Merge branch 'master' into feature/deploy-environment-header

Read a .comfy_environment file from the ComfyUI base directory to
determine the deployment environment (e.g. standalone, portable, desktop).
Defaults to 'local_git' when the file is absent.

The value is sent as an X-Comfy-Deploy-Env header on all requests to
api.comfy.org, allowing the API to differentiate between environment types.

The .comfy_environment file is gitignored so launchers/installers can
write it without affecting the repository.

Co-authored-by: Amp <amp@ampcode.com>
Amp-Thread-ID: https://ampcode.com/threads/T-019d939e-6b4d-738b-8d1a-ac7cbf6736a4

.ci/windows_intel_base_files/run_intel_gpu.bat

@@ -0,0 +1,2 @@
+.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
+pause

.github/workflows/release-stable-all.yml

@@ -20,29 +20,12 @@ jobs:
       git_tag: ${{ inputs.git_tag }}
       cache_tag: "cu130"
       python_minor: "13"
-      python_patch: "11"
+      python_patch: "12"
       rel_name: "nvidia"
       rel_extra_name: ""
       test_release: true
     secrets: inherit
 
-  release_nvidia_cu128:
-    permissions:
-      contents: "write"
-      packages: "write"
-      pull-requests: "read"
-    name: "Release NVIDIA cu128"
-    uses: ./.github/workflows/stable-release.yml
-    with:
-      git_tag: ${{ inputs.git_tag }}
-      cache_tag: "cu128"
-      python_minor: "12"
-      python_patch: "10"
-      rel_name: "nvidia"
-      rel_extra_name: "_cu128"
-      test_release: true
-    secrets: inherit
-
   release_nvidia_cu126:
     permissions:
       contents: "write"
@@ -76,3 +59,20 @@ jobs:
       rel_extra_name: ""
       test_release: false
     secrets: inherit
+
+  release_xpu:
+    permissions:
+      contents: "write"
+      packages: "write"
+      pull-requests: "read"
+    name: "Release Intel XPU"
+    uses: ./.github/workflows/stable-release.yml
+    with:
+      git_tag: ${{ inputs.git_tag }}
+      cache_tag: "xpu"
+      python_minor: "13"
+      python_patch: "12"
+      rel_name: "intel"
+      rel_extra_name: ""
+      test_release: true
+    secrets: inherit

.gitignore

@@ -24,3 +24,4 @@ web_custom_versions/
 openapi.yaml
 filtered-openapi.yaml
 uv.lock
+.comfy_environment

QUANTIZATION.md

@@ -139,9 +139,9 @@ Example:
   "_quantization_metadata": {
     "format_version": "1.0",
     "layers": {
-      "model.layers.0.mlp.up_proj": "float8_e4m3fn",
-      "model.layers.0.mlp.down_proj": "float8_e4m3fn",
-      "model.layers.1.mlp.up_proj": "float8_e4m3fn"
+      "model.layers.0.mlp.up_proj": {"format": "float8_e4m3fn"},
+      "model.layers.0.mlp.down_proj": {"format": "float8_e4m3fn"},
+      "model.layers.1.mlp.up_proj": {"format": "float8_e4m3fn"}
     }
   }
 }
@@ -165,4 +165,4 @@ Activation quantization (e.g., for FP8 Tensor Core operations) requires `input_s
 3. **Compute scales**: Derive `input_scale` from collected statistics
 4. **Store in checkpoint**: Save `input_scale` parameters alongside weights
 
-The calibration dataset should be representative of your target use case. For diffusion models, this typically means a diverse set of prompts and generation parameters.
+The calibration dataset should be representative of your target use case. For diffusion models, this typically means a diverse set of prompts and generation parameters.

README.md

@@ -61,6 +61,7 @@ See what ComfyUI can do with the [newer template workflows](https://comfy.org/wo
 
 ## Features
 - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.
+- NOTE: There are many more models supported than the list below, if you want to see what is supported see our templates list inside ComfyUI.
 - Image Models
    - SD1.x, SD2.x ([unCLIP](https://comfyanonymous.github.io/ComfyUI_examples/unclip/))
    - [SDXL](https://comfyanonymous.github.io/ComfyUI_examples/sdxl/), [SDXL Turbo](https://comfyanonymous.github.io/ComfyUI_examples/sdturbo/)
@@ -136,7 +137,7 @@ ComfyUI follows a weekly release cycle targeting Monday but this regularly chang
    - Builds a new release using the latest stable core version
 
 3. **[ComfyUI Frontend](https://github.com/Comfy-Org/ComfyUI_frontend)**
-   - Weekly frontend updates are merged into the core repository
+   - Every 2+ weeks frontend updates are merged into the core repository
    - Features are frozen for the upcoming core release
    - Development continues for the next release cycle
 
@@ -232,7 +233,7 @@ Put your VAE in: models/vae
 
 AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version:
 
-```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.1```
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2```
 
 This is the command to install the nightly with ROCm 7.2 which might have some performance improvements:
 
@@ -275,7 +276,7 @@ Nvidia users should install stable pytorch using this command:
 
 This is the command to install pytorch nightly instead which might have performance improvements.
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu130```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu132```
 
 #### Troubleshooting
 

blueprints/Image Channels.json

@@ -1 +1,322 @@
-{"revision": 0, "last_node_id": 29, "last_link_id": 0, "nodes": [{"id": 29, "type": "4c9d6ea4-b912-40e5-8766-6793a9758c53", "pos": [1970, -230], "size": [180, 86], "flags": {}, "order": 5, "mode": 0, "inputs": [{"label": "image", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": null}], "outputs": [{"label": "R", "localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": []}, {"label": "G", "localized_name": "IMAGE1", "name": "IMAGE1", "type": "IMAGE", "links": []}, {"label": "B", "localized_name": "IMAGE2", "name": "IMAGE2", "type": "IMAGE", "links": []}, {"label": "A", "localized_name": "IMAGE3", "name": "IMAGE3", "type": "IMAGE", "links": []}], "title": "Image Channels", "properties": {"proxyWidgets": []}, "widgets_values": []}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "4c9d6ea4-b912-40e5-8766-6793a9758c53", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 28, "lastLinkId": 39, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Image Channels", "inputNode": {"id": -10, "bounding": [1820, -185, 120, 60]}, "outputNode": {"id": -20, "bounding": [2460, -215, 120, 120]}, "inputs": [{"id": "3522932b-2d86-4a1f-a02a-cb29f3a9d7fe", "name": "images.image0", "type": "IMAGE", "linkIds": [39], "localized_name": "images.image0", "label": "image", "pos": [1920, -165]}], "outputs": [{"id": "605cb9c3-b065-4d9b-81d2-3ec331889b2b", "name": "IMAGE0", "type": "IMAGE", "linkIds": [26], "localized_name": "IMAGE0", "label": "R", "pos": [2480, -195]}, {"id": "fb44a77e-0522-43e9-9527-82e7465b3596", "name": "IMAGE1", "type": "IMAGE", "linkIds": [27], "localized_name": "IMAGE1", "label": "G", "pos": [2480, -175]}, {"id": "81460ee6-0131-402a-874f-6bf3001fc4ff", "name": "IMAGE2", "type": "IMAGE", "linkIds": [28], "localized_name": "IMAGE2", "label": "B", "pos": [2480, -155]}, {"id": "ae690246-80d4-4951-b1d9-9306d8a77417", "name": "IMAGE3", "type": "IMAGE", "linkIds": [29], "localized_name": "IMAGE3", "label": "A", "pos": [2480, -135]}], "widgets": [], "nodes": [{"id": 23, "type": "GLSLShader", "pos": [2000, -330], "size": [400, 172], "flags": {}, "order": 0, "mode": 0, "inputs": [{"label": "image", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": 39}, {"localized_name": "fragment_shader", "name": "fragment_shader", "type": "STRING", "widget": {"name": "fragment_shader"}, "link": null}, {"localized_name": "size_mode", "name": "size_mode", "type": "COMFY_DYNAMICCOMBO_V3", "widget": {"name": "size_mode"}, "link": null}, {"label": "image1", "localized_name": "images.image1", "name": "images.image1", "shape": 7, "type": "IMAGE", "link": null}], "outputs": [{"label": "R", "localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": [26]}, {"label": "G", "localized_name": "IMAGE1", "name": "IMAGE1", "type": "IMAGE", "links": [27]}, {"label": "B", "localized_name": "IMAGE2", "name": "IMAGE2", "type": "IMAGE", "links": [28]}, {"label": "A", "localized_name": "IMAGE3", "name": "IMAGE3", "type": "IMAGE", "links": [29]}], "properties": {"Node name for S&R": "GLSLShader"}, "widgets_values": ["#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\nlayout(location = 1) out vec4 fragColor1;\nlayout(location = 2) out vec4 fragColor2;\nlayout(location = 3) out vec4 fragColor3;\n\nvoid main() {\n  vec4 color = texture(u_image0, v_texCoord);\n  // Output each channel as grayscale to separate render targets\n  fragColor0 = vec4(vec3(color.r), 1.0);  // Red channel\n  fragColor1 = vec4(vec3(color.g), 1.0);  // Green channel\n  fragColor2 = vec4(vec3(color.b), 1.0);  // Blue channel\n  fragColor3 = vec4(vec3(color.a), 1.0);  // Alpha channel\n}\n", "from_input"]}], "groups": [], "links": [{"id": 39, "origin_id": -10, "origin_slot": 0, "target_id": 23, "target_slot": 0, "type": "IMAGE"}, {"id": 26, "origin_id": 23, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "IMAGE"}, {"id": 27, "origin_id": 23, "origin_slot": 1, "target_id": -20, "target_slot": 1, "type": "IMAGE"}, {"id": 28, "origin_id": 23, "origin_slot": 2, "target_id": -20, "target_slot": 2, "type": "IMAGE"}, {"id": 29, "origin_id": 23, "origin_slot": 3, "target_id": -20, "target_slot": 3, "type": "IMAGE"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Image Tools/Color adjust"}]}}
+{
+  "revision": 0,
+  "last_node_id": 29,
+  "last_link_id": 0,
+  "nodes": [
+    {
+      "id": 29,
+      "type": "4c9d6ea4-b912-40e5-8766-6793a9758c53",
+      "pos": [
+        1970,
+        -230
+      ],
+      "size": [
+        180,
+        86
+      ],
+      "flags": {},
+      "order": 5,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "image",
+          "localized_name": "images.image0",
+          "name": "images.image0",
+          "type": "IMAGE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "R",
+          "localized_name": "IMAGE0",
+          "name": "IMAGE0",
+          "type": "IMAGE",
+          "links": []
+        },
+        {
+          "label": "G",
+          "localized_name": "IMAGE1",
+          "name": "IMAGE1",
+          "type": "IMAGE",
+          "links": []
+        },
+        {
+          "label": "B",
+          "localized_name": "IMAGE2",
+          "name": "IMAGE2",
+          "type": "IMAGE",
+          "links": []
+        },
+        {
+          "label": "A",
+          "localized_name": "IMAGE3",
+          "name": "IMAGE3",
+          "type": "IMAGE",
+          "links": []
+        }
+      ],
+      "title": "Image Channels",
+      "properties": {
+        "proxyWidgets": []
+      },
+      "widgets_values": []
+    }
+  ],
+  "links": [],
+  "version": 0.4,
+  "definitions": {
+    "subgraphs": [
+      {
+        "id": "4c9d6ea4-b912-40e5-8766-6793a9758c53",
+        "version": 1,
+        "state": {
+          "lastGroupId": 0,
+          "lastNodeId": 28,
+          "lastLinkId": 39,
+          "lastRerouteId": 0
+        },
+        "revision": 0,
+        "config": {},
+        "name": "Image Channels",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            1820,
+            -185,
+            120,
+            60
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            2460,
+            -215,
+            120,
+            120
+          ]
+        },
+        "inputs": [
+          {
+            "id": "3522932b-2d86-4a1f-a02a-cb29f3a9d7fe",
+            "name": "images.image0",
+            "type": "IMAGE",
+            "linkIds": [
+              39
+            ],
+            "localized_name": "images.image0",
+            "label": "image",
+            "pos": [
+              1920,
+              -165
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "605cb9c3-b065-4d9b-81d2-3ec331889b2b",
+            "name": "IMAGE0",
+            "type": "IMAGE",
+            "linkIds": [
+              26
+            ],
+            "localized_name": "IMAGE0",
+            "label": "R",
+            "pos": [
+              2480,
+              -195
+            ]
+          },
+          {
+            "id": "fb44a77e-0522-43e9-9527-82e7465b3596",
+            "name": "IMAGE1",
+            "type": "IMAGE",
+            "linkIds": [
+              27
+            ],
+            "localized_name": "IMAGE1",
+            "label": "G",
+            "pos": [
+              2480,
+              -175
+            ]
+          },
+          {
+            "id": "81460ee6-0131-402a-874f-6bf3001fc4ff",
+            "name": "IMAGE2",
+            "type": "IMAGE",
+            "linkIds": [
+              28
+            ],
+            "localized_name": "IMAGE2",
+            "label": "B",
+            "pos": [
+              2480,
+              -155
+            ]
+          },
+          {
+            "id": "ae690246-80d4-4951-b1d9-9306d8a77417",
+            "name": "IMAGE3",
+            "type": "IMAGE",
+            "linkIds": [
+              29
+            ],
+            "localized_name": "IMAGE3",
+            "label": "A",
+            "pos": [
+              2480,
+              -135
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 23,
+            "type": "GLSLShader",
+            "pos": [
+              2000,
+              -330
+            ],
+            "size": [
+              400,
+              172
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "label": "image",
+                "localized_name": "images.image0",
+                "name": "images.image0",
+                "type": "IMAGE",
+                "link": 39
+              },
+              {
+                "localized_name": "fragment_shader",
+                "name": "fragment_shader",
+                "type": "STRING",
+                "widget": {
+                  "name": "fragment_shader"
+                },
+                "link": null
+              },
+              {
+                "localized_name": "size_mode",
+                "name": "size_mode",
+                "type": "COMFY_DYNAMICCOMBO_V3",
+                "widget": {
+                  "name": "size_mode"
+                },
+                "link": null
+              },
+              {
+                "label": "image1",
+                "localized_name": "images.image1",
+                "name": "images.image1",
+                "shape": 7,
+                "type": "IMAGE",
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "label": "R",
+                "localized_name": "IMAGE0",
+                "name": "IMAGE0",
+                "type": "IMAGE",
+                "links": [
+                  26
+                ]
+              },
+              {
+                "label": "G",
+                "localized_name": "IMAGE1",
+                "name": "IMAGE1",
+                "type": "IMAGE",
+                "links": [
+                  27
+                ]
+              },
+              {
+                "label": "B",
+                "localized_name": "IMAGE2",
+                "name": "IMAGE2",
+                "type": "IMAGE",
+                "links": [
+                  28
+                ]
+              },
+              {
+                "label": "A",
+                "localized_name": "IMAGE3",
+                "name": "IMAGE3",
+                "type": "IMAGE",
+                "links": [
+                  29
+                ]
+              }
+            ],
+            "properties": {
+              "Node name for S&R": "GLSLShader"
+            },
+            "widgets_values": [
+              "#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\nlayout(location = 1) out vec4 fragColor1;\nlayout(location = 2) out vec4 fragColor2;\nlayout(location = 3) out vec4 fragColor3;\n\nvoid main() {\n  vec4 color = texture(u_image0, v_texCoord);\n  // Output each channel as grayscale to separate render targets\n  fragColor0 = vec4(vec3(color.r), 1.0);  // Red channel\n  fragColor1 = vec4(vec3(color.g), 1.0);  // Green channel\n  fragColor2 = vec4(vec3(color.b), 1.0);  // Blue channel\n  fragColor3 = vec4(vec3(color.a), 1.0);  // Alpha channel\n}\n",
+              "from_input"
+            ]
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 39,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 23,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 26,
+            "origin_id": 23,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 27,
+            "origin_id": 23,
+            "origin_slot": 1,
+            "target_id": -20,
+            "target_slot": 1,
+            "type": "IMAGE"
+          },
+          {
+            "id": 28,
+            "origin_id": 23,
+            "origin_slot": 2,
+            "target_id": -20,
+            "target_slot": 2,
+            "type": "IMAGE"
+          },
+          {
+            "id": 29,
+            "origin_id": 23,
+            "origin_slot": 3,
+            "target_id": -20,
+            "target_slot": 3,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {
+          "workflowRendererVersion": "LG"
+        },
+        "category": "Image Tools/Color adjust"
+      }
+    ]
+  }
+}

