Merge d2c287a368 into c5ecd231a2

Signed-off-by: bigcat88 <bigcat88@icloud.com>

.github/workflows/openapi-lint.yml

@@ -0,0 +1,31 @@
+name: OpenAPI Lint
+
+on:
+  pull_request:
+    paths:
+      - 'openapi.yaml'
+      - '.spectral.yaml'
+      - '.github/workflows/openapi-lint.yml'
+
+permissions:
+  contents: read
+
+jobs:
+  spectral:
+    name: Run Spectral
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Set up Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: '20'
+
+      - name: Install Spectral
+        run: npm install -g @stoplight/spectral-cli@6
+
+      - name: Lint openapi.yaml
+        run: spectral lint openapi.yaml --ruleset .spectral.yaml --fail-severity=error

.github/workflows/stable-release.yml

@@ -145,6 +145,8 @@ jobs:
           cp -r ComfyUI/.ci/windows_${{ inputs.rel_name }}_base_files/* ./
           cp ../update_comfyui_and_python_dependencies.bat ./update/
 
+          echo 'local-portable' > ComfyUI/.comfy_environment
+
           cd ..
 
           "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma2 -mx=9 -mfb=128 -md=768m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable

.gitignore

@@ -23,3 +23,4 @@ web_custom_versions/
 .DS_Store
 filtered-openapi.yaml
 uv.lock
+.comfy_environment

.spectral.yaml

@@ -0,0 +1,91 @@
+extends:
+  - spectral:oas
+
+# Severity levels: error, warn, info, hint, off
+# Rules from the built-in "spectral:oas" ruleset are active by default.
+# Below we tune severity and add custom rules for our conventions.
+#
+# This ruleset mirrors Comfy-Org/cloud/.spectral.yaml so specs across the
+# organization are linted against a single consistent standard.
+
+rules:
+  # -----------------------------------------------------------------------
+  # Built-in rule severity overrides
+  # -----------------------------------------------------------------------
+  operation-operationId: error
+  operation-description: warn
+  operation-tag-defined: error
+  info-contact: off
+  info-description: warn
+  no-eval-in-markdown: error
+  no-$ref-siblings: error
+
+  # -----------------------------------------------------------------------
+  # Custom rules: naming conventions
+  # -----------------------------------------------------------------------
+
+  # Property names should be snake_case
+  property-name-snake-case:
+    description: Property names must be snake_case
+    severity: warn
+    given: "$.components.schemas.*.properties[*]~"
+    then:
+      function: pattern
+      functionOptions:
+        match: "^[a-z][a-z0-9]*(_[a-z0-9]+)*$"
+
+  # Operation IDs should be camelCase
+  operation-id-camel-case:
+    description: Operation IDs must be camelCase
+    severity: warn
+    given: "$.paths.*.*.operationId"
+    then:
+      function: pattern
+      functionOptions:
+        match: "^[a-z][a-zA-Z0-9]*$"
+
+  # -----------------------------------------------------------------------
+  # Custom rules: response conventions
+  # -----------------------------------------------------------------------
+
+  # Error responses (4xx, 5xx) should use a consistent shape
+  error-response-schema:
+    description: Error responses should reference a standard error schema
+    severity: hint
+    given: "$.paths.*.*.responses[?(@property >= '400' && @property < '600')].content['application/json'].schema"
+    then:
+      field: "$ref"
+      function: truthy
+
+  # All 2xx responses with JSON body should have a schema
+  response-schema-defined:
+    description: Success responses with JSON content should define a schema
+    severity: warn
+    given: "$.paths.*.*.responses[?(@property >= '200' && @property < '300')].content['application/json']"
+    then:
+      field: schema
+      function: truthy
+
+  # -----------------------------------------------------------------------
+  # Custom rules: best practices
+  # -----------------------------------------------------------------------
+
+  # Path parameters must have a description
+  path-param-description:
+    description: Path parameters should have a description
+    severity: warn
+    given:
+      - "$.paths.*.parameters[?(@.in == 'path')]"
+      - "$.paths.*.*.parameters[?(@.in == 'path')]"
+    then:
+      field: description
+      function: truthy
+
+  # Schemas should have a description
+  schema-description:
+    description: Component schemas should have a description
+    severity: hint
+    given: "$.components.schemas.*"
+    then:
+      field: description
+      function: truthy

README.md

@@ -1,7 +1,7 @@
 <div align="center">
 
 # ComfyUI
-**The most powerful and modular visual AI engine and application.**
+**The most powerful and modular AI engine for content creation.**
 
 
 [![Website][website-shield]][website-url]
@@ -31,10 +31,16 @@
 [github-downloads-latest-shield]: https://img.shields.io/github/downloads/comfyanonymous/ComfyUI/latest/total?style=flat&label=downloads%40latest
 [github-downloads-link]: https://github.com/comfyanonymous/ComfyUI/releases
 
-![ComfyUI Screenshot](https://github.com/user-attachments/assets/7ccaf2c1-9b72-41ae-9a89-5688c94b7abe)
+<img width="1590" height="795" alt="ComfyUI Screenshot" src="https://github.com/user-attachments/assets/36e065e0-bfae-4456-8c7f-8369d5ea48a2" />
+<br>
 </div>
 
-ComfyUI lets you design and execute advanced stable diffusion pipelines using a graph/nodes/flowchart based interface. Available on Windows, Linux, and macOS.
+ComfyUI is the AI creation engine for visual professionals who demand control over every model, every parameter, and every output. Its powerful and modular node graph interface empowers creatives to generate images, videos, 3D models, audio, and more...
+- ComfyUI natively supports the latest open-source state of the art models.
+- API nodes provide access to the best closed source models such as Nano Banana, Seedance, Hunyuan3D, etc.
+- It is available on Windows, Linux, and macOS, locally with our desktop application or on our cloud.
+- The most sophisticated workflows can be exposed through a simple UI thanks to App Mode.
+- It integrates seamlessly into production pipelines with our API endpoints.
 
 ## Get Started
 
@@ -77,6 +83,7 @@ See what ComfyUI can do with the [newer template workflows](https://comfy.org/wo
    - [Hunyuan Image 2.1](https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_image/)
    - [Flux 2](https://comfyanonymous.github.io/ComfyUI_examples/flux2/)
    - [Z Image](https://comfyanonymous.github.io/ComfyUI_examples/z_image/)
+   - Ernie 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)
@@ -126,7 +133,7 @@ Workflow examples can be found on the [Examples page](https://comfyanonymous.git
 ComfyUI follows a weekly release cycle targeting Monday but this regularly changes because of model releases or large changes to the codebase. There are three interconnected repositories:
 
 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
-   - Releases a new stable version (e.g., v0.7.0) roughly every week.
+   - Releases a new major stable version (e.g., v0.7.0) roughly every 2 weeks.
    - Starting from v0.4.0 patch versions will be used for fixes backported onto the current stable release.
    - Minor versions will be used for releases off the master branch.
    - Patch versions may still be used for releases on the master branch in cases where a backport would not make sense.

app/frontend_management.py

@@ -27,7 +27,7 @@ def frontend_install_warning_message():
     return f"""
 {get_missing_requirements_message()}
 
-This error is happening because the ComfyUI frontend is no longer shipped as part of the main repo but as a pip package instead.
+The ComfyUI frontend is shipped in a pip package so it needs to be updated separately from the ComfyUI code.
 """.strip()
 
 def parse_version(version: str) -> tuple[int, int, int]:

app/node_replace_manager.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 
+import logging
+
 from aiohttp import web
 
 from typing import TYPE_CHECKING, TypedDict
@@ -31,8 +33,22 @@ class NodeReplaceManager:
         self._replacements: dict[str, list[NodeReplace]] = {}
 
     def register(self, node_replace: NodeReplace):
-        """Register a node replacement mapping."""
-        self._replacements.setdefault(node_replace.old_node_id, []).append(node_replace)
+        """Register a node replacement mapping.
+
+        Idempotent: if a replacement with the same (old_node_id, new_node_id)
+        is already registered, the duplicate is ignored. This prevents stale
+        entries from accumulating when custom nodes are reloaded in the same
+        process (e.g. via ComfyUI-Manager).
+        """
+        existing = self._replacements.setdefault(node_replace.old_node_id, [])
+        for entry in existing:
+            if entry.new_node_id == node_replace.new_node_id:
+                logging.debug(
+                    "Node replacement %s -> %s already registered, ignoring duplicate.",
+                    node_replace.old_node_id, node_replace.new_node_id,
+                )
+                return
+        existing.append(node_replace)
 
     def get_replacement(self, old_node_id: str) -> list[NodeReplace] | None:
         """Get replacements for an old node ID."""

app/user_manager.py

@@ -28,8 +28,8 @@ def get_file_info(path: str, relative_to: str) -> FileInfo:
     return {
         "path": os.path.relpath(path, relative_to).replace(os.sep, '/'),
         "size": os.path.getsize(path),
-        "modified": os.path.getmtime(path),
-        "created": os.path.getctime(path)
+        "modified": int(os.path.getmtime(path) * 1000),
+        "created": int(os.path.getctime(path) * 1000),
     }
 
 

comfy/background_removal/birefnet.json

@@ -0,0 +1,7 @@
+{
+    "model_type": "birefnet",
+    "image_std": [1.0, 1.0, 1.0],
+    "image_mean": [0.0, 0.0, 0.0],
+    "image_size": 1024,
+    "resize_to_original": true
+}

comfy/background_removal/birefnet.py

@@ -0,0 +1,689 @@
+import torch
+import comfy.ops
+import numpy as np
+import torch.nn as nn
+from functools import partial
+import torch.nn.functional as F
+from torchvision.ops import deform_conv2d
+from comfy.ldm.modules.attention import optimized_attention_for_device
+
+CXT = [3072, 1536, 768, 384][1:][::-1][-3:]
+
+class Attention(nn.Module):
+    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, device=None, dtype=None, operations=None):
+        super().__init__()
+
+        self.dim = dim
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.q = operations.Linear(dim, dim, bias=qkv_bias, device=device, dtype=dtype)
+        self.kv = operations.Linear(dim, dim * 2, bias=qkv_bias, device=device, dtype=dtype)
+        self.proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        optimized_attention = optimized_attention_for_device(x.device, mask=False, small_input=True)
+        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
+        kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        k, v = kv[0], kv[1]
+
+        x = optimized_attention(
+            q, k, v, heads=self.num_heads, skip_output_reshape=True, skip_reshape=True
+        ).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+
+        return x
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, device=None, dtype=None, operations=None):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = operations.Linear(in_features, hidden_features, device=device, dtype=dtype)
+        self.act = nn.GELU()
+        self.fc2 = operations.Linear(hidden_features, out_features, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.fc2(x)
+        return x
+
+
+def window_partition(x, window_size):
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, device=None, dtype=None, operations=None):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim ** -0.5
+
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads, device=device, dtype=dtype))
+
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w], indexing='ij'))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, device=device, dtype=dtype)
+        self.proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        B_, N, C = x.shape
+        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.long().view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        return x
+
+
+class SwinTransformerBlock(nn.Module):
+    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, qk_scale=None,
+                 norm_layer=nn.LayerNorm, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+
+        self.norm1 = norm_layer(dim, device=device, dtype=dtype)
+        self.attn = WindowAttention(
+            dim, window_size=(self.window_size, self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, qk_scale=qk_scale, device=device, dtype=dtype, operations=operations)
+
+        self.norm2 = norm_layer(dim, device=device, dtype=dtype)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, device=device, dtype=dtype, operations=operations)
+
+        self.H = None
+        self.W = None
+
+    def forward(self, x, mask_matrix):
+        B, L, C = x.shape
+        H, W = self.H, self.W
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(B, H, W, C)
+
+        pad_l = pad_t = 0
+        pad_r = (self.window_size - W % self.window_size) % self.window_size
+        pad_b = (self.window_size - H % self.window_size) % self.window_size
+        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
+        _, Hp, Wp, _ = x.shape
+
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+            attn_mask = mask_matrix
+        else:
+            shifted_x = x
+            attn_mask = None
+
+        x_windows = window_partition(shifted_x, self.window_size)
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)
+
+        attn_windows = self.attn(x_windows, mask=attn_mask)
+
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C
+
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+
+        if pad_r > 0 or pad_b > 0:
+            x = x[:, :H, :W, :].contiguous()
+
+        x = x.view(B, H * W, C)
+
+        x = shortcut + x
+        x = x + self.mlp(self.norm2(x))
+
+        return x
+
+
+class PatchMerging(nn.Module):
+    def __init__(self, dim, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.dim = dim
+        self.reduction = operations.Linear(4 * dim, 2 * dim, bias=False, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(4 * dim, device=device, dtype=dtype)
+
+    def forward(self, x, H, W):
+        B, L, C = x.shape
+        x = x.view(B, H, W, C)
+
+        # padding
+        pad_input = (H % 2 == 1) or (W % 2 == 1)
+        if pad_input:
+            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    def __init__(self,
+                 dim,
+                 depth,
+                 num_heads,
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.window_size = window_size
+        self.shift_size = window_size // 2
+        self.depth = depth
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                norm_layer=norm_layer,
+                device=device, dtype=dtype, operations=operations)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(dim=dim, device=device, dtype=dtype, operations=operations)
+        else:
+            self.downsample = None
+
+    def forward(self, x, H, W):
+        Hp = int(np.ceil(H / self.window_size)) * self.window_size
+        Wp = int(np.ceil(W / self.window_size)) * self.window_size
+        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        for blk in self.blocks:
+            blk.H, blk.W = H, W
+            x = blk(x, attn_mask)
+        if self.downsample is not None:
+            x_down = self.downsample(x, H, W)
+            Wh, Ww = (H + 1) // 2, (W + 1) // 2
+            return x, H, W, x_down, Wh, Ww
+        else:
+            return x, H, W, x, H, W
+
+
+class PatchEmbed(nn.Module):
+    def __init__(self, patch_size=4, in_channels=3, embed_dim=96, norm_layer=None, device=None, dtype=None, operations=None):
+        super().__init__()
+        patch_size = (patch_size, patch_size)
+        self.patch_size = patch_size
+
+        self.in_channels = in_channels
+        self.embed_dim = embed_dim
+
+        self.proj = operations.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size, device=device, dtype=dtype)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim, device=device, dtype=dtype)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        _, _, H, W = x.size()
+        if W % self.patch_size[1] != 0:
+            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
+        if H % self.patch_size[0] != 0:
+            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))
+
+        x = self.proj(x)  # B C Wh Ww
+        if self.norm is not None:
+            Wh, Ww = x.size(2), x.size(3)
+            x = x.flatten(2).transpose(1, 2)
+            x = self.norm(x)
+            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)
+
+        return x
+
+
+class SwinTransformer(nn.Module):
+    def __init__(self,
+                 pretrain_img_size=224,
+                 patch_size=4,
+                 in_channels=3,
+                 embed_dim=96,
+                 depths=[2, 2, 6, 2],
+                 num_heads=[3, 6, 12, 24],
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 patch_norm=True,
+                 out_indices=(0, 1, 2, 3),
+                 frozen_stages=-1,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+
+        norm_layer = partial(operations.LayerNorm, device=device, dtype=dtype)
+        self.pretrain_img_size = pretrain_img_size
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.patch_norm = patch_norm
+        self.out_indices = out_indices
+        self.frozen_stages = frozen_stages
+
+        self.patch_embed = PatchEmbed(
+            patch_size=patch_size, in_channels=in_channels, embed_dim=embed_dim,
+            device=device, dtype=dtype, operations=operations,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2 ** i_layer),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                norm_layer=norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                device=device, dtype=dtype, operations=operations)
+            self.layers.append(layer)
+
+        num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
+        self.num_features = num_features
+
+        for i_layer in out_indices:
+            layer = norm_layer(num_features[i_layer])
+            layer_name = f'norm{i_layer}'
+            self.add_module(layer_name, layer)
+
+
+    def forward(self, x):
+        x = self.patch_embed(x)
+
+        Wh, Ww = x.size(2), x.size(3)
+
+        outs = []
+        x = x.flatten(2).transpose(1, 2)
+        for i in range(self.num_layers):
+            layer = self.layers[i]
+            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)
+
+            if i in self.out_indices:
+                norm_layer = getattr(self, f'norm{i}')
+                x_out = norm_layer(x_out)
+
+                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
+                outs.append(out)
+
+        return tuple(outs)
+
+class DeformableConv2d(nn.Module):
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size=3,
+                 stride=1,
+                 padding=1,
+                 bias=False, device=None, dtype=None, operations=None):
+
+        super(DeformableConv2d, self).__init__()
+
+        kernel_size = kernel_size if type(kernel_size) is tuple else (kernel_size, kernel_size)
+        self.stride = stride if type(stride) is tuple else (stride, stride)
+        self.padding = padding
+
+        self.offset_conv = operations.Conv2d(in_channels,
+                                     2 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True, device=device, dtype=dtype)
+
+        self.modulator_conv = operations.Conv2d(in_channels,
+                                     1 * kernel_size[0] * kernel_size[1],
+                                     kernel_size=kernel_size,
+                                     stride=stride,
+                                     padding=self.padding,
+                                     bias=True, device=device, dtype=dtype)
+
+        self.regular_conv = operations.Conv2d(in_channels,
+                                      out_channels=out_channels,
+                                      kernel_size=kernel_size,
+                                      stride=stride,
+                                      padding=self.padding,
+                                      bias=bias, device=device, dtype=dtype)
+
+    def forward(self, x):
+        offset = self.offset_conv(x)
+        modulator = 2. * torch.sigmoid(self.modulator_conv(x))
+        weight, bias, offload_info = comfy.ops.cast_bias_weight(self.regular_conv, x, offloadable=True)
+
+        x = deform_conv2d(
+            input=x,
+            offset=offset,
+            weight=weight,
+            bias=None,
+            padding=self.padding,
+            mask=modulator,
+            stride=self.stride,
+        )
+        comfy.ops.uncast_bias_weight(self.regular_conv, weight, bias, offload_info)
+        return x
+
+class BasicDecBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, inter_channels=64, device=None, dtype=None, operations=None):
+        super(BasicDecBlk, self).__init__()
+        inter_channels = 64
+        self.conv_in = operations.Conv2d(in_channels, inter_channels, 3, 1, padding=1, device=device, dtype=dtype)
+        self.relu_in = nn.ReLU(inplace=True)
+        self.dec_att = ASPPDeformable(in_channels=inter_channels, device=device, dtype=dtype, operations=operations)
+        self.conv_out = operations.Conv2d(inter_channels, out_channels, 3, 1, padding=1, device=device, dtype=dtype)
+        self.bn_in = operations.BatchNorm2d(inter_channels, device=device, dtype=dtype)
+        self.bn_out = operations.BatchNorm2d(out_channels, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        x = self.bn_in(x)
+        x = self.relu_in(x)
+        x = self.dec_att(x)
+        x = self.conv_out(x)
+        x = self.bn_out(x)
+        return x
+
+
+class BasicLatBlk(nn.Module):
+    def __init__(self, in_channels=64, out_channels=64, device=None, dtype=None, operations=None):
+        super(BasicLatBlk, self).__init__()
+        self.conv = operations.Conv2d(in_channels, out_channels, 1, 1, 0, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.conv(x)
+        return x
+
+
+class _ASPPModuleDeformable(nn.Module):
+    def __init__(self, in_channels, planes, kernel_size, padding, device, dtype, operations):
+        super(_ASPPModuleDeformable, self).__init__()
+        self.atrous_conv = DeformableConv2d(in_channels, planes, kernel_size=kernel_size,
+                                            stride=1, padding=padding, bias=False, device=device, dtype=dtype, operations=operations)
+        self.bn = operations.BatchNorm2d(planes, device=device, dtype=dtype)
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x = self.atrous_conv(x)
+        x = self.bn(x)
+
+        return self.relu(x)
+
+
+class ASPPDeformable(nn.Module):
+    def __init__(self, in_channels, out_channels=None, parallel_block_sizes=[1, 3, 7], device=None, dtype=None, operations=None):
+        super(ASPPDeformable, self).__init__()
+        self.down_scale = 1
+        if out_channels is None:
+            out_channels = in_channels
+        self.in_channelster = 256 // self.down_scale
+
+        self.aspp1 = _ASPPModuleDeformable(in_channels, self.in_channelster, 1, padding=0, device=device, dtype=dtype, operations=operations)
+        self.aspp_deforms = nn.ModuleList([
+            _ASPPModuleDeformable(in_channels, self.in_channelster, conv_size, padding=int(conv_size//2), device=device, dtype=dtype, operations=operations)
+              for conv_size in parallel_block_sizes
+        ])
+
+        self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)),
+                                             operations.Conv2d(in_channels, self.in_channelster, 1, stride=1, bias=False, device=device, dtype=dtype),
+                                             operations.BatchNorm2d(self.in_channelster, device=device, dtype=dtype),
+                                             nn.ReLU(inplace=True))
+        self.conv1 = operations.Conv2d(self.in_channelster * (2 + len(self.aspp_deforms)), out_channels, 1, bias=False, device=device, dtype=dtype)
+        self.bn1 = operations.BatchNorm2d(out_channels, device=device, dtype=dtype)
+        self.relu = nn.ReLU(inplace=True)
+
+    def forward(self, x):
+        x1 = self.aspp1(x)
+        x_aspp_deforms = [aspp_deform(x) for aspp_deform in self.aspp_deforms]
+        x5 = self.global_avg_pool(x)
+        x5 = F.interpolate(x5, size=x1.size()[2:], mode='bilinear', align_corners=True)
+        x = torch.cat((x1, *x_aspp_deforms, x5), dim=1)
+
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+
+        return x
+
+class BiRefNet(nn.Module):
+    def __init__(self, config=None, dtype=None, device=None, operations=None):
+        super(BiRefNet, self).__init__()
+        self.bb = SwinTransformer(embed_dim=192, depths=[2, 2, 18, 2], num_heads=[6, 12, 24, 48], window_size=12, device=device, dtype=dtype, operations=operations)
+
+        channels = [1536, 768, 384, 192]
+        channels = [c * 2 for c in channels]
+        self.cxt = channels[1:][::-1][-3:]
+        self.squeeze_module = nn.Sequential(*[
+            BasicDecBlk(channels[0]+sum(self.cxt), channels[0], device=device, dtype=dtype, operations=operations)
+            for _ in range(1)
+        ])
+
+        self.decoder = Decoder(channels, device=device, dtype=dtype, operations=operations)
+
+    def forward_enc(self, x):
+        x1, x2, x3, x4 = self.bb(x)
+        B, C, H, W = x.shape
+        x1_, x2_, x3_, x4_ = self.bb(F.interpolate(x, size=(H//2, W//2), mode='bilinear', align_corners=True))
+        x1 = torch.cat([x1, F.interpolate(x1_, size=x1.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+        x2 = torch.cat([x2, F.interpolate(x2_, size=x2.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+        x3 = torch.cat([x3, F.interpolate(x3_, size=x3.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+        x4 = torch.cat([x4, F.interpolate(x4_, size=x4.shape[2:], mode='bilinear', align_corners=True)], dim=1)
+        x4 = torch.cat(
+            (
+                *[
+                    F.interpolate(x1, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                    F.interpolate(x2, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                    F.interpolate(x3, size=x4.shape[2:], mode='bilinear', align_corners=True),
+                ][-len(CXT):],
+                x4
+            ),
+            dim=1
+        )
+        return (x1, x2, x3, x4)
+
+    def forward_ori(self, x):
+        (x1, x2, x3, x4) = self.forward_enc(x)
+        x4 = self.squeeze_module(x4)
+        features = [x, x1, x2, x3, x4]
+        scaled_preds = self.decoder(features)
+        return scaled_preds
+
+    def forward(self, pixel_values, intermediate_output=None):
+        scaled_preds = self.forward_ori(pixel_values)
+        return scaled_preds
+
+
+class Decoder(nn.Module):
+    def __init__(self, channels, device, dtype, operations):
+        super(Decoder, self).__init__()
+        # factory kwargs
+        fk = {"device":device, "dtype":dtype, "operations":operations}
+        DecoderBlock = partial(BasicDecBlk, **fk)
+        LateralBlock = partial(BasicLatBlk, **fk)
+        DBlock = partial(SimpleConvs, **fk)
+
+        self.split = True
+        N_dec_ipt = 64
+        ic = 64
+        ipt_cha_opt = 1
+        self.ipt_blk5 = DBlock(2**10*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+        self.ipt_blk4 = DBlock(2**8*3 if self.split else 3, [N_dec_ipt, channels[0]//8][ipt_cha_opt], inter_channels=ic)
+        self.ipt_blk3 = DBlock(2**6*3 if self.split else 3, [N_dec_ipt, channels[1]//8][ipt_cha_opt], inter_channels=ic)
+        self.ipt_blk2 = DBlock(2**4*3 if self.split else 3, [N_dec_ipt, channels[2]//8][ipt_cha_opt], inter_channels=ic)
+        self.ipt_blk1 = DBlock(2**0*3 if self.split else 3, [N_dec_ipt, channels[3]//8][ipt_cha_opt], inter_channels=ic)
+
+        self.decoder_block4 = DecoderBlock(channels[0]+([N_dec_ipt, channels[0]//8][ipt_cha_opt]), channels[1])
+        self.decoder_block3 = DecoderBlock(channels[1]+([N_dec_ipt, channels[0]//8][ipt_cha_opt]), channels[2])
+        self.decoder_block2 = DecoderBlock(channels[2]+([N_dec_ipt, channels[1]//8][ipt_cha_opt]), channels[3])
+        self.decoder_block1 = DecoderBlock(channels[3]+([N_dec_ipt, channels[2]//8][ipt_cha_opt]), channels[3]//2)
+
+        fk = {"device":device, "dtype":dtype}
+
+        self.conv_out1 = nn.Sequential(operations.Conv2d(channels[3]//2+([N_dec_ipt, channels[3]//8][ipt_cha_opt]), 1, 1, 1, 0, **fk))
+
+        self.lateral_block4 = LateralBlock(channels[1], channels[1])
+        self.lateral_block3 = LateralBlock(channels[2], channels[2])
+        self.lateral_block2 = LateralBlock(channels[3], channels[3])
+
+        self.conv_ms_spvn_4 = operations.Conv2d(channels[1], 1, 1, 1, 0, **fk)
+        self.conv_ms_spvn_3 = operations.Conv2d(channels[2], 1, 1, 1, 0, **fk)
+        self.conv_ms_spvn_2 = operations.Conv2d(channels[3], 1, 1, 1, 0, **fk)
+
+        _N = 16
+
+        self.gdt_convs_4 = nn.Sequential(operations.Conv2d(channels[0] // 2, _N, 3, 1, 1, **fk), operations.BatchNorm2d(_N, **fk), nn.ReLU(inplace=True))
+        self.gdt_convs_3 = nn.Sequential(operations.Conv2d(channels[1] // 2, _N, 3, 1, 1, **fk), operations.BatchNorm2d(_N, **fk), nn.ReLU(inplace=True))
+        self.gdt_convs_2 = nn.Sequential(operations.Conv2d(channels[2] // 2, _N, 3, 1, 1, **fk), operations.BatchNorm2d(_N, **fk), nn.ReLU(inplace=True))
+
+        [setattr(self, f"gdt_convs_pred_{i}", nn.Sequential(operations.Conv2d(_N, 1, 1, 1, 0, **fk))) for i in range(2, 5)]
+        [setattr(self, f"gdt_convs_attn_{i}", nn.Sequential(operations.Conv2d(_N, 1, 1, 1, 0, **fk))) for i in range(2, 5)]
+
+    def get_patches_batch(self, x, p):
+        _size_h, _size_w = p.shape[2:]
+        patches_batch = []
+        for idx in range(x.shape[0]):
+            columns_x = torch.split(x[idx], split_size_or_sections=_size_w, dim=-1)
+            patches_x = []
+            for column_x in columns_x:
+                patches_x += [p.unsqueeze(0) for p in torch.split(column_x, split_size_or_sections=_size_h, dim=-2)]
+            patch_sample = torch.cat(patches_x, dim=1)
+            patches_batch.append(patch_sample)
+        return torch.cat(patches_batch, dim=0)
+
+    def forward(self, features):
+        x, x1, x2, x3, x4 = features
+
+        patches_batch = self.get_patches_batch(x, x4) if self.split else x
+        x4 = torch.cat((x4, self.ipt_blk5(F.interpolate(patches_batch, size=x4.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p4 = self.decoder_block4(x4)
+        p4_gdt = self.gdt_convs_4(p4)
+        gdt_attn_4 = self.gdt_convs_attn_4(p4_gdt).sigmoid()
+        p4 = p4 * gdt_attn_4
+        _p4 = F.interpolate(p4, size=x3.shape[2:], mode='bilinear', align_corners=True)
+        _p3 = _p4 + self.lateral_block4(x3)
+
+        patches_batch = self.get_patches_batch(x, _p3) if self.split else x
+        _p3 = torch.cat((_p3, self.ipt_blk4(F.interpolate(patches_batch, size=x3.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p3 = self.decoder_block3(_p3)
+
+        p3_gdt = self.gdt_convs_3(p3)
+        gdt_attn_3 = self.gdt_convs_attn_3(p3_gdt).sigmoid()
+        p3 = p3 * gdt_attn_3
+        _p3 = F.interpolate(p3, size=x2.shape[2:], mode='bilinear', align_corners=True)
+        _p2 = _p3 + self.lateral_block3(x2)
+
+        patches_batch = self.get_patches_batch(x, _p2) if self.split else x
+        _p2 = torch.cat((_p2, self.ipt_blk3(F.interpolate(patches_batch, size=x2.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p2 = self.decoder_block2(_p2)
+
+        p2_gdt = self.gdt_convs_2(p2)
+        gdt_attn_2 = self.gdt_convs_attn_2(p2_gdt).sigmoid()
+        p2 = p2 * gdt_attn_2
+
+        _p2 = F.interpolate(p2, size=x1.shape[2:], mode='bilinear', align_corners=True)
+        _p1 = _p2 + self.lateral_block2(x1)
+
+        patches_batch = self.get_patches_batch(x, _p1) if self.split else x
+        _p1 = torch.cat((_p1, self.ipt_blk2(F.interpolate(patches_batch, size=x1.shape[2:], mode='bilinear', align_corners=True))), 1)
+        _p1 = self.decoder_block1(_p1)
+        _p1 = F.interpolate(_p1, size=x.shape[2:], mode='bilinear', align_corners=True)
+
+        patches_batch = self.get_patches_batch(x, _p1) if self.split else x
+        _p1 = torch.cat((_p1, self.ipt_blk1(F.interpolate(patches_batch, size=x.shape[2:], mode='bilinear', align_corners=True))), 1)
+        p1_out = self.conv_out1(_p1)
+        return p1_out
+
+
+class SimpleConvs(nn.Module):
+    def __init__(
+        self, in_channels: int, out_channels: int, inter_channels=64, device=None, dtype=None, operations=None
+    ) -> None:
+        super().__init__()
+        self.conv1 = operations.Conv2d(in_channels, inter_channels, 3, 1, 1, device=device, dtype=dtype)
+        self.conv_out = operations.Conv2d(inter_channels, out_channels, 3, 1, 1, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.conv_out(self.conv1(x))

comfy/bg_removal_model.py

@@ -0,0 +1,78 @@
+from .utils import load_torch_file
+import os
+import json
+import torch
+import logging
+
+import comfy.ops
+import comfy.model_patcher
+import comfy.model_management
+import comfy.clip_model
+import comfy.background_removal.birefnet
+
+BG_REMOVAL_MODELS = {
+    "birefnet": comfy.background_removal.birefnet.BiRefNet
+}
+
+class BackgroundRemovalModel():
+    def __init__(self, json_config):
+        with open(json_config) as f:
+            config = json.load(f)
+
+        self.image_size = config.get("image_size", 1024)
+        self.image_mean = config.get("image_mean", [0.0, 0.0, 0.0])
+        self.image_std = config.get("image_std", [1.0, 1.0, 1.0])
+        self.model_type = config.get("model_type", "birefnet")
+        self.config = config.copy()
+        model_class = BG_REMOVAL_MODELS.get(self.model_type)
+
+        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 = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast)
+        self.model.eval()
+
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
+
+    def get_sd(self):
+        return self.model.state_dict()
+
+    def encode_image(self, image):
+        comfy.model_management.load_model_gpu(self.patcher)
+        H, W = image.shape[1], image.shape[2]
+        pixel_values = comfy.clip_model.clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=False)
+        out = self.model(pixel_values=pixel_values)
+        out = torch.nn.functional.interpolate(out, size=(H, W), mode="bicubic", antialias=False)
+
+        mask = out.sigmoid().to(device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype())
+        if mask.ndim == 3:
+            mask = mask.unsqueeze(0)
+        if mask.shape[1] != 1:
+            mask = mask.movedim(-1, 1)
+
+        return mask
+
+
+def load_background_removal_model(sd):
+    if "bb.layers.1.blocks.0.attn.relative_position_index" in sd:
+        json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "background_removal"), "birefnet.json")
+    else:
+        return None
+
+    bg_model = BackgroundRemovalModel(json_config)
+    m, u = bg_model.load_sd(sd)
+    if len(m) > 0:
+        logging.warning("missing background removal: {}".format(m))
+    u = set(u)
+    keys = list(sd.keys())
+    for k in keys:
+        if k not in u:
+            sd.pop(k)
+    return bg_model
+
+def load(ckpt_path):
+    sd = load_torch_file(ckpt_path)
+    return load_background_removal_model(sd)

comfy/cli_args.py

@@ -91,6 +91,7 @@ parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE"
 
 parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
 parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
+parser.add_argument("--enable-triton-backend", action="store_true", help="ComfyUI will enable the use of Triton backend in comfy-kitchen. Is disabled at launch by default.")
 
