {len(approx_output_change_rates)}: {approx_output_change_rates}")
+        total_steps = len(args[3])-1
+        logging.info(f"{easycache.name} - skipped {easycache.total_steps_skipped}/{total_steps} steps ({total_steps/(total_steps-easycache.total_steps_skipped):.2f}x speedup).")
+        easycache.reset()
+        guider.model_options = orig_model_options
+
+
+class EasyCacheHolder:
+    def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False):
+        self.name = "EasyCache"
+        self.reuse_threshold = reuse_threshold
+        self.start_percent = start_percent
+        self.end_percent = end_percent
+        self.subsample_factor = subsample_factor
+        self.offload_cache_diff = offload_cache_diff
+        self.verbose = verbose
+        # timestep values
+        self.start_t = 0.0
+        self.end_t = 0.0
+        # control values
+        self.relative_transformation_rate: float = None
+        self.cumulative_change_rate = 0.0
+        self.initial_step = True
+        self.skip_current_step = False
+        # cache values
+        self.first_cond_uuid = None
+        self.x_prev_subsampled: torch.Tensor = None
+        self.output_prev_subsampled: torch.Tensor = None
+        self.output_prev_norm: torch.Tensor = None
+        self.uuid_cache_diffs: dict[UUID, torch.Tensor] = {}
+        self.output_change_rates = []
+        self.approx_output_change_rates = []
+        self.total_steps_skipped = 0
+        # 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()
+
+    def should_do_easycache(self, timestep: float) -> bool:
+        return (timestep[0] <= self.start_t).item()
+
+    def has_x_prev_subsampled(self) -> bool:
+        return self.x_prev_subsampled is not None
+
+    def has_output_prev_subsampled(self) -> bool:
+        return self.output_prev_subsampled is not None
+
+    def has_output_prev_norm(self) -> bool:
+        return self.output_prev_norm is not None
+
+    def has_relative_transformation_rate(self) -> bool:
+        return self.relative_transformation_rate is not None
+
+    def prepare_timesteps(self, model_sampling):
+        self.start_t = model_sampling.percent_to_sigma(self.start_percent)
+        self.end_t = model_sampling.percent_to_sigma(self.end_percent)
+        return self
+
+    def subsample(self, x: torch.Tensor, uuids: list[UUID], clone: bool = True) -> torch.Tensor:
+        batch_offset = x.shape[0] // len(uuids)
+        uuid_idx = uuids.index(self.first_cond_uuid)
+        if self.subsample_factor > 1:
+            to_return = x[uuid_idx*batch_offset:(uuid_idx+1)*batch_offset, ..., ::self.subsample_factor, ::self.subsample_factor]
+            if clone:
+                return to_return.clone()
+            return to_return
+        to_return = x[uuid_idx*batch_offset:(uuid_idx+1)*batch_offset, ...]
+        if clone:
+            return to_return.clone()
+        return to_return
+
+    def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID]):
+        if self.first_cond_uuid in uuids:
+            self.total_steps_skipped += 1
+        batch_offset = x.shape[0] // len(uuids)
+        for i, uuid in enumerate(uuids):
+            # if cached dims don't match x dims, cut off excess and hope for the best (cosmos world2video)
+            if x.shape[1:] != self.uuid_cache_diffs[uuid].shape[1:]:
+                if not self.allow_mismatch:
+                    raise ValueError(f"Cached dims {self.uuid_cache_diffs[uuid].shape} don't match x dims {x.shape} - this is no good")
+                slicing = []
+                skip_this_dim = True
+                for dim_u, dim_x in zip(self.uuid_cache_diffs[uuid].shape, x.shape):
+                    if skip_this_dim:
+                        skip_this_dim = False
+                        continue
+                    if dim_u != dim_x:
+                        if self.cut_from_start:
+                            slicing.append(slice(dim_x-dim_u, None))
+                        else:
+                            slicing.append(slice(None, dim_u))
+                    else:
+                        slicing.append(slice(None))
+                slicing = [slice(i*batch_offset,(i+1)*batch_offset)] + slicing
+                x = x[slicing]
+            x += self.uuid_cache_diffs[uuid].to(x.device)
+        return x
+
+    def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID]):
+        # if output dims don't match x dims, cut off excess and hope for the best (cosmos world2video)
+        if output.shape[1:] != x.shape[1:]:
+            if not self.allow_mismatch:
+                raise ValueError(f"Output dims {output.shape} don't match x dims {x.shape} - this is no good")
+            slicing = []
+            skip_dim = True
+            for dim_o, dim_x in zip(output.shape, x.shape):
+                if not skip_dim and dim_o != dim_x:
+                    if self.cut_from_start:
+                        slicing.append(slice(dim_x-dim_o, None))
+                    else:
+                        slicing.append(slice(None, dim_o))
+                else:
+                    slicing.append(slice(None))
+                skip_dim = False
+            x = x[slicing]
+        diff = output - x
+        batch_offset = diff.shape[0] // len(uuids)
+        for i, uuid in enumerate(uuids):
+            self.uuid_cache_diffs[uuid] = diff[i*batch_offset:(i+1)*batch_offset, ...]
+
+    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
+        self.initial_step = True
+        self.skip_current_step = False
+        self.output_change_rates = []
+        self.first_cond_uuid = 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
+        del self.uuid_cache_diffs
+        self.uuid_cache_diffs = {}
+        self.total_steps_skipped = 0
+        self.state_metadata = None
+        return self
+
+    def clone(self):
+        return EasyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose)
+
+
+class EasyCacheNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="EasyCache",
+            display_name="EasyCache",
+            description="Native EasyCache implementation.",
+            category="advanced/debug/model",
+            is_experimental=True,
+            inputs=[