blueprints/Prompt Enhance.json

@@ -1 +1,278 @@
-{"revision": 0, "last_node_id": 15, "last_link_id": 0, "nodes": [{"id": 15, "type": "24d8bbfd-39d4-4774-bff0-3de40cc7a471", "pos": [-1490, 2040], "size": [400, 260], "flags": {}, "order": 0, "mode": 0, "inputs": [{"name": "prompt", "type": "STRING", "widget": {"name": "prompt"}, "link": null}, {"label": "reference images", "name": "images", "type": "IMAGE", "link": null}], "outputs": [{"name": "STRING", "type": "STRING", "links": null}], "title": "Prompt Enhance", "properties": {"proxyWidgets": [["-1", "prompt"]], "cnr_id": "comfy-core", "ver": "0.14.1"}, "widgets_values": [""]}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "24d8bbfd-39d4-4774-bff0-3de40cc7a471", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 15, "lastLinkId": 14, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Prompt Enhance", "inputNode": {"id": -10, "bounding": [-2170, 2110, 138.876953125, 80]}, "outputNode": {"id": -20, "bounding": [-640, 2110, 120, 60]}, "inputs": [{"id": "aeab7216-00e0-4528-a09b-bba50845c5a6", "name": "prompt", "type": "STRING", "linkIds": [11], "pos": [-2051.123046875, 2130]}, {"id": "7b73fd36-aa31-4771-9066-f6c83879994b", "name": "images", "type": "IMAGE", "linkIds": [14], "label": "reference images", "pos": [-2051.123046875, 2150]}], "outputs": [{"id": "c7b0d930-68a1-48d1-b496-0519e5837064", "name": "STRING", "type": "STRING", "linkIds": [13], "pos": [-620, 2130]}], "widgets": [], "nodes": [{"id": 11, "type": "GeminiNode", "pos": [-1560, 1990], "size": [470, 470], "flags": {}, "order": 0, "mode": 0, "inputs": [{"localized_name": "images", "name": "images", "shape": 7, "type": "IMAGE", "link": 14}, {"localized_name": "audio", "name": "audio", "shape": 7, "type": "AUDIO", "link": null}, {"localized_name": "video", "name": "video", "shape": 7, "type": "VIDEO", "link": null}, {"localized_name": "files", "name": "files", "shape": 7, "type": "GEMINI_INPUT_FILES", "link": null}, {"localized_name": "prompt", "name": "prompt", "type": "STRING", "widget": {"name": "prompt"}, "link": 11}, {"localized_name": "model", "name": "model", "type": "COMBO", "widget": {"name": "model"}, "link": null}, {"localized_name": "seed", "name": "seed", "type": "INT", "widget": {"name": "seed"}, "link": null}, {"localized_name": "system_prompt", "name": "system_prompt", "shape": 7, "type": "STRING", "widget": {"name": "system_prompt"}, "link": null}], "outputs": [{"localized_name": "STRING", "name": "STRING", "type": "STRING", "links": [13]}], "properties": {"cnr_id": "comfy-core", "ver": "0.14.1", "Node name for S&R": "GeminiNode"}, "widgets_values": ["", "gemini-3-pro-preview", 42, "randomize", "You are an expert in prompt writing.\nBased on the input, rewrite the user's input into a detailed prompt.\nincluding camera settings, lighting, composition, and style.\nReturn the prompt only"], "color": "#432", "bgcolor": "#653"}], "groups": [], "links": [{"id": 11, "origin_id": -10, "origin_slot": 0, "target_id": 11, "target_slot": 4, "type": "STRING"}, {"id": 13, "origin_id": 11, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "STRING"}, {"id": 14, "origin_id": -10, "origin_slot": 1, "target_id": 11, "target_slot": 0, "type": "IMAGE"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Text generation/Prompt enhance"}]}, "extra": {}}
+{
+  "revision": 0,
+  "last_node_id": 15,
+  "last_link_id": 0,
+  "nodes": [
+    {
+      "id": 15,
+      "type": "24d8bbfd-39d4-4774-bff0-3de40cc7a471",
+      "pos": [
+        -1490,
+        2040
+      ],
+      "size": [
+        400,
+        260
+      ],
+      "flags": {},
+      "order": 0,
+      "mode": 0,
+      "inputs": [
+        {
+          "name": "prompt",
+          "type": "STRING",
+          "widget": {
+            "name": "prompt"
+          },
+          "link": null
+        },
+        {
+          "label": "reference images",
+          "name": "images",
+          "type": "IMAGE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "name": "STRING",
+          "type": "STRING",
+          "links": null
+        }
+      ],
+      "title": "Prompt Enhance",
+      "properties": {
+        "proxyWidgets": [
+          [
+            "-1",
+            "prompt"
+          ]
+        ],
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1"
+      },
+      "widgets_values": [
+        ""
+      ]
+    }
+  ],
+  "links": [],
+  "version": 0.4,
+  "definitions": {
+    "subgraphs": [
+      {
+        "id": "24d8bbfd-39d4-4774-bff0-3de40cc7a471",
+        "version": 1,
+        "state": {
+          "lastGroupId": 0,
+          "lastNodeId": 15,
+          "lastLinkId": 14,
+          "lastRerouteId": 0
+        },
+        "revision": 0,
+        "config": {},
+        "name": "Prompt Enhance",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            -2170,
+            2110,
+            138.876953125,
+            80
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            -640,
+            2110,
+            120,
+            60
+          ]
+        },
+        "inputs": [
+          {
+            "id": "aeab7216-00e0-4528-a09b-bba50845c5a6",
+            "name": "prompt",
+            "type": "STRING",
+            "linkIds": [
+              11
+            ],
+            "pos": [
+              -2051.123046875,
+              2130
+            ]
+          },
+          {
+            "id": "7b73fd36-aa31-4771-9066-f6c83879994b",
+            "name": "images",
+            "type": "IMAGE",
+            "linkIds": [
+              14
+            ],
+            "label": "reference images",
+            "pos": [
+              -2051.123046875,
+              2150
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "c7b0d930-68a1-48d1-b496-0519e5837064",
+            "name": "STRING",
+            "type": "STRING",
+            "linkIds": [
+              13
+            ],
+            "pos": [
+              -620,
+              2130
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 11,
+            "type": "GeminiNode",
+            "pos": [
+              -1560,
+              1990
+            ],
+            "size": [
+              470,
+              470
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "shape": 7,
+                "type": "IMAGE",
+                "link": 14
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "shape": 7,
+                "type": "AUDIO",
+                "link": null
+              },
+              {
+                "localized_name": "video",
+                "name": "video",
+                "shape": 7,
+                "type": "VIDEO",
+                "link": null
+              },
+              {
+                "localized_name": "files",
+                "name": "files",
+                "shape": 7,
+                "type": "GEMINI_INPUT_FILES",
+                "link": null
+              },
+              {
+                "localized_name": "prompt",
+                "name": "prompt",
+                "type": "STRING",
+                "widget": {
+                  "name": "prompt"
+                },
+                "link": 11
+              },
+              {
+                "localized_name": "model",
+                "name": "model",
+                "type": "COMBO",
+                "widget": {
+                  "name": "model"
+                },
+                "link": null
+              },
+              {
+                "localized_name": "seed",
+                "name": "seed",
+                "type": "INT",
+                "widget": {
+                  "name": "seed"
+                },
+                "link": null
+              },
+              {
+                "localized_name": "system_prompt",
+                "name": "system_prompt",
+                "shape": 7,
+                "type": "STRING",
+                "widget": {
+                  "name": "system_prompt"
+                },
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "STRING",
+                "name": "STRING",
+                "type": "STRING",
+                "links": [
+                  13
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.14.1",
+              "Node name for S&R": "GeminiNode"
+            },
+            "widgets_values": [
+              "",
+              "gemini-3-pro-preview",
+              42,
+              "randomize",
+              "You are an expert in prompt writing.\nBased on the input, rewrite the user's input into a detailed prompt.\nincluding camera settings, lighting, composition, and style.\nReturn the prompt only"
+            ],
+            "color": "#432",
+            "bgcolor": "#653"
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 11,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 11,
+            "target_slot": 4,
+            "type": "STRING"
+          },
+          {
+            "id": 13,
+            "origin_id": 11,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "STRING"
+          },
+          {
+            "id": 14,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 11,
+            "target_slot": 0,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {
+          "workflowRendererVersion": "LG"
+        },
+        "category": "Text generation/Prompt enhance"
+      }
+    ]
+  },
+  "extra": {}
+}