 class LatentPreviewMethod(enum.Enum):
     NoPreviews = "none"
@@ -237,6 +238,8 @@ database_default_path = os.path.abspath(
 )
 parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.")
 parser.add_argument("--enable-assets", action="store_true", help="Enable the assets system (API routes, database synchronization, and background scanning).")
+parser.add_argument("--feature-flag", type=str, action='append', default=[], metavar="KEY[=VALUE]", help="Set a server feature flag. Use KEY=VALUE to set an explicit value, or bare KEY to set it to true. Can be specified multiple times. Boolean values (true/false) and numbers are auto-converted. Examples: --feature-flag show_signin_button=true  or  --feature-flag show_signin_button")
+parser.add_argument("--list-feature-flags", action="store_true", help="Print the registry of known CLI-settable feature flags as JSON and exit.")
 
 if comfy.options.args_parsing:
     args = parser.parse_args()

comfy/context_windows.py

@@ -63,7 +63,11 @@ class IndexListContextWindow(ContextWindowABC):
             dim = self.dim
         if dim == 0 and full.shape[dim] == 1:
             return full
-        idx = tuple([slice(None)] * dim + [self.index_list])
+        indices = self.index_list
+        anchor_idx = getattr(self, 'causal_anchor_index', None)
+        if anchor_idx is not None and anchor_idx >= 0:
+            indices = [anchor_idx] + list(indices)
+        idx = tuple([slice(None)] * dim + [indices])
         window = full[idx]
         if retain_index_list:
             idx = tuple([slice(None)] * dim + [retain_index_list])
@@ -113,7 +117,14 @@ def slice_cond(cond_value, window: IndexListContextWindow, x_in: torch.Tensor, d
 
     # skip leading latent positions that have no corresponding conditioning (e.g. reference frames)
     if temporal_offset > 0:
-        indices = [i - temporal_offset for i in window.index_list[temporal_offset:]]
+        anchor_idx = getattr(window, 'causal_anchor_index', None)
+        if anchor_idx is not None and anchor_idx >= 0:
+            # anchor occupies one of the no-cond positions, so skip one fewer from window.index_list
+            skip_count = temporal_offset - 1
+        else:
+            skip_count = temporal_offset
+
+        indices = [i - temporal_offset for i in window.index_list[skip_count:]]
         indices = [i for i in indices if 0 <= i]
     else:
         indices = list(window.index_list)
@@ -150,7 +161,8 @@ class ContextFuseMethod:
 ContextResults = collections.namedtuple("ContextResults", ['window_idx', 'sub_conds_out', 'sub_conds', 'window'])
 class IndexListContextHandler(ContextHandlerABC):
     def __init__(self, context_schedule: ContextSchedule, fuse_method: ContextFuseMethod, context_length: int=1, context_overlap: int=0, context_stride: int=1,
-                 closed_loop: bool=False, dim:int=0, freenoise: bool=False, cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False):
+                 closed_loop: bool=False, dim:int=0, freenoise: bool=False, cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False,
+                 causal_window_fix: bool=True):
         self.context_schedule = context_schedule
         self.fuse_method = fuse_method
         self.context_length = context_length
@@ -162,6 +174,7 @@ class IndexListContextHandler(ContextHandlerABC):
         self.freenoise = freenoise
         self.cond_retain_index_list = [int(x.strip()) for x in cond_retain_index_list.split(",")] if cond_retain_index_list else []
         self.split_conds_to_windows = split_conds_to_windows
+        self.causal_window_fix = causal_window_fix
 
         self.callbacks = {}
 
@@ -318,6 +331,14 @@ class IndexListContextHandler(ContextHandlerABC):
             # allow processing to end between context window executions for faster Cancel
             comfy.model_management.throw_exception_if_processing_interrupted()
 
+            # causal_window_fix: prepend a pre-window frame that will be stripped post-forward
+            anchor_applied = False
+            if self.causal_window_fix:
+                anchor_idx = window.index_list[0] - 1
+                if 0 <= anchor_idx < x_in.size(self.dim):
+                    window.causal_anchor_index = anchor_idx
+                    anchor_applied = True
+
             for callback in comfy.patcher_extension.get_all_callbacks(IndexListCallbacks.EVALUATE_CONTEXT_WINDOWS, self.callbacks):
                 callback(self, model, x_in, conds, timestep, model_options, window_idx, window, model_options, device, first_device)
 
@@ -332,6 +353,12 @@ class IndexListContextHandler(ContextHandlerABC):
             if device is not None:
                 for i in range(len(sub_conds_out)):
                     sub_conds_out[i] = sub_conds_out[i].to(x_in.device)
+
+            # strip causal_window_fix anchor if applied
+            if anchor_applied:
+                for i in range(len(sub_conds_out)):
+                    sub_conds_out[i] = sub_conds_out[i].narrow(self.dim, 1, sub_conds_out[i].shape[self.dim] - 1)
+
             results.append(ContextResults(window_idx, sub_conds_out, sub_conds, window))
         return results
 

comfy/deploy_environment.py

@@ -0,0 +1,34 @@
+import functools
+import logging
+import os
+
+logger = logging.getLogger(__name__)
+
+_DEFAULT_DEPLOY_ENV = "local-git"
+_ENV_FILENAME = ".comfy_environment"
+
+# Resolve the ComfyUI install directory (the parent of this `comfy/` package).
+# We deliberately avoid `folder_paths.base_path` here because that is overridden
+# by the `--base-directory` CLI arg to a user-supplied path, whereas the
+# `.comfy_environment` marker is written by launchers/installers next to the
+# ComfyUI install itself.
+_COMFY_INSTALL_DIR = os.path.dirname(os.path.dirname(os.path.realpath(__file__)))
+
+
+@functools.cache
+def get_deploy_environment() -> str:
+    env_file = os.path.join(_COMFY_INSTALL_DIR, _ENV_FILENAME)
+    try:
+        with open(env_file, encoding="utf-8") as f:
+            # Cap the read so a malformed or maliciously crafted file (e.g.
+            # a single huge line with no newline) can't blow up memory.
+            first_line = f.readline(128).strip()
+            value = "".join(c for c in first_line if 32 <= ord(c) < 127)
+            if value:
+                return value
+    except FileNotFoundError:
+        pass
+    except Exception as e:
+        logger.error("Failed to read %s: %s", env_file, e)
+
+    return _DEFAULT_DEPLOY_ENV

comfy/hooks.py

@@ -93,7 +93,7 @@ class Hook:
         self.hook_scope = hook_scope
         '''Scope of where this hook should apply in terms of the conds used in sampling run.'''
         self.custom_should_register = default_should_register
-        '''Can be overriden with a compatible function to decide if this hook should be registered without the need to override .should_register'''
+        '''Can be overridden with a compatible function to decide if this hook should be registered without the need to override .should_register'''
 
     @property
     def strength(self):

comfy/k_diffusion/sampling.py

@@ -1810,3 +1810,119 @@ def sample_sa_solver(model, x, sigmas, extra_args=None, callback=None, disable=F
 def sample_sa_solver_pece(model, x, sigmas, extra_args=None, callback=None, disable=False, tau_func=None, s_noise=1.0, noise_sampler=None, predictor_order=3, corrector_order=4, simple_order_2=False):
     """Stochastic Adams Solver with PECE (Predict–Evaluate–Correct–Evaluate) mode (NeurIPS 2023)."""
     return sample_sa_solver(model, x, sigmas, extra_args=extra_args, callback=callback, disable=disable, tau_func=tau_func, s_noise=s_noise, noise_sampler=noise_sampler, predictor_order=predictor_order, corrector_order=corrector_order, use_pece=True, simple_order_2=simple_order_2)
+
+
+@torch.no_grad()
+def sample_ar_video(model, x, sigmas, extra_args=None, callback=None, disable=None,
+                    num_frame_per_block=1):
+    """
+    Autoregressive video sampler: block-by-block denoising with KV cache
+    and flow-match re-noising for Causal Forcing / Self-Forcing models.
+
+    Requires a Causal-WAN compatible model (diffusion_model must expose
+    init_kv_caches / init_crossattn_caches) and 5-D latents [B,C,T,H,W].
+
+    All AR-loop parameters are passed via the SamplerARVideo node, not read
+    from the checkpoint or transformer_options.
+    """
+    extra_args = {} if extra_args is None else extra_args
+    model_options = extra_args.get("model_options", {})
+    transformer_options = model_options.get("transformer_options", {})
+
+    if x.ndim != 5:
+        raise ValueError(
+            f"ar_video sampler requires 5-D video latents [B,C,T,H,W], got {x.ndim}-D tensor with shape {x.shape}. "
+            "This sampler is only compatible with autoregressive video models (e.g. Causal-WAN)."
+        )
+
+    inner_model = model.inner_model.inner_model
+    causal_model = inner_model.diffusion_model
+
+    if not (hasattr(causal_model, "init_kv_caches") and hasattr(causal_model, "init_crossattn_caches")):
+        raise TypeError(
+            "ar_video sampler requires a Causal-WAN compatible model whose diffusion_model "
+            "exposes init_kv_caches() and init_crossattn_caches(). The loaded checkpoint "
+            "does not support this interface — choose a different sampler."
+        )
+
+    seed = extra_args.get("seed", 0)
+
+    bs, c, lat_t, lat_h, lat_w = x.shape
+    frame_seq_len = -(-lat_h // 2) * -(-lat_w // 2) # ceiling division
+    num_blocks = -(-lat_t // num_frame_per_block)   # ceiling division
+    device = x.device
+    model_dtype = inner_model.get_dtype()
+
+    kv_caches = causal_model.init_kv_caches(bs, lat_t * frame_seq_len, device, model_dtype)
+    crossattn_caches = causal_model.init_crossattn_caches(bs, device, model_dtype)
+
+    output = torch.zeros_like(x)
+    s_in = x.new_ones([x.shape[0]])
+    current_start_frame = 0
+
+    # I2V: seed KV cache with the initial image latent before the denoising loop
+    initial_latent = transformer_options.get("ar_config", {}).get("initial_latent", None)
+    if initial_latent is not None:
+        initial_latent = inner_model.process_latent_in(initial_latent).to(device=device, dtype=model_dtype)
+        n_init = initial_latent.shape[2]
+        output[:, :, :n_init] = initial_latent
+
+        ar_state = {"start_frame": 0, "kv_caches": kv_caches, "crossattn_caches": crossattn_caches}
+        transformer_options["ar_state"] = ar_state
+        zero_sigma = sigmas.new_zeros([1])
+        _ = model(initial_latent, zero_sigma * s_in, **extra_args)
+
+        current_start_frame = n_init
+        remaining = lat_t - n_init
+        num_blocks = -(-remaining // num_frame_per_block)
+
+    num_sigma_steps = len(sigmas) - 1
+    total_real_steps = num_blocks * num_sigma_steps
+    step_count = 0
+
+    try:
+        for block_idx in trange(num_blocks, disable=disable):
+            bf = min(num_frame_per_block, lat_t - current_start_frame)
+            fs, fe = current_start_frame, current_start_frame + bf
+            noisy_input = x[:, :, fs:fe]
+
+            ar_state = {
+                "start_frame": current_start_frame,
+                "kv_caches": kv_caches,
+                "crossattn_caches": crossattn_caches,
+            }
+            transformer_options["ar_state"] = ar_state
+
+            for i in range(num_sigma_steps):
+                denoised = model(noisy_input, sigmas[i] * s_in, **extra_args)
+
+                if callback is not None:
+                    scaled_i = step_count * num_sigma_steps // total_real_steps
+                    callback({"x": noisy_input, "i": scaled_i, "sigma": sigmas[i],
+                              "sigma_hat": sigmas[i], "denoised": denoised})
+
+                if sigmas[i + 1] == 0:
+                    noisy_input = denoised
+                else:
+                    sigma_next = sigmas[i + 1]
+                    torch.manual_seed(seed + block_idx * 1000 + i)
+                    fresh_noise = torch.randn_like(denoised)
+                    noisy_input = (1.0 - sigma_next) * denoised + sigma_next * fresh_noise
+
+                    for cache in kv_caches:
+                        cache["end"] -= bf * frame_seq_len
+
+                step_count += 1
+
+            output[:, :, fs:fe] = noisy_input
+
+            for cache in kv_caches:
+                cache["end"] -= bf * frame_seq_len
+            zero_sigma = sigmas.new_zeros([1])
+            _ = model(noisy_input, zero_sigma * s_in, **extra_args)
+
+            current_start_frame += bf
+    finally:
+        transformer_options.pop("ar_state", None)
+
+    return output

comfy/latent_formats.py

@@ -9,6 +9,7 @@ class LatentFormat:
     latent_rgb_factors_reshape = None
     taesd_decoder_name = None
     spacial_downscale_ratio = 8
+    temporal_downscale_ratio = 1
 
     def process_in(self, latent):
         return latent * self.scale_factor
@@ -235,6 +236,7 @@ class Flux2(LatentFormat):
 class Mochi(LatentFormat):
     latent_channels = 12
     latent_dimensions = 3
+    temporal_downscale_ratio = 6
 
     def __init__(self):
         self.scale_factor = 1.0
@@ -278,6 +280,7 @@ class LTXV(LatentFormat):
     latent_channels = 128
     latent_dimensions = 3
     spacial_downscale_ratio = 32
+    temporal_downscale_ratio = 8
 
     def __init__(self):
         self.latent_rgb_factors = [
@@ -421,6 +424,7 @@ class LTXAV(LTXV):
 class HunyuanVideo(LatentFormat):
     latent_channels = 16
     latent_dimensions = 3
+    temporal_downscale_ratio = 4
     scale_factor = 0.476986
     latent_rgb_factors = [
         [-0.0395, -0.0331,  0.0445],
@@ -447,6 +451,7 @@ class HunyuanVideo(LatentFormat):
 class Cosmos1CV8x8x8(LatentFormat):
     latent_channels = 16
     latent_dimensions = 3
+    temporal_downscale_ratio = 8
 
     latent_rgb_factors = [
         [ 0.1817,  0.2284,  0.2423],
@@ -472,6 +477,7 @@ class Cosmos1CV8x8x8(LatentFormat):
 class Wan21(LatentFormat):
     latent_channels = 16
     latent_dimensions = 3
+    temporal_downscale_ratio = 4
 
     latent_rgb_factors = [
             [-0.1299, -0.1692,  0.2932],
@@ -734,6 +740,7 @@ class HunyuanVideo15(LatentFormat):
     latent_channels = 32
     latent_dimensions = 3
     spacial_downscale_ratio = 16
+    temporal_downscale_ratio = 4
     scale_factor = 1.03682
     taesd_decoder_name = "lighttaehy1_5"
 
@@ -786,8 +793,27 @@ class ZImagePixelSpace(ChromaRadiance):
     pass
 
 class CogVideoX(LatentFormat):
+    """Latent format for CogVideoX-2b (THUDM/CogVideoX-2b).
+
+    scale_factor matches the vae/config.json scaling_factor for the 2b variant.
+    The 5b-class checkpoints (CogVideoX-5b, CogVideoX-1.5-5B, CogVideoX-Fun-V1.5-*)
+    use a different value; see CogVideoX1_5 below.
+    """
     latent_channels = 16
     latent_dimensions = 3
+    temporal_downscale_ratio = 4
 
     def __init__(self):
         self.scale_factor = 1.15258426
+
+
+class CogVideoX1_5(CogVideoX):
+    """Latent format for 5b-class CogVideoX checkpoints.
+
+    Covers THUDM/CogVideoX-5b, THUDM/CogVideoX-1.5-5B, and the CogVideoX-Fun
+    V1.5-5b family (including VOID inpainting). All of these have
+    scaling_factor=0.7 in their vae/config.json. Auto-selected in
+    supported_models.CogVideoX_T2V based on transformer hidden dim.
+    """
+    def __init__(self):
+        self.scale_factor = 0.7

comfy/ldm/lightricks/av_model.py

@@ -16,6 +16,7 @@ from comfy.ldm.lightricks.model import (
 from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
 from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
 import comfy.ldm.common_dit
+import comfy.model_prefetch
 
 class CompressedTimestep:
     """Store video timestep embeddings in compressed form using per-frame indexing."""
@@ -907,9 +908,11 @@ class LTXAVModel(LTXVModel):
         """Process transformer blocks for LTXAV."""
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        prefetch_queue = comfy.model_prefetch.make_prefetch_queue(list(self.transformer_blocks), vx.device, transformer_options)
 
         # Process transformer blocks
         for i, block in enumerate(self.transformer_blocks):
+            comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, block)
             if ("double_block", i) in blocks_replace:
 
                 def block_wrap(args):
@@ -982,6 +985,8 @@ class LTXAVModel(LTXVModel):
                     a_prompt_timestep=a_prompt_timestep,
                 )
 
+        comfy.model_prefetch.prefetch_queue_pop(prefetch_queue, vx.device, None)
+
         return [vx, ax]
 
     def _process_output(self, x, embedded_timestep, keyframe_idxs, **kwargs):

comfy/ldm/modules/attention.py

@@ -14,6 +14,8 @@ from .sub_quadratic_attention import efficient_dot_product_attention
 
 from comfy import model_management
 
+TORCH_HAS_GQA = model_management.torch_version_numeric >= (2, 5)
+
 if model_management.xformers_enabled():
     import xformers
     import xformers.ops
@@ -150,7 +152,12 @@ def attention_basic(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
         b, _, dim_head = q.shape
         dim_head //= heads
 
-    scale = dim_head ** -0.5
+    if kwargs.get("enable_gqa", False) and q.shape[-3] != k.shape[-3]:
+        n_rep = q.shape[-3] // k.shape[-3]
+        k = k.repeat_interleave(n_rep, dim=-3)
+        v = v.repeat_interleave(n_rep, dim=-3)
+
+    scale = kwargs.get("scale", dim_head ** -0.5)
 
     h = heads
     if skip_reshape:
@@ -219,6 +226,10 @@ def attention_sub_quad(query, key, value, heads, mask=None, attn_precision=None,
         b, _, dim_head = query.shape
         dim_head //= heads
 
+    if "scale" in kwargs:
+        # Pre-scale query to match requested scale (cancels internal 1/sqrt(dim_head))
+        query = query * (kwargs["scale"] * dim_head ** 0.5)
+
     if skip_reshape:
         query = query.reshape(b * heads, -1, dim_head)
         value = value.reshape(b * heads, -1, dim_head)
@@ -290,7 +301,7 @@ def attention_split(q, k, v, heads, mask=None, attn_precision=None, skip_reshape
         b, _, dim_head = q.shape
         dim_head //= heads
 
-    scale = dim_head ** -0.5
+    scale = kwargs.get("scale", dim_head ** -0.5)
 
     if skip_reshape:
          q, k, v = map(
@@ -500,8 +511,13 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
         if mask.ndim == 3:
             mask = mask.unsqueeze(1)
 
+    # Pass through extra SDPA kwargs (scale, enable_gqa) if provided
+    # enable_gqa requires PyTorch 2.5+; older versions use manual KV expansion above
+    sdpa_keys = ("scale", "enable_gqa") if TORCH_HAS_GQA else ("scale",)
+    sdpa_extra = {k: v for k, v in kwargs.items() if k in sdpa_keys}
+
     if SDP_BATCH_LIMIT >= b:
-        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
+        out = comfy.ops.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False, **sdpa_extra)
         if not skip_output_reshape:
             out = (
                 out.transpose(1, 2).reshape(b, -1, heads * dim_head)
@@ -519,7 +535,7 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
                 k[i : i + SDP_BATCH_LIMIT],
                 v[i : i + SDP_BATCH_LIMIT],
                 attn_mask=m,
-                dropout_p=0.0, is_causal=False
+                dropout_p=0.0, is_causal=False, **sdpa_extra
             ).transpose(1, 2).reshape(-1, q.shape[2], heads * dim_head)
     return out
 

comfy/ldm/modules/diffusionmodules/util.py

@@ -140,7 +140,7 @@ def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True):
     alphas = alphacums[ddim_timesteps]
     alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist())
 
-    # according the the formula provided in https://arxiv.org/abs/2010.02502
+    # according to the formula provided in https://arxiv.org/abs/2010.02502
     sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev))
     if verbose:
         logging.info(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}')

comfy/ldm/sam3/detector.py

@@ -561,7 +561,8 @@ class SAM3Model(nn.Module):
         return high_res_masks
 
     def forward_video(self, images, initial_masks, pbar=None, text_prompts=None,
-                       new_det_thresh=0.5, max_objects=0, detect_interval=1):
+                       new_det_thresh=0.5, max_objects=0, detect_interval=1,
+                       target_device=None, target_dtype=None):
         """Track video with optional per-frame text-prompted detection."""
         bb = self.detector.backbone["vision_backbone"]
 
@@ -589,8 +590,10 @@ class SAM3Model(nn.Module):
             return self.tracker.track_video_with_detection(
                 backbone_fn, images, initial_masks, detect_fn,
                 new_det_thresh=new_det_thresh, max_objects=max_objects,
-                detect_interval=detect_interval, backbone_obj=bb, pbar=pbar)
+                detect_interval=detect_interval, backbone_obj=bb, pbar=pbar,
+                target_device=target_device, target_dtype=target_dtype)
         # SAM3 (non-multiplex) — no detection support, requires initial masks
         if initial_masks is None:
             raise ValueError("SAM3 (non-multiplex) requires initial_mask for video tracking")
-        return self.tracker.track_video(backbone_fn, images, initial_masks, pbar=pbar, backbone_obj=bb)
+        return self.tracker.track_video(backbone_fn, images, initial_masks, pbar=pbar, backbone_obj=bb,
+                                         target_device=target_device, target_dtype=target_dtype)

comfy/ldm/sam3/tracker.py

@@ -200,8 +200,13 @@ def pack_masks(masks):
 
 def unpack_masks(packed):
     """Unpack bit-packed [*, H, W//8] uint8 to bool [*, H, W*8]."""
-    shifts = torch.arange(8, device=packed.device)
-    return ((packed.unsqueeze(-1) >> shifts) & 1).view(*packed.shape[:-1], -1).bool()
+    bits = torch.tensor([1, 2, 4, 8, 16, 32, 64, 128], dtype=torch.uint8, device=packed.device)
+    return (packed.unsqueeze(-1) & bits).bool().view(*packed.shape[:-1], -1)
+
+
+def _prep_frame(images, idx, device, dt, size):
+    """Slice CPU full-res frames, transfer to GPU in target dtype, and resize to (size, size)."""
+    return comfy.utils.common_upscale(images[idx].to(device=device, dtype=dt), size, size, "bicubic", crop="disabled")
 
 
 def _compute_backbone(backbone_fn, frame, frame_idx=None):
@@ -1078,16 +1083,19 @@ class SAM3Tracker(nn.Module):
         # SAM3: drop last FPN level
         return vision_feats[:-1], vision_pos[:-1], feat_sizes[:-1]
 
-    def _track_single_object(self, backbone_fn, images, initial_mask, pbar=None):
+    def _track_single_object(self, backbone_fn, images, initial_mask, pbar=None,
+                             target_device=None, target_dtype=None):
         """Track one object, computing backbone per frame to save VRAM."""
         N = images.shape[0]
-        device, dt = images.device, images.dtype
+        device = target_device if target_device is not None else images.device
+        dt = target_dtype if target_dtype is not None else images.dtype
+        size = self.image_size
         output_dict = {"cond_frame_outputs": {}, "non_cond_frame_outputs": {}}
         all_masks = []
 
         for frame_idx in tqdm(range(N), desc="tracking"):
             vision_feats, vision_pos, feat_sizes = self._compute_backbone_frame(
-                backbone_fn, images[frame_idx:frame_idx + 1], frame_idx=frame_idx)
+                backbone_fn, _prep_frame(images, slice(frame_idx, frame_idx + 1), device, dt, size), frame_idx=frame_idx)
             mask_input = None
             if frame_idx == 0:
                 mask_input = F.interpolate(initial_mask.to(device=device, dtype=dt),
@@ -1114,12 +1122,13 @@ class SAM3Tracker(nn.Module):
 
         return torch.cat(all_masks, dim=0)  # [N, 1, H, W]
 
-    def track_video(self, backbone_fn, images, initial_masks, pbar=None, **kwargs):
+    def track_video(self, backbone_fn, images, initial_masks, pbar=None,
+                    target_device=None, target_dtype=None, **kwargs):
         """Track one or more objects across video frames.
 
         Args:
             backbone_fn: callable that returns (sam2_features, sam2_positions, trunk_out) for a frame
-            images: [N, 3, 1008, 1008] video frames
+            images: [N, 3, H, W] CPU full-res video frames (resized per-frame to self.image_size)
             initial_masks: [N_obj, 1, H, W] binary masks for first frame (one per object)
             pbar: optional progress bar
 
@@ -1130,7 +1139,8 @@ class SAM3Tracker(nn.Module):
         per_object = []
         for obj_idx in range(N_obj):
             obj_masks = self._track_single_object(
-                backbone_fn, images, initial_masks[obj_idx:obj_idx + 1], pbar=pbar)
+                backbone_fn, images, initial_masks[obj_idx:obj_idx + 1], pbar=pbar,
+                target_device=target_device, target_dtype=target_dtype)
             per_object.append(obj_masks)
 
         return torch.cat(per_object, dim=1)  # [N, N_obj, H, W]
@@ -1632,11 +1642,18 @@ class SAM31Tracker(nn.Module):
             return det_scores[new_dets].tolist() if det_scores is not None else [0.0] * new_dets.sum().item()
         return []
 
+    INTERNAL_MAX_OBJECTS = 64  # Hard ceiling on accumulated tracks; max_objects=0 or any value above this is clamped here.
+
     def track_video_with_detection(self, backbone_fn, images, initial_masks, detect_fn=None,
                                    new_det_thresh=0.5, max_objects=0, detect_interval=1,
-                                   backbone_obj=None, pbar=None):
+                                   backbone_obj=None, pbar=None, target_device=None, target_dtype=None):
         """Track with optional per-frame detection. Returns [N, max_N_obj, H, W] mask logits."""
-        N, device, dt = images.shape[0], images.device, images.dtype
+        if max_objects <= 0 or max_objects > self.INTERNAL_MAX_OBJECTS:
+            max_objects = self.INTERNAL_MAX_OBJECTS
+        N = images.shape[0]
+        device = target_device if target_device is not None else images.device
+        dt = target_dtype if target_dtype is not None else images.dtype
+        size = self.image_size
         output_dict = {"cond_frame_outputs": {}, "non_cond_frame_outputs": {}}
         all_masks = []
         idev = comfy.model_management.intermediate_device()
@@ -1656,7 +1673,7 @@ class SAM31Tracker(nn.Module):
                 prefetch = True
             except RuntimeError:
                 pass
-        cur_bb = self._compute_backbone_frame(backbone_fn, images[0:1], frame_idx=0)
+        cur_bb = self._compute_backbone_frame(backbone_fn, _prep_frame(images, slice(0, 1), device, dt, size), frame_idx=0)
 
         for frame_idx in tqdm(range(N), desc="tracking"):
             vision_feats, vision_pos, feat_sizes, high_res_prop, trunk_out = cur_bb
@@ -1666,7 +1683,7 @@ class SAM31Tracker(nn.Module):
                 backbone_stream.wait_stream(torch.cuda.current_stream(device))
                 with torch.cuda.stream(backbone_stream):
                     next_bb = self._compute_backbone_frame(
-                        backbone_fn, images[frame_idx + 1:frame_idx + 2], frame_idx=frame_idx + 1)
+                        backbone_fn, _prep_frame(images, slice(frame_idx + 1, frame_idx + 2), device, dt, size), frame_idx=frame_idx + 1)
 