+                io.Model.Input("model", tooltip="The model to add EasyCache to."),
+                io.Float.Input("reuse_threshold", min=0.0, default=0.2, max=3.0, step=0.01, tooltip="The threshold for reusing cached steps."),
+                io.Float.Input("start_percent", min=0.0, default=0.15, max=1.0, step=0.01, tooltip="The relative sampling step to begin use of EasyCache."),
+                io.Float.Input("end_percent", min=0.0, default=0.95, max=1.0, step=0.01, tooltip="The relative sampling step to end use of EasyCache."),
+                io.Boolean.Input("verbose", default=False, tooltip="Whether to log verbose information."),
+            ],
+            outputs=[
+                io.Model.Output(tooltip="The model with EasyCache."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, reuse_threshold: float, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput:
+        model = model.clone()
+        model.model_options["transformer_options"]["easycache"] = EasyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "easycache", easycache_sample_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.CALC_COND_BATCH, "easycache", easycache_calc_cond_batch_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, "easycache", easycache_forward_wrapper)
+        return io.NodeOutput(model)
+
+
+class LazyCacheHolder:
+    def __init__(self, reuse_threshold: float, start_percent: float, end_percent: float, subsample_factor: int, offload_cache_diff: bool, verbose: bool=False):
+        self.name = "LazyCache"
+        self.reuse_threshold = reuse_threshold
+        self.start_percent = start_percent
+        self.end_percent = end_percent
+        self.subsample_factor = subsample_factor
+        self.offload_cache_diff = offload_cache_diff
+        self.verbose = verbose
+        # timestep values
+        self.start_t = 0.0
+        self.end_t = 0.0
+        # control values
+        self.relative_transformation_rate: float = None
+        self.cumulative_change_rate = 0.0
+        self.initial_step = True
+        # cache values
+        self.x_prev_subsampled: torch.Tensor = None
+        self.output_prev_subsampled: torch.Tensor = None
+        self.output_prev_norm: torch.Tensor = None
+        self.cache_diff: torch.Tensor = None
+        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
+
+    def is_past_end_timestep(self, timestep: float) -> bool:
+        return not (timestep[0] > self.end_t).item()
+
+    def should_do_easycache(self, timestep: float) -> bool:
+        return (timestep[0] <= self.start_t).item()
+
+    def has_x_prev_subsampled(self) -> bool:
+        return self.x_prev_subsampled is not None
+
+    def has_output_prev_subsampled(self) -> bool:
+        return self.output_prev_subsampled is not None
+
+    def has_output_prev_norm(self) -> bool:
+        return self.output_prev_norm is not None
+
+    def has_relative_transformation_rate(self) -> bool:
+        return self.relative_transformation_rate is not None
+
+    def prepare_timesteps(self, model_sampling):
+        self.start_t = model_sampling.percent_to_sigma(self.start_percent)
+        self.end_t = model_sampling.percent_to_sigma(self.end_percent)
+        return self
+
+    def subsample(self, x: torch.Tensor, clone: bool = True) -> torch.Tensor:
+        if self.subsample_factor > 1:
+            to_return = x[..., ::self.subsample_factor, ::self.subsample_factor]
+            if clone:
+                return to_return.clone()
+            return to_return
+        if clone:
+            return x.clone()
+        return x
+
+    def apply_cache_diff(self, x: torch.Tensor):
+        self.total_steps_skipped += 1
+        return x + self.cache_diff.to(x.device)
+
+    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
+        self.initial_step = True
+        self.output_change_rates = []
+        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):
+        return LazyCacheHolder(self.reuse_threshold, self.start_percent, self.end_percent, self.subsample_factor, self.offload_cache_diff, self.verbose)
+
+class LazyCacheNode(io.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        return io.Schema(
+            node_id="LazyCache",
+            display_name="LazyCache",
+            description="A homebrew version of EasyCache - even 'easier' version of EasyCache to implement. Overall works worse than EasyCache, but better in some rare cases AND universal compatibility with everything in ComfyUI.",
+            category="advanced/debug/model",
+            is_experimental=True,
+            inputs=[
+                io.Model.Input("model", tooltip="The model to add LazyCache to."),
+                io.Float.Input("reuse_threshold", min=0.0, default=0.2, max=3.0, step=0.01, tooltip="The threshold for reusing cached steps."),
+                io.Float.Input("start_percent", min=0.0, default=0.15, max=1.0, step=0.01, tooltip="The relative sampling step to begin use of LazyCache."),
+                io.Float.Input("end_percent", min=0.0, default=0.95, max=1.0, step=0.01, tooltip="The relative sampling step to end use of LazyCache."),
+                io.Boolean.Input("verbose", default=False, tooltip="Whether to log verbose information."),
+            ],
+            outputs=[
+                io.Model.Output(tooltip="The model with LazyCache."),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model: io.Model.Type, reuse_threshold: float, start_percent: float, end_percent: float, verbose: bool) -> io.NodeOutput:
+        model = model.clone()
+        model.model_options["transformer_options"]["easycache"] = LazyCacheHolder(reuse_threshold, start_percent, end_percent, subsample_factor=8, offload_cache_diff=False, verbose=verbose)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.OUTER_SAMPLE, "lazycache", easycache_sample_wrapper)
+        model.add_wrapper_with_key(comfy.patcher_extension.WrappersMP.PREDICT_NOISE, "lazycache", lazycache_predict_noise_wrapper)
+        return io.NodeOutput(model)
+
+
+class EasyCacheExtension(ComfyExtension):
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            EasyCacheNode,
+            LazyCacheNode,
+        ]
+
+def comfy_entrypoint():
+    return EasyCacheExtension()