blueprints/Sharpen.json

@@ -1 +1,309 @@
-{"revision": 0, "last_node_id": 25, "last_link_id": 0, "nodes": [{"id": 25, "type": "621ba4e2-22a8-482d-a369-023753198b7b", "pos": [4610, -790], "size": [230, 58], "flags": {}, "order": 4, "mode": 0, "inputs": [{"label": "image", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": null}], "outputs": [{"label": "IMAGE", "localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": []}], "title": "Sharpen", "properties": {"proxyWidgets": [["24", "value"]]}, "widgets_values": []}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "621ba4e2-22a8-482d-a369-023753198b7b", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 24, "lastLinkId": 36, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Sharpen", "inputNode": {"id": -10, "bounding": [4090, -825, 120, 60]}, "outputNode": {"id": -20, "bounding": [5150, -825, 120, 60]}, "inputs": [{"id": "37011fb7-14b7-4e0e-b1a0-6a02e8da1fd7", "name": "images.image0", "type": "IMAGE", "linkIds": [34], "localized_name": "images.image0", "label": "image", "pos": [4190, -805]}], "outputs": [{"id": "e9182b3f-635c-4cd4-a152-4b4be17ae4b9", "name": "IMAGE0", "type": "IMAGE", "linkIds": [35], "localized_name": "IMAGE0", "label": "IMAGE", "pos": [5170, -805]}], "widgets": [], "nodes": [{"id": 24, "type": "PrimitiveFloat", "pos": [4280, -1240], "size": [270, 58], "flags": {}, "order": 0, "mode": 0, "inputs": [{"label": "strength", "localized_name": "value", "name": "value", "type": "FLOAT", "widget": {"name": "value"}, "link": null}], "outputs": [{"localized_name": "FLOAT", "name": "FLOAT", "type": "FLOAT", "links": [36]}], "properties": {"Node name for S&R": "PrimitiveFloat", "min": 0, "max": 3, "precision": 2, "step": 0.05}, "widgets_values": [0.5]}, {"id": 23, "type": "GLSLShader", "pos": [4570, -1240], "size": [370, 192], "flags": {}, "order": 1, "mode": 0, "inputs": [{"label": "image0", "localized_name": "images.image0", "name": "images.image0", "type": "IMAGE", "link": 34}, {"label": "image1", "localized_name": "images.image1", "name": "images.image1", "shape": 7, "type": "IMAGE", "link": null}, {"label": "u_float0", "localized_name": "floats.u_float0", "name": "floats.u_float0", "shape": 7, "type": "FLOAT", "link": 36}, {"label": "u_float1", "localized_name": "floats.u_float1", "name": "floats.u_float1", "shape": 7, "type": "FLOAT", "link": null}, {"label": "u_int0", "localized_name": "ints.u_int0", "name": "ints.u_int0", "shape": 7, "type": "INT", "link": null}, {"localized_name": "fragment_shader", "name": "fragment_shader", "type": "STRING", "widget": {"name": "fragment_shader"}, "link": null}, {"localized_name": "size_mode", "name": "size_mode", "type": "COMFY_DYNAMICCOMBO_V3", "widget": {"name": "size_mode"}, "link": null}], "outputs": [{"localized_name": "IMAGE0", "name": "IMAGE0", "type": "IMAGE", "links": [35]}, {"localized_name": "IMAGE1", "name": "IMAGE1", "type": "IMAGE", "links": null}, {"localized_name": "IMAGE2", "name": "IMAGE2", "type": "IMAGE", "links": null}, {"localized_name": "IMAGE3", "name": "IMAGE3", "type": "IMAGE", "links": null}], "properties": {"Node name for S&R": "GLSLShader"}, "widgets_values": ["#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\nuniform vec2 u_resolution;\nuniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nvoid main() {\n    vec2 texel = 1.0 / u_resolution;\n    \n    // Sample center and neighbors\n    vec4 center = texture(u_image0, v_texCoord);\n    vec4 top    = texture(u_image0, v_texCoord + vec2( 0.0, -texel.y));\n    vec4 bottom = texture(u_image0, v_texCoord + vec2( 0.0,  texel.y));\n    vec4 left   = texture(u_image0, v_texCoord + vec2(-texel.x,  0.0));\n    vec4 right  = texture(u_image0, v_texCoord + vec2( texel.x,  0.0));\n    \n    // Edge enhancement (Laplacian)\n    vec4 edges = center * 4.0 - top - bottom - left - right;\n    \n    // Add edges back scaled by strength\n    vec4 sharpened = center + edges * u_float0;\n    \n    fragColor0 = vec4(clamp(sharpened.rgb, 0.0, 1.0), center.a);\n}", "from_input"]}], "groups": [], "links": [{"id": 36, "origin_id": 24, "origin_slot": 0, "target_id": 23, "target_slot": 2, "type": "FLOAT"}, {"id": 34, "origin_id": -10, "origin_slot": 0, "target_id": 23, "target_slot": 0, "type": "IMAGE"}, {"id": 35, "origin_id": 23, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "IMAGE"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Image Tools/Sharpen"}]}}
+{
+  "revision": 0,
+  "last_node_id": 25,
+  "last_link_id": 0,
+  "nodes": [
+    {
+      "id": 25,
+      "type": "621ba4e2-22a8-482d-a369-023753198b7b",
+      "pos": [
+        4610,
+        -790
+      ],
+      "size": [
+        230,
+        58
+      ],
+      "flags": {},
+      "order": 4,
+      "mode": 0,
+      "inputs": [
+        {
+          "label": "image",
+          "localized_name": "images.image0",
+          "name": "images.image0",
+          "type": "IMAGE",
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "label": "IMAGE",
+          "localized_name": "IMAGE0",
+          "name": "IMAGE0",
+          "type": "IMAGE",
+          "links": []
+        }
+      ],
+      "title": "Sharpen",
+      "properties": {
+        "proxyWidgets": [
+          [
+            "24",
+            "value"
+          ]
+        ]
+      },
+      "widgets_values": []
+    }
+  ],
+  "links": [],
+  "version": 0.4,
+  "definitions": {
+    "subgraphs": [
+      {
+        "id": "621ba4e2-22a8-482d-a369-023753198b7b",
+        "version": 1,
+        "state": {
+          "lastGroupId": 0,
+          "lastNodeId": 24,
+          "lastLinkId": 36,
+          "lastRerouteId": 0
+        },
+        "revision": 0,
+        "config": {},
+        "name": "Sharpen",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            4090,
+            -825,
+            120,
+            60
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            5150,
+            -825,
+            120,
+            60
+          ]
+        },
+        "inputs": [
+          {
+            "id": "37011fb7-14b7-4e0e-b1a0-6a02e8da1fd7",
+            "name": "images.image0",
+            "type": "IMAGE",
+            "linkIds": [
+              34
+            ],
+            "localized_name": "images.image0",
+            "label": "image",
+            "pos": [
+              4190,
+              -805
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "e9182b3f-635c-4cd4-a152-4b4be17ae4b9",
+            "name": "IMAGE0",
+            "type": "IMAGE",
+            "linkIds": [
+              35
+            ],
+            "localized_name": "IMAGE0",
+            "label": "IMAGE",
+            "pos": [
+              5170,
+              -805
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 24,
+            "type": "PrimitiveFloat",
+            "pos": [
+              4280,
+              -1240
+            ],
+            "size": [
+              270,
+              58
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "label": "strength",
+                "localized_name": "value",
+                "name": "value",
+                "type": "FLOAT",
+                "widget": {
+                  "name": "value"
+                },
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "FLOAT",
+                "name": "FLOAT",
+                "type": "FLOAT",
+                "links": [
+                  36
+                ]
+              }
+            ],
+            "properties": {
+              "Node name for S&R": "PrimitiveFloat",
+              "min": 0,
+              "max": 3,
+              "precision": 2,
+              "step": 0.05
+            },
+            "widgets_values": [
+              0.5
+            ]
+          },
+          {
+            "id": 23,
+            "type": "GLSLShader",
+            "pos": [
+              4570,
+              -1240
+            ],
+            "size": [
+              370,
+              192
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+              {
+                "label": "image0",
+                "localized_name": "images.image0",
+                "name": "images.image0",
+                "type": "IMAGE",
+                "link": 34
+              },
+              {
+                "label": "image1",
+                "localized_name": "images.image1",
+                "name": "images.image1",
+                "shape": 7,
+                "type": "IMAGE",
+                "link": null
+              },
+              {
+                "label": "u_float0",
+                "localized_name": "floats.u_float0",
+                "name": "floats.u_float0",
+                "shape": 7,
+                "type": "FLOAT",
+                "link": 36
+              },
+              {
+                "label": "u_float1",
+                "localized_name": "floats.u_float1",
+                "name": "floats.u_float1",
+                "shape": 7,
+                "type": "FLOAT",
+                "link": null
+              },
+              {
+                "label": "u_int0",
+                "localized_name": "ints.u_int0",
+                "name": "ints.u_int0",
+                "shape": 7,
+                "type": "INT",
+                "link": null
+              },
+              {
+                "localized_name": "fragment_shader",
+                "name": "fragment_shader",
+                "type": "STRING",
+                "widget": {
+                  "name": "fragment_shader"
+                },
+                "link": null
+              },
+              {
+                "localized_name": "size_mode",
+                "name": "size_mode",
+                "type": "COMFY_DYNAMICCOMBO_V3",
+                "widget": {
+                  "name": "size_mode"
+                },
+                "link": null
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "IMAGE0",
+                "name": "IMAGE0",
+                "type": "IMAGE",
+                "links": [
+                  35
+                ]
+              },
+              {
+                "localized_name": "IMAGE1",
+                "name": "IMAGE1",
+                "type": "IMAGE",
+                "links": null
+              },
+              {
+                "localized_name": "IMAGE2",
+                "name": "IMAGE2",
+                "type": "IMAGE",
+                "links": null
+              },
+              {
+                "localized_name": "IMAGE3",
+                "name": "IMAGE3",
+                "type": "IMAGE",
+                "links": null
+              }
+            ],
+            "properties": {
+              "Node name for S&R": "GLSLShader"
+            },
+            "widgets_values": [
+              "#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\nuniform vec2 u_resolution;\nuniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0\n\nin vec2 v_texCoord;\nlayout(location = 0) out vec4 fragColor0;\n\nvoid main() {\n    vec2 texel = 1.0 / u_resolution;\n    \n    // Sample center and neighbors\n    vec4 center = texture(u_image0, v_texCoord);\n    vec4 top    = texture(u_image0, v_texCoord + vec2( 0.0, -texel.y));\n    vec4 bottom = texture(u_image0, v_texCoord + vec2( 0.0,  texel.y));\n    vec4 left   = texture(u_image0, v_texCoord + vec2(-texel.x,  0.0));\n    vec4 right  = texture(u_image0, v_texCoord + vec2( texel.x,  0.0));\n    \n    // Edge enhancement (Laplacian)\n    vec4 edges = center * 4.0 - top - bottom - left - right;\n    \n    // Add edges back scaled by strength\n    vec4 sharpened = center + edges * u_float0;\n    \n    fragColor0 = vec4(clamp(sharpened.rgb, 0.0, 1.0), center.a);\n}",
+              "from_input"
+            ]
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 36,
+            "origin_id": 24,
+            "origin_slot": 0,
+            "target_id": 23,
+            "target_slot": 2,
+            "type": "FLOAT"
+          },
+          {
+            "id": 34,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 23,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 35,
+            "origin_id": 23,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "IMAGE"
+          }
+        ],
+        "extra": {
+          "workflowRendererVersion": "LG"
+        },
+        "category": "Image Tools/Sharpen"
+      }
+    ]
+  }
+}