             # Per-frame detection with NMS (skip if no detect_fn, or interval/max not met)
             det_masks = torch.empty(0, device=device)
@@ -1687,7 +1704,7 @@ class SAM31Tracker(nn.Module):
                 current_out = self._condition_with_masks(
                     initial_masks.to(device=device, dtype=dt), frame_idx, vision_feats, vision_pos,
                     feat_sizes, high_res_prop, output_dict, N, mux_state, backbone_obj,
-                    images[frame_idx:frame_idx + 1], trunk_out)
+                    _prep_frame(images, slice(frame_idx, frame_idx + 1), device, dt, size), trunk_out)
                 last_occluded = torch.full((mux_state.total_valid_entries,), -1, device=device, dtype=torch.long)
                 obj_scores = [1.0] * mux_state.total_valid_entries
                 if keep_alive is not None:
@@ -1702,7 +1719,7 @@ class SAM31Tracker(nn.Module):
                     current_out = self._condition_with_masks(
                         det_masks, frame_idx, vision_feats, vision_pos, feat_sizes, high_res_prop,
                         output_dict, N, mux_state, backbone_obj,
-                        images[frame_idx:frame_idx + 1], trunk_out, threshold=0.0)
+                        _prep_frame(images, slice(frame_idx, frame_idx + 1), device, dt, size), trunk_out, threshold=0.0)
                     last_occluded = torch.full((mux_state.total_valid_entries,), -1, device=device, dtype=torch.long)
                     obj_scores = det_scores[:mux_state.total_valid_entries].tolist()
                     if keep_alive is not None:
@@ -1718,7 +1735,7 @@ class SAM31Tracker(nn.Module):
                             torch.cuda.current_stream(device).wait_stream(backbone_stream)
                             cur_bb = next_bb
                         else:
-                            cur_bb = self._compute_backbone_frame(backbone_fn, images[frame_idx + 1:frame_idx + 2], frame_idx=frame_idx + 1)
+                            cur_bb = self._compute_backbone_frame(backbone_fn, _prep_frame(images, slice(frame_idx + 1, frame_idx + 2), device, dt, size), frame_idx=frame_idx + 1)
                     continue
             else:
                 N_obj = mux_state.total_valid_entries
@@ -1768,7 +1785,7 @@ class SAM31Tracker(nn.Module):
                     torch.cuda.current_stream(device).wait_stream(backbone_stream)
                     cur_bb = next_bb
                 else:
-                    cur_bb = self._compute_backbone_frame(backbone_fn, images[frame_idx + 1:frame_idx + 2], frame_idx=frame_idx + 1)
+                    cur_bb = self._compute_backbone_frame(backbone_fn, _prep_frame(images, slice(frame_idx + 1, frame_idx + 2), device, dt, size), frame_idx=frame_idx + 1)
 
         if not all_masks or all(m is None for m in all_masks):
             return {"packed_masks": None, "n_frames": N, "scores": []}

comfy/ldm/wan/ar_model.py

@@ -0,0 +1,276 @@
+"""
+CausalWanModel: Wan 2.1 backbone with KV-cached causal self-attention for
+autoregressive (frame-by-frame) video generation via Causal Forcing.
+
+Weight-compatible with the standard WanModel -- same layer names, same shapes.
+The difference is purely in the forward pass: this model processes one temporal
+block at a time and maintains a KV cache across blocks.
+
+Reference: https://github.com/thu-ml/Causal-Forcing
+"""
+
+import torch
+import torch.nn as nn
+
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.math import apply_rope1
+from comfy.ldm.wan.model import (
+    sinusoidal_embedding_1d,
+    repeat_e,
+    WanModel,
+    WanAttentionBlock,
+)
+import comfy.ldm.common_dit
+import comfy.model_management
+
+
+class CausalWanSelfAttention(nn.Module):
+    """Self-attention with KV cache support for autoregressive inference."""
+
+    def __init__(self, dim, num_heads, window_size=(-1, -1), qk_norm=True,
+                 eps=1e-6, operation_settings={}):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.qk_norm = qk_norm
+        self.eps = eps
+
+        ops = operation_settings.get("operations")
+        device = operation_settings.get("device")
+        dtype = operation_settings.get("dtype")
+
+        self.q = ops.Linear(dim, dim, device=device, dtype=dtype)
+        self.k = ops.Linear(dim, dim, device=device, dtype=dtype)
+        self.v = ops.Linear(dim, dim, device=device, dtype=dtype)
+        self.o = ops.Linear(dim, dim, device=device, dtype=dtype)
+        self.norm_q = ops.RMSNorm(dim, eps=eps, elementwise_affine=True, device=device, dtype=dtype) if qk_norm else nn.Identity()
+        self.norm_k = ops.RMSNorm(dim, eps=eps, elementwise_affine=True, device=device, dtype=dtype) if qk_norm else nn.Identity()
+
+    def forward(self, x, freqs, kv_cache=None, transformer_options={}):
+        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
+
+        q = apply_rope1(self.norm_q(self.q(x)).view(b, s, n, d), freqs)
+        k = apply_rope1(self.norm_k(self.k(x)).view(b, s, n, d), freqs)
+        v = self.v(x).view(b, s, n, d)
+
+        if kv_cache is None:
+            x = optimized_attention(
+                q.view(b, s, n * d),
+                k.view(b, s, n * d),
+                v.view(b, s, n * d),
+                heads=self.num_heads,
+                transformer_options=transformer_options,
+            )
+        else:
+            end = kv_cache["end"]
+            new_end = end + s
+
+            # Roped K and plain V go into cache
+            kv_cache["k"][:, end:new_end] = k
+            kv_cache["v"][:, end:new_end] = v
+            kv_cache["end"] = new_end
+
+            x = optimized_attention(
+                q.view(b, s, n * d),
+                kv_cache["k"][:, :new_end].view(b, new_end, n * d),
+                kv_cache["v"][:, :new_end].view(b, new_end, n * d),
+                heads=self.num_heads,
+                transformer_options=transformer_options,
+            )
+
+        x = self.o(x)
+        return x
+
+
+class CausalWanAttentionBlock(WanAttentionBlock):
+    """Transformer block with KV-cached self-attention and cross-attention caching."""
+
+    def __init__(self, cross_attn_type, dim, ffn_dim, num_heads,
+                 window_size=(-1, -1), qk_norm=True, cross_attn_norm=False,
+                 eps=1e-6, operation_settings={}):
+        super().__init__(cross_attn_type, dim, ffn_dim, num_heads,
+                         window_size, qk_norm, cross_attn_norm, eps,
+                         operation_settings=operation_settings)
+        self.self_attn = CausalWanSelfAttention(
+            dim, num_heads, window_size, qk_norm, eps,
+            operation_settings=operation_settings)
+
+    def forward(self, x, e, freqs, context, context_img_len=257,
+                kv_cache=None, crossattn_cache=None, transformer_options={}):
+        if e.ndim < 4:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        else:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e).unbind(2)
+
+        # Self-attention with optional KV cache
+        x = x.contiguous()
+        y = self.self_attn(
+            torch.addcmul(repeat_e(e[0], x), self.norm1(x), 1 + repeat_e(e[1], x)),
+            freqs, kv_cache=kv_cache, transformer_options=transformer_options)
+        x = torch.addcmul(x, y, repeat_e(e[2], x))
+        del y
+
+        # Cross-attention with optional caching
+        if crossattn_cache is not None and crossattn_cache.get("is_init"):
+            q = self.cross_attn.norm_q(self.cross_attn.q(self.norm3(x)))
+            x_ca = optimized_attention(
+                q, crossattn_cache["k"], crossattn_cache["v"],
+                heads=self.num_heads, transformer_options=transformer_options)
+            x = x + self.cross_attn.o(x_ca)
+        else:
+            x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len, transformer_options=transformer_options)
+            if crossattn_cache is not None:
+                crossattn_cache["k"] = self.cross_attn.norm_k(self.cross_attn.k(context))
+                crossattn_cache["v"] = self.cross_attn.v(context)
+                crossattn_cache["is_init"] = True
+
+        # FFN
+        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
+
+
+class CausalWanModel(WanModel):
+    """
+    Wan 2.1 diffusion backbone with causal KV-cache support.
+
+    Same weight structure as WanModel -- loads identical state dicts.
+    Adds forward_block() for frame-by-frame autoregressive inference.
+    """
+
+    def __init__(self,
+                 model_type='t2v',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None):
+        super().__init__(
+            model_type=model_type, 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,
+            wan_attn_block_class=CausalWanAttentionBlock,
+            device=device, dtype=dtype, operations=operations)
+
+    def forward_block(self, x, timestep, context, start_frame,
+                      kv_caches, crossattn_caches, clip_fea=None):
+        """
+        Forward one temporal block for autoregressive inference.
+
+        Args:
+            x: [B, C, block_frames, H, W] input latent for the current block
+            timestep: [B, block_frames] per-frame timesteps
+            context: [B, L, text_dim] raw text embeddings (pre-text_embedding)
+            start_frame: temporal frame index for RoPE offset
+            kv_caches: list of per-layer KV cache dicts
+            crossattn_caches: list of per-layer cross-attention cache dicts
+            clip_fea: optional CLIP features for I2V
+
+        Returns:
+            flow_pred: [B, C_out, block_frames, H, W] flow prediction
+        """
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+        bs, c, t, h, w = x.shape
+
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # Per-frame time embedding
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, timestep.flatten()).to(dtype=x.dtype))
+        e = e.reshape(timestep.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        # Text embedding (reuses crossattn_cache after first block)
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None and self.img_emb is not None:
+            context_clip = self.img_emb(clip_fea)
+            context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        # RoPE for current block's temporal position
+        freqs = self.rope_encode(t, h, w, t_start=start_frame, device=x.device, dtype=x.dtype)
+
+        # Transformer blocks
+        for i, block in enumerate(self.blocks):
+            x = block(x, e=e0, freqs=freqs, context=context,
+                      context_img_len=context_img_len,
+                      kv_cache=kv_caches[i],
+                      crossattn_cache=crossattn_caches[i])
+
+        # Head
+        x = self.head(x, e)
+
+        # Unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x[:, :, :t, :h, :w]
+
+    def init_kv_caches(self, batch_size, max_seq_len, device, dtype):
+        """Create fresh KV caches for all layers."""
+        caches = []
+        for _ in range(self.num_layers):
+            caches.append({
+                "k": torch.zeros(batch_size, max_seq_len, self.num_heads, self.head_dim, device=device, dtype=dtype),
+                "v": torch.zeros(batch_size, max_seq_len, self.num_heads, self.head_dim, device=device, dtype=dtype),
+                "end": 0,
+            })
+        return caches
+
+    def init_crossattn_caches(self, batch_size, device, dtype):
+        """Create fresh cross-attention caches for all layers."""
+        caches = []
+        for _ in range(self.num_layers):
+            caches.append({"is_init": False})
+        return caches
+
+    def reset_kv_caches(self, kv_caches):
+        """Reset KV caches to empty (reuse allocated memory)."""
+        for cache in kv_caches:
+            cache["end"] = 0
+
+    def reset_crossattn_caches(self, crossattn_caches):
+        """Reset cross-attention caches."""
+        for cache in crossattn_caches:
+            cache["is_init"] = False
+
+    @property
+    def head_dim(self):
+        return self.dim // self.num_heads
+
+    def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
+        ar_state = transformer_options.get("ar_state")
+        if ar_state is not None:
+            bs = x.shape[0]
+            block_frames = x.shape[2]
+            t_per_frame = timestep.unsqueeze(1).expand(bs, block_frames)
+            return self.forward_block(
+                x=x, timestep=t_per_frame, context=context,
+                start_frame=ar_state["start_frame"],
+                kv_caches=ar_state["kv_caches"],
+                crossattn_caches=ar_state["crossattn_caches"],
+                clip_fea=clip_fea,
+            )
+
+        return super().forward(x, timestep, context, clip_fea=clip_fea,
+                               time_dim_concat=time_dim_concat,
+                               transformer_options=transformer_options, **kwargs)

comfy/lora.py

@@ -17,6 +17,7 @@
 """
 
 from __future__ import annotations
+import comfy.memory_management
 import comfy.utils
 import comfy.model_management
 import comfy.model_base
@@ -473,3 +474,17 @@ def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, ori
             weight = old_weight
 
     return weight
+
+def prefetch_prepared_value(value, allocate_buffer, stream):
+    if isinstance(value, torch.Tensor):
+        dest = allocate_buffer(comfy.memory_management.vram_aligned_size(value))
+        comfy.model_management.cast_to_gathered([value], dest, non_blocking=True, stream=stream)
+        return comfy.memory_management.interpret_gathered_like([value], dest)[0]
+    elif isinstance(value, weight_adapter.WeightAdapterBase):
+        return type(value)(value.loaded_keys, prefetch_prepared_value(value.weights, allocate_buffer, stream))
+    elif isinstance(value, tuple):
+        return tuple(prefetch_prepared_value(item, allocate_buffer, stream) for item in value)
+    elif isinstance(value, list):
+        return [prefetch_prepared_value(item, allocate_buffer, stream) for item in value]
+
+    return value

comfy/model_base.py

@@ -42,6 +42,7 @@ import comfy.ldm.cosmos.predict2
 import comfy.ldm.lumina.model
 import comfy.ldm.wan.model
 import comfy.ldm.wan.model_animate
+import comfy.ldm.wan.ar_model
 import comfy.ldm.hunyuan3d.model
 import comfy.ldm.hidream.model
 import comfy.ldm.chroma.model
@@ -214,6 +215,11 @@ class BaseModel(torch.nn.Module):
         if "latent_shapes" in extra_conds:
             xc = utils.unpack_latents(xc, extra_conds.pop("latent_shapes"))
 
+        transformer_options = transformer_options.copy()
+        transformer_options["prefetch_dynamic_vbars"] = (
+            self.current_patcher is not None and self.current_patcher.is_dynamic()
+        )
+
         model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds)
         if len(model_output) > 1 and not torch.is_tensor(model_output):
             model_output, _ = utils.pack_latents(model_output)
@@ -1360,6 +1366,13 @@ class WAN21(BaseModel):
         return out
 
 
+class WAN21_CausalAR(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.ar_model.CausalWanModel)
+        self.image_to_video = False
+
+
 class WAN21_Vace(WAN21):
     def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.VaceWanModel)

comfy/model_management.py

@@ -31,6 +31,7 @@ from contextlib import nullcontext
 import comfy.memory_management
 import comfy.utils
 import comfy.quant_ops
+import comfy_aimdo.vram_buffer
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -720,13 +721,15 @@ 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_temp = set()
+    # Order-preserving dedup. A plain set() would randomize iteration order across runs
+    models_temp = {}
     for m in models:
-        models_temp.add(m)
+        models_temp[m] = None
         for mm in m.model_patches_models():
-            models_temp.add(mm)
+            models_temp[mm] = None
 
-    models = models_temp
+    models = list(models_temp)
+    models.reverse()
 
     models_to_load = []
 
@@ -1175,6 +1178,10 @@ stream_counters = {}
 
 STREAM_CAST_BUFFERS = {}
 LARGEST_CASTED_WEIGHT = (None, 0)
+STREAM_AIMDO_CAST_BUFFERS = {}
+LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
+
+DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE = 16 * 1024 ** 3
 
 def get_cast_buffer(offload_stream, device, size, ref):
     global LARGEST_CASTED_WEIGHT
@@ -1208,13 +1215,26 @@ def get_cast_buffer(offload_stream, device, size, ref):
 
     return cast_buffer
 
+def get_aimdo_cast_buffer(offload_stream, device):
+    cast_buffer = STREAM_AIMDO_CAST_BUFFERS.get(offload_stream, None)
+    if cast_buffer is None:
+        cast_buffer = comfy_aimdo.vram_buffer.VRAMBuffer(DEFAULT_AIMDO_CAST_BUFFER_RESERVATION_SIZE, device.index)
+        STREAM_AIMDO_CAST_BUFFERS[offload_stream] = cast_buffer
+
+    return cast_buffer
 def reset_cast_buffers():
     global LARGEST_CASTED_WEIGHT
+    global LARGEST_AIMDO_CASTED_WEIGHT
+
     LARGEST_CASTED_WEIGHT = (None, 0)
-    for offload_stream in STREAM_CAST_BUFFERS:
-        offload_stream.synchronize()
+    LARGEST_AIMDO_CASTED_WEIGHT = (None, 0)
+    for offload_stream in set(STREAM_CAST_BUFFERS) | set(STREAM_AIMDO_CAST_BUFFERS):
+        if offload_stream is not None:
+            offload_stream.synchronize()
     synchronize()
+
     STREAM_CAST_BUFFERS.clear()
+    STREAM_AIMDO_CAST_BUFFERS.clear()
     soft_empty_cache()
 
 def get_offload_stream(device):

comfy/model_patcher.py

@@ -26,6 +26,7 @@ import uuid
 from typing import Callable, Optional
 
 import torch
+import tqdm
 
 import comfy.float
 import comfy.hooks
@@ -121,9 +122,20 @@ class LowVramPatch:
         self.patches = patches
         self.convert_func = convert_func # TODO: remove
         self.set_func = set_func
+        self.prepared_patches = None
+
+    def prepare(self, allocate_buffer, stream):
+        self.prepared_patches = [
+            (patch[0], comfy.lora.prefetch_prepared_value(patch[1], allocate_buffer, stream), patch[2], patch[3], patch[4])
+            for patch in self.patches[self.key]
+        ]
+
+    def clear_prepared(self):
+        self.prepared_patches = None
 
     def __call__(self, weight):
-        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=weight.dtype)
+        patches = self.prepared_patches if self.prepared_patches is not None else self.patches[self.key]
+        return comfy.lora.calculate_weight(patches, weight, self.key, intermediate_dtype=weight.dtype)
 
 LOWVRAM_PATCH_ESTIMATE_MATH_FACTOR = 2
 
@@ -1640,7 +1652,11 @@ class ModelPatcherDynamic(ModelPatcher):
                 self.model.model_loaded_weight_memory += casted_buf.numel() * casted_buf.element_size()
 
             force_load_stat = f" Force pre-loaded {len(self.backup)} weights: {self.model.model_loaded_weight_memory // 1024} KB." if len(self.backup) > 0 else ""
-            logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.{force_load_stat}")
+            log_key = (self.patches_uuid, allocated_size, num_patches, len(self.backup), self.model.model_loaded_weight_memory)
+            in_loop = bool(getattr(tqdm.tqdm, "_instances", None))
+            level = logging.DEBUG if in_loop and getattr(self, "_last_prepare_log_key", None) == log_key else logging.INFO
+            self._last_prepare_log_key = log_key
+            logging.log(level, f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.{force_load_stat}")
 
             self.model.device = device_to
             self.model.current_weight_patches_uuid = self.patches_uuid

comfy/model_prefetch.py

@@ -0,0 +1,66 @@
+import comfy_aimdo.model_vbar
+import comfy.model_management
+import comfy.ops
+
+PREFETCH_QUEUES = []
+
+def cleanup_prefetched_modules(comfy_modules):
+    for s in comfy_modules:
+        prefetch = getattr(s, "_prefetch", None)
+        if prefetch is None:
+            continue
+        for param_key in ("weight", "bias"):
+            lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+            if lowvram_fn is not None:
+                lowvram_fn.clear_prepared()
+        if prefetch["signature"] is not None:
+            comfy_aimdo.model_vbar.vbar_unpin(s._v)
+        delattr(s, "_prefetch")
+
+def cleanup_prefetch_queues():
+    global PREFETCH_QUEUES
+
+    for queue in PREFETCH_QUEUES:
+        for entry in queue:
+            if entry is None or not isinstance(entry, tuple):
+                continue
+            _, prefetch_state = entry
+            comfy_modules = prefetch_state[1]
+            if comfy_modules is not None:
+                cleanup_prefetched_modules(comfy_modules)
+    PREFETCH_QUEUES = []
+
+def prefetch_queue_pop(queue, device, module):
+    if queue is None:
+        return
+
+    consumed = queue.pop(0)
+    if consumed is not None:
+        offload_stream, prefetch_state = consumed
+        if offload_stream is not None:
+            offload_stream.wait_stream(comfy.model_management.current_stream(device))
+        _, comfy_modules = prefetch_state
+        if comfy_modules is not None:
+            cleanup_prefetched_modules(comfy_modules)
+
+    prefetch = queue[0]
+    if prefetch is not None:
+        comfy_modules = []
+        for s in prefetch.modules():
+            if hasattr(s, "_v"):
+                comfy_modules.append(s)
+
+        offload_stream = comfy.ops.cast_modules_with_vbar(comfy_modules, None, device, None, True)
+        comfy.model_management.sync_stream(device, offload_stream)
+        queue[0] = (offload_stream, (prefetch, comfy_modules))
+
+def make_prefetch_queue(queue, device, transformer_options):
+    if (not transformer_options.get("prefetch_dynamic_vbars", False)
+        or comfy.model_management.NUM_STREAMS == 0
+        or comfy.model_management.is_device_cpu(device)
+        or not comfy.model_management.device_supports_non_blocking(device)):
+        return None
+
+    queue = [None] + queue + [None]
+    PREFETCH_QUEUES.append(queue)
+    return queue

comfy/ops.py

@@ -86,38 +86,61 @@ def materialize_meta_param(s, param_keys):
             setattr(s, param_key, torch.nn.Parameter(torch.zeros(param.shape, dtype=param.dtype), requires_grad=param.requires_grad))
 
 
-def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant):
-    #vbar doesn't support CPU weights, but some custom nodes have weird paths
-    #that might switch the layer to the CPU and expect it to work. We have to take
-    #a clone conservatively as we are mmapped and some SFT files are packed misaligned
-    #If you are a custom node author reading this, please move your layer to the GPU
-    #or declare your ModelPatcher as CPU in the first place.
-    if comfy.model_management.is_device_cpu(device):
-        materialize_meta_param(s, ["weight", "bias"])
-        weight = s.weight.to(dtype=dtype, copy=True)
-        if isinstance(weight, QuantizedTensor):
-            weight = weight.dequantize()
-        bias = None
-        if s.bias is not None:
-            bias = s.bias.to(dtype=bias_dtype, copy=True)
-        return weight, bias, (None, None, None)
-
+# FIXME: add n=1 cache hit fast path
+def cast_modules_with_vbar(comfy_modules, dtype, device, bias_dtype, non_blocking):
     offload_stream = None
-    xfer_dest = None
+    cast_buffer = None
+    cast_buffer_offset = 0
+
+    def ensure_offload_stream(module, required_size, check_largest):
+        nonlocal offload_stream
+        nonlocal cast_buffer
+
+        if offload_stream is None:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+        if offload_stream is None or not check_largest or len(comfy_modules) != 1:
+            return
+
+        current_size = 0 if cast_buffer is None else cast_buffer.size()
+        if current_size < required_size and module is comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT[0]:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+            cast_buffer = None
+        if required_size > comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT[1]:
+            comfy.model_management.LARGEST_AIMDO_CASTED_WEIGHT = (module, required_size)
+
+    def get_cast_buffer(buffer_size):
+        nonlocal offload_stream
+        nonlocal cast_buffer
+        nonlocal cast_buffer_offset
+
+        if buffer_size == 0:
+            return None
+
+        if offload_stream is None:
+            return torch.empty((buffer_size,), dtype=torch.uint8, device=device)
+
+        cast_buffer = comfy.model_management.get_aimdo_cast_buffer(offload_stream, device)
+        buffer = comfy_aimdo.torch.aimdo_to_tensor(cast_buffer.get(buffer_size, cast_buffer_offset), device)
+        cast_buffer_offset += buffer_size
+        return buffer
+
+    for s in comfy_modules:
+        signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
+        resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
+        prefetch = {
+            "signature": signature,
+            "resident": resident,
+        }
 
-    signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
-    resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
-    if signature is not None:
         if resident:
-            weight = s._v_weight
-            bias = s._v_bias
-        else:
-            xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
+            s._prefetch = prefetch
+            continue
 
-    if not resident:
         materialize_meta_param(s, ["weight", "bias"])
+        xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device) if signature is not None else None
         cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
         cast_dest = None
+        needs_cast = False
 
         xfer_source = [ s.weight, s.bias ]
 
@@ -129,22 +152,15 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             if data is None:
                 continue
             if data.dtype != geometry.dtype:
+                needs_cast = True
                 cast_dest = xfer_dest
-                if cast_dest is None:
-                    cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device)
                 xfer_dest = None
                 break
 
         dest_size = comfy.memory_management.vram_aligned_size(xfer_source)
-        offload_stream = comfy.model_management.get_offload_stream(device)
-        if xfer_dest is None and offload_stream is not None:
-                xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
-                if xfer_dest is None:
-                    offload_stream = comfy.model_management.get_offload_stream(device)
-                    xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
+        ensure_offload_stream(s, dest_size if xfer_dest is None else 0, True)
         if xfer_dest is None:
-            xfer_dest = torch.empty((dest_size,), dtype=torch.uint8, device=device)
-            offload_stream = None
+            xfer_dest = get_cast_buffer(dest_size)
 
         if signature is None and pin is None:
             comfy.pinned_memory.pin_memory(s)
@@ -157,27 +173,54 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             xfer_source = [ pin ]
         #send it over
         comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=offload_stream)
-        comfy.model_management.sync_stream(device, offload_stream)
 
-        if cast_dest is not None:
+        for param_key in ("weight", "bias"):
+            lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+            if lowvram_fn is not None:
+                ensure_offload_stream(s, cast_buffer_offset, False)
+                lowvram_fn.prepare(lambda size: get_cast_buffer(size), offload_stream)
+
+        prefetch["xfer_dest"] = xfer_dest
+        prefetch["cast_dest"] = cast_dest
+        prefetch["cast_geometry"] = cast_geometry
+        prefetch["needs_cast"] = needs_cast
+        s._prefetch = prefetch
+
+    return offload_stream
+
+
+def resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, want_requant):
+
+    prefetch = getattr(s, "_prefetch", None)
+
+    if prefetch["resident"]:
+        weight = s._v_weight
+        bias = s._v_bias
+    else:
+        xfer_dest = prefetch["xfer_dest"]
+        if prefetch["needs_cast"]:
+            cast_dest = prefetch["cast_dest"] if prefetch["cast_dest"] is not None else torch.empty((comfy.memory_management.vram_aligned_size(prefetch["cast_geometry"]),), dtype=torch.uint8, device=device)
             for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like([s.weight, s.bias ], xfer_dest),
-                                           comfy.memory_management.interpret_gathered_like(cast_geometry, cast_dest)):
+                                           comfy.memory_management.interpret_gathered_like(prefetch["cast_geometry"], cast_dest)):
                 if post_cast is not None:
                     post_cast.copy_(pre_cast)
             xfer_dest = cast_dest
 
-        params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest)
+        params = comfy.memory_management.interpret_gathered_like(prefetch["cast_geometry"], xfer_dest)
         weight = params[0]
         bias = params[1]
-        if signature is not None:
+        if prefetch["signature"] is not None:
             s._v_weight = weight
             s._v_bias = bias
-        s._v_signature=signature
+        s._v_signature = prefetch["signature"]
 
     def post_cast(s, param_key, x, dtype, resident, update_weight):
         lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
         fns = getattr(s, param_key + "_function", [])
 
+        if x is None:
+            return None
+
         orig = x
 
         def to_dequant(tensor, dtype):
@@ -205,14 +248,15 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             x = f(x)
         return x
 
-    update_weight = signature is not None
+    update_weight = prefetch["signature"] is not None
+    weight = post_cast(s, "weight", weight, dtype, prefetch["resident"], update_weight)
+    if bias is not None:
+        bias = post_cast(s, "bias", bias, bias_dtype, prefetch["resident"], update_weight)
 
-    weight = post_cast(s, "weight", weight, dtype, resident, update_weight)
-    if s.bias is not None:
-        bias = post_cast(s, "bias", bias, bias_dtype, resident, update_weight)
+    if prefetch["signature"] is not None:
+        prefetch["resident"] = True
 
-    #FIXME: weird offload return protocol
-    return weight, bias, (offload_stream, device if signature is not None else None, None)
+    return weight, bias
 
 
 def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False, compute_dtype=None, want_requant=False):
@@ -230,10 +274,46 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
         if device is None:
             device = input.device
 
+    def format_return(result, offloadable):
+        weight, bias, offload_stream = result
+        return (weight, bias, offload_stream) if offloadable else (weight, bias)
+
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
 
     if hasattr(s, "_v"):
-        return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant)
+
+        #vbar doesn't support CPU weights, but some custom nodes have weird paths
+        #that might switch the layer to the CPU and expect it to work. We have to take
+        #a clone conservatively as we are mmapped and some SFT files are packed misaligned
+        #If you are a custom node author reading this, please move your layer to the GPU
+        #or declare your ModelPatcher as CPU in the first place.
+        if comfy.model_management.is_device_cpu(device):
+            materialize_meta_param(s, ["weight", "bias"])
+            weight = s.weight.to(dtype=dtype, copy=True)
+            if isinstance(weight, QuantizedTensor):
+                weight = weight.dequantize()
+            bias = s.bias.to(dtype=bias_dtype, copy=True) if s.bias is not None else None
+            return format_return((weight, bias, (None, None, None)), offloadable)
+
+        prefetched = hasattr(s, "_prefetch")
+        offload_stream = None
+        offload_device = None
+        if not prefetched:
+            offload_stream = cast_modules_with_vbar([s], dtype, device, bias_dtype, non_blocking)
+            comfy.model_management.sync_stream(device, offload_stream)
+
+        weight, bias = resolve_cast_module_with_vbar(s, dtype, device, bias_dtype, compute_dtype, want_requant)
+
+        if not prefetched:
+            if getattr(s, "_prefetch")["signature"] is not None:
+                offload_device = device
+            for param_key in ("weight", "bias"):
+                lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
+                if lowvram_fn is not None:
+                    lowvram_fn.clear_prepared()
+            delattr(s, "_prefetch")
+        return format_return((weight, bias, (offload_stream, offload_device, None)), offloadable)
+
 
     if offloadable and (device != s.weight.device or
                         (s.bias is not None and device != s.bias.device)):
@@ -280,11 +360,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
         for f in s.weight_function:
             weight = f(weight)
 
-    if offloadable:
-        return weight, bias, (offload_stream, weight_a, bias_a)
-    else:
-        #Legacy function signature
-        return weight, bias
+    return format_return((weight, bias, (offload_stream, weight_a, bias_a)), offloadable)
 
 
 def uncast_bias_weight(s, weight, bias, offload_stream):
@@ -486,6 +562,25 @@ class disable_weight_init:
             else:
                 return super().forward(*args, **kwargs)
 
+    class BatchNorm2d(torch.nn.BatchNorm2d, CastWeightBiasOp):
+        def reset_parameters(self):
+            return None
+
+        def forward_comfy_cast_weights(self, input):
+            weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            running_mean = self.running_mean.to(device=input.device, dtype=weight.dtype) if self.running_mean is not None else None
+            running_var = self.running_var.to(device=input.device, dtype=weight.dtype) if self.running_var is not None else None
+            x = torch.nn.functional.batch_norm(input, running_mean, running_var, weight, bias, self.training, self.momentum, self.eps)
+            uncast_bias_weight(self, weight, bias, offload_stream)
+            return x
+
+        def forward(self, *args, **kwargs):
+            run_every_op()
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
+                return self.forward_comfy_cast_weights(*args, **kwargs)
+            else:
+                return super().forward(*args, **kwargs)
+
     class LayerNorm(torch.nn.LayerNorm, CastWeightBiasOp):
         def reset_parameters(self):
             return None
@@ -673,6 +768,9 @@ class manual_cast(disable_weight_init):
     class Conv3d(disable_weight_init.Conv3d):
         comfy_cast_weights = True
 
+    class BatchNorm2d(disable_weight_init.BatchNorm2d):
+        comfy_cast_weights = True
+
     class GroupNorm(disable_weight_init.GroupNorm):
         comfy_cast_weights = True
 
@@ -1173,6 +1271,93 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                         self._buffers[key] = fn(buf)
                 return self
 
+        class Embedding(manual_cast.Embedding):
+            def _load_from_state_dict(self, state_dict, prefix, local_metadata,
+                                    strict, missing_keys, unexpected_keys, error_msgs):
+                weight_key = f"{prefix}weight"
+                layer_conf = state_dict.pop(f"{prefix}comfy_quant", None)
+                if layer_conf is not None:
+                    layer_conf = json.loads(layer_conf.numpy().tobytes())
+
+                # Only fp8 makes sense for embeddings (per-row dequant via index select).
+                # Block-scaled formats (NVFP4, MXFP8) can't do per-row lookup efficiently.
+                quant_format = layer_conf.get("format", None) if layer_conf is not None else None
+                if quant_format in ["float8_e4m3fn", "float8_e5m2"] and weight_key in state_dict:
+                    self.quant_format = quant_format
+                    qconfig = QUANT_ALGOS[quant_format]
+                    layout_cls = get_layout_class(qconfig["comfy_tensor_layout"])
+                    weight = state_dict.pop(weight_key)
+                    manually_loaded_keys = [weight_key]
+
+                    scale_key = f"{prefix}weight_scale"
+                    scale = state_dict.pop(scale_key, None)
+                    if scale is not None:
+                        scale = scale.float()
+                        manually_loaded_keys.append(scale_key)
+
+                    params = layout_cls.Params(
+                        scale=scale if scale is not None else torch.ones((), dtype=torch.float32),
+                        orig_dtype=MixedPrecisionOps._compute_dtype,
+                        orig_shape=(self.num_embeddings, self.embedding_dim),
+                    )
+                    self.weight = torch.nn.Parameter(
+                        QuantizedTensor(weight.to(dtype=qconfig["storage_t"]), qconfig["comfy_tensor_layout"], params),
+                        requires_grad=False)
+
+                    super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+                    for k in manually_loaded_keys:
+                        if k in missing_keys:
+                            missing_keys.remove(k)
+                else:
+                    if layer_conf is not None:
+                        state_dict[f"{prefix}comfy_quant"] = torch.tensor(list(json.dumps(layer_conf).encode('utf-8')), dtype=torch.uint8)
+                    super()._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs)
+
+            def state_dict(self, *args, destination=None, prefix="", **kwargs):
+                if destination is not None:
+                    sd = destination
+                else:
+                    sd = {}
+
+                if not hasattr(self, 'weight') or self.weight is None:
+                    return sd
+
+                if isinstance(self.weight, QuantizedTensor):
+                    sd_out = self.weight.state_dict("{}weight".format(prefix))
+                    for k in sd_out:
+                        sd[k] = sd_out[k]
+
+                    quant_conf = {"format": self.quant_format}
+                    sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8)
+                else:
+                    sd["{}weight".format(prefix)] = self.weight
+                return sd
+
+            def forward_comfy_cast_weights(self, input, out_dtype=None):
+                weight = self.weight
+
+                # Optimized path: lookup in fp8, dequantize only the selected rows.
+                if isinstance(weight, QuantizedTensor) and len(self.weight_function) == 0:
+                    qdata, _, offload_stream = cast_bias_weight(self, device=input.device, dtype=weight.dtype, offloadable=True)
+                    if isinstance(qdata, QuantizedTensor):
+                        scale = qdata._params.scale
+                        qdata = qdata._qdata
+                    else:
+                        scale = None
+
+                    x = torch.nn.functional.embedding(
+                        input, qdata, self.padding_idx, self.max_norm,
+                        self.norm_type, self.scale_grad_by_freq, self.sparse)
+                    uncast_bias_weight(self, qdata, None, offload_stream)
+                    target_dtype = out_dtype if out_dtype is not None else weight._params.orig_dtype
+                    x = x.to(dtype=target_dtype)
+                    if scale is not None and scale != 1.0:
+                        x = x * scale.to(dtype=target_dtype)
+                    return x
+
+                # Fallback for non-quantized or weight_function (LoRA) case
+                return super().forward_comfy_cast_weights(input, out_dtype=out_dtype)
+
     return MixedPrecisionOps
 
 def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None):

comfy/quant_ops.py

@@ -1,6 +1,8 @@
 import torch
 import logging
 
+from comfy.cli_args import args
+
 try:
     import comfy_kitchen as ck
     from comfy_kitchen.tensor import (
@@ -21,7 +23,15 @@ try:
             ck.registry.disable("cuda")
             logging.warning("WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations.")
 