comfy_extras/nodes_latent.py

@@ -1,6 +1,7 @@
 import comfy.utils
 import comfy_extras.nodes_post_processing
 import torch
+import nodes
 
 
 def reshape_latent_to(target_shape, latent, repeat_batch=True):
@@ -105,6 +106,73 @@ 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 LatentCut:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": {"samples": ("LATENT",),
+                             "dim": (["x", "y", "t"], ),
+                             "index": ("INT", {"default": 0, "min": -nodes.MAX_RESOLUTION, "max": nodes.MAX_RESOLUTION, "step": 1}),
+                             "amount": ("INT", {"default": 1, "min": 1, "max": nodes.MAX_RESOLUTION, "step": 1})}}
+
+    RETURN_TYPES = ("LATENT",)
+    FUNCTION = "op"
+
+    CATEGORY = "latent/advanced"
+
+    def op(self, samples, dim, index, amount):
+        samples_out = samples.copy()
+
+        s1 = samples["samples"]
+
+        if "x" in dim:
+            dim = s1.ndim - 1
+        elif "y" in dim:
+            dim = s1.ndim - 2
+        elif "t" in dim:
+            dim = s1.ndim - 3
+
+        if index >= 0:
+            index = min(index, s1.shape[dim] - 1)
+            amount = min(s1.shape[dim] - index, amount)
+        else:
+            index = max(index, -s1.shape[dim])
+            amount = min(-index, amount)
+
+        samples_out["samples"] = torch.narrow(s1, dim, index, amount)
+        return (samples_out,)
+
 class LatentBatch:
     @classmethod
     def INPUT_TYPES(s):
@@ -279,6 +347,8 @@ NODE_CLASS_MAPPINGS = {
     "LatentSubtract": LatentSubtract,
     "LatentMultiply": LatentMultiply,
     "LatentInterpolate": LatentInterpolate,
+    "LatentConcat": LatentConcat,
+    "LatentCut": LatentCut,
     "LatentBatch": LatentBatch,
     "LatentBatchSeedBehavior": LatentBatchSeedBehavior,
     "LatentApplyOperation": LatentApplyOperation,

comfy_extras/nodes_lt.py

@@ -166,7 +166,7 @@ class LTXVAddGuide:
         negative = self.add_keyframe_index(negative, frame_idx, guiding_latent, scale_factors)
 
         mask = torch.full(
-            (noise_mask.shape[0], 1, guiding_latent.shape[2], 1, 1),
+            (noise_mask.shape[0], 1, guiding_latent.shape[2], noise_mask.shape[3], noise_mask.shape[4]),
             1.0 - strength,
             dtype=noise_mask.dtype,
             device=noise_mask.device,