blueprints/Video Upscale(GAN x4).json

@@ -1 +1,420 @@
-{"revision": 0, "last_node_id": 13, "last_link_id": 0, "nodes": [{"id": 13, "type": "cf95b747-3e17-46cb-8097-cac60ff9b2e1", "pos": [1120, 330], "size": [240, 58], "flags": {}, "order": 3, "mode": 0, "inputs": [{"localized_name": "video", "name": "video", "type": "VIDEO", "link": null}, {"name": "model_name", "type": "COMBO", "widget": {"name": "model_name"}, "link": null}], "outputs": [{"localized_name": "VIDEO", "name": "VIDEO", "type": "VIDEO", "links": []}], "title": "Video Upscale(GAN x4)", "properties": {"proxyWidgets": [["-1", "model_name"]], "cnr_id": "comfy-core", "ver": "0.14.1"}, "widgets_values": ["RealESRGAN_x4plus.safetensors"]}], "links": [], "version": 0.4, "definitions": {"subgraphs": [{"id": "cf95b747-3e17-46cb-8097-cac60ff9b2e1", "version": 1, "state": {"lastGroupId": 0, "lastNodeId": 13, "lastLinkId": 19, "lastRerouteId": 0}, "revision": 0, "config": {}, "name": "Video Upscale(GAN x4)", "inputNode": {"id": -10, "bounding": [550, 460, 120, 80]}, "outputNode": {"id": -20, "bounding": [1490, 460, 120, 60]}, "inputs": [{"id": "666d633e-93e7-42dc-8d11-2b7b99b0f2a6", "name": "video", "type": "VIDEO", "linkIds": [10], "localized_name": "video", "pos": [650, 480]}, {"id": "2e23a087-caa8-4d65-99e6-662761aa905a", "name": "model_name", "type": "COMBO", "linkIds": [19], "pos": [650, 500]}], "outputs": [{"id": "0c1768ea-3ec2-412f-9af6-8e0fa36dae70", "name": "VIDEO", "type": "VIDEO", "linkIds": [15], "localized_name": "VIDEO", "pos": [1510, 480]}], "widgets": [], "nodes": [{"id": 2, "type": "ImageUpscaleWithModel", "pos": [1110, 450], "size": [320, 46], "flags": {}, "order": 1, "mode": 0, "inputs": [{"localized_name": "upscale_model", "name": "upscale_model", "type": "UPSCALE_MODEL", "link": 1}, {"localized_name": "image", "name": "image", "type": "IMAGE", "link": 14}], "outputs": [{"localized_name": "IMAGE", "name": "IMAGE", "type": "IMAGE", "links": [13]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "ImageUpscaleWithModel"}}, {"id": 11, "type": "CreateVideo", "pos": [1110, 550], "size": [320, 78], "flags": {}, "order": 3, "mode": 0, "inputs": [{"localized_name": "images", "name": "images", "type": "IMAGE", "link": 13}, {"localized_name": "audio", "name": "audio", "shape": 7, "type": "AUDIO", "link": 16}, {"localized_name": "fps", "name": "fps", "type": "FLOAT", "widget": {"name": "fps"}, "link": 12}], "outputs": [{"localized_name": "VIDEO", "name": "VIDEO", "type": "VIDEO", "links": [15]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "CreateVideo"}, "widgets_values": [30]}, {"id": 10, "type": "GetVideoComponents", "pos": [1110, 330], "size": [320, 70], "flags": {}, "order": 2, "mode": 0, "inputs": [{"localized_name": "video", "name": "video", "type": "VIDEO", "link": 10}], "outputs": [{"localized_name": "images", "name": "images", "type": "IMAGE", "links": [14]}, {"localized_name": "audio", "name": "audio", "type": "AUDIO", "links": [16]}, {"localized_name": "fps", "name": "fps", "type": "FLOAT", "links": [12]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "GetVideoComponents"}}, {"id": 1, "type": "UpscaleModelLoader", "pos": [750, 450], "size": [280, 60], "flags": {}, "order": 0, "mode": 0, "inputs": [{"localized_name": "model_name", "name": "model_name", "type": "COMBO", "widget": {"name": "model_name"}, "link": 19}], "outputs": [{"localized_name": "UPSCALE_MODEL", "name": "UPSCALE_MODEL", "type": "UPSCALE_MODEL", "links": [1]}], "properties": {"cnr_id": "comfy-core", "ver": "0.10.0", "Node name for S&R": "UpscaleModelLoader", "models": [{"name": "RealESRGAN_x4plus.safetensors", "url": "https://huggingface.co/Comfy-Org/Real-ESRGAN_repackaged/resolve/main/RealESRGAN_x4plus.safetensors", "directory": "upscale_models"}]}, "widgets_values": ["RealESRGAN_x4plus.safetensors"]}], "groups": [], "links": [{"id": 1, "origin_id": 1, "origin_slot": 0, "target_id": 2, "target_slot": 0, "type": "UPSCALE_MODEL"}, {"id": 14, "origin_id": 10, "origin_slot": 0, "target_id": 2, "target_slot": 1, "type": "IMAGE"}, {"id": 13, "origin_id": 2, "origin_slot": 0, "target_id": 11, "target_slot": 0, "type": "IMAGE"}, {"id": 16, "origin_id": 10, "origin_slot": 1, "target_id": 11, "target_slot": 1, "type": "AUDIO"}, {"id": 12, "origin_id": 10, "origin_slot": 2, "target_id": 11, "target_slot": 2, "type": "FLOAT"}, {"id": 10, "origin_id": -10, "origin_slot": 0, "target_id": 10, "target_slot": 0, "type": "VIDEO"}, {"id": 15, "origin_id": 11, "origin_slot": 0, "target_id": -20, "target_slot": 0, "type": "VIDEO"}, {"id": 19, "origin_id": -10, "origin_slot": 1, "target_id": 1, "target_slot": 0, "type": "COMBO"}], "extra": {"workflowRendererVersion": "LG"}, "category": "Video generation and editing/Enhance video"}]}, "extra": {}}
+{
+  "revision": 0,
+  "last_node_id": 13,
+  "last_link_id": 0,
+  "nodes": [
+    {
+      "id": 13,
+      "type": "cf95b747-3e17-46cb-8097-cac60ff9b2e1",
+      "pos": [
+        1120,
+        330
+      ],
+      "size": [
+        240,
+        58
+      ],
+      "flags": {},
+      "order": 3,
+      "mode": 0,
+      "inputs": [
+        {
+          "localized_name": "video",
+          "name": "video",
+          "type": "VIDEO",
+          "link": null
+        },
+        {
+          "name": "model_name",
+          "type": "COMBO",
+          "widget": {
+            "name": "model_name"
+          },
+          "link": null
+        }
+      ],
+      "outputs": [
+        {
+          "localized_name": "VIDEO",
+          "name": "VIDEO",
+          "type": "VIDEO",
+          "links": []
+        }
+      ],
+      "title": "Video Upscale(GAN x4)",
+      "properties": {
+        "proxyWidgets": [
+          [
+            "-1",
+            "model_name"
+          ]
+        ],
+        "cnr_id": "comfy-core",
+        "ver": "0.14.1"
+      },
+      "widgets_values": [
+        "RealESRGAN_x4plus.safetensors"
+      ]
+    }
+  ],
+  "links": [],
+  "version": 0.4,
+  "definitions": {
+    "subgraphs": [
+      {
+        "id": "cf95b747-3e17-46cb-8097-cac60ff9b2e1",
+        "version": 1,
+        "state": {
+          "lastGroupId": 0,
+          "lastNodeId": 13,
+          "lastLinkId": 19,
+          "lastRerouteId": 0
+        },
+        "revision": 0,
+        "config": {},
+        "name": "Video Upscale(GAN x4)",
+        "inputNode": {
+          "id": -10,
+          "bounding": [
+            550,
+            460,
+            120,
+            80
+          ]
+        },
+        "outputNode": {
+          "id": -20,
+          "bounding": [
+            1490,
+            460,
+            120,
+            60
+          ]
+        },
+        "inputs": [
+          {
+            "id": "666d633e-93e7-42dc-8d11-2b7b99b0f2a6",
+            "name": "video",
+            "type": "VIDEO",
+            "linkIds": [
+              10
+            ],
+            "localized_name": "video",
+            "pos": [
+              650,
+              480
+            ]
+          },
+          {
+            "id": "2e23a087-caa8-4d65-99e6-662761aa905a",
+            "name": "model_name",
+            "type": "COMBO",
+            "linkIds": [
+              19
+            ],
+            "pos": [
+              650,
+              500
+            ]
+          }
+        ],
+        "outputs": [
+          {
+            "id": "0c1768ea-3ec2-412f-9af6-8e0fa36dae70",
+            "name": "VIDEO",
+            "type": "VIDEO",
+            "linkIds": [
+              15
+            ],
+            "localized_name": "VIDEO",
+            "pos": [
+              1510,
+              480
+            ]
+          }
+        ],
+        "widgets": [],
+        "nodes": [
+          {
+            "id": 2,
+            "type": "ImageUpscaleWithModel",
+            "pos": [
+              1110,
+              450
+            ],
+            "size": [
+              320,
+              46
+            ],
+            "flags": {},
+            "order": 1,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "upscale_model",
+                "name": "upscale_model",
+                "type": "UPSCALE_MODEL",
+                "link": 1
+              },
+              {
+                "localized_name": "image",
+                "name": "image",
+                "type": "IMAGE",
+                "link": 14
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "IMAGE",
+                "name": "IMAGE",
+                "type": "IMAGE",
+                "links": [
+                  13
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.10.0",
+              "Node name for S&R": "ImageUpscaleWithModel"
+            }
+          },
+          {
+            "id": 11,
+            "type": "CreateVideo",
+            "pos": [
+              1110,
+              550
+            ],
+            "size": [
+              320,
+              78
+            ],
+            "flags": {},
+            "order": 3,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "link": 13
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "shape": 7,
+                "type": "AUDIO",
+                "link": 16
+              },
+              {
+                "localized_name": "fps",
+                "name": "fps",
+                "type": "FLOAT",
+                "widget": {
+                  "name": "fps"
+                },
+                "link": 12
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "VIDEO",
+                "name": "VIDEO",
+                "type": "VIDEO",
+                "links": [
+                  15
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.10.0",
+              "Node name for S&R": "CreateVideo"
+            },
+            "widgets_values": [
+              30
+            ]
+          },
+          {
+            "id": 10,
+            "type": "GetVideoComponents",
+            "pos": [
+              1110,
+              330
+            ],
+            "size": [
+              320,
+              70
+            ],
+            "flags": {},
+            "order": 2,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "video",
+                "name": "video",
+                "type": "VIDEO",
+                "link": 10
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "images",
+                "name": "images",
+                "type": "IMAGE",
+                "links": [
+                  14
+                ]
+              },
+              {
+                "localized_name": "audio",
+                "name": "audio",
+                "type": "AUDIO",
+                "links": [
+                  16
+                ]
+              },
+              {
+                "localized_name": "fps",
+                "name": "fps",
+                "type": "FLOAT",
+                "links": [
+                  12
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.10.0",
+              "Node name for S&R": "GetVideoComponents"
+            }
+          },
+          {
+            "id": 1,
+            "type": "UpscaleModelLoader",
+            "pos": [
+              750,
+              450
+            ],
+            "size": [
+              280,
+              60
+            ],
+            "flags": {},
+            "order": 0,
+            "mode": 0,
+            "inputs": [
+              {
+                "localized_name": "model_name",
+                "name": "model_name",
+                "type": "COMBO",
+                "widget": {
+                  "name": "model_name"
+                },
+                "link": 19
+              }
+            ],
+            "outputs": [
+              {
+                "localized_name": "UPSCALE_MODEL",
+                "name": "UPSCALE_MODEL",
+                "type": "UPSCALE_MODEL",
+                "links": [
+                  1
+                ]
+              }
+            ],
+            "properties": {
+              "cnr_id": "comfy-core",
+              "ver": "0.10.0",
+              "Node name for S&R": "UpscaleModelLoader",
+              "models": [
+                {
+                  "name": "RealESRGAN_x4plus.safetensors",
+                  "url": "https://huggingface.co/Comfy-Org/Real-ESRGAN_repackaged/resolve/main/RealESRGAN_x4plus.safetensors",
+                  "directory": "upscale_models"
+                }
+              ]
+            },
+            "widgets_values": [
+              "RealESRGAN_x4plus.safetensors"
+            ]
+          }
+        ],
+        "groups": [],
+        "links": [
+          {
+            "id": 1,
+            "origin_id": 1,
+            "origin_slot": 0,
+            "target_id": 2,
+            "target_slot": 0,
+            "type": "UPSCALE_MODEL"
+          },
+          {
+            "id": 14,
+            "origin_id": 10,
+            "origin_slot": 0,
+            "target_id": 2,
+            "target_slot": 1,
+            "type": "IMAGE"
+          },
+          {
+            "id": 13,
+            "origin_id": 2,
+            "origin_slot": 0,
+            "target_id": 11,
+            "target_slot": 0,
+            "type": "IMAGE"
+          },
+          {
+            "id": 16,
+            "origin_id": 10,
+            "origin_slot": 1,
+            "target_id": 11,
+            "target_slot": 1,
+            "type": "AUDIO"
+          },
+          {
+            "id": 12,
+            "origin_id": 10,
+            "origin_slot": 2,
+            "target_id": 11,
+            "target_slot": 2,
+            "type": "FLOAT"
+          },
+          {
+            "id": 10,
+            "origin_id": -10,
+            "origin_slot": 0,
+            "target_id": 10,
+            "target_slot": 0,
+            "type": "VIDEO"
+          },
+          {
+            "id": 15,
+            "origin_id": 11,
+            "origin_slot": 0,
+            "target_id": -20,
+            "target_slot": 0,
+            "type": "VIDEO"
+          },
+          {
+            "id": 19,
+            "origin_id": -10,
+            "origin_slot": 1,
+            "target_id": 1,
+            "target_slot": 0,
+            "type": "COMBO"
+          }
+        ],
+        "extra": {
+          "workflowRendererVersion": "LG"
+        },
+        "category": "Video generation and editing/Enhance video"
+      }
+    ]
+  },
+  "extra": {}
+}

comfy/deploy_environment.py

@@ -0,0 +1,33 @@
+import logging
+import os
+
+import folder_paths
+
+logger = logging.getLogger(__name__)
+
+_DEFAULT_DEPLOY_ENV = "local_git"
+_ENV_FILENAME = ".comfy_environment"
+
+_cached_value: str | None = None
+
+
+def get_deploy_environment() -> str:
+    global _cached_value
+    if _cached_value is not None:
+        return _cached_value
+
+    env_file = os.path.join(folder_paths.base_path, _ENV_FILENAME)
+    try:
+        with open(env_file, encoding="utf-8") as f:
+            first_line = f.readline().strip()
+            value = "".join(c for c in first_line if 32 <= ord(c) < 127)
+            if value:
+                _cached_value = value
+                return _cached_value
+    except FileNotFoundError:
+        pass
+    except Exception as e:
+        logger.warning("Failed to read %s: %s", env_file, e)
+
+    _cached_value = _DEFAULT_DEPLOY_ENV
+    return _cached_value

comfy/ldm/ace/ace_step15.py

@@ -611,6 +611,7 @@ class AceStepDiTModel(nn.Module):
         intermediate_size,
         patch_size,
         audio_acoustic_hidden_dim,
+        condition_dim=None,
         layer_types=None,
         sliding_window=128,
         rms_norm_eps=1e-6,
@@ -640,7 +641,7 @@ class AceStepDiTModel(nn.Module):
 
         self.time_embed = TimestepEmbedding(256, hidden_size, dtype=dtype, device=device, operations=operations)
         self.time_embed_r = TimestepEmbedding(256, hidden_size, dtype=dtype, device=device, operations=operations)
-        self.condition_embedder = Linear(hidden_size, hidden_size, dtype=dtype, device=device)
+        self.condition_embedder = Linear(condition_dim, hidden_size, dtype=dtype, device=device)
 
         if layer_types is None:
             layer_types = ["full_attention"] * num_layers
@@ -1035,6 +1036,9 @@ class AceStepConditionGenerationModel(nn.Module):
         fsq_dim=2048,
         fsq_levels=[8, 8, 8, 5, 5, 5],
         fsq_input_num_quantizers=1,
+        encoder_hidden_size=2048,
+        encoder_intermediate_size=6144,
+        encoder_num_heads=16,
         audio_model=None,
         dtype=None,
         device=None,
@@ -1054,24 +1058,24 @@ class AceStepConditionGenerationModel(nn.Module):
 
         self.decoder = AceStepDiTModel(
             in_channels, hidden_size, num_dit_layers, num_heads, num_kv_heads, head_dim,
-            intermediate_size, patch_size, audio_acoustic_hidden_dim,
+            intermediate_size, patch_size, audio_acoustic_hidden_dim, condition_dim=encoder_hidden_size,
             layer_types=layer_types, sliding_window=sliding_window, rms_norm_eps=rms_norm_eps,
             dtype=dtype, device=device, operations=operations
         )
         self.encoder = AceStepConditionEncoder(
-            text_hidden_dim, timbre_hidden_dim, hidden_size, num_lyric_layers, num_timbre_layers,
-            num_heads, num_kv_heads, head_dim, intermediate_size, rms_norm_eps,
+            text_hidden_dim, timbre_hidden_dim, encoder_hidden_size, num_lyric_layers, num_timbre_layers,
+            encoder_num_heads, num_kv_heads, head_dim, encoder_intermediate_size, rms_norm_eps,
             dtype=dtype, device=device, operations=operations
         )
         self.tokenizer = AceStepAudioTokenizer(
-            audio_acoustic_hidden_dim, hidden_size, pool_window_size, fsq_dim=fsq_dim, fsq_levels=fsq_levels, fsq_input_num_quantizers=fsq_input_num_quantizers, num_layers=num_tokenizer_layers, head_dim=head_dim, rms_norm_eps=rms_norm_eps,
+            audio_acoustic_hidden_dim, encoder_hidden_size, pool_window_size, fsq_dim=fsq_dim, fsq_levels=fsq_levels, fsq_input_num_quantizers=fsq_input_num_quantizers, num_layers=num_tokenizer_layers, head_dim=head_dim, rms_norm_eps=rms_norm_eps,
             dtype=dtype, device=device, operations=operations
         )
         self.detokenizer = AudioTokenDetokenizer(
-            hidden_size, pool_window_size, audio_acoustic_hidden_dim, num_layers=2, head_dim=head_dim,
+            encoder_hidden_size, pool_window_size, audio_acoustic_hidden_dim, num_layers=2, head_dim=head_dim,
             dtype=dtype, device=device, operations=operations
         )
-        self.null_condition_emb = nn.Parameter(torch.empty(1, 1, hidden_size, dtype=dtype, device=device))
+        self.null_condition_emb = nn.Parameter(torch.empty(1, 1, encoder_hidden_size, dtype=dtype, device=device))
 