-    ck.registry.disable("triton")
+    if args.enable_triton_backend:
+        try:
+            import triton
+            logging.info("Found triton %s. Enabling comfy-kitchen triton backend.", triton.__version__)
+        except ImportError as e:
+            logging.error(f"Failed to import triton, Error: {e}, the comfy-kitchen triton backend will not be available.")
+            ck.registry.disable("triton")
+    else:
+        ck.registry.disable("triton")
     for k, v in ck.list_backends().items():
         logging.info(f"Found comfy_kitchen backend {k}: {v}")
 except ImportError as e:

comfy/rmsnorm.py

@@ -3,6 +3,7 @@ import comfy.model_management
 
 RMSNorm = torch.nn.RMSNorm
 
+# Note: torch's fused F.rms_norm is faster but produces slightly different output than manual implementations (rsqrt/reduction rounding).
 def rms_norm(x, weight=None, eps=1e-6):
     if weight is None:
         return torch.nn.functional.rms_norm(x, (x.shape[-1],), eps=eps)

comfy/sampler_helpers.py

@@ -89,7 +89,8 @@ def get_additional_models(conds, dtype):
         gligen += get_models_from_cond(conds[k], "gligen")
         add_models += get_models_from_cond(conds[k], "additional_models")
 
-    control_nets = set(cnets)
+    # Order-preserving dedup. A plain set() would randomize iteration order across runs
+    control_nets = list(dict.fromkeys(cnets))
 
     inference_memory = 0
     control_models = []

comfy/sd.py

@@ -65,6 +65,8 @@ import comfy.text_encoders.ace15
 import comfy.text_encoders.longcat_image
 import comfy.text_encoders.qwen35
 import comfy.text_encoders.ernie
+import comfy.text_encoders.gemma4
+import comfy.text_encoders.cogvideo
 
 import comfy.model_patcher
 import comfy.lora
@@ -1223,6 +1225,7 @@ class CLIPType(Enum):
     NEWBIE = 24
     FLUX2 = 25
     LONGCAT_IMAGE = 26
+    COGVIDEOX = 27
 
 
 
@@ -1271,6 +1274,9 @@ class TEModel(Enum):
     QWEN35_9B = 26
     QWEN35_27B = 27
     MINISTRAL_3_3B = 28
+    GEMMA_4_E4B = 29
+    GEMMA_4_E2B = 30
+    GEMMA_4_31B = 31
 
 
 def detect_te_model(sd):
@@ -1296,6 +1302,12 @@ def detect_te_model(sd):
             return TEModel.BYT5_SMALL_GLYPH
         return TEModel.T5_BASE
     if 'model.layers.0.post_feedforward_layernorm.weight' in sd:
+        if 'model.layers.59.self_attn.q_norm.weight' in sd:
+            return TEModel.GEMMA_4_31B
+        if 'model.layers.41.self_attn.q_norm.weight' in sd and 'model.layers.47.self_attn.q_norm.weight' not in sd:
+            return TEModel.GEMMA_4_E4B
+        if 'model.layers.34.self_attn.q_norm.weight' in sd and 'model.layers.41.self_attn.q_norm.weight' not in sd:
+            return TEModel.GEMMA_4_E2B
         if 'model.layers.47.self_attn.q_norm.weight' in sd:
             return TEModel.GEMMA_3_12B
         if 'model.layers.0.self_attn.q_norm.weight' in sd:
@@ -1418,6 +1430,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
                 clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**t5xxl_detect(clip_data),
                                                                         clip_l=False, clip_g=False, t5=True, llama=False, dtype_llama=None)
                 clip_target.tokenizer = comfy.text_encoders.hidream.HiDreamTokenizer
+            elif clip_type == CLIPType.COGVIDEOX:
+                clip_target.clip = comfy.text_encoders.cogvideo.cogvideo_te(**t5xxl_detect(clip_data))
+                clip_target.tokenizer = comfy.text_encoders.cogvideo.CogVideoXTokenizer
             else: #CLIPType.MOCHI
                 clip_target.clip = comfy.text_encoders.genmo.mochi_te(**t5xxl_detect(clip_data))
                 clip_target.tokenizer = comfy.text_encoders.genmo.MochiT5Tokenizer
@@ -1435,6 +1450,13 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             else:
                 clip_target.clip = comfy.text_encoders.sa_t5.SAT5Model
                 clip_target.tokenizer = comfy.text_encoders.sa_t5.SAT5Tokenizer
+        elif te_model in (TEModel.GEMMA_4_E4B, TEModel.GEMMA_4_E2B, TEModel.GEMMA_4_31B):
+            variant = {TEModel.GEMMA_4_E4B: comfy.text_encoders.gemma4.Gemma4_E4B,
+                       TEModel.GEMMA_4_E2B: comfy.text_encoders.gemma4.Gemma4_E2B,
+                       TEModel.GEMMA_4_31B: comfy.text_encoders.gemma4.Gemma4_31B}[te_model]
+            clip_target.clip = comfy.text_encoders.gemma4.gemma4_te(**llama_detect(clip_data), model_class=variant)
+            clip_target.tokenizer = variant.tokenizer
+            tokenizer_data["tokenizer_json"] = clip_data[0].get("tokenizer_json", None)
         elif te_model == TEModel.GEMMA_2_2B:
             clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.lumina2.LuminaTokenizer

comfy/supported_models.py

@@ -1167,6 +1167,25 @@ class WAN21_T2V(supported_models_base.BASE):
         t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}umt5xxl.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.wan.WanT5Tokenizer, comfy.text_encoders.wan.te(**t5_detect))
 
+class WAN21_CausalAR_T2V(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "t2v",
+        "causal_ar": True,
+    }
+
+    sampling_settings = {
+        "shift": 5.0,
+    }
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        self.unet_config.pop("causal_ar", None)
+
+    def get_model(self, state_dict, prefix="", device=None):
+        return model_base.WAN21_CausalAR(self, device=device)
+
+
 class WAN21_I2V(WAN21_T2V):
     unet_config = {
         "image_model": "wan2.1",
@@ -1853,6 +1872,14 @@ class CogVideoX_T2V(supported_models_base.BASE):
     vae_key_prefix = ["vae."]
     text_encoder_key_prefix = ["text_encoders."]
 
+    def __init__(self, unet_config):
+        # 2b-class (dim=1920, heads=30) uses scale_factor=1.15258426.
+        # 5b-class (dim=3072, heads=48) — incl. CogVideoX-5b, 1.5-5B, and
+        # Fun-V1.5 inpainting — uses scale_factor=0.7 per vae/config.json.
+        if unet_config.get("num_attention_heads", 0) >= 48:
+            self.latent_format = latent_formats.CogVideoX1_5
+        super().__init__(unet_config)
+
     def get_model(self, state_dict, prefix="", device=None):
         # CogVideoX 1.5 (patch_size_t=2) has different training base dimensions for RoPE
         if self.unet_config.get("patch_size_t") is not None:
@@ -1879,6 +1906,20 @@ class CogVideoX_I2V(CogVideoX_T2V):
         out = model_base.CogVideoX(self, image_to_video=True, device=device)
         return out
 
+class CogVideoX_Inpaint(CogVideoX_T2V):
+    unet_config = {
+        "image_model": "cogvideox",
+        "in_channels": 48,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        if self.unet_config.get("patch_size_t") is not None:
+            self.unet_config.setdefault("sample_height", 96)
+            self.unet_config.setdefault("sample_width", 170)
+            self.unet_config.setdefault("sample_frames", 81)
+        out = model_base.CogVideoX(self, image_to_video=True, device=device)
+        return out
+
 
 models = [
     LotusD,
@@ -1929,6 +1970,7 @@ models = [
     ZImage,
     Lumina2,
     WAN22_T2V,
+    WAN21_CausalAR_T2V,
     WAN21_T2V,
     WAN21_I2V,
     WAN21_FunControl2V,
@@ -1958,6 +2000,7 @@ models = [
     ErnieImage,
     SAM3,
     SAM31,
+    CogVideoX_Inpaint,
     CogVideoX_I2V,
     CogVideoX_T2V,
     SVD_img2vid,

comfy/text_encoders/cogvideo.py

@@ -1,6 +1,48 @@
 import comfy.text_encoders.sd3_clip
+from comfy import sd1_clip
 
 
 class CogVideoXT5Tokenizer(comfy.text_encoders.sd3_clip.T5XXLTokenizer):
+    """Inner T5 tokenizer for CogVideoX.
+
+    CogVideoX was trained with T5 embeddings padded to 226 tokens (not 77 like SD3).
+    Used both directly by supported_models.CogVideoX_T2V.clip_target (paired with
+    the raw T5XXLModel) and by the CogVideoXTokenizer outer wrapper below.
+    """
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, min_length=226)
+
+
+class CogVideoXTokenizer(sd1_clip.SD1Tokenizer):
+    """Outer tokenizer wrapper for CLIPLoader (type="cogvideox")."""
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data,
+                         clip_name="t5xxl", tokenizer=CogVideoXT5Tokenizer)
+
+
+class CogVideoXT5XXL(sd1_clip.SD1ClipModel):
+    """Outer T5XXL model wrapper for CLIPLoader (type="cogvideox").
+
+    Wraps the raw T5XXL model in the SD1ClipModel interface so that CLIP.__init__
+    (which reads self.dtypes) works correctly. The inner model is the standard
+    sd3_clip.T5XXLModel (no attention_mask change needed for CogVideoX).
+    """
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, name="t5xxl",
+                         clip_model=comfy.text_encoders.sd3_clip.T5XXLModel,
+                         model_options=model_options)
+
+
+def cogvideo_te(dtype_t5=None, t5_quantization_metadata=None):
+    """Factory that returns a CogVideoXT5XXL class configured with the detected
+    T5 dtype and optional quantization metadata, for use in load_text_encoder_state_dicts.
+    """
+    class CogVideoXTEModel_(CogVideoXT5XXL):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if t5_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata
+            if dtype_t5 is not None:
+                dtype = dtype_t5
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return CogVideoXTEModel_

comfy/text_encoders/llama.py

@@ -521,7 +521,7 @@ class Attention(nn.Module):
             else:
                 present_key_value = (xk, xv, index + num_tokens)
 
-            if sliding_window is not None and xk.shape[2] > sliding_window:
+            if sliding_window is not None and xk.shape[2] > sliding_window and seq_length == 1:
                 xk = xk[:, :, -sliding_window:]
                 xv = xv[:, :, -sliding_window:]
                 attention_mask = attention_mask[..., -sliding_window:] if attention_mask is not None else None
@@ -533,12 +533,12 @@ class Attention(nn.Module):
         return self.o_proj(output), present_key_value
 
 class MLP(nn.Module):
-    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
+    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None, intermediate_size=None):
         super().__init__()
-        ops = ops or nn
-        self.gate_proj = ops.Linear(config.hidden_size, config.intermediate_size, bias=False, device=device, dtype=dtype)
-        self.up_proj = ops.Linear(config.hidden_size, config.intermediate_size, bias=False, device=device, dtype=dtype)
-        self.down_proj = ops.Linear(config.intermediate_size, config.hidden_size, bias=False, device=device, dtype=dtype)
+        intermediate_size = intermediate_size or config.intermediate_size
+        self.gate_proj = ops.Linear(config.hidden_size, intermediate_size, bias=False, device=device, dtype=dtype)
+        self.up_proj = ops.Linear(config.hidden_size, intermediate_size, bias=False, device=device, dtype=dtype)
+        self.down_proj = ops.Linear(intermediate_size, config.hidden_size, bias=False, device=device, dtype=dtype)
         if config.mlp_activation == "silu":
             self.activation = torch.nn.functional.silu
         elif config.mlp_activation == "gelu_pytorch_tanh":
@@ -647,24 +647,25 @@ class TransformerBlockGemma2(nn.Module):
 
         return x, present_key_value
 
+def _make_scaled_embedding(ops, vocab_size, hidden_size, scale, device, dtype):
+    class ScaledEmbedding(ops.Embedding):
+        def forward(self, input_ids, out_dtype=None):
+            return super().forward(input_ids, out_dtype=out_dtype) * scale
+    return ScaledEmbedding(vocab_size, hidden_size, device=device, dtype=dtype)
+
+
 class Llama2_(nn.Module):
     def __init__(self, config, device=None, dtype=None, ops=None):
         super().__init__()
         self.config = config
         self.vocab_size = config.vocab_size
 
-        self.embed_tokens = ops.Embedding(
-            config.vocab_size,
-            config.hidden_size,
-            device=device,
-            dtype=dtype
-        )
         if self.config.transformer_type == "gemma2" or self.config.transformer_type == "gemma3":
             transformer = TransformerBlockGemma2
-            self.normalize_in = True
+            self.embed_tokens = _make_scaled_embedding(ops, config.vocab_size, config.hidden_size, config.hidden_size ** 0.5, device, dtype)
         else:
             transformer = TransformerBlock
-            self.normalize_in = False
+            self.embed_tokens = ops.Embedding(config.vocab_size, config.hidden_size, device=device, dtype=dtype)
 
         self.layers = nn.ModuleList([
             transformer(config, index=i, device=device, dtype=dtype, ops=ops)
@@ -690,15 +691,12 @@ class Llama2_(nn.Module):
                                     self.config.rope_dims,
                                     device=device)
 
-    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None):
+    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, position_ids=None, embeds_info=[], past_key_values=None, input_ids=None):
         if embeds is not None:
             x = embeds
         else:
             x = self.embed_tokens(x, out_dtype=dtype)
 
-        if self.normalize_in:
-            x *= self.config.hidden_size ** 0.5
-
         seq_len = x.shape[1]
         past_len = 0
         if past_key_values is not None and len(past_key_values) > 0:
@@ -850,7 +848,7 @@ class BaseGenerate:
                                     torch.empty([batch, model_config.num_key_value_heads, max_cache_len, model_config.head_dim], device=device, dtype=execution_dtype), 0))
         return past_key_values
 
-    def generate(self, embeds=None, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.9, min_p=0.0, repetition_penalty=1.0, seed=42, stop_tokens=None, initial_tokens=[], execution_dtype=None, min_tokens=0, presence_penalty=0.0):
+    def generate(self, embeds=None, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.9, min_p=0.0, repetition_penalty=1.0, seed=42, stop_tokens=None, initial_tokens=[], execution_dtype=None, min_tokens=0, presence_penalty=0.0, initial_input_ids=None):
         device = embeds.device
 
         if stop_tokens is None:
@@ -875,14 +873,16 @@ class BaseGenerate:
         pbar = comfy.utils.ProgressBar(max_length)
 
         # Generation loop
+        current_input_ids = initial_input_ids
         for step in tqdm(range(max_length), desc="Generating tokens"):
-            x, _, past_key_values = self.model.forward(None, embeds=embeds, attention_mask=None, past_key_values=past_key_values)
+            x, _, past_key_values = self.model.forward(None, embeds=embeds, attention_mask=None, past_key_values=past_key_values, input_ids=current_input_ids)
             logits = self.logits(x)[:, -1]
             next_token = self.sample_token(logits, temperature, top_k, top_p, min_p, repetition_penalty, initial_tokens + generated_token_ids, generator, do_sample=do_sample, presence_penalty=presence_penalty)
             token_id = next_token[0].item()
             generated_token_ids.append(token_id)
 
             embeds = self.model.embed_tokens(next_token).to(execution_dtype)
+            current_input_ids = next_token if initial_input_ids is not None else None
             pbar.update(1)
 
             if token_id in stop_tokens:

comfy/text_encoders/lt.py

@@ -93,8 +93,7 @@ class Gemma3_12BModel(sd1_clip.SDClipModel):
 
     def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, presence_penalty):
         tokens_only = [[t[0] for t in b] for b in tokens]
-        embeds, _, _, embeds_info = self.process_tokens(tokens_only, self.execution_device)
-        comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5)
+        embeds, _, _, _ = self.process_tokens(tokens_only, self.execution_device)
         return self.transformer.generate(embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, stop_tokens=[106], presence_penalty=presence_penalty)  # 106 is <end_of_turn>
 
 class DualLinearProjection(torch.nn.Module):

comfy/text_encoders/lumina2.py

@@ -50,8 +50,7 @@ class Gemma3_4B_Vision_Model(sd1_clip.SDClipModel):
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B_Vision, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
     def process_tokens(self, tokens, device):
-        embeds, _, _, embeds_info = super().process_tokens(tokens, device)
-        comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5)
+        embeds, _, _, _ = super().process_tokens(tokens, device)
         return embeds
 
 class LuminaModel(sd1_clip.SD1ClipModel):

comfy/text_encoders/qwen35.py

@@ -408,8 +408,6 @@ class Qwen35Transformer(Llama2_):
         nn.Module.__init__(self)
         self.config = config
         self.vocab_size = config.vocab_size
-        self.normalize_in = False
-
         self.embed_tokens = ops.Embedding(config.vocab_size, config.hidden_size, device=device, dtype=dtype)
         self.layers = nn.ModuleList([
             Qwen35TransformerBlock(config, index=i, device=device, dtype=dtype, ops=ops)

comfy/utils.py

@@ -1446,10 +1446,3 @@ def deepcopy_list_dict(obj, memo=None):
     memo[obj_id] = res
     return res
 
-def normalize_image_embeddings(embeds, embeds_info, scale_factor):
-    """Normalize image embeddings to match text embedding scale"""
-    for info in embeds_info:
-        if info.get("type") == "image":
-            start_idx = info["index"]
-            end_idx = start_idx + info["size"]
-            embeds[:, start_idx:end_idx, :] /= scale_factor

comfy_api/feature_flags.py

@@ -5,12 +5,95 @@ This module handles capability negotiation between frontend and backend,
 allowing graceful protocol evolution while maintaining backward compatibility.
 """
 
-from typing import Any
+import logging
+from typing import Any, TypedDict
 
 from comfy.cli_args import args
 
+
+class FeatureFlagInfo(TypedDict):
+    type: str
+    default: Any
+    description: str
+
+
+# Registry of known CLI-settable feature flags.
+# Launchers can query this via --list-feature-flags to discover valid flags.
+CLI_FEATURE_FLAG_REGISTRY: dict[str, FeatureFlagInfo] = {
+    "show_signin_button": {
+        "type": "bool",
+        "default": False,
+        "description": "Show the sign-in button in the frontend even when not signed in",
+    },
+}
+
+
+def _coerce_bool(v: str) -> bool:
+    """Strict bool coercion: only 'true'/'false' (case-insensitive).
+
+    Anything else raises ValueError so the caller can warn and drop the flag,
+    rather than silently treating typos like 'ture' or 'yes' as False.
+    """
+    lower = v.lower()
+    if lower == "true":
+        return True
+    if lower == "false":
+        return False
+    raise ValueError(f"expected 'true' or 'false', got {v!r}")
+
+
+_COERCE_FNS: dict[str, Any] = {
+    "bool": _coerce_bool,
+    "int": lambda v: int(v),
+    "float": lambda v: float(v),
+}
+
+
+def _coerce_flag_value(key: str, raw_value: str) -> Any:
+    """Coerce a raw string value using the registry type, or keep as string.
+
+    Returns the raw string if the key is unregistered or the type is unknown.
+    Raises ValueError/TypeError if the key is registered with a known type but
+    the value cannot be coerced; callers are expected to warn and drop the flag.
+    """
+    info = CLI_FEATURE_FLAG_REGISTRY.get(key)
+    if info is None:
+        return raw_value
+    coerce = _COERCE_FNS.get(info["type"])
+    if coerce is None:
+        return raw_value
+    return coerce(raw_value)
+
+
+def _parse_cli_feature_flags() -> dict[str, Any]:
+    """Parse --feature-flag key=value pairs from CLI args into a dict.
+
+    Items without '=' default to the value 'true' (bare flag form).
+    Flags whose value cannot be coerced to the registered type are dropped
+    with a warning, so a typo like '--feature-flag some_bool=ture' does not
+    silently take effect as the wrong value.
+    """
+    result: dict[str, Any] = {}
+    for item in getattr(args, "feature_flag", []):
+        key, sep, raw_value = item.partition("=")
+        key = key.strip()
+        if not key:
+            continue
+        if not sep:
+            raw_value = "true"
+        try:
+            result[key] = _coerce_flag_value(key, raw_value.strip())
+        except (ValueError, TypeError) as e:
+            info = CLI_FEATURE_FLAG_REGISTRY.get(key, {})
+            logging.warning(
+                "Could not coerce --feature-flag %s=%r to %s (%s); dropping flag.",
+                key, raw_value.strip(), info.get("type", "?"), e,
+            )
+    return result
+
+
 # Default server capabilities
-SERVER_FEATURE_FLAGS: dict[str, Any] = {
+_CORE_FEATURE_FLAGS: dict[str, Any] = {
     "supports_preview_metadata": True,
     "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes
     "extension": {"manager": {"supports_v4": True}},
@@ -18,6 +101,11 @@ SERVER_FEATURE_FLAGS: dict[str, Any] = {
     "assets": args.enable_assets,
 }
 
+# CLI-provided flags cannot overwrite core flags
+_cli_flags = {k: v for k, v in _parse_cli_feature_flags().items() if k not in _CORE_FEATURE_FLAGS}
+
+SERVER_FEATURE_FLAGS: dict[str, Any] = {**_CORE_FEATURE_FLAGS, **_cli_flags}
+
 
 def get_connection_feature(
     sockets_metadata: dict[str, dict[str, Any]],

comfy_api/latest/_io.py

@@ -17,6 +17,7 @@ if TYPE_CHECKING:
     from spandrel import ImageModelDescriptor
     from comfy.clip_vision import ClipVisionModel
     from comfy.clip_vision import Output as ClipVisionOutput_
+    from comfy.bg_removal_model import BackgroundRemovalModel
     from comfy.controlnet import ControlNet
     from comfy.hooks import HookGroup, HookKeyframeGroup
     from comfy.model_patcher import ModelPatcher
@@ -395,7 +396,6 @@ class Combo(ComfyTypeIO):
 @comfytype(io_type="COMBO")
 class MultiCombo(ComfyTypeI):
     '''Multiselect Combo input (dropdown for selecting potentially more than one value).'''
-    # TODO: something is wrong with the serialization, frontend does not recognize it as multiselect
     Type = list[str]
     class Input(Combo.Input):
         def __init__(self, id: str, options: list[str], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
@@ -408,12 +408,14 @@ class MultiCombo(ComfyTypeI):
             self.default: list[str]
 
         def as_dict(self):
-            to_return = super().as_dict() | prune_dict({
-                "multi_select": self.multiselect,
-                "placeholder": self.placeholder,
-                "chip": self.chip,
+            # Frontend expects `multi_select` to be an object config (not a boolean).
+            # Keep top-level `multiselect` from Combo.Input for backwards compatibility.
+            return super().as_dict() | prune_dict({
+                "multi_select": prune_dict({
+                    "placeholder": self.placeholder,
+                    "chip": self.chip,
+                }),
             })
-            return to_return
 
 @comfytype(io_type="IMAGE")
 class Image(ComfyTypeIO):
@@ -613,6 +615,11 @@ class Model(ComfyTypeIO):
     if TYPE_CHECKING:
         Type = ModelPatcher
 
+@comfytype(io_type="BACKGROUND_REMOVAL")
+class BackgroundRemoval(ComfyTypeIO):
+    if TYPE_CHECKING:
+        Type = BackgroundRemovalModel
+
 @comfytype(io_type="CLIP_VISION")
 class ClipVision(ComfyTypeIO):
     if TYPE_CHECKING:
@@ -2256,6 +2263,7 @@ __all__ = [
     "ModelPatch",
     "ClipVision",
     "ClipVisionOutput",
+    "BackgroundRemoval",
     "AudioEncoder",
     "AudioEncoderOutput",
     "StyleModel",

comfy_api_nodes/apis/luma.py

@@ -1,15 +1,12 @@
 from __future__ import annotations
 
-
-import torch
-
 from enum import Enum
 from typing import Optional, Union
 
+import torch
 from pydantic import BaseModel, Field, confloat
 
 
-
 class LumaIO:
     LUMA_REF = "LUMA_REF"
     LUMA_CONCEPTS = "LUMA_CONCEPTS"
@@ -183,13 +180,13 @@ class LumaAssets(BaseModel):
 
 
 class LumaImageRef(BaseModel):
-    '''Used for image gen'''
+    """Used for image gen"""
     url: str = Field(..., description='The URL of the image reference')
     weight: confloat(ge=0.0, le=1.0) = Field(..., description='The weight of the image reference')
 
 
 class LumaImageReference(BaseModel):
-    '''Used for video gen'''
+    """Used for video gen"""
     type: Optional[str] = Field('image', description='Input type, defaults to image')
     url: str = Field(..., description='The URL of the image')
 
@@ -251,3 +248,32 @@ class LumaGeneration(BaseModel):
     assets: Optional[LumaAssets] = Field(None, description='The assets of the generation')
     model: str = Field(..., description='The model used for the generation')
     request: Union[LumaGenerationRequest, LumaImageGenerationRequest] = Field(..., description="The request used for the generation")
+
+
+class Luma2ImageRef(BaseModel):
+    url: str | None = None
+    data: str | None = None
+    media_type: str | None = None
+
+
+class Luma2GenerationRequest(BaseModel):
+    prompt: str = Field(..., min_length=1, max_length=6000)
+    model: str | None = None
+    type: str | None = None
+    aspect_ratio: str | None = None
+    style: str | None = None
+    output_format: str | None = None
+    web_search: bool | None = None
+    image_ref: list[Luma2ImageRef] | None = None
+    source: Luma2ImageRef | None = None
+
+
+class Luma2Generation(BaseModel):
+    id: str | None = None
+    type: str | None = None
+    state: str | None = None
+    model: str | None = None
+    created_at: str | None = None
+    output: list[LumaImageReference] | None = None
+    failure_reason: str | None = None
+    failure_code: str | None = None

comfy_api_nodes/apis/openai.py

@@ -56,14 +56,14 @@ class ModelResponseProperties(BaseModel):
     instructions: str | None = Field(None)
     max_output_tokens: int | None = Field(None)
     model: str | None = Field(None)
-    temperature: float | None = Field(1, description="Controls randomness in the response", ge=0.0, le=2.0)
+    temperature: float | None = Field(None, description="Controls randomness in the response", ge=0.0, le=2.0)
     top_p: float | None = Field(
-        1,
+        None,
         description="Controls diversity of the response via nucleus sampling",
         ge=0.0,
         le=1.0,
     )
-    truncation: str | None = Field("disabled", description="Allowed values: 'auto' or 'disabled'")
+    truncation: str | None = Field(None, description="Allowed values: 'auto' or 'disabled'")
 
 
 class ResponseProperties(BaseModel):

comfy_api_nodes/apis/topaz.py

@@ -1,4 +1,4 @@
-from typing import Optional, Union
+from typing import Optional
 
 from pydantic import BaseModel, Field
 
@@ -72,8 +72,11 @@ class VideoEnhancementFilter(BaseModel):
     grain: Optional[float] = Field(None, description="Grain after AI model processing")
     grainSize: Optional[float] = Field(None, description="Size of generated grain")
     recoverOriginalDetailValue: Optional[float] = Field(None, description="Source details into the output video")
-    creativity: Optional[str] = Field(None, description="Creativity level(high, low) for slc-1 only")
+    creativity: float | str | None = Field(None, description="slc-1/slp-2.5: enum (low/middle/high). ast-2: decimal 0.0-1.0.")
     isOptimizedMode: Optional[bool] = Field(None, description="Set to true for Starlight Creative (slc-1) only")
+    prompt: str | None = Field(None, description="Descriptive scene prompt (ast-2 only)")
+    sharp: float | None = Field(None, description="ast-2 pre-enhance sharpness")
+    realism: float | None = Field(None, description="ast-2 realism control")
 
 
 class OutputInformationVideo(BaseModel):
@@ -90,7 +93,7 @@ class Overrides(BaseModel):
 
 class CreateVideoRequest(BaseModel):
     source: CreateVideoRequestSource = Field(...)
-    filters: list[Union[VideoFrameInterpolationFilter, VideoEnhancementFilter]] = Field(...)
+    filters: list[VideoFrameInterpolationFilter | VideoEnhancementFilter] = Field(...)
     output: OutputInformationVideo = Field(...)
     overrides: Overrides = Field(Overrides(isPaidDiffusion=True))
 

comfy_api_nodes/nodes_bytedance.py

@@ -1271,7 +1271,7 @@ PRICE_BADGE_VIDEO = IO.PriceBadge(
 )
 
 
-def _seedance2_text_inputs(resolutions: list[str]):
+def _seedance2_text_inputs(resolutions: list[str], default_ratio: str = "16:9"):
     return [
         IO.String.Input(
             "prompt",
@@ -1287,6 +1287,7 @@ def _seedance2_text_inputs(resolutions: list[str]):
         IO.Combo.Input(
             "ratio",
             options=["16:9", "4:3", "1:1", "3:4", "9:16", "21:9", "adaptive"],
+            default=default_ratio,
             tooltip="Aspect ratio of the output video.",
         ),
         IO.Int.Input(
@@ -1420,8 +1421,14 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode):
                 IO.DynamicCombo.Input(
                     "model",
                     options=[
-                        IO.DynamicCombo.Option("Seedance 2.0", _seedance2_text_inputs(["480p", "720p", "1080p"])),
-                        IO.DynamicCombo.Option("Seedance 2.0 Fast", _seedance2_text_inputs(["480p", "720p"])),
+                        IO.DynamicCombo.Option(
+                            "Seedance 2.0",
+                            _seedance2_text_inputs(["480p", "720p", "1080p"], default_ratio="adaptive"),
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Seedance 2.0 Fast",
+                            _seedance2_text_inputs(["480p", "720p"], default_ratio="adaptive"),
+                        ),
                     ],
                     tooltip="Seedance 2.0 for maximum quality; Seedance 2.0 Fast for speed optimization.",
                 ),
@@ -1588,9 +1595,9 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode):
         return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
 
 
-def _seedance2_reference_inputs(resolutions: list[str]):
+def _seedance2_reference_inputs(resolutions: list[str], default_ratio: str = "16:9"):
     return [
-        *_seedance2_text_inputs(resolutions),
+        *_seedance2_text_inputs(resolutions, default_ratio=default_ratio),
         IO.Autogrow.Input(
             "reference_images",
             template=IO.Autogrow.TemplateNames(
@@ -1668,8 +1675,14 @@ class ByteDance2ReferenceNode(IO.ComfyNode):
                 IO.DynamicCombo.Input(
                     "model",
                     options=[
-                        IO.DynamicCombo.Option("Seedance 2.0", _seedance2_reference_inputs(["480p", "720p", "1080p"])),
-                        IO.DynamicCombo.Option("Seedance 2.0 Fast", _seedance2_reference_inputs(["480p", "720p"])),
+                        IO.DynamicCombo.Option(
+                            "Seedance 2.0",
+                            _seedance2_reference_inputs(["480p", "720p", "1080p"], default_ratio="adaptive"),
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Seedance 2.0 Fast",
+                            _seedance2_reference_inputs(["480p", "720p"], default_ratio="adaptive"),
+                        ),
                     ],
                     tooltip="Seedance 2.0 for maximum quality; Seedance 2.0 Fast for speed optimization.",
                 ),

comfy_api_nodes/nodes_gemini.py

@@ -83,13 +83,16 @@ class GeminiImageModel(str, Enum):
 
 async def create_image_parts(
     cls: type[IO.ComfyNode],
-    images: Input.Image,
+    images: Input.Image | list[Input.Image],
     image_limit: int = 0,
 ) -> list[GeminiPart]:
     image_parts: list[GeminiPart] = []
     if image_limit < 0:
         raise ValueError("image_limit must be greater than or equal to 0 when creating Gemini image parts.")
-    total_images = get_number_of_images(images)
+
+    # Accept either a single (possibly-batched) tensor or a list of them; share URL budget across all.
+    images_list: list[Input.Image] = images if isinstance(images, list) else [images]
+    total_images = sum(get_number_of_images(img) for img in images_list)
     if total_images <= 0:
         raise ValueError("No images provided to create_image_parts; at least one image is required.")
 