comfy_extras/nodes_model_patch.py

@@ -0,0 +1,163 @@
+import torch
+import folder_paths
+import comfy.utils
+import comfy.ops
+import comfy.model_management
+import comfy.ldm.common_dit
+import comfy.latent_formats
+
+
+class BlockWiseControlBlock(torch.nn.Module):
+    # [linear, gelu, linear]
+    def __init__(self, dim: int = 3072, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.x_rms = operations.RMSNorm(dim, eps=1e-6)
+        self.y_rms = operations.RMSNorm(dim, eps=1e-6)
+        self.input_proj = operations.Linear(dim, dim)
+        self.act = torch.nn.GELU()
+        self.output_proj = operations.Linear(dim, dim)
+
+    def forward(self, x, y):
+        x, y = self.x_rms(x), self.y_rms(y)
+        x = self.input_proj(x + y)
+        x = self.act(x)
+        x = self.output_proj(x)
+        return x
+
+
+class QwenImageBlockWiseControlNet(torch.nn.Module):
+    def __init__(
+        self,
+        num_layers: int = 60,
+        in_dim: int = 64,
+        additional_in_dim: int = 0,
+        dim: int = 3072,
+        device=None, dtype=None, operations=None
+    ):
+        super().__init__()
+        self.additional_in_dim = additional_in_dim
+        self.img_in = operations.Linear(in_dim + additional_in_dim, dim, device=device, dtype=dtype)
+        self.controlnet_blocks = torch.nn.ModuleList(
+            [
+                BlockWiseControlBlock(dim, device=device, dtype=dtype, operations=operations)
+                for _ in range(num_layers)
+            ]
+        )
+
+    def process_input_latent_image(self, latent_image):
+        latent_image[:, :16] = comfy.latent_formats.Wan21().process_in(latent_image[:, :16])
+        patch_size = 2
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(latent_image, (1, patch_size, patch_size))
+        orig_shape = hidden_states.shape
+        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
+        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        return self.img_in(hidden_states)
+
+    def control_block(self, img, controlnet_conditioning, block_id):
+        return self.controlnet_blocks[block_id](img, controlnet_conditioning)
+
+
+class ModelPatchLoader:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "name": (folder_paths.get_filename_list("model_patches"), ),
+                              }}
+    RETURN_TYPES = ("MODEL_PATCH",)
+    FUNCTION = "load_model_patch"
+    EXPERIMENTAL = True
+
+    CATEGORY = "advanced/loaders"
+
+    def load_model_patch(self, name):
+        model_patch_path = folder_paths.get_full_path_or_raise("model_patches", name)
+        sd = comfy.utils.load_torch_file(model_patch_path, safe_load=True)
+        dtype = comfy.utils.weight_dtype(sd)
+        # TODO: this node will work with more types of model patches
+        additional_in_dim = sd["img_in.weight"].shape[1] - 64
+        model = QwenImageBlockWiseControlNet(additional_in_dim=additional_in_dim, device=comfy.model_management.unet_offload_device(), dtype=dtype, 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())
+        return (model,)
+
+
+class DiffSynthCnetPatch:
+    def __init__(self, model_patch, vae, image, strength, mask=None):
+        self.model_patch = model_patch
+        self.vae = vae
+        self.image = image
+        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)
+        if self.model_patch.model.additional_in_dim > 0:
+            if self.mask is None:
+                mask_ = torch.ones_like(latent_image)[:, :self.model_patch.model.additional_in_dim // 4]
+            else:
+                mask_ = comfy.utils.common_upscale(self.mask.mean(dim=1, keepdim=True), latent_image.shape[-1], latent_image.shape[-2], "bilinear", "none")
+
+            return torch.cat([latent_image, mask_], dim=1)
+        else:
+            return latent_image
+
+    def __call__(self, kwargs):
+        x = kwargs.get("x")
+        img = kwargs.get("img")
+        block_index = kwargs.get("block_index")
+        spacial_compression = self.vae.spacial_compression_encode()
+        if self.encoded_image is None or self.encoded_image_size != (x.shape[-2] * spacial_compression, x.shape[-1] * 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[:, :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
+
+    def to(self, device_or_dtype):
+        if isinstance(device_or_dtype, torch.device):
+            self.encoded_image = self.encoded_image.to(device_or_dtype)
+        return self
+
+    def models(self):
+        return [self.model_patch]
+
+class QwenImageDiffsynthControlnet:
+    @classmethod
+    def INPUT_TYPES(s):
+        return {"required": { "model": ("MODEL",),
+                              "model_patch": ("MODEL_PATCH",),
+                              "vae": ("VAE",),
+                              "image": ("IMAGE",),
+                              "strength": ("FLOAT", {"default": 1.0, "min": -10.0, "max": 10.0, "step": 0.01}),
+                              },
+                "optional": {"mask": ("MASK",)}}
+    RETURN_TYPES = ("MODEL",)
+    FUNCTION = "diffsynth_controlnet"
+    EXPERIMENTAL = True
+
+    CATEGORY = "advanced/loaders/qwen"
+
+    def diffsynth_controlnet(self, model, model_patch, vae, image, strength, mask=None):
+        model_patched = model.clone()
+        image = image[:, :, :, :3]
+        if mask is not None:
+            if mask.ndim == 3:
+                mask = mask.unsqueeze(1)
+            if mask.ndim == 4:
+                mask = mask.unsqueeze(2)
+            mask = 1.0 - mask
+
+        model_patched.set_model_double_block_patch(DiffSynthCnetPatch(model_patch, vae, image, strength, mask))
+        return (model_patched,)
+
+
+NODE_CLASS_MAPPINGS = {
+    "ModelPatchLoader": ModelPatchLoader,
+    "QwenImageDiffsynthControlnet": QwenImageDiffsynthControlnet,
+}

comfy_extras/nodes_qwen.py

@@ -1,25 +1,6 @@
 import node_helpers
 import comfy.utils
-
-PREFERRED_QWENIMAGE_RESOLUTIONS = [
-    (672, 1568),
-    (688, 1504),
-    (720, 1456),
-    (752, 1392),
-    (800, 1328),
-    (832, 1248),
-    (880, 1184),
-    (944, 1104),
-    (1024, 1024),
-    (1104, 944),
-    (1184, 880),
-    (1248, 832),
-    (1328, 800),
-    (1392, 752),
-    (1456, 720),
-    (1504, 688),
-    (1568, 672),
-]
+import math
 