     def prepare_condition(
         self,

comfy/ldm/ernie/model.py

@@ -0,0 +1,303 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention
+import comfy.model_management
+
+def rope(pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
+    assert dim % 2 == 0
+    if not comfy.model_management.supports_fp64(pos.device):
+        device = torch.device("cpu")
+    else:
+        device = pos.device
+
+    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=device) / dim
+    omega = 1.0 / (theta**scale)
+    out = torch.einsum("...n,d->...nd", pos.to(device), omega)
+    out = torch.stack([torch.cos(out), torch.sin(out)], dim=0)
+    return out.to(dtype=torch.float32, device=pos.device)
+
+def apply_rotary_emb(x_in: torch.Tensor, freqs_cis: torch.Tensor) -> torch.Tensor:
+    rot_dim = freqs_cis.shape[-1]
+    x, x_pass = x_in[..., :rot_dim], x_in[..., rot_dim:]
+    cos_ = freqs_cis[0]
+    sin_ = freqs_cis[1]
+    x1, x2 = x.chunk(2, dim=-1)
+    x_rotated = torch.cat((-x2, x1), dim=-1)
+    return torch.cat((x * cos_ + x_rotated * sin_, x_pass), dim=-1)
+
+class ErnieImageEmbedND3(nn.Module):
+    def __init__(self, dim: int, theta: int, axes_dim: tuple):
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        self.axes_dim = list(axes_dim)
+
+    def forward(self, ids: torch.Tensor) -> torch.Tensor:
+        emb = torch.cat([rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(3)], dim=-1)
+        emb = emb.unsqueeze(3)  # [2, B, S, 1, head_dim//2]
+        return torch.stack([emb, emb], dim=-1).reshape(*emb.shape[:-1], -1)  # [B, S, 1, head_dim]
+
+class ErnieImagePatchEmbedDynamic(nn.Module):
+    def __init__(self, in_channels: int, embed_dim: int, patch_size: int, operations, device=None, dtype=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.proj = operations.Conv2d(in_channels, embed_dim, kernel_size=patch_size, stride=patch_size, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        batch_size, dim, height, width = x.shape
+        return x.reshape(batch_size, dim, height * width).transpose(1, 2).contiguous()
+
+class Timesteps(nn.Module):
+    def __init__(self, num_channels: int, flip_sin_to_cos: bool = False):
+        super().__init__()
+        self.num_channels = num_channels
+        self.flip_sin_to_cos = flip_sin_to_cos
+
+    def forward(self, timesteps: torch.Tensor) -> torch.Tensor:
+        half_dim = self.num_channels // 2
+        exponent = -math.log(10000) * torch.arange(half_dim, dtype=torch.float32, device=timesteps.device) / half_dim
+        emb = torch.exp(exponent)
+        emb = timesteps[:, None].float() * emb[None, :]
+        if self.flip_sin_to_cos:
+            emb = torch.cat([torch.cos(emb), torch.sin(emb)], dim=-1)
+        else:
+            emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
+        return emb
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_channels: int, time_embed_dim: int, operations, device=None, dtype=None):
+        super().__init__()
+        Linear = operations.Linear
+        self.linear_1 = Linear(in_channels, time_embed_dim, bias=True, device=device, dtype=dtype)
+        self.act = nn.SiLU()
+        self.linear_2 = Linear(time_embed_dim, time_embed_dim, bias=True, device=device, dtype=dtype)
+
+    def forward(self, sample: torch.Tensor) -> torch.Tensor:
+        sample = self.linear_1(sample)
+        sample = self.act(sample)
+        sample = self.linear_2(sample)
+        return sample
+
+class ErnieImageAttention(nn.Module):
+    def __init__(self, query_dim: int, heads: int, dim_head: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        self.heads = heads
+        self.head_dim = dim_head
+        self.inner_dim = heads * dim_head
+
+        Linear = operations.Linear
+        RMSNorm = operations.RMSNorm
+
+        self.to_q = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+        self.to_k = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+        self.to_v = Linear(query_dim, self.inner_dim, bias=False, device=device, dtype=dtype)
+
+        self.norm_q = RMSNorm(dim_head, eps=eps, elementwise_affine=True, device=device, dtype=dtype)
+        self.norm_k = RMSNorm(dim_head, eps=eps, elementwise_affine=True, device=device, dtype=dtype)
+
+        self.to_out = nn.ModuleList([Linear(self.inner_dim, query_dim, bias=False, device=device, dtype=dtype)])
+
+    def forward(self, x: torch.Tensor, attention_mask: torch.Tensor = None, image_rotary_emb: torch.Tensor = None) -> torch.Tensor:
+        B, S, _ = x.shape
+
+        q_flat = self.to_q(x)
+        k_flat = self.to_k(x)
+        v_flat = self.to_v(x)
+
+        query = q_flat.view(B, S, self.heads, self.head_dim)
+        key = k_flat.view(B, S, self.heads, self.head_dim)
+
+        query = self.norm_q(query)
+        key = self.norm_k(key)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        query, key = query.to(x.dtype), key.to(x.dtype)
+
+        q_flat = query.reshape(B, S, -1)
+        k_flat = key.reshape(B, S, -1)
+
+        hidden_states = optimized_attention(q_flat, k_flat, v_flat, self.heads, mask=attention_mask)
+
+        return self.to_out[0](hidden_states)
+
+class ErnieImageFeedForward(nn.Module):
+    def __init__(self, hidden_size: int, ffn_hidden_size: int, operations, device=None, dtype=None):
+        super().__init__()
+        Linear = operations.Linear
+        self.gate_proj = Linear(hidden_size, ffn_hidden_size, bias=False, device=device, dtype=dtype)
+        self.up_proj = Linear(hidden_size, ffn_hidden_size, bias=False, device=device, dtype=dtype)
+        self.linear_fc2 = Linear(ffn_hidden_size, hidden_size, bias=False, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.linear_fc2(self.up_proj(x) * F.gelu(self.gate_proj(x)))
+
+class ErnieImageSharedAdaLNBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, ffn_hidden_size: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        RMSNorm = operations.RMSNorm
+
+        self.adaLN_sa_ln = RMSNorm(hidden_size, eps=eps, device=device, dtype=dtype)
+        self.self_attention = ErnieImageAttention(
+            query_dim=hidden_size,
+            dim_head=hidden_size // num_heads,
+            heads=num_heads,
+            eps=eps,
+            operations=operations,
+            device=device,
+            dtype=dtype
+        )
+        self.adaLN_mlp_ln = RMSNorm(hidden_size, eps=eps, device=device, dtype=dtype)
+        self.mlp = ErnieImageFeedForward(hidden_size, ffn_hidden_size, operations=operations, device=device, dtype=dtype)
+
+    def forward(self, x, rotary_pos_emb, temb, attention_mask=None):
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = temb
+
+        residual = x
+        x_norm = self.adaLN_sa_ln(x)
+        x_norm = (x_norm.float() * (1 + scale_msa.float()) + shift_msa.float()).to(x.dtype)
+
+        attn_out = self.self_attention(x_norm, attention_mask=attention_mask, image_rotary_emb=rotary_pos_emb)
+        x = residual + (gate_msa.float() * attn_out.float()).to(x.dtype)
+
+        residual = x
+        x_norm = self.adaLN_mlp_ln(x)
+        x_norm = (x_norm.float() * (1 + scale_mlp.float()) + shift_mlp.float()).to(x.dtype)
+
+        return residual + (gate_mlp.float() * self.mlp(x_norm).float()).to(x.dtype)
+
+class ErnieImageAdaLNContinuous(nn.Module):
+    def __init__(self, hidden_size: int, eps: float = 1e-6, operations=None, device=None, dtype=None):
+        super().__init__()
+        LayerNorm = operations.LayerNorm
+        Linear = operations.Linear
+        self.norm = LayerNorm(hidden_size, elementwise_affine=False, eps=eps, device=device, dtype=dtype)
+        self.linear = Linear(hidden_size, hidden_size * 2, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, conditioning: torch.Tensor) -> torch.Tensor:
+        scale, shift = self.linear(conditioning).chunk(2, dim=-1)
+        x = self.norm(x)
+        x = x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+        return x
+
+class ErnieImageModel(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int = 4096,
+        num_attention_heads: int = 32,
+        num_layers: int = 36,
+        ffn_hidden_size: int = 12288,
+        in_channels: int = 128,
+        out_channels: int = 128,
+        patch_size: int = 1,
+        text_in_dim: int = 3072,
+        rope_theta: int = 256,
+        rope_axes_dim: tuple = (32, 48, 48),
+        eps: float = 1e-6,
+        qk_layernorm: bool = True,
+        device=None,
+        dtype=None,
+        operations=None,
+        **kwargs
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.hidden_size = hidden_size
+        self.num_heads = num_attention_heads
+        self.head_dim = hidden_size // num_attention_heads
+        self.patch_size = patch_size
+        self.out_channels = out_channels
+
+        Linear = operations.Linear
+
+        self.x_embedder = ErnieImagePatchEmbedDynamic(in_channels, hidden_size, patch_size, operations, device, dtype)
+        self.text_proj = Linear(text_in_dim, hidden_size, bias=False, device=device, dtype=dtype) if text_in_dim != hidden_size else None
+
+        self.time_proj = Timesteps(hidden_size, flip_sin_to_cos=False)
+        self.time_embedding = TimestepEmbedding(hidden_size, hidden_size, operations, device, dtype)
+
+        self.pos_embed = ErnieImageEmbedND3(dim=self.head_dim, theta=rope_theta, axes_dim=rope_axes_dim)
+
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            Linear(hidden_size, 6 * hidden_size, device=device, dtype=dtype)
+        )
+
+        self.layers = nn.ModuleList([
+            ErnieImageSharedAdaLNBlock(hidden_size, num_attention_heads, ffn_hidden_size, eps, operations, device, dtype)
+            for _ in range(num_layers)
+        ])
+
+        self.final_norm = ErnieImageAdaLNContinuous(hidden_size, eps, operations, device, dtype)
+        self.final_linear = Linear(hidden_size, patch_size * patch_size * out_channels, device=device, dtype=dtype)
+
+    def forward(self, x, timesteps, context, **kwargs):
+        device, dtype = x.device, x.dtype
+        B, C, H, W = x.shape
+        p, Hp, Wp = self.patch_size, H // self.patch_size, W // self.patch_size
+        N_img = Hp * Wp
+
+        img_bsh = self.x_embedder(x)
+
+        text_bth = context
+        if self.text_proj is not None and text_bth.numel() > 0:
+            text_bth = self.text_proj(text_bth)
+        Tmax = text_bth.shape[1]
+
+        hidden_states = torch.cat([img_bsh, text_bth], dim=1)
+
+        text_ids = torch.zeros((B, Tmax, 3), device=device, dtype=torch.float32)
+        text_ids[:, :, 0] = torch.linspace(0, Tmax - 1, steps=Tmax, device=x.device, dtype=torch.float32)
+        index = float(Tmax)
+
+        transformer_options = kwargs.get("transformer_options", {})
+        rope_options = transformer_options.get("rope_options", None)
+
+        h_len, w_len = float(Hp), float(Wp)
+        h_offset, w_offset = 0.0, 0.0
+
+        if rope_options is not None:
+            h_len = (h_len - 1.0) * rope_options.get("scale_y", 1.0) + 1.0
+            w_len = (w_len - 1.0) * rope_options.get("scale_x", 1.0) + 1.0
+            index += rope_options.get("shift_t", 0.0)
+            h_offset += rope_options.get("shift_y", 0.0)
+            w_offset += rope_options.get("shift_x", 0.0)
+
+        image_ids = torch.zeros((Hp, Wp, 3), device=device, dtype=torch.float32)
+        image_ids[:, :, 0] = image_ids[:, :, 1] + index
+        image_ids[:, :, 1] = image_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=Hp, device=device, dtype=torch.float32).unsqueeze(1)
+        image_ids[:, :, 2] = image_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=Wp, device=device, dtype=torch.float32).unsqueeze(0)
+
+        image_ids = image_ids.view(1, N_img, 3).expand(B, -1, -1)
+
+        rotary_pos_emb = self.pos_embed(torch.cat([image_ids, text_ids], dim=1)).to(x.dtype)
+        del image_ids, text_ids
+
+        sample = self.time_proj(timesteps).to(dtype)
+        c = self.time_embedding(sample)
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = [
+            t.unsqueeze(1).contiguous() for t in self.adaLN_modulation(c).chunk(6, dim=-1)
+        ]
+
+        temb = [shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp]
+        for layer in self.layers:
+            hidden_states = layer(hidden_states, rotary_pos_emb, temb)
+
+        hidden_states = self.final_norm(hidden_states, c).type_as(hidden_states)
+
+        patches = self.final_linear(hidden_states)[:, :N_img, :]
+        output = (
+            patches.view(B, Hp, Wp, p, p, self.out_channels)
+            .permute(0, 5, 1, 3, 2, 4)
+            .contiguous()
+            .view(B, self.out_channels, H, W)
+        )
+
+        return output

comfy/ldm/flux/math.py

@@ -16,7 +16,7 @@ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor, mask=None, transforme
 
 def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     assert dim % 2 == 0
-    if comfy.model_management.is_device_mps(pos.device) or comfy.model_management.is_intel_xpu() or comfy.model_management.is_directml_enabled():
+    if not comfy.model_management.supports_fp64(pos.device):
         device = torch.device("cpu")
     else:
         device = pos.device

comfy/ldm/models/autoencoder.py

@@ -155,6 +155,7 @@ class AutoencodingEngineLegacy(AutoencodingEngine):
     def __init__(self, embed_dim: int, **kwargs):
         self.max_batch_size = kwargs.pop("max_batch_size", None)
         ddconfig = kwargs.pop("ddconfig")
+        decoder_ddconfig = kwargs.pop("decoder_ddconfig", ddconfig)
         super().__init__(
             encoder_config={
                 "target": "comfy.ldm.modules.diffusionmodules.model.Encoder",
@@ -162,7 +163,7 @@ class AutoencodingEngineLegacy(AutoencodingEngine):
             },
             decoder_config={
                 "target": "comfy.ldm.modules.diffusionmodules.model.Decoder",
-                "params": ddconfig,
+                "params": decoder_ddconfig,
             },
             **kwargs,
         )