@@ -98,10 +101,18 @@ async def create_image_parts(
 
     # Number of images we'll send as URLs (fileData)
     num_url_images = min(effective_max, 10)  # Vertex API max number of image links
+    upload_kwargs: dict = {"wait_label": "Uploading reference images"}
+    if effective_max > num_url_images:
+        # Split path (e.g. 11+ images): suppress per-image counter to avoid a confusing dual-fraction label.
+        upload_kwargs = {
+            "wait_label": f"Uploading reference images ({num_url_images}+)",
+            "show_batch_index": False,
+        }
     reference_images_urls = await upload_images_to_comfyapi(
         cls,
-        images,
+        images_list,
         max_images=num_url_images,
+        **upload_kwargs,
     )
     for reference_image_url in reference_images_urls:
         image_parts.append(
@@ -112,15 +123,22 @@ async def create_image_parts(
                 )
             )
         )
-    for idx in range(num_url_images, effective_max):
-        image_parts.append(
-            GeminiPart(
-                inlineData=GeminiInlineData(
-                    mimeType=GeminiMimeType.image_png,
-                    data=tensor_to_base64_string(images[idx]),
+    if effective_max > num_url_images:
+        flat: list[torch.Tensor] = []
+        for tensor in images_list:
+            if len(tensor.shape) == 4:
+                flat.extend(tensor[i] for i in range(tensor.shape[0]))
+            else:
+                flat.append(tensor)
+        for idx in range(num_url_images, effective_max):
+            image_parts.append(
+                GeminiPart(
+                    inlineData=GeminiInlineData(
+                        mimeType=GeminiMimeType.image_png,
+                        data=tensor_to_base64_string(flat[idx]),
+                    )
                 )
             )
-        )
     return image_parts
 
 
@@ -891,10 +909,6 @@ class GeminiNanoBanana2(IO.ComfyNode):
                         "9:16",
                         "16:9",
                         "21:9",
-                        # "1:4",
-                        # "4:1",
-                        # "8:1",
-                        # "1:8",
                     ],
                     default="auto",
                     tooltip="If set to 'auto', matches your input image's aspect ratio; "
@@ -902,12 +916,7 @@ class GeminiNanoBanana2(IO.ComfyNode):
                 ),
                 IO.Combo.Input(
                     "resolution",
-                    options=[
-                        # "512px",
-                        "1K",
-                        "2K",
-                        "4K",
-                    ],
+                    options=["1K", "2K", "4K"],
                     tooltip="Target output resolution. For 2K/4K the native Gemini upscaler is used.",
                 ),
                 IO.Combo.Input(
@@ -956,6 +965,7 @@ class GeminiNanoBanana2(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=GEMINI_IMAGE_2_PRICE_BADGE,
+            is_deprecated=True,
         )
 
     @classmethod
@@ -1016,6 +1026,197 @@ class GeminiNanoBanana2(IO.ComfyNode):
         )
 
 
+def _nano_banana_2_v2_model_inputs():
+    return [
+        IO.Combo.Input(
+            "aspect_ratio",
+            options=[
+                "auto",
+                "1:1",
+                "2:3",
+                "3:2",
+                "3:4",
+                "4:3",
+                "4:5",
+                "5:4",
+                "9:16",
+                "16:9",
+                "21:9",
+                "1:4",
+                "4:1",
+                "8:1",
+                "1:8",
+            ],
+            default="auto",
+            tooltip="If set to 'auto', matches your input image's aspect ratio; "
+            "if no image is provided, a 16:9 square is usually generated.",
+        ),
+        IO.Combo.Input(
+            "resolution",
+            options=["1K", "2K", "4K"],
+            tooltip="Target output resolution. For 2K/4K the native Gemini upscaler is used.",
+        ),
+        IO.Combo.Input(
+            "thinking_level",
+            options=["MINIMAL", "HIGH"],
+        ),
+        IO.Autogrow.Input(
+            "images",
+            template=IO.Autogrow.TemplateNames(
+                IO.Image.Input("image"),
+                names=[f"image_{i}" for i in range(1, 15)],
+                min=0,
+            ),
+            tooltip="Optional reference image(s). Up to 14 images total.",
+        ),
+        IO.Custom("GEMINI_INPUT_FILES").Input(
+            "files",
+            optional=True,
+            tooltip="Optional file(s) to use as context for the model. "
+                    "Accepts inputs from the Gemini Generate Content Input Files node.",
+        ),
+    ]
+
+
+class GeminiNanoBanana2V2(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="GeminiNanoBanana2V2",
+            display_name="Nano Banana 2",
+            category="api node/image/Gemini",
+            description="Generate or edit images synchronously via Google Vertex API.",
+            inputs=[
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="Text prompt describing the image to generate or the edits to apply. "
+                    "Include any constraints, styles, or details the model should follow.",
+                    default="",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "Nano Banana 2 (Gemini 3.1 Flash Image)",
+                            _nano_banana_2_v2_model_inputs(),
+                        ),
+                    ],
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=42,
+                    min=0,
+                    max=0xFFFFFFFFFFFFFFFF,
+                    control_after_generate=True,
+                    tooltip="When the 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.",
+                ),
+                IO.Combo.Input(
+                    "response_modalities",
+                    options=["IMAGE", "IMAGE+TEXT"],
+                    advanced=True,
+                ),
+                IO.String.Input(
+                    "system_prompt",
+                    multiline=True,
+                    default=GEMINI_IMAGE_SYS_PROMPT,
+                    optional=True,
+                    tooltip="Foundational instructions that dictate an AI's behavior.",
+                    advanced=True,
+                ),
+            ],
+            outputs=[
+                IO.Image.Output(),
+                IO.String.Output(),
+                IO.Image.Output(
+                    display_name="thought_image",
+                    tooltip="First image from the model's thinking process. "
+                    "Only available with thinking_level HIGH and IMAGE+TEXT modality.",
+                ),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.resolution"]),
+                expr="""
+                (
+                  $r := $lookup(widgets, "model.resolution");
+                  $prices := {"1k": 0.0696, "2k": 0.1014, "4k": 0.154};
+                  {"type":"usd","usd": $lookup($prices, $r), "format":{"suffix":"/Image","approximate":true}}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        prompt: str,
+        model: dict,
+        seed: int,
+        response_modalities: str,
+        system_prompt: str = "",
+    ) -> IO.NodeOutput:
+        validate_string(prompt, strip_whitespace=True, min_length=1)
+        model_choice = model["model"]
+        if model_choice == "Nano Banana 2 (Gemini 3.1 Flash Image)":
+            model_id = "gemini-3.1-flash-image-preview"
+        else:
+            model_id = model_choice
+
+        images = model.get("images") or {}
+        parts: list[GeminiPart] = [GeminiPart(text=prompt)]
+        if images:
+            image_tensors: list[Input.Image] = [t for t in images.values() if t is not None]
+            if image_tensors:
+                if sum(get_number_of_images(t) for t in image_tensors) > 14:
+                    raise ValueError("The current maximum number of supported images is 14.")
+                parts.extend(await create_image_parts(cls, image_tensors))
+        files = model.get("files")
+        if files is not None:
+            parts.extend(files)
+
+        image_config = GeminiImageConfig(imageSize=model["resolution"])
+        if model["aspect_ratio"] != "auto":
+            image_config.aspectRatio = model["aspect_ratio"]
+
+        gemini_system_prompt = None
+        if system_prompt:
+            gemini_system_prompt = GeminiSystemInstructionContent(parts=[GeminiTextPart(text=system_prompt)], role=None)
+
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/vertexai/gemini/{model_id}", method="POST"),
+            data=GeminiImageGenerateContentRequest(
+                contents=[
+                    GeminiContent(role=GeminiRole.user, parts=parts),
+                ],
+                generationConfig=GeminiImageGenerationConfig(
+                    responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]),
+                    imageConfig=image_config,
+                    thinkingConfig=GeminiThinkingConfig(thinkingLevel=model["thinking_level"]),
+                ),
+                systemInstruction=gemini_system_prompt,
+            ),
+            response_model=GeminiGenerateContentResponse,
+            price_extractor=calculate_tokens_price,
+        )
+        return IO.NodeOutput(
+            await get_image_from_response(response),
+            get_text_from_response(response),
+            await get_image_from_response(response, thought=True),
+        )
+
+
 class GeminiExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -1024,6 +1225,7 @@ class GeminiExtension(ComfyExtension):
             GeminiImage,
             GeminiImage2,
             GeminiNanoBanana2,
+            GeminiNanoBanana2V2,
             GeminiInputFiles,
         ]
 

comfy_api_nodes/nodes_grok.py

@@ -54,7 +54,12 @@ class GrokImageNode(IO.ComfyNode):
             inputs=[
                 IO.Combo.Input(
                     "model",
-                    options=["grok-imagine-image-pro", "grok-imagine-image", "grok-imagine-image-beta"],
+                    options=[
+                        "grok-imagine-image-quality",
+                        "grok-imagine-image-pro",
+                        "grok-imagine-image",
+                        "grok-imagine-image-beta",
+                    ],
                 ),
                 IO.String.Input(
                     "prompt",
@@ -100,6 +105,13 @@ class GrokImageNode(IO.ComfyNode):
                     "actual results are nondeterministic regardless of seed.",
                 ),
                 IO.Combo.Input("resolution", options=["1K", "2K"], optional=True),
+                IO.String.Input(
+                    "xai_api_key",
+                    default="",
+                    tooltip="Your xAI API Key (optional). If provided, it will bypass Comfy org limits.",
+                    optional=True,
+                    advanced=True,
+                ),
             ],
             outputs=[
                 IO.Image.Output(),
@@ -111,10 +123,12 @@ class GrokImageNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["model", "number_of_images"]),
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "number_of_images", "resolution"]),
                 expr="""
                 (
-                  $rate := $contains(widgets.model, "pro") ? 0.07 : 0.02;
+                  $rate := widgets.model = "grok-imagine-image-quality"
+                    ? (widgets.resolution = "1k" ? 0.05 : 0.07)
+                    : ($contains(widgets.model, "pro") ? 0.07 : 0.02);
                   {"type":"usd","usd": $rate * widgets.number_of_images}
                 )
                 """,
@@ -130,11 +144,23 @@ class GrokImageNode(IO.ComfyNode):
         number_of_images: int,
         seed: int,
         resolution: str = "1K",
+        xai_api_key: str = "",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
+
+        xai_api_key = xai_api_key.strip()
+        if xai_api_key.lower().startswith("bearer "):
+            xai_api_key = xai_api_key[7:].strip()
+
+        path = "/proxy/xai/v1/images/generations"
+        headers = None
+        if xai_api_key:
+            path = "https://api.x.ai/v1/images/generations"
+            headers = {"Authorization": f"Bearer {xai_api_key}"}
+
         response = await sync_op(
             cls,
-            ApiEndpoint(path="/proxy/xai/v1/images/generations", method="POST"),
+            ApiEndpoint(path=path, method="POST", headers=headers),
             data=ImageGenerationRequest(
                 model=model,
                 prompt=prompt,
@@ -167,7 +193,12 @@ class GrokImageEditNode(IO.ComfyNode):
             inputs=[
                 IO.Combo.Input(
                     "model",
-                    options=["grok-imagine-image-pro", "grok-imagine-image", "grok-imagine-image-beta"],
+                    options=[
+                        "grok-imagine-image-quality",
+                        "grok-imagine-image-pro",
+                        "grok-imagine-image",
+                        "grok-imagine-image-beta",
+                    ],
                 ),
                 IO.Image.Input("image", display_name="images"),
                 IO.String.Input(
@@ -217,6 +248,13 @@ class GrokImageEditNode(IO.ComfyNode):
                     optional=True,
                     tooltip="Only allowed when multiple images are connected to the image input.",
                 ),
+                IO.String.Input(
+                    "xai_api_key",
+                    default="",
+                    tooltip="Your xAI API Key (optional). If provided, it will bypass Comfy org limits.",
+                    optional=True,
+                    advanced=True,
+                ),
             ],
             outputs=[
                 IO.Image.Output(),
@@ -228,11 +266,19 @@ class GrokImageEditNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["model", "number_of_images"]),
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "number_of_images", "resolution"]),
                 expr="""
                 (
-                  $rate := $contains(widgets.model, "pro") ? 0.07 : 0.02;
-                  {"type":"usd","usd": 0.002 + $rate * widgets.number_of_images}
+                  $isQualityModel := widgets.model = "grok-imagine-image-quality";
+                  $isPro := $contains(widgets.model, "pro");
+                  $rate := $isQualityModel
+                    ? (widgets.resolution = "1k" ? 0.05 : 0.07)
+                    : ($isPro ? 0.07 : 0.02);
+                  $base := $isQualityModel ? 0.01 : 0.002;
+                  $output := $rate * widgets.number_of_images;
+                  $isPro
+                    ? {"type":"usd","usd": $base + $output}
+                    : {"type":"range_usd","min_usd": $base + $output, "max_usd": 3 * $base + $output}
                 )
                 """,
             ),
@@ -248,6 +294,7 @@ class GrokImageEditNode(IO.ComfyNode):
         number_of_images: int,
         seed: int,
         aspect_ratio: str = "auto",
+        xai_api_key: str = "",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
         if model == "grok-imagine-image-pro":
@@ -259,9 +306,20 @@ class GrokImageEditNode(IO.ComfyNode):
             raise ValueError(
                 "Custom aspect ratio is only allowed when multiple images are connected to the image input."
             )
+
+        xai_api_key = xai_api_key.strip()
+        if xai_api_key.lower().startswith("bearer "):
+            xai_api_key = xai_api_key[7:].strip()
+
+        path = "/proxy/xai/v1/images/edits"
+        headers = None
+        if xai_api_key:
+            path = "https://api.x.ai/v1/images/edits"
+            headers = {"Authorization": f"Bearer {xai_api_key}"}
+
         response = await sync_op(
             cls,
-            ApiEndpoint(path="/proxy/xai/v1/images/edits", method="POST"),
+            ApiEndpoint(path=path, method="POST", headers=headers),
             data=ImageEditRequest(
                 model=model,
                 images=[InputUrlObject(url=f"data:image/png;base64,{tensor_to_base64_string(i)}") for i in image],
@@ -330,6 +388,13 @@ class GrokVideoNode(IO.ComfyNode):
                     "actual results are nondeterministic regardless of seed.",
                 ),
                 IO.Image.Input("image", optional=True),
+                IO.String.Input(
+                    "xai_api_key",
+                    default="",
+                    tooltip="Your xAI API Key (optional). If provided, it will bypass Comfy org limits.",
+                    optional=True,
+                    advanced=True,
+                ),
             ],
             outputs=[
                 IO.Video.Output(),
@@ -362,6 +427,7 @@ class GrokVideoNode(IO.ComfyNode):
         duration: int,
         seed: int,
         image: Input.Image | None = None,
+        xai_api_key: str = "",
     ) -> IO.NodeOutput:
         if model == "grok-imagine-video-beta":
             model = "grok-imagine-video"
@@ -371,9 +437,20 @@ class GrokVideoNode(IO.ComfyNode):
                 raise ValueError("Only one input image is supported.")
             image_url = InputUrlObject(url=f"data:image/png;base64,{tensor_to_base64_string(image)}")
         validate_string(prompt, strip_whitespace=True, min_length=1)
+
+        xai_api_key = xai_api_key.strip()
+        if xai_api_key.lower().startswith("bearer "):
+            xai_api_key = xai_api_key[7:].strip()
+
+        path = "/proxy/xai/v1/videos/generations"
+        headers = None
+        if xai_api_key:
+            path = "https://api.x.ai/v1/videos/generations"
+            headers = {"Authorization": f"Bearer {xai_api_key}"}
+
         initial_response = await sync_op(
             cls,
-            ApiEndpoint(path="/proxy/xai/v1/videos/generations", method="POST"),
+            ApiEndpoint(path=path, method="POST", headers=headers),
             data=VideoGenerationRequest(
                 model=model,
                 image=image_url,
@@ -385,9 +462,13 @@ class GrokVideoNode(IO.ComfyNode):
             ),
             response_model=VideoGenerationResponse,
         )
+        poll_path = f"/proxy/xai/v1/videos/{initial_response.request_id}"
+        if xai_api_key:
+            poll_path = f"https://api.x.ai/v1/videos/{initial_response.request_id}"
+
         response = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/xai/v1/videos/{initial_response.request_id}"),
+            ApiEndpoint(path=poll_path, headers=headers),
             status_extractor=lambda r: r.status if r.status is not None else "complete",
             response_model=VideoStatusResponse,
             price_extractor=_extract_grok_price,
@@ -423,6 +504,13 @@ class GrokVideoEditNode(IO.ComfyNode):
                     tooltip="Seed to determine if node should re-run; "
                     "actual results are nondeterministic regardless of seed.",
                 ),
+                IO.String.Input(
+                    "xai_api_key",
+                    default="",
+                    tooltip="Your xAI API Key (optional). If provided, it will bypass Comfy org limits.",
+                    optional=True,
+                    advanced=True,
+                ),
             ],
             outputs=[
                 IO.Video.Output(),
@@ -445,6 +533,7 @@ class GrokVideoEditNode(IO.ComfyNode):
         prompt: str,
         video: Input.Video,
         seed: int,
+        xai_api_key: str = "",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
         validate_video_duration(video, min_duration=1, max_duration=8.7)
@@ -452,9 +541,20 @@ class GrokVideoEditNode(IO.ComfyNode):
         video_size = get_fs_object_size(video_stream)
         if video_size > 50 * 1024 * 1024:
             raise ValueError(f"Video size ({video_size / 1024 / 1024:.1f}MB) exceeds 50MB limit.")
+
+        xai_api_key = xai_api_key.strip()
+        if xai_api_key.lower().startswith("bearer "):
+            xai_api_key = xai_api_key[7:].strip()
+
+        path = "/proxy/xai/v1/videos/edits"
+        headers = None
+        if xai_api_key:
+            path = "https://api.x.ai/v1/videos/edits"
+            headers = {"Authorization": f"Bearer {xai_api_key}"}
+
         initial_response = await sync_op(
             cls,
-            ApiEndpoint(path="/proxy/xai/v1/videos/edits", method="POST"),
+            ApiEndpoint(path=path, method="POST", headers=headers),
             data=VideoEditRequest(
                 model=model,
                 video=InputUrlObject(url=await upload_video_to_comfyapi(cls, video)),
@@ -463,9 +563,13 @@ class GrokVideoEditNode(IO.ComfyNode):
             ),
             response_model=VideoGenerationResponse,
         )
+        poll_path = f"/proxy/xai/v1/videos/{initial_response.request_id}"
+        if xai_api_key:
+            poll_path = f"https://api.x.ai/v1/videos/{initial_response.request_id}"
+
         response = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/xai/v1/videos/{initial_response.request_id}"),
+            ApiEndpoint(path=poll_path, headers=headers),
             status_extractor=lambda r: r.status if r.status is not None else "complete",
             response_model=VideoStatusResponse,
             price_extractor=_extract_grok_price,
@@ -539,6 +643,13 @@ class GrokVideoReferenceNode(IO.ComfyNode):
                     tooltip="Seed to determine if node should re-run; "
                     "actual results are nondeterministic regardless of seed.",
                 ),
+                IO.String.Input(
+                    "xai_api_key",
+                    default="",
+                    tooltip="Your xAI API Key (optional). If provided, it will bypass Comfy org limits.",
+                    optional=True,
+                    advanced=True,
+                ),
             ],
             outputs=[
                 IO.Video.Output(),
@@ -573,8 +684,18 @@ class GrokVideoReferenceNode(IO.ComfyNode):
         prompt: str,
         model: dict,
         seed: int,
+        xai_api_key: str = "",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
+
+        xai_api_key = xai_api_key.strip()
+        if xai_api_key.lower().startswith("bearer "):
+            xai_api_key = xai_api_key[7:].strip()
+
+        # We must use proxy to upload images temporarily even if they provide their own key for video generation
+        # because the API requires URLs and we use our proxy for image hosting during the request.
+        # Wait, if they are providing their own key to our backend for generation,
+        # `upload_images_to_comfyapi` relies on `comfyapi`. This is fine.
         ref_image_urls = await upload_images_to_comfyapi(
             cls,
             list(model["reference_images"].values()),
@@ -582,9 +703,16 @@ class GrokVideoReferenceNode(IO.ComfyNode):
             wait_label="Uploading base images",
             max_images=7,
         )
+
+        path = "/proxy/xai/v1/videos/generations"
+        headers = None
+        if xai_api_key:
+            path = "https://api.x.ai/v1/videos/generations"
+            headers = {"Authorization": f"Bearer {xai_api_key}"}
+
         initial_response = await sync_op(
             cls,
-            ApiEndpoint(path="/proxy/xai/v1/videos/generations", method="POST"),
+            ApiEndpoint(path=path, method="POST", headers=headers),
             data=VideoGenerationRequest(
                 model=model["model"],
                 reference_images=[InputUrlObject(url=i) for i in ref_image_urls],
@@ -596,9 +724,13 @@ class GrokVideoReferenceNode(IO.ComfyNode):
             ),
             response_model=VideoGenerationResponse,
         )
+        poll_path = f"/proxy/xai/v1/videos/{initial_response.request_id}"
+        if xai_api_key:
+            poll_path = f"https://api.x.ai/v1/videos/{initial_response.request_id}"
+
         response = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/xai/v1/videos/{initial_response.request_id}"),
+            ApiEndpoint(path=poll_path, headers=headers),
             status_extractor=lambda r: r.status if r.status is not None else "complete",
             response_model=VideoStatusResponse,
             price_extractor=_extract_grok_video_price,
@@ -653,6 +785,13 @@ class GrokVideoExtendNode(IO.ComfyNode):
                     tooltip="Seed to determine if node should re-run; "
                     "actual results are nondeterministic regardless of seed.",
                 ),
+                IO.String.Input(
+                    "xai_api_key",
+                    default="",
+                    tooltip="Your xAI API Key (optional). If provided, it will bypass Comfy org limits.",
+                    optional=True,
+                    advanced=True,
+                ),
             ],
             outputs=[
                 IO.Video.Output(),
@@ -685,15 +824,27 @@ class GrokVideoExtendNode(IO.ComfyNode):
         video: Input.Video,
         model: dict,
         seed: int,
+        xai_api_key: str = "",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
         validate_video_duration(video, min_duration=2, max_duration=15)
         video_size = get_fs_object_size(video.get_stream_source())
         if video_size > 50 * 1024 * 1024:
             raise ValueError(f"Video size ({video_size / 1024 / 1024:.1f}MB) exceeds 50MB limit.")
+
+        xai_api_key = xai_api_key.strip()
+        if xai_api_key.lower().startswith("bearer "):
+            xai_api_key = xai_api_key[7:].strip()
+
+        path = "/proxy/xai/v1/videos/extensions"
+        headers = None
+        if xai_api_key:
+            path = "https://api.x.ai/v1/videos/extensions"
+            headers = {"Authorization": f"Bearer {xai_api_key}"}
+
         initial_response = await sync_op(
             cls,
-            ApiEndpoint(path="/proxy/xai/v1/videos/extensions", method="POST"),
+            ApiEndpoint(path=path, method="POST", headers=headers),
             data=VideoExtensionRequest(
                 prompt=prompt,
                 video=InputUrlObject(url=await upload_video_to_comfyapi(cls, video)),
@@ -701,9 +852,13 @@ class GrokVideoExtendNode(IO.ComfyNode):
             ),
             response_model=VideoGenerationResponse,
         )
+        poll_path = f"/proxy/xai/v1/videos/{initial_response.request_id}"
+        if xai_api_key:
+            poll_path = f"https://api.x.ai/v1/videos/{initial_response.request_id}"
+
         response = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/xai/v1/videos/{initial_response.request_id}"),
+            ApiEndpoint(path=poll_path, headers=headers),
             status_extractor=lambda r: r.status if r.status is not None else "complete",
             response_model=VideoStatusResponse,
             price_extractor=_extract_grok_video_price,

comfy_api_nodes/nodes_kling.py

@@ -2787,11 +2787,15 @@ class MotionControl(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["mode"]),
+                depends_on=IO.PriceBadgeDepends(widgets=["mode", "model"]),
                 expr="""
                 (
-                  $prices := {"std": 0.07, "pro": 0.112};
-                  {"type":"usd","usd": $lookup($prices, widgets.mode), "format":{"suffix":"/second"}}
+                  $prices := {
+                    "kling-v3": {"std": 0.126, "pro": 0.168},
+                    "kling-v2-6": {"std": 0.07, "pro": 0.112}
+                  };
+                  $modelPrices := $lookup($prices, widgets.model);
+                  {"type":"usd","usd": $lookup($modelPrices, widgets.mode), "format":{"suffix":"/second"}}
                 )
                 """,
             ),

comfy_api_nodes/nodes_luma.py

@@ -1,10 +1,11 @@
-from typing import Optional
-
 import torch
 from typing_extensions import override
 
-from comfy_api.latest import IO, ComfyExtension
+from comfy_api.latest import IO, ComfyExtension, Input
 from comfy_api_nodes.apis.luma import (
+    Luma2Generation,
+    Luma2GenerationRequest,
+    Luma2ImageRef,
     LumaAspectRatio,
     LumaCharacterRef,
     LumaConceptChain,
@@ -30,6 +31,7 @@ from comfy_api_nodes.util import (
     download_url_to_video_output,
     poll_op,
     sync_op,
+    upload_image_to_comfyapi,
     upload_images_to_comfyapi,
     validate_string,
 )
@@ -212,9 +214,9 @@ class LumaImageGenerationNode(IO.ComfyNode):
         aspect_ratio: str,
         seed,
         style_image_weight: float,
-        image_luma_ref: Optional[LumaReferenceChain] = None,
-        style_image: Optional[torch.Tensor] = None,
-        character_image: Optional[torch.Tensor] = None,
+        image_luma_ref: LumaReferenceChain | None = None,
+        style_image: torch.Tensor | None = None,
+        character_image: torch.Tensor | None = None,
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=3)
         # handle image_luma_ref
@@ -434,7 +436,7 @@ class LumaTextToVideoGenerationNode(IO.ComfyNode):
         duration: str,
         loop: bool,
         seed,
-        luma_concepts: Optional[LumaConceptChain] = None,
+        luma_concepts: LumaConceptChain | None = None,
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=False, min_length=3)
         duration = duration if model != LumaVideoModel.ray_1_6 else None
@@ -533,7 +535,6 @@ class LumaImageToVideoGenerationNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=PRICE_BADGE_VIDEO,
-
         )
 
     @classmethod
@@ -644,6 +645,293 @@ PRICE_BADGE_VIDEO = IO.PriceBadge(
 )
 
 
+def _luma2_uni1_common_inputs(max_image_refs: int) -> list:
+    return [
+        IO.Combo.Input(
+            "style",
+            options=["auto", "manga"],
+            default="auto",
+            tooltip="Style preset. 'auto' picks based on the prompt; "
+            "'manga' applies a manga/anime aesthetic and requires a portrait "
+            "aspect ratio (2:3, 9:16, 1:2, 1:3).",
+        ),
+        IO.Boolean.Input(
+            "web_search",
+            default=False,
+            tooltip="Search the web for visual references before generating.",
+        ),
+        IO.Autogrow.Input(
+            "image_ref",
+            template=IO.Autogrow.TemplateNames(
+                IO.Image.Input("image"),
+                names=[f"image_{i}" for i in range(1, max_image_refs + 1)],
+                min=0,
+            ),
+            optional=True,
+            tooltip=f"Up to {max_image_refs} reference images for style/content guidance.",
+        ),
+    ]
+
+
+async def _luma2_upload_image_refs(
+    cls: type[IO.ComfyNode],
+    refs: dict | None,
+    max_count: int,
+) -> list[Luma2ImageRef] | None:
+    if not refs:
+        return None
+    out: list[Luma2ImageRef] = []
+    for key in refs:
+        url = await upload_image_to_comfyapi(cls, refs[key])
+        out.append(Luma2ImageRef(url=url))
+    if len(out) > max_count:
+        raise ValueError(f"Maximum {max_count} reference images are allowed.")
+    return out or None
+
+
+async def _luma2_submit_and_poll(
+    cls: type[IO.ComfyNode],
+    request: Luma2GenerationRequest,
+) -> Input.Image:
+    initial = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/luma_2/generations", method="POST"),
+        response_model=Luma2Generation,
+        data=request,
+    )
+    if not initial.id:
+        raise RuntimeError("Luma 2 API did not return a generation id.")
+    final = await poll_op(
+        cls,
+        ApiEndpoint(path=f"/proxy/luma_2/generations/{initial.id}", method="GET"),
+        response_model=Luma2Generation,
+        status_extractor=lambda r: r.state,
+        progress_extractor=lambda r: None,
+    )
+    if not final.output:
+        msg = final.failure_reason or "no output returned"
+        raise RuntimeError(f"Luma 2 generation failed: {msg}")
+    url = final.output[0].url
+    if not url:
+        raise RuntimeError("Luma 2 generation completed without an output URL.")
+    return await download_url_to_image_tensor(url)
+
+
+class LumaImageNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="LumaImageNode2",
+            display_name="Luma UNI-1 Image",
+            category="api node/image/Luma",
+            description="Generate images from text using the Luma UNI-1 model.",
+            inputs=[
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Text description of the desired image. 1–6000 characters.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "uni-1",
+                            [
+                                IO.Combo.Input(
+                                    "aspect_ratio",
+                                    options=[
+                                        "auto",
+                                        "3:1",
+                                        "2:1",
+                                        "16:9",
+                                        "3:2",
+                                        "1:1",
+                                        "2:3",
+                                        "9:16",
+                                        "1:2",
+                                        "1:3",
+                                    ],
+                                    default="auto",
+                                    tooltip="Output image aspect ratio. 'auto' lets "
+                                    "the model pick based on the prompt.",
+                                ),
+                                *_luma2_uni1_common_inputs(max_image_refs=9),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "uni-1-max",
+                            [
+                                IO.Combo.Input(
+                                    "aspect_ratio",
+                                    options=[
+                                        "auto",
+                                        "3:1",
+                                        "2:1",
+                                        "16:9",
+                                        "3:2",
+                                        "1:1",
+                                        "2:3",
+                                        "9:16",
+                                        "1:2",
+                                        "1:3",
+                                    ],
+                                    default="auto",
+                                    tooltip="Output image aspect ratio. 'auto' lets "
+                                    "the model pick based on the prompt.",
+                                ),
+                                *_luma2_uni1_common_inputs(max_image_refs=9),
+                            ],
+                        ),
+                    ],
+                    tooltip="Model to use for generation.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+            ],
+            outputs=[IO.Image.Output()],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model"], input_groups=["model.image_ref"]),
+                expr="""
+                (
+                  $m := widgets.model;
+                  $refs := $lookup(inputGroups, "model.image_ref");
+                  $base := $m = "uni-1-max" ? 0.1 : 0.0404;
+                  {"type":"usd","usd": $round($base + 0.003 * $refs, 4)}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        prompt: str,
+        model: dict,
+        seed: int,
+    ) -> IO.NodeOutput:
+        validate_string(prompt, min_length=1, max_length=6000)
+        aspect_ratio = model["aspect_ratio"]
+        style = model["style"]
+        allowed_manga_ratios = {"2:3", "9:16", "1:2", "1:3"}
+        if style == "manga" and aspect_ratio != "auto" and aspect_ratio not in allowed_manga_ratios:
+            raise ValueError(
+                f"'manga' style requires a portrait aspect ratio "
+                f"({', '.join(sorted(allowed_manga_ratios))}) or 'auto'; got '{aspect_ratio}'."
+            )
+        request = Luma2GenerationRequest(
+            prompt=prompt,
+            model=model["model"],
+            type="image",
+            aspect_ratio=aspect_ratio if aspect_ratio != "auto" else None,
+            style=style if style != "auto" else None,
+            output_format="png",
+            web_search=model["web_search"],
+            image_ref=await _luma2_upload_image_refs(cls, model.get("image_ref"), max_count=9),
+        )
+        return IO.NodeOutput(await _luma2_submit_and_poll(cls, request))
+
+
+class LumaImageEditNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="LumaImageEditNode2",
+            display_name="Luma UNI-1 Image Edit",
+            category="api node/image/Luma",
+            description="Edit an existing image with a text prompt using the Luma UNI-1 model.",
+            inputs=[
+                IO.Image.Input(
+                    "source",
+                    tooltip="Source image to edit.",
+                ),
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Description of the desired edit. 1–6000 characters.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "uni-1",
+                            _luma2_uni1_common_inputs(max_image_refs=8),
+                        ),
+                        IO.DynamicCombo.Option(
+                            "uni-1-max",
+                            _luma2_uni1_common_inputs(max_image_refs=8),
+                        ),
+                    ],
+                    tooltip="Model to use for editing.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+            ],
+            outputs=[IO.Image.Output()],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model"], input_groups=["model.image_ref"]),
+                expr="""
+                (
+                  $m := widgets.model;
+                  $refs := $lookup(inputGroups, "model.image_ref");
+                  $base := $m = "uni-1-max" ? 0.103 : 0.0434;
+                  {"type":"usd","usd": $round($base + 0.003 * $refs, 4)}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        source: Input.Image,
+        prompt: str,
+        model: dict,
+        seed: int,
+    ) -> IO.NodeOutput:
+        validate_string(prompt, min_length=1, max_length=6000)
+        request = Luma2GenerationRequest(
+            prompt=prompt,
+            model=model["model"],
+            type="image_edit",
+            source=Luma2ImageRef(url=await upload_image_to_comfyapi(cls, source)),
+            style=model["style"] if model["style"] != "auto" else None,
+            output_format="png",
+            web_search=model["web_search"],
+            image_ref=await _luma2_upload_image_refs(cls, model.get("image_ref"), max_count=8),
+        )
+        return IO.NodeOutput(await _luma2_submit_and_poll(cls, request))
+
+
 class LumaExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -654,6 +942,8 @@ class LumaExtension(ComfyExtension):
             LumaImageToVideoGenerationNode,
             LumaReferenceNode,
             LumaConceptsNode,
+            LumaImageNode,
+            LumaImageEditNode,
         ]
 