 
 class TextEncodeQwenImageEdit:
@@ -42,13 +23,17 @@ class TextEncodeQwenImageEdit:
         if image is None:
             images = []
         else:
-            images = [image]
+            samples = image.movedim(-1, 1)
+            total = int(1024 * 1024)
+
+            scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2]))
+            width = round(samples.shape[3] * scale_by)
+            height = round(samples.shape[2] * scale_by)
+
+            s = comfy.utils.common_upscale(samples, width, height, "area", "disabled")
+            image = s.movedim(1, -1)
+            images = [image[:, :, :, :3]]
             if vae is not None:
-                width = image.shape[2]
-                height = image.shape[1]
-                aspect_ratio = width / height
-                _, width, height = min((abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_QWENIMAGE_RESOLUTIONS)
-                image = comfy.utils.common_upscale(image.movedim(-1, 1), width, height, "lanczos", "center").movedim(1, -1)
                 ref_latent = vae.encode(image[:, :, :, :3])
 
         tokens = clip.tokenize(prompt, images=images)

comfy_extras/nodes_string.py

@@ -1,77 +1,91 @@
 import re
+from typing_extensions import override
 
-from comfy.comfy_types.node_typing import IO
+from comfy_api.latest import ComfyExtension, io
 
-class StringConcatenate():
+
+class StringConcatenate(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string_a": (IO.STRING, {"multiline": True}),
-                "string_b": (IO.STRING, {"multiline": True}),
-                "delimiter": (IO.STRING, {"multiline": False, "default": ""})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringConcatenate",
+            display_name="Concatenate",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string_a", multiline=True),
+                io.String.Input("string_b", multiline=True),
+                io.String.Input("delimiter", multiline=False, default=""),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string_a, string_b, delimiter, **kwargs):
-        return delimiter.join((string_a, string_b)),
-
-class StringSubstring():
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "start": (IO.INT, {}),
-                "end": (IO.INT, {}),
-            }
-        }
+    def execute(cls, string_a, string_b, delimiter):
+        return io.NodeOutput(delimiter.join((string_a, string_b)))
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
 
-    def execute(self, string, start, end, **kwargs):
-        return string[start:end],
-
-class StringLength():
+class StringSubstring(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringSubstring",
+            display_name="Substring",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.Int.Input("start"),
+                io.Int.Input("end"),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.INT,)
-    RETURN_NAMES = ("length",)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, **kwargs):
-        length = len(string)
-
-        return length,
-
-class CaseConverter():
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "mode": (IO.COMBO, {"options": ["UPPERCASE", "lowercase", "Capitalize", "Title Case"]})
-            }
-        }
+    def execute(cls, string, start, end):
+        return io.NodeOutput(string[start:end])
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
 
-    def execute(self, string, mode, **kwargs):
+class StringLength(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringLength",
+            display_name="Length",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+            ],
+            outputs=[
+                io.Int.Output(display_name="length"),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, string):
+        return io.NodeOutput(len(string))
+
+
+class CaseConverter(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="CaseConverter",
+            display_name="Case Converter",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.Combo.Input("mode", options=["UPPERCASE", "lowercase", "Capitalize", "Title Case"]),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, string, mode):
         if mode == "UPPERCASE":
             result = string.upper()
         elif mode == "lowercase":
@@ -83,24 +97,27 @@ class CaseConverter():
         else:
             result = string
 
-        return result,
+        return io.NodeOutput(result)
 
 
-class StringTrim():
+class StringTrim(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "mode": (IO.COMBO, {"options": ["Both", "Left", "Right"]})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringTrim",
+            display_name="Trim",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.Combo.Input("mode", options=["Both", "Left", "Right"]),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, mode, **kwargs):
+    @classmethod
+    def execute(cls, string, mode):
         if mode == "Both":
             result = string.strip()
         elif mode == "Left":
@@ -110,70 +127,78 @@ class StringTrim():
         else:
             result = string
 
-        return result,
+        return io.NodeOutput(result)
 
-class StringReplace():
+
+class StringReplace(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "find": (IO.STRING, {"multiline": True}),
-                "replace": (IO.STRING, {"multiline": True})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringReplace",
+            display_name="Replace",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.String.Input("find", multiline=True),
+                io.String.Input("replace", multiline=True),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, find, replace, **kwargs):
-        result = string.replace(find, replace)
-        return result,
-
-
-class StringContains():
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "substring": (IO.STRING, {"multiline": True}),
-                "case_sensitive": (IO.BOOLEAN, {"default": True})
-            }
-        }
+    def execute(cls, string, find, replace):
+        return io.NodeOutput(string.replace(find, replace))
 
-    RETURN_TYPES = (IO.BOOLEAN,)
-    RETURN_NAMES = ("contains",)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
 
-    def execute(self, string, substring, case_sensitive, **kwargs):
+class StringContains(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringContains",
+            display_name="Contains",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.String.Input("substring", multiline=True),
+                io.Boolean.Input("case_sensitive", default=True),
+            ],
+            outputs=[
+                io.Boolean.Output(display_name="contains"),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, string, substring, case_sensitive):
         if case_sensitive:
             contains = substring in string
         else:
             contains = substring.lower() in string.lower()
 
-        return contains,
+        return io.NodeOutput(contains)
 