comfy/ldm/modules/encoders/noise_aug_modules.py

@@ -3,12 +3,9 @@ from ..diffusionmodules.openaimodel import Timestep
 import torch
 
 class CLIPEmbeddingNoiseAugmentation(ImageConcatWithNoiseAugmentation):
-    def __init__(self, *args, clip_stats_path=None, timestep_dim=256, **kwargs):
+    def __init__(self, *args, timestep_dim=256, **kwargs):
         super().__init__(*args, **kwargs)
-        if clip_stats_path is None:
-            clip_mean, clip_std = torch.zeros(timestep_dim), torch.ones(timestep_dim)
-        else:
-            clip_mean, clip_std = torch.load(clip_stats_path, map_location="cpu")
+        clip_mean, clip_std = torch.zeros(timestep_dim), torch.ones(timestep_dim)
         self.register_buffer("data_mean", clip_mean[None, :], persistent=False)
         self.register_buffer("data_std", clip_std[None, :], persistent=False)
         self.time_embed = Timestep(timestep_dim)

comfy/ldm/modules/sdpose.py

@@ -90,7 +90,7 @@ class HeatmapHead(torch.nn.Module):
                 origin_max = np.max(hm[k])
                 dr = np.zeros((H + 2 * border, W + 2 * border), dtype=np.float32)
                 dr[border:-border, border:-border] = hm[k].copy()
-                dr = gaussian_filter(dr, sigma=2.0)
+                dr = gaussian_filter(dr, sigma=2.0, truncate=2.5)
                 hm[k] = dr[border:-border, border:-border].copy()
                 cur_max = np.max(hm[k])
                 if cur_max > 0:

comfy/ldm/rt_detr/rtdetr_v4.py

@@ -0,0 +1,725 @@
+from collections import OrderedDict
+from typing import List
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision
+import comfy.model_management
+from comfy.ldm.modules.attention import optimized_attention_for_device
+
+COCO_CLASSES = [
+    'person','bicycle','car','motorcycle','airplane','bus','train','truck','boat',
+    'traffic light','fire hydrant','stop sign','parking meter','bench','bird','cat',
+    'dog','horse','sheep','cow','elephant','bear','zebra','giraffe','backpack',
+    'umbrella','handbag','tie','suitcase','frisbee','skis','snowboard','sports ball',
+    'kite','baseball bat','baseball glove','skateboard','surfboard','tennis racket',
+    'bottle','wine glass','cup','fork','knife','spoon','bowl','banana','apple',
+    'sandwich','orange','broccoli','carrot','hot dog','pizza','donut','cake','chair',
+    'couch','potted plant','bed','dining table','toilet','tv','laptop','mouse',
+    'remote','keyboard','cell phone','microwave','oven','toaster','sink',
+    'refrigerator','book','clock','vase','scissors','teddy bear','hair drier','toothbrush',
+]
+
+# ---------------------------------------------------------------------------
+# HGNetv2 backbone
+# ---------------------------------------------------------------------------
+
+class ConvBNAct(nn.Module):
+    """Conv→BN→ReLU.  padding='same' adds asymmetric zero-pad (stem)."""
+    def __init__(self, ic, oc, k=3, s=1, groups=1, use_act=True, device=None, dtype=None, operations=None):
+        super().__init__()
+
+        self.conv = operations.Conv2d(ic, oc, k, s, (k - 1) // 2, groups=groups, bias=False, device=device, dtype=dtype)
+        self.bn   = nn.BatchNorm2d(oc, device=device, dtype=dtype)
+        self.act  = nn.ReLU() if use_act else nn.Identity()
+
+    def forward(self, x):
+        return self.act(self.bn(self.conv(x)))
+
+class LightConvBNAct(nn.Module):
+    def __init__(self, ic, oc, k, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.conv1 = ConvBNAct(ic, oc, 1, use_act=False, device=device, dtype=dtype, operations=operations)
+        self.conv2 = ConvBNAct(oc, oc, k, groups=oc, use_act=True, device=device, dtype=dtype, operations=operations)
+
+    def forward(self, x):
+        return self.conv2(self.conv1(x))
+
+class _StemBlock(nn.Module):
+    def __init__(self, ic, mc, oc, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.stem1  = ConvBNAct(ic,    mc,    3, 2, device=device, dtype=dtype, operations=operations)
+        # stem2a/stem2b use kernel=2, stride=1, no internal padding;
+        # padding is applied manually in forward (matching PaddlePaddle original)
+        self.stem2a = ConvBNAct(mc,    mc//2, 2, 1, device=device, dtype=dtype, operations=operations)
+        self.stem2b = ConvBNAct(mc//2, mc,    2, 1, device=device, dtype=dtype, operations=operations)
+        self.stem3  = ConvBNAct(mc*2,  mc,    3, 2, device=device, dtype=dtype, operations=operations)
+        self.stem4  = ConvBNAct(mc,    oc,    1, device=device, dtype=dtype, operations=operations)
+        self.pool   = nn.MaxPool2d(2, 1, ceil_mode=True)
+
+    def forward(self, x):
+        x  = self.stem1(x)
+        x  = F.pad(x, (0, 1, 0, 1))   # pad before pool and stem2a
+        x2 = self.stem2a(x)
+        x2 = F.pad(x2, (0, 1, 0, 1))  # pad before stem2b
+        x2 = self.stem2b(x2)
+        x1 = self.pool(x)
+        return self.stem4(self.stem3(torch.cat([x1, x2], 1)))
+
+
+class _HG_Block(nn.Module):
+    def __init__(self, ic, mc, oc, layer_num, k=3, residual=False, light=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.residual = residual
+        if light:
+            self.layers = nn.ModuleList(
+                [LightConvBNAct(ic if i == 0 else mc, mc, k, device=device, dtype=dtype, operations=operations) for i in range(layer_num)])
+        else:
+            self.layers = nn.ModuleList(
+                [ConvBNAct(ic if i == 0 else mc, mc, k, device=device, dtype=dtype, operations=operations) for i in range(layer_num)])
+        total = ic + layer_num * mc
+
+        self.aggregation = nn.Sequential(
+            ConvBNAct(total,   oc // 2, 1, device=device, dtype=dtype, operations=operations),
+            ConvBNAct(oc // 2, oc,      1, device=device, dtype=dtype, operations=operations))
+
+    def forward(self, x):
+        identity = x
+        outs = [x]
+        for layer in self.layers:
+            x = layer(x)
+            outs.append(x)
+        x = self.aggregation(torch.cat(outs, 1))
+        return x + identity if self.residual else x
+
+
+class _HG_Stage(nn.Module):
+    # config order: ic, mc, oc, num_blocks, downsample, light, k, layer_num
+    def __init__(self, ic, mc, oc, num_blocks, downsample=True, light=False, k=3, layer_num=6, device=None, dtype=None, operations=None):
+        super().__init__()
+        if downsample:
+            self.downsample = ConvBNAct(ic, ic, 3, 2, groups=ic, use_act=False, device=device, dtype=dtype, operations=operations)
+        else:
+            self.downsample = nn.Identity()
+        self.blocks = nn.Sequential(*[
+            _HG_Block(ic if i == 0 else oc, mc, oc, layer_num,
+                      k=k, residual=(i != 0), light=light, device=device, dtype=dtype, operations=operations)
+            for i in range(num_blocks)
+        ])
+
+    def forward(self, x):
+        return self.blocks(self.downsample(x))
+
+
+class HGNetv2(nn.Module):
+    # B5 config: stem=[3,32,64], stages=[ic, mc, oc, blocks, down, light, k, layers]
+    _STAGE_CFGS = [[64,  64,  128,  1, False, False, 3, 6],
+                   [128, 128, 512,  2, True,  False, 3, 6],
+                   [512, 256, 1024, 5, True,  True,  5, 6],
+                   [1024,512, 2048, 2, True,  True,  5, 6]]
+
+    def __init__(self, return_idx=(1, 2, 3), device=None, dtype=None, operations=None):
+        super().__init__()
+        self.stem   = _StemBlock(3, 32, 64, device=device, dtype=dtype, operations=operations)
+        self.stages = nn.ModuleList([_HG_Stage(*cfg, device=device, dtype=dtype, operations=operations) for cfg in self._STAGE_CFGS])
+        self.return_idx  = list(return_idx)
+        self.out_channels = [self._STAGE_CFGS[i][2] for i in return_idx]
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        x = self.stem(x)
+        outs = []
+        for i, stage in enumerate(self.stages):
+            x = stage(x)
+            if i in self.return_idx:
+                outs.append(x)
+        return outs
+
+
+# ---------------------------------------------------------------------------
+# Encoder — HybridEncoder  (dfine version: RepNCSPELAN4 + SCDown PAN)
+# ---------------------------------------------------------------------------
+
+class ConvNormLayer(nn.Module):
+    """Conv→act (expects pre-fused BN weights)."""
+    def __init__(self, ic, oc, k, s, g=1, padding=None, act=None, device=None, dtype=None, operations=None):
+        super().__init__()
+        p = (k - 1) // 2 if padding is None else padding
+        self.conv = operations.Conv2d(ic, oc, k, s, p, groups=g, bias=True, device=device, dtype=dtype)
+        self.act  = nn.SiLU() if act == 'silu' else nn.Identity()
+
+    def forward(self, x):
+        return self.act(self.conv(x))
+
+
+class VGGBlock(nn.Module):
+    """Rep-VGG block (expects pre-fused weights)."""
+    def __init__(self, ic, oc, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.conv = operations.Conv2d(ic, oc, 3, 1, padding=1, bias=True, device=device, dtype=dtype)
+        self.act  = nn.SiLU()
+
+    def forward(self, x):
+        return self.act(self.conv(x))
+
+
+class CSPLayer(nn.Module):
+    def __init__(self, ic, oc, num_blocks=3, expansion=1.0, act='silu', device=None, dtype=None, operations=None):
+        super().__init__()
+        h = int(oc * expansion)
+        self.conv1 = ConvNormLayer(ic, h, 1, 1, act=act, device=device, dtype=dtype, operations=operations)
+        self.conv2 = ConvNormLayer(ic, h, 1, 1, act=act, device=device, dtype=dtype, operations=operations)
+        self.bottlenecks = nn.Sequential(*[VGGBlock(h, h, device=device, dtype=dtype, operations=operations) for _ in range(num_blocks)])
+        self.conv3 = ConvNormLayer(h, oc, 1, 1, act=act, device=device, dtype=dtype, operations=operations) if h != oc else nn.Identity()
+
+    def forward(self, x):
+        return self.conv3(self.bottlenecks(self.conv1(x)) + self.conv2(x))
+
+
+class RepNCSPELAN4(nn.Module):
+    """CSP-ELAN block — the FPN/PAN block in RTv4's HybridEncoder."""
+    def __init__(self, c1, c2, c3, c4, n=3, act='silu', device=None, dtype=None, operations=None):
+        super().__init__()
+        self.c = c3 // 2
+        self.cv1 = ConvNormLayer(c1, c3, 1, 1, act=act, device=device, dtype=dtype, operations=operations)
+        self.cv2 = nn.Sequential(CSPLayer(c3 // 2, c4, n, 1.0, act=act, device=device, dtype=dtype, operations=operations), ConvNormLayer(c4, c4, 3, 1, act=act, device=device, dtype=dtype, operations=operations))
+        self.cv3 = nn.Sequential(CSPLayer(c4, c4, n, 1.0, act=act, device=device, dtype=dtype, operations=operations), ConvNormLayer(c4, c4, 3, 1, act=act, device=device, dtype=dtype, operations=operations))
+        self.cv4 = ConvNormLayer(c3 + 2 * c4, c2, 1, 1, act=act, device=device, dtype=dtype, operations=operations)
+
+    def forward(self, x):
+        y = list(self.cv1(x).split((self.c, self.c), 1))
+        y.extend(m(y[-1]) for m in [self.cv2, self.cv3])
+        return self.cv4(torch.cat(y, 1))
+
+
+class SCDown(nn.Module):
+    """Separable conv downsampling used in HybridEncoder PAN bottom-up path."""
+    def __init__(self, ic, oc, k, s, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.cv1 = ConvNormLayer(ic, oc, 1, 1, device=device, dtype=dtype, operations=operations)
+        self.cv2 = ConvNormLayer(oc, oc, k, s, g=oc, device=device, dtype=dtype, operations=operations)
+
+    def forward(self, x):
+        return self.cv2(self.cv1(x))
+
+
+class SelfAttention(nn.Module):
+    def __init__(self, embed_dim, num_heads, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.head_dim  = embed_dim // num_heads
+        self.q_proj   = operations.Linear(embed_dim, embed_dim, device=device, dtype=dtype)
+        self.k_proj   = operations.Linear(embed_dim, embed_dim, device=device, dtype=dtype)
+        self.v_proj   = operations.Linear(embed_dim, embed_dim, device=device, dtype=dtype)
+        self.out_proj = operations.Linear(embed_dim, embed_dim, device=device, dtype=dtype)
+
+    def forward(self, query, key, value, attn_mask=None):
+        optimized_attention = optimized_attention_for_device(query.device, False, small_input=True)
+        q, k, v = self.q_proj(query), self.k_proj(key), self.v_proj(value)
+        out = optimized_attention(q, k, v, heads=self.num_heads, mask=attn_mask)
+        return self.out_proj(out)
+
+
+class _TransformerEncoderLayer(nn.Module):
+    """Single AIFI encoder layer (pre- or post-norm, GELU by default)."""
+    def __init__(self, d_model, nhead, dim_feedforward, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.self_attn  = SelfAttention(d_model, nhead, device=device, dtype=dtype, operations=operations)
+        self.linear1    = operations.Linear(d_model, dim_feedforward, device=device, dtype=dtype)
+        self.linear2    = operations.Linear(dim_feedforward, d_model, device=device, dtype=dtype)
+        self.norm1      = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.norm2      = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.activation = nn.GELU()
+
+    def forward(self, src, src_mask=None, pos_embed=None):
+        q = k = src if pos_embed is None else src + pos_embed
+        src2 = self.self_attn(q, k, value=src, attn_mask=src_mask)
+        src = self.norm1(src + src2)
+        src2 = self.linear2(self.activation(self.linear1(src)))
+        return self.norm2(src + src2)
+
+
+class _TransformerEncoder(nn.Module):
+    """Thin wrapper so state-dict keys are  encoder.0.layers.N.*"""
+    def __init__(self, num_layers, d_model, nhead, dim_feedforward, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.layers = nn.ModuleList([
+            _TransformerEncoderLayer(d_model, nhead, dim_feedforward, device=device, dtype=dtype, operations=operations)
+            for _ in range(num_layers)
+        ])
+
+    def forward(self, src, src_mask=None, pos_embed=None):
+        for layer in self.layers:
+            src = layer(src, src_mask=src_mask, pos_embed=pos_embed)
+        return src
+
+
+class HybridEncoder(nn.Module):
+    def __init__(self, in_channels=(512, 1024, 2048), feat_strides=(8, 16, 32), hidden_dim=256, nhead=8, dim_feedforward=2048, use_encoder_idx=(2,), num_encoder_layers=1,
+                 pe_temperature=10000, expansion=1.0, depth_mult=1.0, act='silu', eval_spatial_size=(640, 640), device=None, dtype=None, operations=None):
+        super().__init__()
+        self.in_channels       = list(in_channels)
+        self.feat_strides      = list(feat_strides)
+        self.hidden_dim        = hidden_dim
+        self.use_encoder_idx   = list(use_encoder_idx)
+        self.pe_temperature    = pe_temperature
+        self.eval_spatial_size = eval_spatial_size
+        self.out_channels      = [hidden_dim] * len(in_channels)
+        self.out_strides       = list(feat_strides)
+
+        # channel projection (expects pre-fused weights)
+        self.input_proj = nn.ModuleList([
+            nn.Sequential(OrderedDict([('conv', operations.Conv2d(ch, hidden_dim, 1, bias=True, device=device, dtype=dtype))]))
+            for ch in in_channels
+        ])
+
+        # AIFI transformer — use _TransformerEncoder so keys are  encoder.0.layers.N.*
+        self.encoder = nn.ModuleList([
+            _TransformerEncoder(num_encoder_layers, hidden_dim, nhead, dim_feedforward, device=device, dtype=dtype, operations=operations)
+            for _ in range(len(use_encoder_idx))
+        ])
+
+        nb  = round(3 * depth_mult)
+        exp = expansion
+
+        # top-down FPN  (dfine: lateral conv has no act)
+        self.lateral_convs = nn.ModuleList(
+            [ConvNormLayer(hidden_dim, hidden_dim, 1, 1, device=device, dtype=dtype, operations=operations)
+             for _ in range(len(in_channels) - 1)])
+        self.fpn_blocks = nn.ModuleList(
+            [RepNCSPELAN4(hidden_dim * 2, hidden_dim, hidden_dim * 2, round(exp * hidden_dim // 2), nb, act=act, device=device, dtype=dtype, operations=operations)
+             for _ in range(len(in_channels) - 1)])
+
+        # bottom-up PAN  (dfine: nn.Sequential(SCDown) — keeps checkpoint key  .0.cv1/.0.cv2)
+        self.downsample_convs = nn.ModuleList(
+            [nn.Sequential(SCDown(hidden_dim, hidden_dim, 3, 2, device=device, dtype=dtype, operations=operations))
+             for _ in range(len(in_channels) - 1)])
+        self.pan_blocks = nn.ModuleList(
+            [RepNCSPELAN4(hidden_dim * 2, hidden_dim, hidden_dim * 2, round(exp * hidden_dim // 2), nb, act=act, device=device, dtype=dtype, operations=operations)
+             for _ in range(len(in_channels) - 1)])
+
+        # cache positional embeddings for fixed spatial size
+        if eval_spatial_size:
+            for idx in self.use_encoder_idx:
+                stride = self.feat_strides[idx]
+                pe = self._build_pe(eval_spatial_size[1] // stride,
+                                    eval_spatial_size[0] // stride,
+                                    hidden_dim, pe_temperature)
+                setattr(self, f'pos_embed{idx}', pe)
+
+    @staticmethod
+    def _build_pe(w, h, dim=256, temp=10000.):
+        assert dim % 4 == 0
+        gw = torch.arange(w, dtype=torch.float32)
+        gh = torch.arange(h, dtype=torch.float32)
+        gw, gh = torch.meshgrid(gw, gh, indexing='ij')
+        pdim  = dim // 4
+        omega = 1. / (temp ** (torch.arange(pdim, dtype=torch.float32) / pdim))
+        ow = gw.flatten()[:, None] @ omega[None]
+        oh = gh.flatten()[:, None] @ omega[None]
+        return torch.cat([ow.sin(), ow.cos(), oh.sin(), oh.cos()], 1)[None]
+
+    def forward(self, feats: List[torch.Tensor]) -> List[torch.Tensor]:
+        proj = [self.input_proj[i](f) for i, f in enumerate(feats)]
+
+        for i, enc_idx in enumerate(self.use_encoder_idx):
+            h, w = proj[enc_idx].shape[2:]
+            src  = proj[enc_idx].flatten(2).permute(0, 2, 1)
+            pe = getattr(self, f'pos_embed{enc_idx}').to(device=src.device, dtype=src.dtype)
+            for layer in self.encoder[i].layers:
+                src = layer(src, pos_embed=pe)
+            proj[enc_idx] = src.permute(0, 2, 1).reshape(-1, self.hidden_dim, h, w).contiguous()
+
+        n = len(self.in_channels)
+        inner = [proj[-1]]
+        for k in range(n - 1, 0, -1):
+            j = n - 1 - k
+            top = self.lateral_convs[j](inner[0])
+            inner[0] = top
+            up = F.interpolate(top, scale_factor=2., mode='nearest')
+            inner.insert(0, self.fpn_blocks[j](torch.cat([up, proj[k - 1]], 1)))
+
+        outs = [inner[0]]
+        for k in range(n - 1):
+            outs.append(self.pan_blocks[k](
+                torch.cat([self.downsample_convs[k](outs[-1]), inner[k + 1]], 1)))
+        return outs
+
+
+# ---------------------------------------------------------------------------
+# Decoder — DFINETransformer
+# ---------------------------------------------------------------------------
+
+def _deformable_attn_v2(value: list, spatial_shapes, sampling_locations: torch.Tensor, attention_weights: torch.Tensor, num_points_list: List[int]) -> torch.Tensor:
+    """
+    value            : list of per-level tensors  [bs*n_head, c, h_l, w_l]
+    sampling_locations: [bs, Lq, n_head, sum(pts), 2]  in [0,1]
+    attention_weights : [bs, Lq, n_head, sum(pts)]
+    """
+    _, c = value[0].shape[:2]      # bs*n_head, c
+    _, Lq, n_head, _, _ = sampling_locations.shape
+    bs = sampling_locations.shape[0]
+    n_h = n_head
+
+    grids = (2 * sampling_locations - 1)          # [bs, Lq, n_head, sum_pts, 2]
+    grids = grids.permute(0, 2, 1, 3, 4).flatten(0, 1)  # [bs*n_head, Lq, sum_pts, 2]
+    grids_per_lvl = grids.split(num_points_list, dim=2)  # list of [bs*n_head, Lq, pts_l, 2]
+
+    sampled = []
+    for lvl, (h, w) in enumerate(spatial_shapes):
+        val_l = value[lvl].reshape(bs * n_h, c, h, w)
+        sv = F.grid_sample(val_l, grids_per_lvl[lvl], mode='bilinear', padding_mode='zeros', align_corners=False)
+        sampled.append(sv) # sv: [bs*n_head, c, Lq, pts_l]
+