 

comfy_api_nodes/nodes_openai.py

@@ -39,16 +39,18 @@ STARTING_POINT_ID_PATTERN = r"<starting_point_id:(.*)>"
 
 
 class SupportedOpenAIModel(str, Enum):
-    o4_mini = "o4-mini"
-    o1 = "o1"
-    o3 = "o3"
-    o1_pro = "o1-pro"
-    gpt_4_1 = "gpt-4.1"
-    gpt_4_1_mini = "gpt-4.1-mini"
-    gpt_4_1_nano = "gpt-4.1-nano"
+    gpt_5_5_pro = "gpt-5.5-pro"
+    gpt_5_5 = "gpt-5.5"
     gpt_5 = "gpt-5"
     gpt_5_mini = "gpt-5-mini"
     gpt_5_nano = "gpt-5-nano"
+    gpt_4_1 = "gpt-4.1"
+    gpt_4_1_mini = "gpt-4.1-mini"
+    gpt_4_1_nano = "gpt-4.1-nano"
+    o4_mini = "o4-mini"
+    o3 = "o3"
+    o1_pro = "o1-pro"
+    o1 = "o1"
 
 
 async def validate_and_cast_response(response, timeout: int = None) -> torch.Tensor:
@@ -739,6 +741,16 @@ class OpenAIChatNode(IO.ComfyNode):
                     "usd": [0.002, 0.008],
                     "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" }
                   }
+                  : $contains($m, "gpt-5.5-pro") ? {
+                    "type": "list_usd",
+                    "usd": [0.03, 0.18],
+                    "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" }
+                  }
+                  : $contains($m, "gpt-5.5") ? {
+                    "type": "list_usd",
+                    "usd": [0.005, 0.03],
+                    "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" }
+                  }
                   : $contains($m, "gpt-5-nano") ? {
                     "type": "list_usd",
                     "usd": [0.00005, 0.0004],

comfy_api_nodes/nodes_sora.py

@@ -33,7 +33,7 @@ class OpenAIVideoSora2(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="OpenAIVideoSora2",
-            display_name="OpenAI Sora - Video (Deprecated)",
+            display_name="OpenAI Sora - Video (DEPRECATED)",
             category="api node/video/Sora",
             description=(
                 "OpenAI video and audio generation.\n\n"

comfy_api_nodes/nodes_topaz.py

@@ -36,11 +36,15 @@ from comfy_api_nodes.util import (
 )
 
 UPSCALER_MODELS_MAP = {
+    "Astra 2": "ast-2",
     "Starlight (Astra) Fast": "slf-1",
     "Starlight (Astra) Creative": "slc-1",
     "Starlight Precise 2.5": "slp-2.5",
 }
 
+AST2_MAX_FRAMES = 9000
+AST2_MAX_FRAMES_WITH_PROMPT = 450
+
 
 class TopazImageEnhance(IO.ComfyNode):
     @classmethod
@@ -230,13 +234,20 @@ class TopazVideoEnhance(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="TopazVideoEnhance",
-            display_name="Topaz Video Enhance",
+            display_name="Topaz Video Enhance (Legacy)",
             category="api node/video/Topaz",
             description="Breathe new life into video with powerful upscaling and recovery technology.",
             inputs=[
                 IO.Video.Input("video"),
                 IO.Boolean.Input("upscaler_enabled", default=True),
-                IO.Combo.Input("upscaler_model", options=list(UPSCALER_MODELS_MAP.keys())),
+                IO.Combo.Input(
+                    "upscaler_model",
+                    options=[
+                        "Starlight (Astra) Fast",
+                        "Starlight (Astra) Creative",
+                        "Starlight Precise 2.5",
+                    ],
+                ),
                 IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),
                 IO.Combo.Input(
                     "upscaler_creativity",
@@ -304,6 +315,7 @@ class TopazVideoEnhance(IO.ComfyNode):
                 IO.Hidden.unique_id,
             ],
             is_api_node=True,
+            is_deprecated=True,
         )
 
     @classmethod
@@ -457,12 +469,357 @@ class TopazVideoEnhance(IO.ComfyNode):
         return IO.NodeOutput(await download_url_to_video_output(final_response.download.url))
 
 
+class TopazVideoEnhanceV2(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="TopazVideoEnhanceV2",
+            display_name="Topaz Video Enhance",
+            category="api node/video/Topaz",
+            description="Breathe new life into video with powerful upscaling and recovery technology.",
+            inputs=[
+                IO.Video.Input("video"),
+                IO.DynamicCombo.Input(
+                    "upscaler_model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "Astra 2",
+                            [
+                                IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),
+                                IO.Float.Input(
+                                    "creativity",
+                                    default=0.5,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Creative strength of the upscale.",
+                                ),
+                                IO.String.Input(
+                                    "prompt",
+                                    multiline=True,
+                                    default="",
+                                    tooltip="Optional descriptive (not instructive) scene prompt."
+                                    f"Capping input at {AST2_MAX_FRAMES_WITH_PROMPT} frames (~15s @ 30fps) when set.",
+                                ),
+                                IO.Float.Input(
+                                    "sharp",
+                                    default=0.5,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Pre-enhance sharpness: "
+                                    "0.0=Gaussian blur, 0.5=passthrough (default), 1.0=USM sharpening.",
+                                    advanced=True,
+                                ),
+                                IO.Float.Input(
+                                    "realism",
+                                    default=0.0,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Pulls output toward photographic realism."
+                                    "Leave at 0 for the model default.",
+                                    advanced=True,
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Starlight (Astra) Fast",
+                            [IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Starlight (Astra) Creative",
+                            [
+                                IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),
+                                IO.Combo.Input(
+                                    "creativity",
+                                    options=["low", "middle", "high"],
+                                    default="low",
+                                    tooltip="Creative strength of the upscale.",
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "Starlight Precise 2.5",
+                            [IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"])],
+                        ),
+                        IO.DynamicCombo.Option("Disabled", []),
+                    ],
+                ),
+                IO.DynamicCombo.Input(
+                    "interpolation_model",
+                    options=[
+                        IO.DynamicCombo.Option("Disabled", []),
+                        IO.DynamicCombo.Option(
+                            "apo-8",
+                            [
+                                IO.Int.Input(
+                                    "interpolation_frame_rate",
+                                    default=60,
+                                    min=15,
+                                    max=240,
+                                    display_mode=IO.NumberDisplay.number,
+                                    tooltip="Output frame rate.",
+                                ),
+                                IO.Int.Input(
+                                    "interpolation_slowmo",
+                                    default=1,
+                                    min=1,
+                                    max=16,
+                                    display_mode=IO.NumberDisplay.number,
+                                    tooltip="Slow-motion factor applied to the input video. "
+                                    "For example, 2 makes the output twice as slow and doubles the duration.",
+                                    advanced=True,
+                                ),
+                                IO.Boolean.Input(
+                                    "interpolation_duplicate",
+                                    default=False,
+                                    tooltip="Analyze the input for duplicate frames and remove them.",
+                                    advanced=True,
+                                ),
+                                IO.Float.Input(
+                                    "interpolation_duplicate_threshold",
+                                    default=0.01,
+                                    min=0.001,
+                                    max=0.1,
+                                    step=0.001,
+                                    display_mode=IO.NumberDisplay.number,
+                                    tooltip="Detection sensitivity for duplicate frames.",
+                                    advanced=True,
+                                ),
+                            ],
+                        ),
+                    ],
+                ),
+                IO.Combo.Input(
+                    "dynamic_compression_level",
+                    options=["Low", "Mid", "High"],
+                    default="Low",
+                    tooltip="CQP level.",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=[
+                    "upscaler_model",
+                    "upscaler_model.upscaler_resolution",
+                    "interpolation_model",
+                ]),
+                expr="""
+                (
+                  $model := $lookup(widgets, "upscaler_model");
+                  $res := $lookup(widgets, "upscaler_model.upscaler_resolution");
+                  $interp := $lookup(widgets, "interpolation_model");
+                  $is4k := $contains($res, "4k");
+                  $hasInterp := $interp != "disabled";
+                  $rates := {
+                    "starlight (astra) fast":     {"hd": 0.43, "uhd": 0.85},
+                    "starlight precise 2.5":      {"hd": 0.70, "uhd": 1.54},
+                    "astra 2":                    {"hd": 1.72, "uhd": 2.85},
+                    "starlight (astra) creative": {"hd": 2.25, "uhd": 3.99}
+                  };
+                  $surcharge := $is4k ? 0.28 : 0.14;
+                  $entry := $lookup($rates, $model);
+                  $base := $is4k ? $entry.uhd : $entry.hd;
+                  $hi := $base + ($hasInterp ? $surcharge : 0);
+                  $model = "disabled"
+                    ? {"type":"text","text":"Interpolation only"}
+                    : ($hasInterp
+                        ? {"type":"text","text":"~" & $string($base) & "–" & $string($hi) & " credits/src frame"}
+                        : {"type":"text","text":"~" & $string($base) & " credits/src frame"})
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        video: Input.Video,
+        upscaler_model: dict,
+        interpolation_model: dict,
+        dynamic_compression_level: str = "Low",
+    ) -> IO.NodeOutput:
+        upscaler_choice = upscaler_model["upscaler_model"]
+        interpolation_choice = interpolation_model["interpolation_model"]
+        if upscaler_choice == "Disabled" and interpolation_choice == "Disabled":
+            raise ValueError("There is nothing to do: both upscaling and interpolation are disabled.")
+        validate_container_format_is_mp4(video)
+        src_width, src_height = video.get_dimensions()
+        src_frame_rate = int(video.get_frame_rate())
+        duration_sec = video.get_duration()
+        src_video_stream = video.get_stream_source()
+        target_width = src_width
+        target_height = src_height
+        target_frame_rate = src_frame_rate
+        filters = []
+        if upscaler_choice != "Disabled":
+            if "1080p" in upscaler_model["upscaler_resolution"]:
+                target_pixel_p = 1080
+                max_long_side = 1920
+            else:
+                target_pixel_p = 2160
+                max_long_side = 3840
+            ar = src_width / src_height
+            if src_width >= src_height:
+                # Landscape or Square; Attempt to set height to target (e.g., 2160), calculate width
+                target_height = target_pixel_p
+                target_width = int(target_height * ar)
+                # Check if width exceeds standard bounds (for ultra-wide e.g., 21:9 ARs)
+                if target_width > max_long_side:
+                    target_width = max_long_side
+                    target_height = int(target_width / ar)
+            else:
+                # Portrait; Attempt to set width to target (e.g., 2160), calculate height
+                target_width = target_pixel_p
+                target_height = int(target_width / ar)
+                # Check if height exceeds standard bounds
+                if target_height > max_long_side:
+                    target_height = max_long_side
+                    target_width = int(target_height * ar)
+            if target_width % 2 != 0:
+                target_width += 1
+            if target_height % 2 != 0:
+                target_height += 1
+            model_id = UPSCALER_MODELS_MAP[upscaler_choice]
+            if model_id == "slc-1":
+                filters.append(
+                    VideoEnhancementFilter(
+                        model=model_id,
+                        creativity=upscaler_model["creativity"],
+                        isOptimizedMode=True,
+                    )
+                )
+            elif model_id == "ast-2":
+                n_frames = video.get_frame_count()
+                ast2_prompt = (upscaler_model["prompt"] or "").strip()
+                if ast2_prompt and n_frames > AST2_MAX_FRAMES_WITH_PROMPT:
+                    raise ValueError(
+                        f"Astra 2 with a prompt is limited to {AST2_MAX_FRAMES_WITH_PROMPT} input frames "
+                        f"(~15s @ 30fps); video has {n_frames}. Clear the prompt or shorten the clip."
+                    )
+                if n_frames > AST2_MAX_FRAMES:
+                    raise ValueError(f"Astra 2 is limited to {AST2_MAX_FRAMES} input frames; video has {n_frames}.")
+                realism = upscaler_model["realism"]
+                filters.append(
+                    VideoEnhancementFilter(
+                        model=model_id,
+                        creativity=upscaler_model["creativity"],
+                        prompt=(ast2_prompt or None),
+                        sharp=upscaler_model["sharp"],
+                        realism=(realism if realism > 0 else None),
+                    )
+                )
+            else:
+                filters.append(VideoEnhancementFilter(model=model_id))
+        if interpolation_choice != "Disabled":
+            target_frame_rate = interpolation_model["interpolation_frame_rate"]
+            filters.append(
+                VideoFrameInterpolationFilter(
+                    model=interpolation_choice,
+                    slowmo=interpolation_model["interpolation_slowmo"],
+                    fps=interpolation_model["interpolation_frame_rate"],
+                    duplicate=interpolation_model["interpolation_duplicate"],
+                    duplicate_threshold=interpolation_model["interpolation_duplicate_threshold"],
+                ),
+            )
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/topaz/video/", method="POST"),
+            response_model=CreateVideoResponse,
+            data=CreateVideoRequest(
+                source=CreateVideoRequestSource(
+                    container="mp4",
+                    size=get_fs_object_size(src_video_stream),
+                    duration=int(duration_sec),
+                    frameCount=video.get_frame_count(),
+                    frameRate=src_frame_rate,
+                    resolution=Resolution(width=src_width, height=src_height),
+                ),
+                filters=filters,
+                output=OutputInformationVideo(
+                    resolution=Resolution(width=target_width, height=target_height),
+                    frameRate=target_frame_rate,
+                    audioCodec="AAC",
+                    audioTransfer="Copy",
+                    dynamicCompressionLevel=dynamic_compression_level,
+                ),
+            ),
+            wait_label="Creating task",
+            final_label_on_success="Task created",
+        )
+        upload_res = await sync_op(
+            cls,
+            ApiEndpoint(
+                path=f"/proxy/topaz/video/{initial_res.requestId}/accept",
+                method="PATCH",
+            ),
+            response_model=VideoAcceptResponse,
+            wait_label="Preparing upload",
+            final_label_on_success="Upload started",
+        )
+        if len(upload_res.urls) > 1:
+            raise NotImplementedError(
+                "Large files are not currently supported. Please open an issue in the ComfyUI repository."
+            )
+        async with aiohttp.ClientSession(headers={"Content-Type": "video/mp4"}) as session:
+            if isinstance(src_video_stream, BytesIO):
+                src_video_stream.seek(0)
+                async with session.put(upload_res.urls[0], data=src_video_stream, raise_for_status=True) as res:
+                    upload_etag = res.headers["Etag"]
+            else:
+                with builtins.open(src_video_stream, "rb") as video_file:
+                    async with session.put(upload_res.urls[0], data=video_file, raise_for_status=True) as res:
+                        upload_etag = res.headers["Etag"]
+        await sync_op(
+            cls,
+            ApiEndpoint(
+                path=f"/proxy/topaz/video/{initial_res.requestId}/complete-upload",
+                method="PATCH",
+            ),
+            response_model=VideoCompleteUploadResponse,
+            data=VideoCompleteUploadRequest(
+                uploadResults=[
+                    VideoCompleteUploadRequestPart(
+                        partNum=1,
+                        eTag=upload_etag,
+                    ),
+                ],
+            ),
+            wait_label="Finalizing upload",
+            final_label_on_success="Upload completed",
+        )
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/topaz/video/{initial_res.requestId}/status"),
+            response_model=VideoStatusResponse,
+            status_extractor=lambda x: x.status,
+            progress_extractor=lambda x: getattr(x, "progress", 0),
+            price_extractor=lambda x: (x.estimates.cost[0] * 0.08 if x.estimates and x.estimates.cost[0] else None),
+            poll_interval=10.0,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(final_response.download.url))
+
+
 class TopazExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
         return [
             TopazImageEnhance,
             TopazVideoEnhance,
+            TopazVideoEnhanceV2,
         ]
 
 

comfy_api_nodes/util/client.py

@@ -19,6 +19,8 @@ from comfy import utils
 from comfy_api.latest import IO
 from server import PromptServer
 
+from comfy.deploy_environment import get_deploy_environment
+
 from . import request_logger
 from ._helpers import (
     default_base_url,
@@ -624,6 +626,7 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool):
         payload_headers = {"Accept": "*/*"} if expect_binary else {"Accept": "application/json"}
         if not parsed_url.scheme and not parsed_url.netloc:  # is URL relative?
             payload_headers.update(get_auth_header(cfg.node_cls))
+            payload_headers["Comfy-Env"] = get_deploy_environment()
         if cfg.endpoint.headers:
             payload_headers.update(cfg.endpoint.headers)
 

comfy_extras/frame_interpolation_models/film_net.py

@@ -199,6 +199,9 @@ class FILMNet(nn.Module):
     def get_dtype(self):
         return self.extract.extract_sublevels.convs[0][0].conv.weight.dtype
 
+    def memory_used_forward(self, shape, dtype):
+        return 1700 * shape[1] * shape[2] * dtype.itemsize
+
     def _build_warp_grids(self, H, W, device):
         """Pre-compute warp grids for all pyramid levels."""
         if (H, W) in self._warp_grids:

comfy_extras/frame_interpolation_models/ifnet.py

@@ -74,6 +74,9 @@ class IFNet(nn.Module):
     def get_dtype(self):
         return self.encode.cnn0.weight.dtype
 
+    def memory_used_forward(self, shape, dtype):
+        return 300 * shape[1] * shape[2] * dtype.itemsize
+
     def _build_warp_grids(self, H, W, device):
         if (H, W) in self._warp_grids:
             return

comfy_extras/nodes_ace.py

@@ -42,7 +42,7 @@ class TextEncodeAceStepAudio15(IO.ComfyNode):
                 IO.Int.Input("bpm", default=120, min=10, max=300),
                 IO.Float.Input("duration", default=120.0, min=0.0, max=2000.0, step=0.1),
                 IO.Combo.Input("timesignature", options=['2', '3', '4', '6']),
-                IO.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]),
+                IO.Combo.Input("language", options=['ar', 'az', 'bg', 'bn', 'ca', 'cs', 'da', 'de', 'el', 'en', 'es', 'fa', 'fi', 'fr', 'he', 'hi', 'hr', 'ht', 'hu', 'id', 'is', 'it', 'ja', 'ko', 'la', 'lt', 'ms', 'ne', 'nl', 'no', 'pa', 'pl', 'pt', 'ro', 'ru', 'sa', 'sk', 'sr', 'sv', 'sw', 'ta', 'te', 'th', 'tl', 'tr', 'uk', 'ur', 'vi', 'yue', 'zh', 'unknown'], default='en'),
                 IO.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]),
                 IO.Boolean.Input("generate_audio_codes", default=True, tooltip="Enable the LLM that generates audio codes. This can be slow but will increase the quality of the generated audio. Turn this off if you are giving the model an audio reference.", advanced=True),
                 IO.Float.Input("cfg_scale", default=2.0, min=0.0, max=100.0, step=0.1, advanced=True),

comfy_extras/nodes_advanced_samplers.py

@@ -92,7 +92,7 @@ class SamplerEulerCFGpp(io.ComfyNode):
         return io.Schema(
             node_id="SamplerEulerCFGpp",
             display_name="SamplerEulerCFG++",
-            category="_for_testing",  # "sampling/custom_sampling/samplers"
+            category="experimental",  # "sampling/custom_sampling/samplers"
             inputs=[
                 io.Combo.Input("version", options=["regular", "alternative"], advanced=True),
             ],

comfy_extras/nodes_ar_video.py

@@ -0,0 +1,136 @@
+"""
+ComfyUI nodes for autoregressive video generation (Causal Forcing, Self-Forcing, etc.).
+  - EmptyARVideoLatent: create 5D [B, C, T, H, W] video latent tensors
+  - SamplerARVideo: SAMPLER for the block-by-block autoregressive denoising loop
+  - ARVideoI2V: image-to-video conditioning for AR models (seeds KV cache with start image)
+"""
+
+import torch
+from typing_extensions import override
+
+import comfy.model_management
+import comfy.samplers
+import comfy.utils
+from comfy_api.latest import ComfyExtension, io
+
+
+class EmptyARVideoLatent(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="EmptyARVideoLatent",
+            category="latent/video",
+            inputs=[
+                io.Int.Input("width", default=832, min=16, max=8192, step=16),
+                io.Int.Input("height", default=480, min=16, max=8192, step=16),
+                io.Int.Input("length", default=81, min=1, max=1024, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=64),
+            ],
+            outputs=[
+                io.Latent.Output(display_name="LATENT"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, width, height, length, batch_size) -> io.NodeOutput:
+        lat_t = ((length - 1) // 4) + 1
+        latent = torch.zeros(
+            [batch_size, 16, lat_t, height // 8, width // 8],
+            device=comfy.model_management.intermediate_device(),
+        )
+        return io.NodeOutput({"samples": latent})
+
+
+class SamplerARVideo(io.ComfyNode):
+    """Sampler for autoregressive video models (Causal Forcing, Self-Forcing).
+
+    All AR-loop parameters are owned by this node so they live in the workflow.
+    Add new widgets here as the AR sampler grows new options.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="SamplerARVideo",
+            display_name="Sampler AR Video",
+            category="sampling/custom_sampling/samplers",
+            inputs=[
+                io.Int.Input(
+                    "num_frame_per_block",
+                    default=1, min=1, max=64,
+                    tooltip="Frames per autoregressive block. 1 = framewise, "
+                            "3 = chunkwise. Must match the checkpoint's training mode.",
+                ),
+            ],
+            outputs=[io.Sampler.Output()],
+        )
+
+    @classmethod
+    def execute(cls, num_frame_per_block) -> io.NodeOutput:
+        extra_options = {
+            "num_frame_per_block": num_frame_per_block,
+        }
+        return io.NodeOutput(comfy.samplers.ksampler("ar_video", extra_options))
+
+
+class ARVideoI2V(io.ComfyNode):
+    """Image-to-video setup for AR video models (Causal Forcing, Self-Forcing).
+
+    VAE-encodes the start image and stores it in the model's transformer_options
+    so that sample_ar_video can seed the KV cache before denoising.
+    Uses the same T2V model checkpoint -- no separate I2V architecture needed.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="ARVideoI2V",
+            category="conditioning/video_models",
+            inputs=[
+                io.Model.Input("model"),
+                io.Vae.Input("vae"),
+                io.Image.Input("start_image"),
+                io.Int.Input("width", default=832, min=16, max=8192, step=16),
+                io.Int.Input("height", default=480, min=16, max=8192, step=16),
+                io.Int.Input("length", default=81, min=1, max=1024, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=64),
+            ],
+            outputs=[
+                io.Model.Output(display_name="MODEL"),
+                io.Latent.Output(display_name="LATENT"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model, vae, start_image, width, height, length, batch_size) -> io.NodeOutput:
+        start_image = comfy.utils.common_upscale(
+            start_image[:1].movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+
+        initial_latent = vae.encode(start_image[:, :, :, :3])
+
+        m = model.clone()
+        to = m.model_options.setdefault("transformer_options", {})
+        ar_cfg = to.setdefault("ar_config", {})
+        ar_cfg["initial_latent"] = initial_latent
+
+        lat_t = ((length - 1) // 4) + 1
+        latent = torch.zeros(
+            [batch_size, 16, lat_t, height // 8, width // 8],
+            device=comfy.model_management.intermediate_device(),
+        )
+        return io.NodeOutput(m, {"samples": latent})
+
+
+class ARVideoExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            EmptyARVideoLatent,
+            SamplerARVideo,
+            ARVideoI2V,
+        ]
+
+
+async def comfy_entrypoint() -> ARVideoExtension:
+    return ARVideoExtension()

comfy_extras/nodes_attention_multiply.py

@@ -25,7 +25,7 @@ class UNetSelfAttentionMultiply(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="UNetSelfAttentionMultiply",
-            category="_for_testing/attention_experiments",
+            category="experimental/attention_experiments",
             inputs=[
                 io.Model.Input("model"),
                 io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01, advanced=True),
@@ -48,7 +48,7 @@ class UNetCrossAttentionMultiply(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="UNetCrossAttentionMultiply",
-            category="_for_testing/attention_experiments",
+            category="experimental/attention_experiments",
             inputs=[
                 io.Model.Input("model"),
                 io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01, advanced=True),
@@ -72,7 +72,7 @@ class CLIPAttentionMultiply(io.ComfyNode):
         return io.Schema(
             node_id="CLIPAttentionMultiply",
             search_aliases=["clip attention scale", "text encoder attention"],
-            category="_for_testing/attention_experiments",
+            category="experimental/attention_experiments",
             inputs=[
                 io.Clip.Input("clip"),
                 io.Float.Input("q", default=1.0, min=0.0, max=10.0, step=0.01, advanced=True),
@@ -106,7 +106,7 @@ class UNetTemporalAttentionMultiply(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="UNetTemporalAttentionMultiply",
-            category="_for_testing/attention_experiments",
+            category="experimental/attention_experiments",
             inputs=[
                 io.Model.Input("model"),
                 io.Float.Input("self_structural", default=1.0, min=0.0, max=10.0, step=0.01, advanced=True),

comfy_extras/nodes_audio_encoder.py

@@ -10,6 +10,7 @@ class AudioEncoderLoader(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="AudioEncoderLoader",
+            display_name="Load Audio Encoder",
             category="loaders",
             inputs=[
                 io.Combo.Input(

comfy_extras/nodes_bg_removal.py

@@ -0,0 +1,60 @@
+import folder_paths
+from typing_extensions import override
+from comfy_api.latest import ComfyExtension, IO
+from comfy.bg_removal_model import load
+
+
+class LoadBackgroundRemovalModel(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        files = folder_paths.get_filename_list("background_removal")
+        return IO.Schema(
+            node_id="LoadBackgroundRemovalModel",
+            display_name="Load Background Removal Model",
+            category="loaders",
+            inputs=[
+                IO.Combo.Input("bg_removal_name", options=sorted(files), tooltip="The model used to remove backgrounds from images"),
+            ],
+            outputs=[
+                IO.BackgroundRemoval.Output("bg_model")
+            ]
+        )
+    @classmethod
+    def execute(cls, bg_removal_name):
+        path = folder_paths.get_full_path_or_raise("background_removal", bg_removal_name)
+        bg = load(path)
+        if bg is None:
+            raise RuntimeError("ERROR: background model file is invalid and does not contain a valid background removal model.")
+        return IO.NodeOutput(bg)
+
+class RemoveBackground(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="RemoveBackground",
+            display_name="Remove Background",
+            category="image/background removal",
+            inputs=[
+                IO.Image.Input("image", tooltip="Input image to remove the background from"),
+                IO.BackgroundRemoval.Input("bg_removal_model", tooltip="Background removal model used to generate the mask")
+            ],
+            outputs=[
+                IO.Mask.Output("mask", tooltip="Generated foreground mask")
+            ]
+        )
+    @classmethod
+    def execute(cls, image, bg_removal_model):
+        mask = bg_removal_model.encode_image(image)
+        return IO.NodeOutput(mask)
+
+class BackgroundRemovalExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            LoadBackgroundRemovalModel,
+            RemoveBackground
+        ]
+
+
+async def comfy_entrypoint() -> BackgroundRemovalExtension:
+    return BackgroundRemovalExtension()

comfy_extras/nodes_camera_trajectory.py

@@ -153,7 +153,7 @@ class WanCameraEmbedding(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="WanCameraEmbedding",
-            category="camera",
+            category="conditioning/video_models",
             inputs=[
                 io.Combo.Input(
                     "camera_pose",

comfy_extras/nodes_compositing.py

@@ -202,14 +202,11 @@ class JoinImageWithAlpha(io.ComfyNode):
 
     @classmethod
     def execute(cls, image: torch.Tensor, alpha: torch.Tensor) -> io.NodeOutput:
-        batch_size = min(len(image), len(alpha))
-        out_images = []
-
-        alpha = 1.0 - resize_mask(alpha, image.shape[1:])
-        for i in range(batch_size):
-           out_images.append(torch.cat((image[i][:,:,:3], alpha[i].unsqueeze(2)), dim=2))
-
-        return io.NodeOutput(torch.stack(out_images))
+        batch_size = max(len(image), len(alpha))
+        alpha = 1.0 - resize_mask(alpha.to(image), image.shape[1:])
+        alpha = comfy.utils.repeat_to_batch_size(alpha, batch_size)
+        image = comfy.utils.repeat_to_batch_size(image, batch_size)
+        return io.NodeOutput(torch.cat((image[..., :3], alpha.unsqueeze(-1)), dim=-1))
 
 
 class CompositingExtension(ComfyExtension):

comfy_extras/nodes_cond.py

@@ -8,7 +8,7 @@ class CLIPTextEncodeControlnet(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="CLIPTextEncodeControlnet",
-            category="_for_testing/conditioning",
+            category="experimental/conditioning",
             inputs=[
                 io.Clip.Input("clip"),
                 io.Conditioning.Input("conditioning"),
@@ -35,7 +35,7 @@ class T5TokenizerOptions(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="T5TokenizerOptions",
-            category="_for_testing/conditioning",
+            category="experimental/conditioning",
             inputs=[
                 io.Clip.Input("clip"),
                 io.Int.Input("min_padding", default=0, min=0, max=10000, step=1, advanced=True),

comfy_extras/nodes_context_windows.py

@@ -10,7 +10,7 @@ class ContextWindowsManualNode(io.ComfyNode):
         return io.Schema(
             node_id="ContextWindowsManual",
             display_name="Context Windows (Manual)",
-            category="context",
+            category="model_patches",
             description="Manually set context windows.",
             inputs=[
                 io.Model.Input("model", tooltip="The model to apply context windows to during sampling."),
@@ -29,6 +29,7 @@ class ContextWindowsManualNode(io.ComfyNode):
                 io.Boolean.Input("freenoise", default=False, tooltip="Whether to apply FreeNoise noise shuffling, improves window blending."),
                 io.String.Input("cond_retain_index_list", default="", tooltip="List of latent indices to retain in the conditioning tensors for each window, for example setting this to '0' will use the initial start image for each window."),
                 io.Boolean.Input("split_conds_to_windows", default=False, tooltip="Whether to split multiple conditionings (created by ConditionCombine) to each window based on region index."),
+                io.Boolean.Input("causal_window_fix", default=True, tooltip="Whether to add a causal fix frame to non-0-indexed context windows."),
             ],
             outputs=[
                 io.Model.Output(tooltip="The model with context windows applied during sampling."),
@@ -38,7 +39,7 @@ class ContextWindowsManualNode(io.ComfyNode):
 