 
-class StringCompare():
+class StringCompare(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string_a": (IO.STRING, {"multiline": True}),
-                "string_b": (IO.STRING, {"multiline": True}),
-                "mode": (IO.COMBO, {"options": ["Starts With", "Ends With", "Equal"]}),
-                "case_sensitive": (IO.BOOLEAN, {"default": True})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="StringCompare",
+            display_name="Compare",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string_a", multiline=True),
+                io.String.Input("string_b", multiline=True),
+                io.Combo.Input("mode", options=["Starts With", "Ends With", "Equal"]),
+                io.Boolean.Input("case_sensitive", default=True),
+            ],
+            outputs=[
+                io.Boolean.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.BOOLEAN,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string_a, string_b, mode, case_sensitive, **kwargs):
+    @classmethod
+    def execute(cls, string_a, string_b, mode, case_sensitive):
         if case_sensitive:
             a = string_a
             b = string_b
@@ -182,31 +207,34 @@ class StringCompare():
             b = string_b.lower()
 
         if mode == "Equal":
-            return a == b,
+            return io.NodeOutput(a == b)
         elif mode == "Starts With":
-            return a.startswith(b),
+            return io.NodeOutput(a.startswith(b))
         elif mode == "Ends With":
-            return a.endswith(b),
+            return io.NodeOutput(a.endswith(b))
 
-class RegexMatch():
+
+class RegexMatch(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "regex_pattern": (IO.STRING, {"multiline": True}),
-                "case_insensitive": (IO.BOOLEAN, {"default": True}),
-                "multiline": (IO.BOOLEAN, {"default": False}),
-                "dotall": (IO.BOOLEAN, {"default": False})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="RegexMatch",
+            display_name="Regex Match",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.String.Input("regex_pattern", multiline=True),
+                io.Boolean.Input("case_insensitive", default=True),
+                io.Boolean.Input("multiline", default=False),
+                io.Boolean.Input("dotall", default=False),
+            ],
+            outputs=[
+                io.Boolean.Output(display_name="matches"),
+            ]
+        )
 
-    RETURN_TYPES = (IO.BOOLEAN,)
-    RETURN_NAMES = ("matches",)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, regex_pattern, case_insensitive, multiline, dotall, **kwargs):
+    @classmethod
+    def execute(cls, string, regex_pattern, case_insensitive, multiline, dotall):
         flags = 0
 
         if case_insensitive:
@@ -223,29 +251,32 @@ class RegexMatch():
         except re.error:
             result = False
 
-        return result,
+        return io.NodeOutput(result)
 
 
-class RegexExtract():
+class RegexExtract(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "regex_pattern": (IO.STRING, {"multiline": True}),
-                "mode": (IO.COMBO, {"options": ["First Match", "All Matches", "First Group", "All Groups"]}),
-                "case_insensitive": (IO.BOOLEAN, {"default": True}),
-                "multiline": (IO.BOOLEAN, {"default": False}),
-                "dotall": (IO.BOOLEAN, {"default": False}),
-                "group_index": (IO.INT, {"default": 1, "min": 0, "max": 100})
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="RegexExtract",
+            display_name="Regex Extract",
+            category="utils/string",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.String.Input("regex_pattern", multiline=True),
+                io.Combo.Input("mode", options=["First Match", "All Matches", "First Group", "All Groups"]),
+                io.Boolean.Input("case_insensitive", default=True),
+                io.Boolean.Input("multiline", default=False),
+                io.Boolean.Input("dotall", default=False),
+                io.Int.Input("group_index", default=1, min=0, max=100),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, regex_pattern, mode, case_insensitive, multiline, dotall, group_index, **kwargs):
+    @classmethod
+    def execute(cls, string, regex_pattern, mode, case_insensitive, multiline, dotall, group_index):
         join_delimiter = "\n"
 
         flags = 0
@@ -294,32 +325,33 @@ class RegexExtract():
         except re.error:
             result = ""
 
-        return result,
+        return io.NodeOutput(result)
 
 
-class RegexReplace():
-    DESCRIPTION = "Find and replace text using regex patterns."
+class RegexReplace(io.ComfyNode):
     @classmethod
-    def INPUT_TYPES(s):
-        return {
-            "required": {
-                "string": (IO.STRING, {"multiline": True}),
-                "regex_pattern": (IO.STRING, {"multiline": True}),
-                "replace": (IO.STRING, {"multiline": True}),
-            },
-            "optional": {
-                "case_insensitive": (IO.BOOLEAN, {"default": True}),
-                "multiline": (IO.BOOLEAN, {"default": False}),
-                "dotall": (IO.BOOLEAN, {"default": False, "tooltip": "When enabled, the dot (.) character will match any character including newline characters. When disabled, dots won't match newlines."}),
-                "count": (IO.INT, {"default": 0, "min": 0, "max": 100, "tooltip": "Maximum number of replacements to make. Set to 0 to replace all occurrences (default). Set to 1 to replace only the first match, 2 for the first two matches, etc."}),
-            }
-        }
+    def define_schema(cls):
+        return io.Schema(
+            node_id="RegexReplace",
+            display_name="Regex Replace",
+            category="utils/string",
+            description="Find and replace text using regex patterns.",
+            inputs=[
+                io.String.Input("string", multiline=True),
+                io.String.Input("regex_pattern", multiline=True),
+                io.String.Input("replace", multiline=True),
+                io.Boolean.Input("case_insensitive", default=True, optional=True),
+                io.Boolean.Input("multiline", default=False, optional=True),
+                io.Boolean.Input("dotall", default=False, optional=True, tooltip="When enabled, the dot (.) character will match any character including newline characters. When disabled, dots won't match newlines."),
+                io.Int.Input("count", default=0, min=0, max=100, optional=True, tooltip="Maximum number of replacements to make. Set to 0 to replace all occurrences (default). Set to 1 to replace only the first match, 2 for the first two matches, etc."),
+            ],
+            outputs=[
+                io.String.Output(),
+            ]
+        )
 