     @classmethod
     def execute(cls, model: io.Model.Type, context_length: int, context_overlap: int, context_schedule: str, context_stride: int, closed_loop: bool, fuse_method: str, dim: int, freenoise: bool,
-                cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False) -> io.Model:
+                cond_retain_index_list: list[int]=[], split_conds_to_windows: bool=False, causal_window_fix: bool=True) -> io.Model:
         model = model.clone()
         model.model_options["context_handler"] = comfy.context_windows.IndexListContextHandler(
             context_schedule=comfy.context_windows.get_matching_context_schedule(context_schedule),
@@ -50,7 +51,8 @@ class ContextWindowsManualNode(io.ComfyNode):
             dim=dim,
             freenoise=freenoise,
             cond_retain_index_list=cond_retain_index_list,
-            split_conds_to_windows=split_conds_to_windows
+            split_conds_to_windows=split_conds_to_windows,
+            causal_window_fix=causal_window_fix,
         )
         # make memory usage calculation only take into account the context window latents
         comfy.context_windows.create_prepare_sampling_wrapper(model)

comfy_extras/nodes_custom_sampler.py

@@ -984,7 +984,7 @@ class AddNoise(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="AddNoise",
-            category="_for_testing/custom_sampling/noise",
+            category="experimental/custom_sampling/noise",
             is_experimental=True,
             inputs=[
                 io.Model.Input("model"),
@@ -1034,7 +1034,7 @@ class ManualSigmas(io.ComfyNode):
         return io.Schema(
             node_id="ManualSigmas",
             search_aliases=["custom noise schedule", "define sigmas"],
-            category="_for_testing/custom_sampling",
+            category="experimental/custom_sampling",
             is_experimental=True,
             inputs=[
                 io.String.Input("sigmas", default="1, 0.5", multiline=False)

comfy_extras/nodes_differential_diffusion.py

@@ -13,7 +13,7 @@ class DifferentialDiffusion(io.ComfyNode):
             node_id="DifferentialDiffusion",
             search_aliases=["inpaint gradient", "variable denoise strength"],
             display_name="Differential Diffusion",
-            category="_for_testing",
+            category="experimental",
             inputs=[
                 io.Model.Input("model"),
                 io.Float.Input(

comfy_extras/nodes_flux.py

@@ -102,7 +102,7 @@ class FluxDisableGuidance(io.ComfyNode):
     append = execute  # TODO: remove
 
 
-PREFERED_KONTEXT_RESOLUTIONS = [
+PREFERRED_KONTEXT_RESOLUTIONS = [
     (672, 1568),
     (688, 1504),
     (720, 1456),
@@ -143,7 +143,7 @@ class FluxKontextImageScale(io.ComfyNode):
         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 PREFERED_KONTEXT_RESOLUTIONS)
+        _, width, height = min((abs(aspect_ratio - w / h), w, h) for w, h in PREFERRED_KONTEXT_RESOLUTIONS)
         image = comfy.utils.common_upscale(image.movedim(-1, 1), width, height, "lanczos", "center").movedim(1, -1)
         return io.NodeOutput(image)
 

comfy_extras/nodes_frame_interpolation.py

@@ -37,7 +37,7 @@ class FrameInterpolationModelLoader(io.ComfyNode):
         model = cls._detect_and_load(sd)
         dtype = torch.float16 if model_management.should_use_fp16(model_management.get_torch_device()) else torch.float32
         model.eval().to(dtype)
-        patcher = comfy.model_patcher.ModelPatcher(
+        patcher = comfy.model_patcher.CoreModelPatcher(
             model,
             load_device=model_management.get_torch_device(),
             offload_device=model_management.unet_offload_device(),
@@ -78,7 +78,7 @@ class FrameInterpolate(io.ComfyNode):
         return io.Schema(
             node_id="FrameInterpolate",
             display_name="Frame Interpolate",
-            category="image/video",
+            category="video",
             search_aliases=["rife", "film", "frame interpolation", "slow motion", "interpolate frames", "vfi"],
             inputs=[
                 FrameInterpolationModel.Input("interp_model"),
@@ -98,16 +98,13 @@ class FrameInterpolate(io.ComfyNode):
         if num_frames < 2 or multiplier < 2:
             return io.NodeOutput(images)
 
-        model_management.load_model_gpu(interp_model)
         device = interp_model.load_device
         dtype = interp_model.model_dtype()
         inference_model = interp_model.model
-
-        # Free VRAM for inference activations (model weights + ~20x a single frame's worth)
-        H, W = images.shape[1], images.shape[2]
-        activation_mem = H * W * 3 * images.element_size() * 20
-        model_management.free_memory(activation_mem, device)
+        activation_mem = inference_model.memory_used_forward(images.shape, dtype)
+        model_management.load_models_gpu([interp_model], memory_required=activation_mem)
         align = getattr(inference_model, "pad_align", 1)
+        H, W = images.shape[1], images.shape[2]
 
         # Prepare a single padded frame on device for determining output dimensions
         def prepare_frame(idx):

comfy_extras/nodes_fresca.py

@@ -60,7 +60,7 @@ class FreSca(io.ComfyNode):
             node_id="FreSca",
             search_aliases=["frequency guidance"],
             display_name="FreSca",
-            category="_for_testing",
+            category="experimental",
             description="Applies frequency-dependent scaling to the guidance",
             inputs=[
                 io.Model.Input("model"),

comfy_extras/nodes_hunyuan.py

@@ -131,6 +131,8 @@ class HunyuanVideo15SuperResolution(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="HunyuanVideo15SuperResolution",
+            display_name="Hunyuan Video 1.5 Super Resolution",
+            category="conditioning/video_models",
             inputs=[
                 io.Conditioning.Input("positive"),
                 io.Conditioning.Input("negative"),
@@ -381,6 +383,8 @@ class HunyuanRefinerLatent(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="HunyuanRefinerLatent",
+            display_name="Hunyuan Latent Refiner",
+            category="conditioning/video_models",
             inputs=[
                 io.Conditioning.Input("positive"),
                 io.Conditioning.Input("negative"),

comfy_extras/nodes_hunyuan3d.py

@@ -40,7 +40,7 @@ class Hunyuan3Dv2Conditioning(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="Hunyuan3Dv2Conditioning",
-            category="conditioning/video_models",
+            category="conditioning/3d_models",
             inputs=[
                 IO.ClipVisionOutput.Input("clip_vision_output"),
             ],
@@ -65,7 +65,7 @@ class Hunyuan3Dv2ConditioningMultiView(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="Hunyuan3Dv2ConditioningMultiView",
-            category="conditioning/video_models",
+            category="conditioning/3d_models",
             inputs=[
                 IO.ClipVisionOutput.Input("front", optional=True),
                 IO.ClipVisionOutput.Input("left", optional=True),
@@ -424,6 +424,7 @@ class VoxelToMeshBasic(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="VoxelToMeshBasic",
+            display_name="Voxel to Mesh (Basic)",
             category="3d",
             inputs=[
                 IO.Voxel.Input("voxel"),
@@ -453,6 +454,7 @@ class VoxelToMesh(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="VoxelToMesh",
+            display_name="Voxel to Mesh",
             category="3d",
             inputs=[
                 IO.Voxel.Input("voxel"),

comfy_extras/nodes_hypernetwork.py

@@ -102,6 +102,7 @@ class HypernetworkLoader(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="HypernetworkLoader",
+            display_name="Load Hypernetwork",
             category="loaders",
             inputs=[
                 IO.Model.Input("model"),

comfy_extras/nodes_image_compare.py

@@ -11,7 +11,7 @@ class ImageCompare(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="ImageCompare",
-            display_name="Image Compare",
+            display_name="Compare Images",
             description="Compares two images side by side with a slider.",
             category="image",
             essentials_category="Image Tools",

comfy_extras/nodes_images.py

@@ -24,7 +24,7 @@ class ImageCrop(IO.ComfyNode):
         return IO.Schema(
             node_id="ImageCrop",
             search_aliases=["trim"],
-            display_name="Image Crop (Deprecated)",
+            display_name="Crop Image (DEPRECATED)",
             category="image/transform",
             is_deprecated=True,
             essentials_category="Image Tools",
@@ -56,7 +56,7 @@ class ImageCropV2(IO.ComfyNode):
         return IO.Schema(
             node_id="ImageCropV2",
             search_aliases=["trim"],
-            display_name="Image Crop",
+            display_name="Crop Image",
             category="image/transform",
             essentials_category="Image Tools",
             has_intermediate_output=True,
@@ -109,6 +109,7 @@ class RepeatImageBatch(IO.ComfyNode):
         return IO.Schema(
             node_id="RepeatImageBatch",
             search_aliases=["duplicate image", "clone image"],
+            display_name="Repeat Image Batch",
             category="image/batch",
             inputs=[
                 IO.Image.Input("image"),
@@ -131,6 +132,7 @@ class ImageFromBatch(IO.ComfyNode):
         return IO.Schema(
             node_id="ImageFromBatch",
             search_aliases=["select image", "pick from batch", "extract image"],
+            display_name="Get Image from Batch",
             category="image/batch",
             inputs=[
                 IO.Image.Input("image"),
@@ -157,7 +159,8 @@ class ImageAddNoise(IO.ComfyNode):
         return IO.Schema(
             node_id="ImageAddNoise",
             search_aliases=["film grain"],
-            category="image",
+            display_name="Add Noise to Image",
+            category="image/postprocessing",
             inputs=[
                 IO.Image.Input("image"),
                 IO.Int.Input(
@@ -259,7 +262,7 @@ class ImageStitch(IO.ComfyNode):
         return IO.Schema(
             node_id="ImageStitch",
             search_aliases=["combine images", "join images", "concatenate images", "side by side"],
-            display_name="Image Stitch",
+            display_name="Stitch Images",
             description="Stitches image2 to image1 in the specified direction.\n"
             "If image2 is not provided, returns image1 unchanged.\n"
             "Optional spacing can be added between images.",
@@ -434,6 +437,7 @@ class ResizeAndPadImage(IO.ComfyNode):
         return IO.Schema(
             node_id="ResizeAndPadImage",
             search_aliases=["fit to size"],
+            display_name="Resize And Pad Image",
             category="image/transform",
             inputs=[
                 IO.Image.Input("image"),
@@ -485,6 +489,7 @@ class SaveSVGNode(IO.ComfyNode):
         return IO.Schema(
             node_id="SaveSVGNode",
             search_aliases=["export vector", "save vector graphics"],
+            display_name="Save SVG",
             description="Save SVG files on disk.",
             category="image/save",
             inputs=[
@@ -591,7 +596,7 @@ class ImageRotate(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="ImageRotate",
-            display_name="Image Rotate",
+            display_name="Rotate Image",
             search_aliases=["turn", "flip orientation"],
             category="image/transform",
             essentials_category="Image Tools",
@@ -624,6 +629,7 @@ class ImageFlip(IO.ComfyNode):
         return IO.Schema(
             node_id="ImageFlip",
             search_aliases=["mirror", "reflect"],
+            display_name="Flip Image",
             category="image/transform",
             inputs=[
                 IO.Image.Input("image"),
@@ -650,6 +656,7 @@ class ImageScaleToMaxDimension(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="ImageScaleToMaxDimension",
+            display_name="Scale Image to Max Dimension",
             category="image/upscaling",
             inputs=[
                 IO.Image.Input("image"),
@@ -709,7 +716,7 @@ class SplitImageToTileList(IO.ComfyNode):
     def get_grid_coords(width, height, tile_width, tile_height, overlap):
         coords = []
         stride_x = round(max(tile_width * 0.25, tile_width - overlap))
-        stride_y = round(max(tile_width * 0.25, tile_height - overlap))
+        stride_y = round(max(tile_height * 0.25, tile_height - overlap))
 
         y = 0
         while y < height:

comfy_extras/nodes_lora_extract.py

@@ -91,7 +91,7 @@ class LoraSave(io.ComfyNode):
             node_id="LoraSave",
             search_aliases=["export lora"],
             display_name="Extract and Save Lora",
-            category="_for_testing",
+            category="experimental",
             inputs=[
                 io.String.Input("filename_prefix", default="loras/ComfyUI_extracted_lora"),
                 io.Int.Input("rank", default=8, min=1, max=4096, step=1, advanced=True),

comfy_extras/nodes_lt.py

@@ -106,12 +106,12 @@ class LTXVImgToVideoInplace(io.ComfyNode):
         if bypass:
             return (latent,)
 
-        samples = latent["samples"]
+        samples = latent["samples"].clone()
         _, height_scale_factor, width_scale_factor = (
             vae.downscale_index_formula
         )
 
-        batch, _, latent_frames, latent_height, latent_width = samples.shape
+        _, _, _, latent_height, latent_width = samples.shape
         width = latent_width * width_scale_factor
         height = latent_height * height_scale_factor
 
@@ -124,11 +124,7 @@ class LTXVImgToVideoInplace(io.ComfyNode):
 
         samples[:, :, :t.shape[2]] = t
 
-        conditioning_latent_frames_mask = torch.ones(
-            (batch, 1, latent_frames, 1, 1),
-            dtype=torch.float32,
-            device=samples.device,
-        )
+        conditioning_latent_frames_mask = get_noise_mask(latent)
         conditioning_latent_frames_mask[:, :, :t.shape[2]] = 1.0 - strength
 
         return io.NodeOutput({"samples": samples, "noise_mask": conditioning_latent_frames_mask})
@@ -236,7 +232,7 @@ class LTXVAddGuide(io.ComfyNode):
     def encode(cls, vae, latent_width, latent_height, images, scale_factors):
         time_scale_factor, width_scale_factor, height_scale_factor = scale_factors
         images = images[:(images.shape[0] - 1) // time_scale_factor * time_scale_factor + 1]
-        pixels = comfy.utils.common_upscale(images.movedim(-1, 1), latent_width * width_scale_factor, latent_height * height_scale_factor, "bilinear", crop="disabled").movedim(1, -1)
+        pixels = comfy.utils.common_upscale(images.movedim(-1, 1), latent_width * width_scale_factor, latent_height * height_scale_factor, "bilinear", crop="center").movedim(1, -1)
         encode_pixels = pixels[:, :, :, :3]
         t = vae.encode(encode_pixels)
         return encode_pixels, t
@@ -594,7 +590,8 @@ class LTXVPreprocess(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="LTXVPreprocess",
-            category="image",
+            display_name="LTXV Preprocess",
+            category="video/preprocessors",
             inputs=[
                 io.Image.Input("image"),
                 io.Int.Input(

comfy_extras/nodes_lt_audio.py

@@ -147,7 +147,6 @@ class LTXVEmptyLatentAudio(io.ComfyNode):
 
         z_channels = audio_vae.latent_channels
         audio_freq = audio_vae.first_stage_model.latent_frequency_bins
-        sampling_rate = int(audio_vae.first_stage_model.sample_rate)
 
         num_audio_latents = audio_vae.first_stage_model.num_of_latents_from_frames(frames_number, frame_rate)
 
@@ -159,7 +158,6 @@ class LTXVEmptyLatentAudio(io.ComfyNode):
         return io.NodeOutput(
             {
                 "samples": audio_latents,
-                "sample_rate": sampling_rate,
                 "type": "audio",
             }
         )

comfy_extras/nodes_mahiro.py

@@ -11,7 +11,7 @@ class Mahiro(io.ComfyNode):
         return io.Schema(
             node_id="Mahiro",
             display_name="Positive-Biased Guidance",
-            category="_for_testing",
+            category="experimental",
             description="Modify the guidance to scale more on the 'direction' of the positive prompt rather than the difference between the negative prompt.",
             inputs=[
                 io.Model.Input("model"),

comfy_extras/nodes_mask.py

@@ -40,10 +40,21 @@ def composite(destination, source, x, y, mask = None, multiplier = 8, resize_sou
 
     inverse_mask = torch.ones_like(mask) - mask
 
-    source_portion = mask * source[..., :visible_height, :visible_width]
-    destination_portion = inverse_mask  * destination[..., top:bottom, left:right]
+    source_rgb = source[:, :3, :visible_height, :visible_width]
+    dest_slice = destination[..., top:bottom, left:right]
+
+    if destination.shape[1] == 4:
+        if torch.max(dest_slice) == 0:
+            destination[:, :3, top:bottom, left:right] = source_rgb
+            destination[:, 3:4, top:bottom, left:right] = mask
+        else:
+            destination[:, :3, top:bottom, left:right] = (mask * source_rgb) + (inverse_mask * dest_slice[:, :3])
+            destination[:, 3:4, top:bottom, left:right] = torch.max(mask, dest_slice[:, 3:4])
+    else:
+        source_portion = mask * source_rgb
+        destination_portion = inverse_mask * dest_slice
+        destination[..., top:bottom, left:right] = source_portion + destination_portion
 
-    destination[..., top:bottom, left:right] = source_portion + destination_portion
     return destination
 
 class LatentCompositeMasked(IO.ComfyNode):
@@ -80,21 +91,27 @@ class ImageCompositeMasked(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="ImageCompositeMasked",
-            search_aliases=["paste image", "overlay", "layer"],
+            search_aliases=["overlay", "layer", "paste image", "images composition"],
+            display_name="Image Composite Masked",
             category="image",
             inputs=[
-                IO.Image.Input("destination"),
                 IO.Image.Input("source"),
                 IO.Int.Input("x", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1),
                 IO.Int.Input("y", default=0, min=0, max=nodes.MAX_RESOLUTION, step=1),
                 IO.Boolean.Input("resize_source", default=False),
+                IO.Image.Input("destination", optional=True),
                 IO.Mask.Input("mask", optional=True),
             ],
             outputs=[IO.Image.Output()],
         )
 
     @classmethod
-    def execute(cls, destination, source, x, y, resize_source, mask = None) -> IO.NodeOutput:
+    def execute(cls, source, x, y, resize_source, destination = None, mask = None) -> IO.NodeOutput:
+        if destination is None: # transparent rgba
+            B, H, W, C = source.shape
+            destination = torch.zeros((B, H, W, 4), dtype=source.dtype, device=source.device)
+            if C == 3:
+                source = torch.nn.functional.pad(source, (0, 1), value=1.0)
         destination, source = node_helpers.image_alpha_fix(destination, source)
         destination = destination.clone().movedim(-1, 1)
         output = composite(destination, source.movedim(-1, 1), x, y, mask, 1, resize_source).movedim(1, -1)
@@ -201,6 +218,7 @@ class InvertMask(IO.ComfyNode):
         return IO.Schema(
             node_id="InvertMask",
             search_aliases=["reverse mask", "flip mask"],
+            display_name="Invert Mask",
             category="mask",
             inputs=[
                 IO.Mask.Input("mask"),
@@ -222,6 +240,7 @@ class CropMask(IO.ComfyNode):
         return IO.Schema(
             node_id="CropMask",
             search_aliases=["cut mask", "extract mask region", "mask slice"],
+            display_name="Crop Mask",
             category="mask",
             inputs=[
                 IO.Mask.Input("mask"),
@@ -247,7 +266,8 @@ class MaskComposite(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="MaskComposite",
-            search_aliases=["combine masks", "blend masks", "layer masks"],
+            search_aliases=["combine masks", "blend masks", "layer masks", "masks composition"],
+            display_name="Combine Masks",
             category="mask",
             inputs=[
                 IO.Mask.Input("destination"),
@@ -298,6 +318,7 @@ class FeatherMask(IO.ComfyNode):
         return IO.Schema(
             node_id="FeatherMask",
             search_aliases=["soft edge mask", "blur mask edges", "gradient mask edge"],
+            display_name="Feather Mask",
             category="mask",
             inputs=[
                 IO.Mask.Input("mask"),
@@ -376,7 +397,6 @@ class GrowMask(IO.ComfyNode):
 
     expand_mask = execute  # TODO: remove
 
-
 class ThresholdMask(IO.ComfyNode):
     @classmethod
     def define_schema(cls):

comfy_extras/nodes_math.py

@@ -70,7 +70,7 @@ class MathExpressionNode(io.ComfyNode):
         return io.Schema(
             node_id="ComfyMathExpression",
             display_name="Math Expression",
-            category="math",
+            category="logic",
             search_aliases=[
                 "expression", "formula", "calculate", "calculator",
                 "eval", "math",

comfy_extras/nodes_morphology.py

@@ -59,7 +59,8 @@ class ImageRGBToYUV(io.ComfyNode):
         return io.Schema(
             node_id="ImageRGBToYUV",
             search_aliases=["color space conversion"],
-            category="image/batch",
+            display_name="Image RGB to YUV",
+            category="image/color",
             inputs=[
                 io.Image.Input("image"),
             ],
@@ -81,7 +82,8 @@ class ImageYUVToRGB(io.ComfyNode):
         return io.Schema(
             node_id="ImageYUVToRGB",
             search_aliases=["color space conversion"],
-            category="image/batch",
+            display_name="Image YUV to RGB",
+            category="image/color",
             inputs=[
                 io.Image.Input("Y"),
                 io.Image.Input("U"),

comfy_extras/nodes_number_convert.py

@@ -21,7 +21,7 @@ class NumberConvertNode(io.ComfyNode):
         return io.Schema(
             node_id="ComfyNumberConvert",
             display_name="Number Convert",
-            category="math",
+            category="utils",
             search_aliases=[
                 "int to float", "float to int", "number convert",
                 "int2float", "float2int", "cast", "parse number",

comfy_extras/nodes_perpneg.py

@@ -24,8 +24,8 @@ class PerpNeg(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="PerpNeg",
-            display_name="Perp-Neg (DEPRECATED by PerpNegGuider)",
-            category="_for_testing",
+            display_name="Perp-Neg (DEPRECATED by Perp-Neg Guider)",
+            category="experimental",
             inputs=[
                 io.Model.Input("model"),
                 io.Conditioning.Input("empty_conditioning"),
@@ -127,7 +127,8 @@ class PerpNegGuider(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="PerpNegGuider",
-            category="_for_testing",
+            display_name="Perp-Neg Guider",
+            category="experimental",
             inputs=[
                 io.Model.Input("model"),
                 io.Conditioning.Input("positive"),

comfy_extras/nodes_photomaker.py

@@ -123,7 +123,7 @@ class PhotoMakerLoader(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="PhotoMakerLoader",
-            category="_for_testing/photomaker",
+            category="experimental/photomaker",
             inputs=[
                 io.Combo.Input("photomaker_model_name", options=folder_paths.get_filename_list("photomaker")),
             ],
@@ -149,7 +149,7 @@ class PhotoMakerEncode(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="PhotoMakerEncode",
-            category="_for_testing/photomaker",
+            category="experimental/photomaker",
             inputs=[
                 io.Photomaker.Input("photomaker"),
                 io.Image.Input("image"),

comfy_extras/nodes_post_processing.py

@@ -20,7 +20,8 @@ class Blend(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="ImageBlend",
-            display_name="Image Blend",
+            search_aliases=["mix images"],
+            display_name="Blend Images",
             category="image/postprocessing",
             essentials_category="Image Tools",
             inputs=[
@@ -115,6 +116,7 @@ class Quantize(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="ImageQuantize",
+            display_name="Quantize Image",
             category="image/postprocessing",
             inputs=[
                 io.Image.Input("image"),
@@ -180,6 +182,7 @@ class Sharpen(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="ImageSharpen",
+            display_name="Sharpen Image",
             category="image/postprocessing",
             inputs=[
                 io.Image.Input("image"),
@@ -224,6 +227,7 @@ class ImageScaleToTotalPixels(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="ImageScaleToTotalPixels",
+            display_name="Scale Image to Total Pixels",
             category="image/upscaling",
             inputs=[
                 io.Image.Input("image"),
@@ -434,7 +438,7 @@ class ResizeImageMaskNode(io.ComfyNode):
             node_id="ResizeImageMaskNode",
             display_name="Resize Image/Mask",
             description="Resize an image or mask using various scaling methods.",
-            category="transform",
+            category="image/transform",
             search_aliases=["resize", "resize image", "resize mask", "scale", "scale image", "scale mask", "image resize", "change size", "dimensions", "shrink", "enlarge"],
             inputs=[
                 io.MatchType.Input("input", template=template),
@@ -568,7 +572,7 @@ class BatchImagesNode(io.ComfyNode):
         return io.Schema(
             node_id="BatchImagesNode",
             display_name="Batch Images",
-            category="image",
+            category="image/batch",
             essentials_category="Image Tools",
             search_aliases=["batch", "image batch", "batch images", "combine images", "merge images", "stack images"],
             inputs=[
@@ -666,12 +670,13 @@ class ColorTransfer(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="ColorTransfer",
+            display_name="Color Transfer",
             category="image/postprocessing",
             description="Match the colors of one image to another using various algorithms.",
             search_aliases=["color match", "color grading", "color correction", "match colors", "color transform", "mkl", "reinhard", "histogram"],
             inputs=[
                 io.Image.Input("image_target", tooltip="Image(s) to apply the color transform to."),
-                io.Image.Input("image_ref", optional=True, tooltip="Reference image(s) to match colors to. If not provided, processing is skipped"),
+                io.Image.Input("image_ref", tooltip="Reference image(s) to match colors to."),
                 io.Combo.Input("method", options=['reinhard_lab', 'mkl_lab', 'histogram'],),
                 io.DynamicCombo.Input("source_stats",
                     tooltip="per_frame: each frame matched to image_ref individually. uniform: pool stats across all source frames as baseline, match to image_ref. target_frame: use one chosen frame as the baseline for the transform to image_ref, applied uniformly to all frames (preserves relative differences)",

comfy_extras/nodes_primitive.py

@@ -9,7 +9,8 @@ class String(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="PrimitiveString",
-            display_name="String",
+            search_aliases=["text", "string", "text box", "prompt"],
+            display_name="Text String",
             category="utils/primitive",
             inputs=[
                 io.String.Input("value"),
@@ -27,7 +28,8 @@ class StringMultiline(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="PrimitiveStringMultiline",
-            display_name="String (Multiline)",
+            search_aliases=["text", "string", "text multiline", "string multiline", "text box", "prompt"],
+            display_name="Text String (Multiline)",
             category="utils/primitive",
             essentials_category="Basics",
             inputs=[
@@ -49,7 +51,7 @@ class Int(io.ComfyNode):
             display_name="Int",
             category="utils/primitive",
             inputs=[
-                io.Int.Input("value", min=-sys.maxsize, max=sys.maxsize, control_after_generate=True),
+                io.Int.Input("value", min=-sys.maxsize, max=sys.maxsize, control_after_generate=io.ControlAfterGenerate.fixed),
             ],
             outputs=[io.Int.Output()],
         )

comfy_extras/nodes_rtdetr.py

@@ -15,7 +15,7 @@ class RTDETR_detect(io.ComfyNode):
         return io.Schema(
             node_id="RTDETR_detect",
             display_name="RT-DETR Detect",
-            category="detection/",
+            category="detection",
             search_aliases=["bbox", "bounding box", "object detection", "coco"],
             inputs=[
                 io.Model.Input("model", display_name="model"),
@@ -71,7 +71,7 @@ class DrawBBoxes(io.ComfyNode):
         return io.Schema(
             node_id="DrawBBoxes",
             display_name="Draw BBoxes",
-            category="detection/",
+            category="detection",
             search_aliases=["bbox", "bounding box", "object detection", "rt_detr", "visualize detections", "coco"],
             inputs=[
                 io.Image.Input("image", optional=True),

comfy_extras/nodes_sag.py

@@ -113,7 +113,7 @@ class SelfAttentionGuidance(io.ComfyNode):
         return io.Schema(
             node_id="SelfAttentionGuidance",
             display_name="Self-Attention Guidance",
-            category="_for_testing",
+            category="experimental",
             inputs=[
                 io.Model.Input("model"),
                 io.Float.Input("scale", default=0.5, min=-2.0, max=5.0, step=0.01),

comfy_extras/nodes_sam3.py

@@ -93,7 +93,7 @@ class SAM3_Detect(io.ComfyNode):
         return io.Schema(
             node_id="SAM3_Detect",
             display_name="SAM3 Detect",
-            category="detection/",
+            category="detection",
             search_aliases=["sam3", "segment anything", "open vocabulary", "text detection", "segment"],
             inputs=[
                 io.Model.Input("model", display_name="model"),
@@ -265,15 +265,15 @@ class SAM3_VideoTrack(io.ComfyNode):
         return io.Schema(
             node_id="SAM3_VideoTrack",
             display_name="SAM3 Video Track",
-            category="detection/",
+            category="detection",
             search_aliases=["sam3", "video", "track", "propagate"],
             inputs=[
                 io.Image.Input("images", display_name="images", tooltip="Video frames as batched images"),
                 io.Model.Input("model", display_name="model"),
                 io.Mask.Input("initial_mask", display_name="initial_mask", optional=True, tooltip="Mask(s) for the first frame to track (one per object)"),
                 io.Conditioning.Input("conditioning", display_name="conditioning", optional=True, tooltip="Text conditioning for detecting new objects during tracking"),
-                io.Float.Input("detection_threshold", display_name="detection_threshold", default=0.5, min=0.0, max=1.0, step=0.01, tooltip="Score threshold for text-prompted detection"),
-                io.Int.Input("max_objects", display_name="max_objects", default=0, min=0, tooltip="Max tracked objects (0=unlimited). Initial masks count toward this limit."),
+                io.Float.Input("detection_threshold", display_name="detection_threshold", default=0.5, min=0.0, max=1.0, step=0.01, tooltip="Score threshold for text-prompted detection."),
+                io.Int.Input("max_objects", display_name="max_objects", default=4, min=0, max=64, tooltip="Max tracked objects. Initial masks count toward this limit. 0 uses the internal cap of 64."),
                 io.Int.Input("detect_interval", display_name="detect_interval", default=1, min=1, tooltip="Run detection every N frames (1=every frame). Higher values save compute."),
             ],
             outputs=[
@@ -290,8 +290,7 @@ class SAM3_VideoTrack(io.ComfyNode):
         dtype = model.model.get_dtype()
         sam3_model = model.model.diffusion_model
 
-        frames = images[..., :3].movedim(-1, 1)
-        frames_in = comfy.utils.common_upscale(frames, 1008, 1008, "bilinear", crop="disabled").to(device=device, dtype=dtype)
+        frames_in = images[..., :3].movedim(-1, 1)
 
         init_masks = None
         if initial_mask is not None:
@@ -308,7 +307,7 @@ class SAM3_VideoTrack(io.ComfyNode):
         result = sam3_model.forward_video(
             images=frames_in, initial_masks=init_masks, pbar=pbar, text_prompts=text_prompts,
             new_det_thresh=detection_threshold, max_objects=max_objects,
-            detect_interval=detect_interval)
+            detect_interval=detect_interval, target_device=device, target_dtype=dtype)
         result["orig_size"] = (H, W)
         return io.NodeOutput(result)
 
@@ -321,7 +320,7 @@ class SAM3_TrackPreview(io.ComfyNode):
         return io.Schema(
             node_id="SAM3_TrackPreview",
             display_name="SAM3 Track Preview",
-            category="detection/",
+            category="detection",
             inputs=[
                 SAM3TrackData.Input("track_data", display_name="track_data"),
                 io.Image.Input("images", display_name="images", optional=True),
@@ -449,14 +448,18 @@ class SAM3_TrackPreview(io.ComfyNode):
                     cx = (bool_masks * grid_x).sum(dim=(-1, -2)) // area
                     has = area > 1
                     scores = track_data.get("scores", [])
+                    label_scale = max(3, H // 240) # Scale font with resolutio
+                    size_caps = (area.float().sqrt() / 15).clamp_(min=1).long().tolist() #cap per-object so the number doesn't dwarf small masks
                     for obj_idx in range(N_obj):
                         if has[obj_idx]:
                             _cx, _cy = int(cx[obj_idx]), int(cy[obj_idx])
                             color = cls.COLORS[obj_idx % len(cls.COLORS)]
-                            SAM3_TrackPreview._draw_number_gpu(frame_gpu, obj_idx, _cx, _cy, color)
+                            obj_scale = min(label_scale, size_caps[obj_idx])
+                            score_scale = max(1, obj_scale * 2 // 3)
+                            SAM3_TrackPreview._draw_number_gpu(frame_gpu, obj_idx, _cx, _cy, color, scale=obj_scale)
                             if obj_idx < len(scores) and scores[obj_idx] < 1.0:
                                 SAM3_TrackPreview._draw_number_gpu(frame_gpu, int(scores[obj_idx] * 100),
-                                                                   _cx, _cy + 5 * 3 + 3, color, scale=2)
+                                                                   _cx, _cy + 5 * obj_scale + 3, color, scale=score_scale)
                     frame_cpu.copy_(frame_gpu.clamp_(0, 1).mul_(255).byte())
                 else:
                     frame_cpu.copy_(frame.clamp_(0, 1).mul_(255).byte())
@@ -475,7 +478,7 @@ class SAM3_TrackToMask(io.ComfyNode):
         return io.Schema(
             node_id="SAM3_TrackToMask",
             display_name="SAM3 Track to Mask",
-            category="detection/",
+            category="detection",
             inputs=[
                 SAM3TrackData.Input("track_data", display_name="track_data"),
                 io.String.Input("object_indices", display_name="object_indices", default="",
@@ -507,9 +510,10 @@ class SAM3_TrackToMask(io.ComfyNode):
         if not indices:
             return io.NodeOutput(torch.zeros(N, H, W, device=comfy.model_management.intermediate_device()))
 
-        selected = packed[:, indices]
-        binary = unpack_masks(selected)  # [N, len(indices), Hm, Wm] bool
-        union = binary.any(dim=1, keepdim=True).float()
+        union_packed = packed[:, indices[0]].clone()
+        for i in indices[1:]:
+            union_packed |= packed[:, i]
+        union = unpack_masks(union_packed).unsqueeze(1).float()  # [N, 1, Hm, Wm]
         mask_out = F.interpolate(union, size=(H, W), mode="bilinear", align_corners=False)[:, 0]
         return io.NodeOutput(mask_out)
 

comfy_extras/nodes_stable_cascade.py

@@ -119,7 +119,7 @@ class StableCascade_SuperResolutionControlnet(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StableCascade_SuperResolutionControlnet",
-            category="_for_testing/stable_cascade",
+            category="experimental/stable_cascade",
             is_experimental=True,
             inputs=[
                 io.Image.Input("image"),

comfy_extras/nodes_string.py

@@ -10,9 +10,9 @@ class StringConcatenate(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringConcatenate",
-            display_name="Text Concatenate",
-            category="utils/string",
-            search_aliases=["Concatenate", "text concat", "join text", "merge text", "combine strings", "concat", "concatenate", "append text", "combine text", "string"],
+            search_aliases=["concatenate", "text concat", "join text", "merge text", "combine strings", "string concat", "append text", "combine text"],
+            display_name="Concatenate Text",
+            category="text",
             inputs=[
                 io.String.Input("string_a", multiline=True),
                 io.String.Input("string_b", multiline=True),
@@ -33,9 +33,9 @@ class StringSubstring(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringSubstring",
-            search_aliases=["Substring", "extract text", "text portion"],
-            display_name="Text Substring",
-            category="utils/string",
+            search_aliases=["substring", "extract text", "text portion"],
+            display_name="Substring",
+            category="text",
             inputs=[
                 io.String.Input("string", multiline=True),
                 io.Int.Input("start"),
@@ -58,7 +58,7 @@ class StringLength(io.ComfyNode):
             node_id="StringLength",
             search_aliases=["character count", "text size", "string length"],
             display_name="Text Length",
-            category="utils/string",
+            category="text",
             inputs=[
                 io.String.Input("string", multiline=True),
             ],
@@ -77,9 +77,9 @@ class CaseConverter(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="CaseConverter",
-            search_aliases=["Case Converter", "text case", "uppercase", "lowercase", "capitalize"],
-            display_name="Text Case Converter",
-            category="utils/string",
+            search_aliases=["case converter", "text case", "uppercase", "lowercase", "capitalize"],
+            display_name="Convert Text Case",
+            category="text",
             inputs=[
                 io.String.Input("string", multiline=True),
                 io.Combo.Input("mode", options=["UPPERCASE", "lowercase", "Capitalize", "Title Case"]),
@@ -110,9 +110,9 @@ class StringTrim(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringTrim",
-            search_aliases=["Trim", "clean whitespace", "remove whitespace", "strip"],
-            display_name="Text Trim",
-            category="utils/string",
+            search_aliases=["trim", "clean whitespace", "remove whitespace", "remove spaces","strip"],
+            display_name="Trim Text",
+            category="text",
             inputs=[
                 io.String.Input("string", multiline=True),
                 io.Combo.Input("mode", options=["Both", "Left", "Right"]),
@@ -141,9 +141,9 @@ class StringReplace(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringReplace",
-            search_aliases=["Replace", "find and replace", "substitute", "swap text"],
-            display_name="Text Replace",
-            category="utils/string",
+            search_aliases=["replace", "find and replace", "substitute", "swap text"],
+            display_name="Replace Text",
+            category="text",
             inputs=[
                 io.String.Input("string", multiline=True),
                 io.String.Input("find", multiline=True),
@@ -164,9 +164,9 @@ class StringContains(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringContains",
-            search_aliases=["Contains", "text includes", "string includes"],
-            display_name="Text Contains",
-            category="utils/string",
+            search_aliases=["contains", "text includes", "string includes"],
+            display_name="Contains Text",
+            category="text",
             inputs=[
                 io.String.Input("string", multiline=True),
                 io.String.Input("substring", multiline=True),
@@ -192,9 +192,9 @@ class StringCompare(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringCompare",
-            search_aliases=["Compare", "text match", "string equals", "starts with", "ends with"],
-            display_name="Text Compare",
-            category="utils/string",
+            search_aliases=["compare", "text match", "string equals", "starts with", "ends with"],
+            display_name="Compare Text",
+            category="text",
             inputs=[
                 io.String.Input("string_a", multiline=True),
                 io.String.Input("string_b", multiline=True),
@@ -228,9 +228,9 @@ class RegexMatch(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="RegexMatch",
-            search_aliases=["Regex Match", "regex", "pattern match", "text contains", "string match"],
-            display_name="Text Match",
-            category="utils/string",
+            search_aliases=["regex match", "regex", "pattern match", "text contains", "string match"],
+            display_name="Match Text",
+            category="text",
             inputs=[
                 io.String.Input("string", multiline=True),
                 io.String.Input("regex_pattern", multiline=True),
@@ -269,9 +269,9 @@ class RegexExtract(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="RegexExtract",
-            search_aliases=["Regex Extract", "regex", "pattern extract", "text parser", "parse text"],
-            display_name="Text Extract Substring",
-            category="utils/string",
+            search_aliases=["regex extract", "regex", "pattern extract", "text parser", "parse text"],
+            display_name="Extract Text",
+            category="text",
             inputs=[
                 io.String.Input("string", multiline=True),
                 io.String.Input("regex_pattern", multiline=True),
@@ -344,9 +344,9 @@ class RegexReplace(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="RegexReplace",
-            search_aliases=["Regex Replace", "regex", "pattern replace", "regex replace", "substitution"],
-            display_name="Text Replace (Regex)",
-            category="utils/string",
+            search_aliases=["regex replace", "regex", "pattern replace", "substitution"],
+            display_name="Replace Text (Regex)",
+            category="text",
             description="Find and replace text using regex patterns.",
             inputs=[
                 io.String.Input("string", multiline=True),
@@ -381,8 +381,8 @@ class JsonExtractString(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="JsonExtractString",
-            display_name="Extract String from JSON",
-            category="utils/string",
+            display_name="Extract Text from JSON",
+            category="text",
             search_aliases=["json", "extract json", "parse json", "json value", "read json"],
             inputs=[
                 io.String.Input("json_string", multiline=True),

comfy_extras/nodes_textgen.py

@@ -26,12 +26,15 @@ class TextGenerate(io.ComfyNode):
 
         return io.Schema(
             node_id="TextGenerate",
-            category="textgen",
+            display_name="Generate Text",
+            category="text",
             search_aliases=["LLM", "gemma"],
             inputs=[
                 io.Clip.Input("clip"),
                 io.String.Input("prompt", multiline=True, dynamic_prompts=True, default=""),
                 io.Image.Input("image", optional=True),
+                io.Image.Input("video", optional=True, tooltip="Video frames as image batch. Assumed to be 24 FPS; subsampled to 1 FPS internally."),
+                io.Audio.Input("audio", optional=True),
                 io.Int.Input("max_length", default=256, min=1, max=2048),
                 io.DynamicCombo.Input("sampling_mode", options=sampling_options, display_name="Sampling Mode"),
                 io.Boolean.Input("thinking", optional=True, default=False, tooltip="Operate in thinking mode if the model supports it."),
@@ -43,9 +46,9 @@ class TextGenerate(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True) -> io.NodeOutput:
+    def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True, video=None, audio=None) -> io.NodeOutput:
 
-        tokens = clip.tokenize(prompt, image=image, skip_template=not use_default_template, min_length=1, thinking=thinking)
+        tokens = clip.tokenize(prompt, image=image, skip_template=not use_default_template, min_length=1, thinking=thinking, video=video, audio=audio)
 
         # Get sampling parameters from dynamic combo
         do_sample = sampling_mode.get("sampling_mode") == "on"
@@ -70,7 +73,8 @@ class TextGenerate(io.ComfyNode):
             seed=seed
         )
 
-        generated_text = clip.decode(generated_ids, skip_special_tokens=True)
+        generated_text = clip.decode(generated_ids)
+
         return io.NodeOutput(generated_text)
 
 
@@ -154,6 +158,7 @@ class TextGenerateLTX2Prompt(TextGenerate):
         parent_schema = super().define_schema()
         return io.Schema(
             node_id="TextGenerateLTX2Prompt",
+            display_name="Generate LTX2 Prompt",
             category=parent_schema.category,
             inputs=parent_schema.inputs,
             outputs=parent_schema.outputs,
@@ -161,12 +166,12 @@ class TextGenerateLTX2Prompt(TextGenerate):
         )
 
     @classmethod
-    def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True) -> io.NodeOutput:
+    def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True, video=None, audio=None) -> io.NodeOutput:
         if image is None:
             formatted_prompt = f"<start_of_turn>system\n{LTX2_T2V_SYSTEM_PROMPT.strip()}<end_of_turn>\n<start_of_turn>user\nUser Raw Input Prompt: {prompt}.<end_of_turn>\n<start_of_turn>model\n"
         else:
             formatted_prompt = f"<start_of_turn>system\n{LTX2_I2V_SYSTEM_PROMPT.strip()}<end_of_turn>\n<start_of_turn>user\n\n<image_soft_token>\n\nUser Raw Input Prompt: {prompt}.<end_of_turn>\n<start_of_turn>model\n"
-        return super().execute(clip, formatted_prompt, max_length, sampling_mode, image, thinking, use_default_template)
+        return super().execute(clip, formatted_prompt, max_length, sampling_mode, image=image, thinking=thinking, use_default_template=use_default_template, video=video, audio=audio)
 
 
 class TextgenExtension(ComfyExtension):

comfy_extras/nodes_torch_compile.py

@@ -10,7 +10,7 @@ class TorchCompileModel(io.ComfyNode):
     def define_schema(cls) -> io.Schema:
         return io.Schema(
             node_id="TorchCompileModel",
-            category="_for_testing",
+            category="experimental",
             inputs=[
                 io.Model.Input("model"),
                 io.Combo.Input(

comfy_extras/nodes_train.py

@@ -1361,7 +1361,7 @@ class SaveLoRA(io.ComfyNode):
             node_id="SaveLoRA",
             search_aliases=["export lora"],
             display_name="Save LoRA Weights",
-            category="loaders",
+            category="advanced/model_merging",
             is_experimental=True,
             is_output_node=True,
             inputs=[

comfy_extras/nodes_video.py

@@ -17,7 +17,8 @@ class SaveWEBM(io.ComfyNode):
         return io.Schema(
             node_id="SaveWEBM",
             search_aliases=["export webm"],
-            category="image/video",
+            display_name="Save WEBM",
+            category="video",
             is_experimental=True,
             inputs=[
                 io.Image.Input("images"),
@@ -72,7 +73,7 @@ class SaveVideo(io.ComfyNode):
             node_id="SaveVideo",
             search_aliases=["export video"],
             display_name="Save Video",
-            category="image/video",
+            category="video",
             essentials_category="Basics",
             description="Saves the input images to your ComfyUI output directory.",
             inputs=[
@@ -121,7 +122,7 @@ class CreateVideo(io.ComfyNode):
             node_id="CreateVideo",
             search_aliases=["images to video"],
             display_name="Create Video",
-            category="image/video",
+            category="video",
             description="Create a video from images.",
             inputs=[
                 io.Image.Input("images", tooltip="The images to create a video from."),
@@ -146,7 +147,7 @@ class GetVideoComponents(io.ComfyNode):
             node_id="GetVideoComponents",
             search_aliases=["extract frames", "split video", "video to images", "demux"],
             display_name="Get Video Components",
-            category="image/video",
+            category="video",
             description="Extracts all components from a video: frames, audio, and framerate.",
             inputs=[
                 io.Video.Input("video", tooltip="The video to extract components from."),
@@ -174,7 +175,7 @@ class LoadVideo(io.ComfyNode):
             node_id="LoadVideo",
             search_aliases=["import video", "open video", "video file"],
             display_name="Load Video",
-            category="image/video",
+            category="video",
             essentials_category="Basics",
             inputs=[
                 io.Combo.Input("file", options=sorted(files), upload=io.UploadType.video),
@@ -216,7 +217,7 @@ class VideoSlice(io.ComfyNode):
                 "frame load cap",
                 "start time",
             ],
-            category="image/video",
+            category="video",
             essentials_category="Video Tools",
             inputs=[
                 io.Video.Input("video"),

comfy_extras/nodes_video_model.py

@@ -15,7 +15,7 @@ class ImageOnlyCheckpointLoader:
     RETURN_TYPES = ("MODEL", "CLIP_VISION", "VAE")
     FUNCTION = "load_checkpoint"
 
-    CATEGORY = "loaders/video_models"
+    CATEGORY = "loaders"
 
     def load_checkpoint(self, ckpt_name, output_vae=True, output_clip=True):
         ckpt_path = folder_paths.get_full_path_or_raise("checkpoints", ckpt_name)

comfy_extras/nodes_void.py

@@ -0,0 +1,483 @@
+import logging
+
+import torch
+
+import comfy
+import comfy.model_management
+import comfy.model_patcher
+import comfy.samplers
+import comfy.utils
+import folder_paths
+import node_helpers
+import nodes
+from comfy.utils import model_trange as trange
+from comfy_api.latest import ComfyExtension, io
+from torchvision.models.optical_flow import raft_large
+from typing_extensions import override
+
+
+from comfy_extras.void_noise_warp import RaftOpticalFlow, get_noise_from_video
+
+OpticalFlow = io.Custom("OPTICAL_FLOW")
+
+TEMPORAL_COMPRESSION = 4
+PATCH_SIZE_T = 2
+
+
+def _valid_void_length(length: int) -> int:
+    """Round ``length`` down to a value that produces an even latent_t.
+
+    VOID / CogVideoX-Fun-V1.5 uses patch_size_t=2, so the VAE-encoded latent
+    must have an even temporal dimension. If latent_t is odd, the transformer
+    pad_to_patch_size circular-wraps an extra latent frame onto the end; after
+    the post-transformer crop the last real latent frame has been influenced
+    by the wrapped phantom frame, producing visible jitter and "disappearing"
+    subjects near the end of the decoded video. Rounding down fixes this.
+    """
+    latent_t = ((length - 1) // TEMPORAL_COMPRESSION) + 1
+    if latent_t % PATCH_SIZE_T == 0:
+        return length
+    # Round latent_t down to the nearest multiple of PATCH_SIZE_T, then invert
+    # the ((length - 1) // TEMPORAL_COMPRESSION) + 1 formula. Floor at 1 frame
+    # so we never return a non-positive length.
+    target_latent_t = max(PATCH_SIZE_T, (latent_t // PATCH_SIZE_T) * PATCH_SIZE_T)
+    return (target_latent_t - 1) * TEMPORAL_COMPRESSION + 1
+
+
+class OpticalFlowLoader(io.ComfyNode):
+    """Load an optical flow model from ``models/optical_flow/``.
+
+    Only torchvision's RAFT-large format is recognized today (the model used
+    by VOIDWarpedNoise).  The checkpoint must be placed under
+    ``models/optical_flow/`` — ComfyUI never downloads optical-flow weights
+    at runtime.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="OpticalFlowLoader",
+            display_name="Load Optical Flow Model",
+            category="loaders",
+            inputs=[
+                io.Combo.Input(
+                    "model_name",
+                    options=folder_paths.get_filename_list("optical_flow"),
+                    tooltip=(
+                        "Optical flow model to load.  Files must be placed in the "
+                        "'optical_flow' folder.  Today only torchvision's "
+                        "raft_large.pth is supported."
+                    ),
+                ),
+            ],
+            outputs=[
+                OpticalFlow.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, model_name) -> io.NodeOutput:
+
+        model_path = folder_paths.get_full_path_or_raise("optical_flow", model_name)
+        sd = comfy.utils.load_torch_file(model_path, safe_load=True)
+
+        has_raft_keys = (
+            any(k.startswith("feature_encoder.") for k in sd)
+            and any(k.startswith("context_encoder.") for k in sd)
+            and any(k.startswith("update_block.") for k in sd)
+        )
+        if not has_raft_keys:
+            raise ValueError(
+                "Unrecognized optical flow model format: expected a torchvision "
+                "RAFT-large state dict with 'feature_encoder.', 'context_encoder.' "
+                "and 'update_block.' prefixes."
+            )
+
+        model = raft_large(weights=None, progress=False)
+        model.load_state_dict(sd)
+        model.eval().to(torch.float32)
+
+        patcher = comfy.model_patcher.ModelPatcher(
+            model,
+            load_device=comfy.model_management.get_torch_device(),
+            offload_device=comfy.model_management.unet_offload_device(),
+        )
+        return io.NodeOutput(patcher)
+
+
+class VOIDQuadmaskPreprocess(io.ComfyNode):
+    """Preprocess a quadmask video for VOID inpainting.
+
+    Quantizes mask values to four semantic levels, inverts, and normalizes:
+      0   -> primary object to remove
+      63  -> overlap of primary + affected
+      127 -> affected region (interactions)
+      255 -> background (keep)
+
+    After inversion and normalization, the output mask has values in [0, 1]
+    with four discrete levels: 1.0 (remove), ~0.75, ~0.50, 0.0 (keep).
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDQuadmaskPreprocess",
+            category="mask/video",
+            inputs=[
+                io.Mask.Input("mask"),
+                io.Int.Input("dilate_width", default=0, min=0, max=50, step=1,
+                             tooltip="Dilation radius for the primary mask region (0 = no dilation)"),
+            ],
+            outputs=[
+                io.Mask.Output(display_name="quadmask"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, mask, dilate_width=0) -> io.NodeOutput:
+        m = mask.clone()
+
+        if m.max() <= 1.0:
+            m = m * 255.0
+
+        if dilate_width > 0 and m.ndim >= 3:
+            binary = (m < 128).float()
+            kernel_size = dilate_width * 2 + 1
+            if binary.ndim == 3:
+                binary = binary.unsqueeze(1)
+            dilated = torch.nn.functional.max_pool2d(
+                binary, kernel_size=kernel_size, stride=1, padding=dilate_width
+            )
+            if dilated.ndim == 4:
+                dilated = dilated.squeeze(1)
+            m = torch.where(dilated > 0.5, torch.zeros_like(m), m)
+
+        m = torch.where(m <= 31, torch.zeros_like(m), m)
+        m = torch.where((m > 31) & (m <= 95), torch.full_like(m, 63), m)
+        m = torch.where((m > 95) & (m <= 191), torch.full_like(m, 127), m)
+        m = torch.where(m > 191, torch.full_like(m, 255), m)
+
+        m = (255.0 - m) / 255.0
+
+        return io.NodeOutput(m)
+
+
+class VOIDInpaintConditioning(io.ComfyNode):
+    """Build VOID inpainting conditioning for CogVideoX.
+
+    Encodes the processed quadmask and masked source video through the VAE,
+    producing a 32-channel concat conditioning (16ch mask + 16ch masked video)
+    that gets concatenated with the 16ch noise latent by the model.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDInpaintConditioning",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Image.Input("video", tooltip="Source video frames [T, H, W, 3]"),
+                io.Mask.Input("quadmask", tooltip="Preprocessed quadmask from VOIDQuadmaskPreprocess [T, H, W]"),
+                io.Int.Input("width", default=672, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("height", default=384, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("length", default=45, min=1, max=nodes.MAX_RESOLUTION, step=1,
+                             tooltip="Number of pixel frames to process. For CogVideoX-Fun-V1.5 "
+                                     "(patch_size_t=2), latent_t must be even — lengths that "
+                                     "produce odd latent_t are rounded down (e.g. 49 → 45)."),
+                io.Int.Input("batch_size", default=1, min=1, max=64),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, video, quadmask,
+                width, height, length, batch_size) -> io.NodeOutput:
+
+        adjusted_length = _valid_void_length(length)
+        if adjusted_length != length:
+            logging.warning(
+                "VOIDInpaintConditioning: rounding length %d down to %d so that "
+                "latent_t is even (required by CogVideoX-Fun-V1.5 patch_size_t=2). "
+                "Using odd latent_t causes the last frame to be corrupted by "
+                "circular padding.", length, adjusted_length,
+            )
+            length = adjusted_length
+
+        latent_t = ((length - 1) // TEMPORAL_COMPRESSION) + 1
+        latent_h = height // 8
+        latent_w = width // 8
+
+        vid = video[:length]
+        vid = comfy.utils.common_upscale(
+            vid.movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+
+        qm = quadmask[:length]
+        if qm.ndim == 3:
+            qm = qm.unsqueeze(-1)
+        qm = comfy.utils.common_upscale(
+            qm.movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+        if qm.ndim == 4 and qm.shape[-1] == 1:
+            qm = qm.squeeze(-1)
+
+        mask_condition = qm
+        if mask_condition.ndim == 3:
+            mask_condition_3ch = mask_condition.unsqueeze(-1).expand(-1, -1, -1, 3)
+        else:
+            mask_condition_3ch = mask_condition
+
+        inverted_mask_3ch = 1.0 - mask_condition_3ch
+        masked_video = vid[:, :, :, :3] * (1.0 - mask_condition_3ch)
+
+        mask_latents = vae.encode(inverted_mask_3ch)
+        masked_video_latents = vae.encode(masked_video)
+
+        def _match_temporal(lat, target_t):
+            if lat.shape[2] > target_t:
+                return lat[:, :, :target_t]
+            elif lat.shape[2] < target_t:
+                pad = target_t - lat.shape[2]
+                return torch.cat([lat, lat[:, :, -1:].repeat(1, 1, pad, 1, 1)], dim=2)
+            return lat
+
+        mask_latents = _match_temporal(mask_latents, latent_t)
+        masked_video_latents = _match_temporal(masked_video_latents, latent_t)
+
+        inpaint_latents = torch.cat([mask_latents, masked_video_latents], dim=1)
+
+        # No explicit scaling needed here: the model's CogVideoX.concat_cond()
+        # applies process_latent_in (×latent_format.scale_factor) to each 16-ch
+        # block of the stored conditioning. For 5b-class checkpoints (incl. the
+        # VOID/CogVideoX-Fun-V1.5 inpainting model) that scale_factor is auto-
+        # selected as 0.7 in supported_models.CogVideoX_T2V, which matches the
+        # diffusers vae/config.json scaling_factor VOID was trained with.
+
+        positive = node_helpers.conditioning_set_values(
+            positive, {"concat_latent_image": inpaint_latents}
+        )
+        negative = node_helpers.conditioning_set_values(
+            negative, {"concat_latent_image": inpaint_latents}
+        )
+
+        noise_latent = torch.zeros(
+            [batch_size, 16, latent_t, latent_h, latent_w],
+            device=comfy.model_management.intermediate_device()
+        )
+
+        return io.NodeOutput(positive, negative, {"samples": noise_latent})
+
+
+class VOIDWarpedNoise(io.ComfyNode):
+    """Generate optical-flow warped noise for VOID Pass 2 refinement.
+
+    Takes the Pass 1 output video and produces temporally-correlated noise
+    by warping Gaussian noise along optical flow vectors. This noise is used
+    as the initial latent for Pass 2, resulting in better temporal consistency.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDWarpedNoise",
+            category="latent/video",
+            inputs=[
+                OpticalFlow.Input(
+                    "optical_flow",
+                    tooltip="Optical flow model from OpticalFlowLoader (RAFT-large).",
+                ),
+                io.Image.Input("video", tooltip="Pass 1 output video frames [T, H, W, 3]"),
+                io.Int.Input("width", default=672, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("height", default=384, min=16, max=nodes.MAX_RESOLUTION, step=8),
+                io.Int.Input("length", default=45, min=1, max=nodes.MAX_RESOLUTION, step=1,
+                             tooltip="Number of pixel frames. Rounded down to make latent_t "
+                                     "even (patch_size_t=2 requirement), e.g. 49 → 45."),
+                io.Int.Input("batch_size", default=1, min=1, max=64),
+            ],
+            outputs=[
+                io.Latent.Output(display_name="warped_noise"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, optical_flow, video, width, height, length, batch_size) -> io.NodeOutput:
+
+        adjusted_length = _valid_void_length(length)
+        if adjusted_length != length:
+            logging.warning(
+                "VOIDWarpedNoise: rounding length %d down to %d so that "
+                "latent_t is even (required by CogVideoX-Fun-V1.5 patch_size_t=2).",
+                length, adjusted_length,
+            )
+            length = adjusted_length
+
+        latent_t = ((length - 1) // TEMPORAL_COMPRESSION) + 1
+        latent_h = height // 8
+        latent_w = width // 8
+
+        # RAFT + noise warp is real compute, not an "intermediate" buffer, so
+        # we want the actual torch device (CUDA/MPS).  The final latent is
+        # moved back to intermediate_device() before returning to match the
+        # rest of the ComfyUI pipeline.
+        device = comfy.model_management.get_torch_device()
+
+        comfy.model_management.load_model_gpu(optical_flow)
+        raft = RaftOpticalFlow(optical_flow.model, device=device)
+
+        vid = video[:length].to(device)
+        vid = comfy.utils.common_upscale(
+            vid.movedim(-1, 1), width, height, "bilinear", "center"
+        ).movedim(1, -1)
+        vid_uint8 = (vid.clamp(0, 1) * 255).to(torch.uint8)
+
+        FRAME = 2**-1
+        FLOW = 2**3
+        LATENT_SCALE = 8
+
+        warped = get_noise_from_video(
+            vid_uint8,
+            raft,
+            noise_channels=16,
+            resize_frames=FRAME,
+            resize_flow=FLOW,
+            downscale_factor=round(FRAME * FLOW) * LATENT_SCALE,
+            device=device,
+        )
+
+        if warped.shape[0] != latent_t:
+            indices = torch.linspace(0, warped.shape[0] - 1, latent_t,
+                                     device=device).long()
+            warped = warped[indices]
+
+        if warped.shape[1] != latent_h or warped.shape[2] != latent_w:
+            # (T, H, W, C) → (T, C, H, W) → bilinear resize → back
+            warped = warped.permute(0, 3, 1, 2)
+            warped = torch.nn.functional.interpolate(
+                warped, size=(latent_h, latent_w),
+                mode="bilinear", align_corners=False,
+            )
+            warped = warped.permute(0, 2, 3, 1)
+
+        # (T, H, W, C) → (B, C, T, H, W)
+        warped_tensor = warped.permute(3, 0, 1, 2).unsqueeze(0)
+        if batch_size > 1:
+            warped_tensor = warped_tensor.repeat(batch_size, 1, 1, 1, 1)
+
+        warped_tensor = warped_tensor.to(comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": warped_tensor})
+
+
+class Noise_FromLatent:
+    """Wraps a pre-computed LATENT tensor as a NOISE source."""
+    def __init__(self, latent_dict):
+        self.seed = 0
+        self._samples = latent_dict["samples"]
+
+    def generate_noise(self, input_latent):
+        return self._samples.clone().cpu()
+
+
+class VOIDWarpedNoiseSource(io.ComfyNode):
+    """Convert a LATENT (e.g. from VOIDWarpedNoise) into a NOISE source
+    for use with SamplerCustomAdvanced."""
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDWarpedNoiseSource",
+            category="sampling/custom_sampling/noise",
+            inputs=[
+                io.Latent.Input("warped_noise",
+                    tooltip="Warped noise latent from VOIDWarpedNoise"),
+            ],
+            outputs=[io.Noise.Output()],
+        )
+
+    @classmethod
+    def execute(cls, warped_noise) -> io.NodeOutput:
+        return io.NodeOutput(Noise_FromLatent(warped_noise))
+
+
+class VOID_DDIM(comfy.samplers.Sampler):
+    """DDIM sampler for VOID inpainting models.
+
+    VOID was trained with the diffusers CogVideoXDDIMScheduler which operates in
+    alpha-space (input std ≈ 1). The standard KSampler applies noise_scaling that
+    multiplies by sqrt(1+sigma^2) ≈ 4500x, which is incompatible with VOID's
+    training. This sampler skips noise_scaling and implements the DDIM update rule
+    directly using sigma-to-alpha conversion.
+    """
+
+    def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
+        x = noise.to(torch.float32)
+        model_options = extra_args.get("model_options", {})
+        seed = extra_args.get("seed", None)
+        s_in = x.new_ones([x.shape[0]])
+
+        for i in trange(len(sigmas) - 1, disable=disable_pbar):
+            sigma = sigmas[i]
+            sigma_next = sigmas[i + 1]
+
+            denoised = model_wrap(x, sigma * s_in, model_options=model_options, seed=seed)
+
+            if callback is not None:
+                callback(i, denoised, x, len(sigmas) - 1)
+
+            if sigma_next == 0:
+                x = denoised
+            else:
+                alpha_t = 1.0 / (1.0 + sigma ** 2)
+                alpha_prev = 1.0 / (1.0 + sigma_next ** 2)
+
+                pred_eps = (x - (alpha_t ** 0.5) * denoised) / (1.0 - alpha_t) ** 0.5
+                x = (alpha_prev ** 0.5) * denoised + (1.0 - alpha_prev) ** 0.5 * pred_eps
+
+        return x
+
+
+class VOIDSampler(io.ComfyNode):
+    """VOID DDIM sampler for use with SamplerCustom / SamplerCustomAdvanced.
+
+    Required for VOID inpainting models. Implements the same DDIM loop that VOID
+    was trained with (diffusers CogVideoXDDIMScheduler), without the noise_scaling
+    that the standard KSampler applies. Use with RandomNoise or VOIDWarpedNoiseSource.
+    """
+
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="VOIDSampler",
+            category="sampling/custom_sampling/samplers",
+            inputs=[],
+            outputs=[io.Sampler.Output()],
+        )
+
+    @classmethod
+    def execute(cls) -> io.NodeOutput:
+        return io.NodeOutput(VOID_DDIM())
+
+    get_sampler = execute
+
+
+class VOIDExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            OpticalFlowLoader,
+            VOIDQuadmaskPreprocess,
+            VOIDInpaintConditioning,
+            VOIDWarpedNoise,
+            VOIDWarpedNoiseSource,
+            VOIDSampler,
+        ]
+
+
+async def comfy_entrypoint() -> VOIDExtension:
+    return VOIDExtension()