-    RETURN_TYPES = (IO.STRING,)
-    FUNCTION = "execute"
-    CATEGORY = "utils/string"
-
-    def execute(self, string, regex_pattern, replace, case_insensitive=True, multiline=False, dotall=False, count=0, **kwargs):
+    @classmethod
+    def execute(cls, string, regex_pattern, replace, case_insensitive=True, multiline=False, dotall=False, count=0):
         flags = 0
 
         if case_insensitive:
@@ -329,32 +361,25 @@ class RegexReplace():
         if dotall:
             flags |= re.DOTALL
         result = re.sub(regex_pattern, replace, string, count=count, flags=flags)
-        return result,
+        return io.NodeOutput(result)
 
-NODE_CLASS_MAPPINGS = {
-    "StringConcatenate": StringConcatenate,
-    "StringSubstring": StringSubstring,
-    "StringLength": StringLength,
-    "CaseConverter": CaseConverter,
-    "StringTrim": StringTrim,
-    "StringReplace": StringReplace,
-    "StringContains": StringContains,
-    "StringCompare": StringCompare,
-    "RegexMatch": RegexMatch,
-    "RegexExtract": RegexExtract,
-    "RegexReplace": RegexReplace,
-}
 
-NODE_DISPLAY_NAME_MAPPINGS = {
-    "StringConcatenate": "Concatenate",
-    "StringSubstring": "Substring",
-    "StringLength": "Length",
-    "CaseConverter": "Case Converter",
-    "StringTrim": "Trim",
-    "StringReplace": "Replace",
-    "StringContains": "Contains",
-    "StringCompare": "Compare",
-    "RegexMatch": "Regex Match",
-    "RegexExtract": "Regex Extract",
-    "RegexReplace": "Regex Replace",
-}
+class StringExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            StringConcatenate,
+            StringSubstring,
+            StringLength,
+            CaseConverter,
+            StringTrim,
+            StringReplace,
+            StringContains,
+            StringCompare,
+            RegexMatch,
+            RegexExtract,
+            RegexReplace,
+        ]
+
+async def comfy_entrypoint() -> StringExtension:
+    return StringExtension()

comfy_extras/nodes_wan.py

@@ -139,16 +139,21 @@ class Wan22FunControlToVideo(io.ComfyNode):
 
     @classmethod
     def execute(cls, positive, negative, vae, width, height, length, batch_size, ref_image=None, start_image=None, control_video=None) -> io.NodeOutput:
-        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
-        concat_latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
-        concat_latent = comfy.latent_formats.Wan21().process_out(concat_latent)
+        spacial_scale = vae.spacial_compression_encode()
+        latent_channels = vae.latent_channels
+        latent = torch.zeros([batch_size, latent_channels, ((length - 1) // 4) + 1, height // spacial_scale, width // spacial_scale], device=comfy.model_management.intermediate_device())
+        concat_latent = torch.zeros([batch_size, latent_channels, ((length - 1) // 4) + 1, height // spacial_scale, width // spacial_scale], device=comfy.model_management.intermediate_device())
+        if latent_channels == 48:
+            concat_latent = comfy.latent_formats.Wan22().process_out(concat_latent)
+        else:
+            concat_latent = comfy.latent_formats.Wan21().process_out(concat_latent)
         concat_latent = concat_latent.repeat(1, 2, 1, 1, 1)
         mask = torch.ones((1, 1, latent.shape[2] * 4, latent.shape[-2], latent.shape[-1]))
 
         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)
             concat_latent_image = vae.encode(start_image[:, :, :, :3])
-            concat_latent[:,16:,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]]
+            concat_latent[:,latent_channels:,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]]
             mask[:, :, :start_image.shape[0] + 3] = 0.0
 
         ref_latent = None
@@ -159,11 +164,11 @@ class Wan22FunControlToVideo(io.ComfyNode):
         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)
             concat_latent_image = vae.encode(control_video[:, :, :, :3])
-            concat_latent[:,:16,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]]
+            concat_latent[:,:latent_channels,:concat_latent_image.shape[2]] = concat_latent_image[:,:,:concat_latent.shape[2]]
 
         mask = mask.view(1, mask.shape[2] // 4, 4, mask.shape[3], mask.shape[4]).transpose(1, 2)
-        positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent, "concat_mask": mask, "concat_mask_index": 16})
-        negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent, "concat_mask": mask, "concat_mask_index": 16})
+        positive = node_helpers.conditioning_set_values(positive, {"concat_latent_image": concat_latent, "concat_mask": mask, "concat_mask_index": latent_channels})
+        negative = node_helpers.conditioning_set_values(negative, {"concat_latent_image": concat_latent, "concat_mask": mask, "concat_mask_index": latent_channels})
 
         if ref_latent is not None:
             positive = node_helpers.conditioning_set_values(positive, {"reference_latents": [ref_latent]}, append=True)
@@ -201,7 +206,8 @@ class WanFirstLastFrameToVideo(io.ComfyNode):
 
     @classmethod
     def execute(cls, positive, negative, vae, width, height, length, batch_size, start_image=None, end_image=None, clip_vision_start_image=None, clip_vision_end_image=None) -> io.NodeOutput:
-        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+        spacial_scale = vae.spacial_compression_encode()
+        latent = torch.zeros([batch_size, vae.latent_channels, ((length - 1) // 4) + 1, height // spacial_scale, width // spacial_scale], 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)
         if end_image is not None:
@@ -786,6 +792,229 @@ 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
+
+
+def wan_sound_to_video(positive, negative, vae, width, height, length, batch_size, frame_offset=0, ref_image=None, audio_encoder_output=None, control_video=None, ref_motion=None, ref_motion_latent=None):
+    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)
+        batch_frames = latent_t * 4
+        audio_embed_bucket, num_repeat = get_audio_embed_bucket_fps(feat, fps=fps, batch_frames=batch_frames, 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)
+
+        audio_embed_bucket = audio_embed_bucket[:, :, :, frame_offset:frame_offset + batch_frames]
+        if audio_embed_bucket.shape[3] > 0:
+            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})
+            frame_offset += batch_frames
+
+    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_latent = vae.encode(ref_motion[:, :, :, :3])
+
+    if ref_motion_latent is not None:
+        ref_motion_latent = ref_motion_latent[:, :, -19:]
+        positive = node_helpers.conditioning_set_values(positive, {"reference_motion": ref_motion_latent})
+        negative = node_helpers.conditioning_set_values(negative, {"reference_motion": ref_motion_latent})
+
+    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 positive, negative, out_latent, frame_offset
+
+
+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:
+        positive, negative, out_latent, frame_offset = wan_sound_to_video(positive, negative, vae, width, height, length, batch_size, ref_image=ref_image, audio_encoder_output=audio_encoder_output,
+                                                                          control_video=control_video, ref_motion=ref_motion)
+        return io.NodeOutput(positive, negative, out_latent)
+
+
+class WanSoundImageToVideoExtend(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanSoundImageToVideoExtend",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("length", default=77, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Latent.Input("video_latent"),
+                io.AudioEncoderOutput.Input("audio_encoder_output", optional=True),
+                io.Image.Input("ref_image", optional=True),
+                io.Image.Input("control_video", 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, length, video_latent, ref_image=None, audio_encoder_output=None, control_video=None) -> io.NodeOutput:
+        video_latent = video_latent["samples"]
+        width = video_latent.shape[-1] * 8
+        height = video_latent.shape[-2] * 8
+        batch_size = video_latent.shape[0]
+        frame_offset = video_latent.shape[-3] * 4
+        positive, negative, out_latent, frame_offset = wan_sound_to_video(positive, negative, vae, width, height, length, batch_size, frame_offset=frame_offset, ref_image=ref_image, audio_encoder_output=audio_encoder_output,
+                                                                          control_video=control_video, ref_motion=None, ref_motion_latent=video_latent)
+        return io.NodeOutput(positive, negative, out_latent)
+
+
 class Wan22ImageToVideoLatent(io.ComfyNode):
     @classmethod
     def define_schema(cls):
@@ -844,6 +1073,8 @@ class WanExtension(ComfyExtension):
             TrimVideoLatent,
             WanCameraImageToVideo,
             WanPhantomSubjectToVideo,
+            WanSoundImageToVideo,
+            WanSoundImageToVideoExtend,
             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__ = "0.3.50"
+__version__ = "0.3.55"

folder_paths.py

@@ -46,6 +46,10 @@ folder_names_and_paths["photomaker"] = ([os.path.join(models_dir, "photomaker")]
 
 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")

main.py

@@ -112,6 +112,7 @@ import gc
 
 
 if os.name == "nt":
+    os.environ['MIMALLOC_PURGE_DELAY'] = '0'
     logging.getLogger("xformers").addFilter(lambda record: 'A matching Triton is not available' not in record.getMessage())
 
 if __name__ == "__main__":

nodes.py

@@ -2323,6 +2323,9 @@ async def init_builtin_extra_nodes():
         "nodes_tcfg.py",
         "nodes_context_windows.py",
         "nodes_qwen.py",
+        "nodes_model_patch.py",
+        "nodes_easycache.py",
+        "nodes_audio_encoder.py",
     ]
 
     import_failed = []
@@ -2352,6 +2355,7 @@ async def init_builtin_api_nodes():
         "nodes_moonvalley.py",
         "nodes_rodin.py",
         "nodes_gemini.py",
+        "nodes_vidu.py",
     ]
 
     if not await load_custom_node(os.path.join(api_nodes_dir, "canary.py"), module_parent="comfy_api_nodes"):

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.3.50"
+version = "0.3.55"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.9"

requirements.txt

@@ -1,5 +1,5 @@
-comfyui-frontend-package==1.25.9
-comfyui-workflow-templates==0.1.60
+comfyui-frontend-package==1.25.11
+comfyui-workflow-templates==0.1.70
 comfyui-embedded-docs==0.2.6
 torch
 torchsde