+    attn = attention_weights.permute(0, 2, 1, 3)  # [bs, n_head, Lq, sum_pts]
+    attn = attn.flatten(0, 1).unsqueeze(1)         # [bs*n_head, 1, Lq, sum_pts]
+    out  = (torch.cat(sampled, -1) * attn).sum(-1) # [bs*n_head, c, Lq]
+    out  = out.reshape(bs, n_h * c, Lq)
+    return out.permute(0, 2, 1)                    # [bs, Lq, hidden]
+
+
+class MSDeformableAttention(nn.Module):
+    def __init__(self, embed_dim=256, num_heads=8, num_levels=3, num_points=4, offset_scale=0.5, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.embed_dim, self.num_heads = embed_dim, num_heads
+        self.head_dim  = embed_dim // num_heads
+        pts = num_points if isinstance(num_points, list) else [num_points] * num_levels
+        self.num_points_list = pts
+        self.offset_scale    = offset_scale
+        total = num_heads * sum(pts)
+        self.register_buffer('num_points_scale', torch.tensor([1. / n for n in pts for _ in range(n)], dtype=torch.float32))
+        self.sampling_offsets  = operations.Linear(embed_dim, total * 2, device=device, dtype=dtype)
+        self.attention_weights = operations.Linear(embed_dim, total, device=device, dtype=dtype)
+
+    def forward(self, query, ref_pts, value, spatial_shapes):
+        bs, Lq = query.shape[:2]
+        offsets = self.sampling_offsets(query).reshape(
+            bs, Lq, self.num_heads, sum(self.num_points_list), 2)
+        attn_w  = F.softmax(
+            self.attention_weights(query).reshape(
+                bs, Lq, self.num_heads, sum(self.num_points_list)), -1)
+        scale   = self.num_points_scale.to(query).unsqueeze(-1)
+        offset  = offsets * scale * ref_pts[:, :, None, :, 2:] * self.offset_scale
+        locs    = ref_pts[:, :, None, :, :2] + offset  # [bs, Lq, n_head, sum_pts, 2]
+        return _deformable_attn_v2(value, spatial_shapes, locs, attn_w, self.num_points_list)
+
+
+class Gate(nn.Module):
+    def __init__(self, d_model, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.gate = operations.Linear(2 * d_model, 2 * d_model, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+
+    def forward(self, x1, x2):
+        g1, g2 = torch.sigmoid(self.gate(torch.cat([x1, x2], -1))).chunk(2, -1)
+        return self.norm(g1 * x1 + g2 * x2)
+
+
+class MLP(nn.Module):
+    def __init__(self, in_dim, hidden_dim, out_dim, num_layers, device=None, dtype=None, operations=None):
+        super().__init__()
+        dims = [in_dim] + [hidden_dim] * (num_layers - 1) + [out_dim]
+        self.layers = nn.ModuleList(operations.Linear(dims[i], dims[i + 1], device=device, dtype=dtype) for i in range(num_layers))
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = nn.SiLU()(layer(x)) if i < len(self.layers) - 1 else layer(x)
+        return x
+
+
+class TransformerDecoderLayer(nn.Module):
+    def __init__(self, d_model=256, nhead=8, dim_feedforward=1024, num_levels=3, num_points=4, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.self_attn  = SelfAttention(d_model, nhead, device=device, dtype=dtype, operations=operations)
+        self.norm1      = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.cross_attn = MSDeformableAttention(d_model, nhead, num_levels, num_points, device=device, dtype=dtype, operations=operations)
+        self.gateway    = Gate(d_model, device=device, dtype=dtype, operations=operations)
+        self.linear1    = operations.Linear(d_model, dim_feedforward, device=device, dtype=dtype)
+        self.activation = nn.ReLU()
+        self.linear2    = operations.Linear(dim_feedforward, d_model, device=device, dtype=dtype)
+        self.norm3      = operations.LayerNorm(d_model, device=device, dtype=dtype)
+
+    def forward(self, target, ref_pts, value, spatial_shapes, attn_mask=None, query_pos=None):
+        q = k = target if query_pos is None else target + query_pos
+        t2 = self.self_attn(q, k, value=target, attn_mask=attn_mask)
+        target = self.norm1(target + t2)
+        t2 = self.cross_attn(
+            target if query_pos is None else target + query_pos,
+            ref_pts, value, spatial_shapes)
+        target = self.gateway(target, t2)
+        t2 = self.linear2(self.activation(self.linear1(target)))
+        target = self.norm3((target + t2).clamp(-65504, 65504))
+        return target
+
+
+# ---------------------------------------------------------------------------
+# FDR utilities
+# ---------------------------------------------------------------------------
+
+def weighting_function(reg_max, up, reg_scale):
+    """Non-uniform weighting function W(n) for FDR box regression."""
+    ub1 = (abs(up[0]) * abs(reg_scale)).item()
+    ub2 = ub1 * 2
+    step = (ub1 + 1) ** (2 / (reg_max - 2))
+    left  = [-(step ** i) + 1 for i in range(reg_max // 2 - 1, 0, -1)]
+    right = [ (step ** i) - 1 for i in range(1, reg_max // 2)]
+    vals  = [-ub2] + left + [0] + right + [ub2]
+    return torch.tensor(vals, dtype=up.dtype, device=up.device)
+
+
+def distance2bbox(points, distance, reg_scale):
+    """Decode edge-distances → cxcywh boxes."""
+    rs = abs(reg_scale).to(dtype=points.dtype)
+    x1 = points[..., 0] - (0.5 * rs + distance[..., 0]) * (points[..., 2] / rs)
+    y1 = points[..., 1] - (0.5 * rs + distance[..., 1]) * (points[..., 3] / rs)
+    x2 = points[..., 0] + (0.5 * rs + distance[..., 2]) * (points[..., 2] / rs)
+    y2 = points[..., 1] + (0.5 * rs + distance[..., 3]) * (points[..., 3] / rs)
+    x0, y0, x1_, y1_ = (x1 + x2) / 2, (y1 + y2) / 2, x2 - x1, y2 - y1
+    return torch.stack([x0, y0, x1_, y1_], -1)
+
+
+class Integral(nn.Module):
+    """Sum Pr(n)·W(n) over the distribution bins."""
+    def __init__(self, reg_max=32):
+        super().__init__()
+        self.reg_max = reg_max
+
+    def forward(self, x, project):
+        shape = x.shape
+        x = F.softmax(x.reshape(-1, self.reg_max + 1), 1)
+        x = F.linear(x, project.to(device=x.device, dtype=x.dtype)).reshape(-1, 4)
+        return x.reshape(list(shape[:-1]) + [-1])
+
+
+class LQE(nn.Module):
+    """Location Quality Estimator — refines class scores using corner distribution."""
+    def __init__(self, k=4, hidden_dim=64, num_layers=2, reg_max=32, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.k, self.reg_max = k, reg_max
+        self.reg_conf = MLP(4 * (k + 1), hidden_dim, 1, num_layers, device=device, dtype=dtype, operations=operations)
+
+    def forward(self, scores, pred_corners):
+        B, L, _ = pred_corners.shape
+        prob     = F.softmax(pred_corners.reshape(B, L, 4, self.reg_max + 1), -1)
+        topk, _  = prob.topk(self.k, -1)
+        stat     = torch.cat([topk, topk.mean(-1, keepdim=True)], -1)
+        return scores + self.reg_conf(stat.reshape(B, L, -1))
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(self, hidden_dim, nhead, dim_feedforward, num_levels, num_points, num_layers, reg_max, reg_scale, up, eval_idx=-1, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.hidden_dim = hidden_dim
+        self.num_layers = num_layers
+        self.nhead      = nhead
+        self.eval_idx   = eval_idx if eval_idx >= 0 else num_layers + eval_idx
+        self.up, self.reg_scale, self.reg_max = up, reg_scale, reg_max
+        self.layers = nn.ModuleList([
+            TransformerDecoderLayer(hidden_dim, nhead, dim_feedforward, num_levels, num_points, device=device, dtype=dtype, operations=operations)
+            for _ in range(self.eval_idx + 1)
+        ])
+        self.lqe_layers = nn.ModuleList([LQE(4, 64, 2, reg_max, device=device, dtype=dtype, operations=operations) for _ in range(self.eval_idx + 1)])
+        self.register_buffer('project', weighting_function(reg_max, up, reg_scale))
+
+    def _value_op(self, memory, spatial_shapes):
+        """Reshape memory to per-level value tensors for deformable attention."""
+        c = self.hidden_dim // self.nhead
+        split = [h * w for h, w in spatial_shapes]
+        val = memory.reshape(memory.shape[0], memory.shape[1], self.nhead, c) # memory: [bs, sum(h*w), hidden_dim]
+        # → [bs, n_head, c, sum_hw]
+        val = val.permute(0, 2, 3, 1).flatten(0, 1)  # [bs*n_head, c, sum_hw]
+        return val.split(split, dim=-1)  # list of [bs*n_head, c, h_l*w_l]
+
+    def forward(self, target, ref_pts_unact, memory, spatial_shapes, bbox_head, score_head, query_pos_head, pre_bbox_head, integral):
+        val_split_flat = self._value_op(memory, spatial_shapes) # pre-split value for deformable attention
+
+        # reshape to [bs*n_head, c, h_l, w_l]
+        value = []
+        for lvl, (h, w) in enumerate(spatial_shapes):
+            v = val_split_flat[lvl]   # [bs*n_head, c, h*w]
+            value.append(v.reshape(v.shape[0], v.shape[1], h, w))
+
+        ref_pts  = F.sigmoid(ref_pts_unact)
+        output   = target
+        output_detach = pred_corners_undetach = 0
+
+        dec_bboxes, dec_logits = [], []
+
+        for i, layer in enumerate(self.layers):
+            ref_input    = ref_pts.unsqueeze(2)           # [bs, Lq, 1, 4]
+            query_pos    = query_pos_head(ref_pts).clamp(-10, 10)
+            output       = layer(output, ref_input, value, spatial_shapes, query_pos=query_pos)
+
+            if i == 0:
+                ref_unact = ref_pts.clamp(1e-5, 1 - 1e-5)
+                ref_unact = torch.log(ref_unact / (1 - ref_unact))
+                pre_bboxes = F.sigmoid(pre_bbox_head(output) + ref_unact)
+                ref_pts_initial = pre_bboxes.detach()
+
+            pred_corners = bbox_head[i](output + output_detach) + pred_corners_undetach
+            inter_ref_bbox = distance2bbox(ref_pts_initial, integral(pred_corners, self.project), self.reg_scale)
+
+            if i == self.eval_idx:
+                scores = score_head[i](output)
+                scores = self.lqe_layers[i](scores, pred_corners)
+                dec_bboxes.append(inter_ref_bbox)
+                dec_logits.append(scores)
+                break
+
+            pred_corners_undetach = pred_corners
+            ref_pts        = inter_ref_bbox.detach()
+            output_detach  = output.detach()
+
+        return torch.stack(dec_bboxes), torch.stack(dec_logits)
+
+
+class DFINETransformer(nn.Module):
+    def __init__(self, num_classes=80, hidden_dim=256, num_queries=300, feat_channels=[256, 256, 256], feat_strides=[8, 16, 32],
+                 num_levels=3, num_points=[3, 6, 3], nhead=8, num_layers=6, dim_feedforward=1024, eval_idx=-1, eps=1e-2, reg_max=32,
+                 reg_scale=8.0, eval_spatial_size=(640, 640), device=None, dtype=None, operations=None):
+        super().__init__()
+        assert len(feat_strides) == len(feat_channels)
+        self.hidden_dim  = hidden_dim
+        self.num_queries = num_queries
+        self.num_levels  = num_levels
+        self.eps         = eps
+        self.eval_spatial_size = eval_spatial_size
+
+        self.feat_strides = list(feat_strides)
+        for i in range(num_levels - len(feat_strides)):
+            self.feat_strides.append(feat_strides[-1] * 2 ** (i + 1))
+
+        # input projection (expects pre-fused weights)
+        self.input_proj = nn.ModuleList()
+        for ch in feat_channels:
+            if ch == hidden_dim:
+                self.input_proj.append(nn.Identity())
+            else:
+                self.input_proj.append(nn.Sequential(OrderedDict([
+                    ('conv', operations.Conv2d(ch, hidden_dim, 1, bias=True, device=device, dtype=dtype))])))
+        in_ch = feat_channels[-1]
+        for i in range(num_levels - len(feat_channels)):
+            self.input_proj.append(nn.Sequential(OrderedDict([
+                ('conv', operations.Conv2d(in_ch if i == 0 else hidden_dim,
+                                           hidden_dim, 3, 2, 1, bias=True, device=device, dtype=dtype))])))
+            in_ch = hidden_dim
+
+        # FDR parameters (non-trainable placeholders, set from config)
+        self.up        = nn.Parameter(torch.tensor([0.5]),      requires_grad=False)
+        self.reg_scale = nn.Parameter(torch.tensor([reg_scale]), requires_grad=False)
+
+        pts = num_points if isinstance(num_points, (list, tuple)) else [num_points] * num_levels
+        self.decoder = TransformerDecoder(hidden_dim, nhead, dim_feedforward, num_levels, pts,
+                                          num_layers, reg_max, self.reg_scale, self.up, eval_idx, device=device, dtype=dtype, operations=operations)
+
+        self.query_pos_head = MLP(4, 2 * hidden_dim, hidden_dim, 2, device=device, dtype=dtype, operations=operations)
+        self.enc_output     = nn.Sequential(OrderedDict([
+            ('proj', operations.Linear(hidden_dim, hidden_dim, device=device, dtype=dtype)),
+            ('norm', operations.LayerNorm(hidden_dim, device=device, dtype=dtype))]))
+        self.enc_score_head = operations.Linear(hidden_dim, num_classes, device=device, dtype=dtype)
+        self.enc_bbox_head  = MLP(hidden_dim, hidden_dim, 4, 3, device=device, dtype=dtype, operations=operations)
+
+        self.eval_idx_ = eval_idx if eval_idx >= 0 else num_layers + eval_idx
+        self.dec_score_head = nn.ModuleList(
+            [operations.Linear(hidden_dim, num_classes, device=device, dtype=dtype) for _ in range(self.eval_idx_ + 1)])
+        self.pre_bbox_head  = MLP(hidden_dim, hidden_dim, 4, 3, device=device, dtype=dtype, operations=operations)
+        self.dec_bbox_head  = nn.ModuleList(
+            [MLP(hidden_dim, hidden_dim, 4 * (reg_max + 1), 3, device=device, dtype=dtype, operations=operations)
+             for _ in range(self.eval_idx_ + 1)])
+        self.integral = Integral(reg_max)
+
+        if eval_spatial_size:
+            # Register as buffers so checkpoint values override the freshly-computed defaults
+            anchors, valid_mask = self._gen_anchors()
+            self.register_buffer('anchors', anchors)
+            self.register_buffer('valid_mask', valid_mask)
+
+    def _gen_anchors(self, spatial_shapes=None, grid_size=0.05, dtype=torch.float32, device='cpu'):
+        if spatial_shapes is None:
+            h0, w0 = self.eval_spatial_size
+            spatial_shapes = [[int(h0 / s), int(w0 / s)] for s in self.feat_strides]
+        anchors = []
+        for lvl, (h, w) in enumerate(spatial_shapes):
+            gy, gx = torch.meshgrid(torch.arange(h), torch.arange(w), indexing='ij')
+            gxy = (torch.stack([gx, gy], -1).float() + 0.5) / torch.tensor([w, h], dtype=dtype)
+            wh  = torch.ones_like(gxy) * grid_size * (2. ** lvl)
+            anchors.append(torch.cat([gxy, wh], -1).reshape(-1, h * w, 4))
+        anchors    = torch.cat(anchors, 1).to(device)
+        valid_mask = ((anchors > self.eps) & (anchors < 1 - self.eps)).all(-1, keepdim=True)
+        anchors    = torch.log(anchors / (1 - anchors))
+        anchors    = torch.where(valid_mask, anchors, torch.full_like(anchors, float('inf')))
+        return anchors, valid_mask
+
+    def _encoder_input(self, feats: List[torch.Tensor]):
+        proj = [self.input_proj[i](f) for i, f in enumerate(feats)]
+        for i in range(len(feats), self.num_levels):
+            proj.append(self.input_proj[i](feats[-1] if i == len(feats) else proj[-1]))
+        flat, shapes = [], []
+        for f in proj:
+            _, _, h, w = f.shape
+            flat.append(f.flatten(2).permute(0, 2, 1))
+            shapes.append([h, w])
+        return torch.cat(flat, 1), shapes
+
+    def _decoder_input(self, memory: torch.Tensor):
+        anchors, valid_mask = self.anchors.to(memory), self.valid_mask
+        if memory.shape[0] > 1:
+            anchors = anchors.repeat(memory.shape[0], 1, 1)
+
+        mem      = valid_mask.to(memory) * memory
+        out_mem  = self.enc_output(mem)
+        logits   = self.enc_score_head(out_mem)
+        _, idx   = torch.topk(logits.max(-1).values, self.num_queries, dim=-1)
+        idx_e    = idx.unsqueeze(-1)
+        topk_mem = out_mem.gather(1, idx_e.expand(-1, -1, out_mem.shape[-1]))
+        topk_anc = anchors.gather(1, idx_e.expand(-1, -1, anchors.shape[-1]))
+        topk_ref = self.enc_bbox_head(topk_mem) + topk_anc
+        return topk_mem.detach(), topk_ref.detach()
+
+    def forward(self, feats: List[torch.Tensor]):
+        memory, shapes = self._encoder_input(feats)
+        content, ref   = self._decoder_input(memory)
+        out_bboxes, out_logits = self.decoder(
+            content, ref, memory, shapes,
+            self.dec_bbox_head, self.dec_score_head,
+            self.query_pos_head, self.pre_bbox_head, self.integral)
+        return {'pred_logits': out_logits[-1], 'pred_boxes': out_bboxes[-1]}
+
+
+# ---------------------------------------------------------------------------
+# Main model
+# ---------------------------------------------------------------------------
+
+class RTv4(nn.Module):
+    def __init__(self, num_classes=80, num_queries=300, enc_h=256, dec_h=256, enc_ff=2048, dec_ff=1024, feat_strides=[8, 16, 32], device=None, dtype=None, operations=None, **kwargs):
+        super().__init__()
+        self.device = device
+        self.dtype = dtype
+        self.operations = operations
+
+        self.backbone = HGNetv2(device=device, dtype=dtype, operations=operations)
+        self.encoder  = HybridEncoder(hidden_dim=enc_h, dim_feedforward=enc_ff, device=device, dtype=dtype, operations=operations)
+        self.decoder  = DFINETransformer(num_classes=num_classes, hidden_dim=dec_h, num_queries=num_queries,
+            feat_channels=[enc_h] * len(feat_strides), feat_strides=feat_strides, dim_feedforward=dec_ff, device=device, dtype=dtype, operations=operations)
+
+        self.num_classes = num_classes
+        self.num_queries = num_queries
+        self.load_device = comfy.model_management.get_torch_device()
+
+    def _forward(self, x: torch.Tensor):
+        return self.decoder(self.encoder(self.backbone(x)))
+
+    def postprocess(self, outputs, orig_size: tuple = (640, 640)) -> List[dict]:
+        logits = outputs['pred_logits']
+        boxes  = torchvision.ops.box_convert(outputs['pred_boxes'], 'cxcywh', 'xyxy')
+        boxes  = boxes * torch.tensor(orig_size, device=boxes.device, dtype=boxes.dtype).repeat(1, 2).unsqueeze(1)
+        scores = F.sigmoid(logits)
+        scores, idx = torch.topk(scores.flatten(1), self.num_queries, dim=-1)
+        labels = idx % self.num_classes
+        boxes  = boxes.gather(1, (idx // self.num_classes).unsqueeze(-1).expand(-1, -1, 4))
+        return [{'labels': lbl, 'boxes': b, 'scores': s} for lbl, b, s in zip(labels, boxes, scores)]
+
+    def forward(self, x: torch.Tensor, orig_size: tuple = (640, 640), **kwargs):
+        outputs = self._forward(x.to(device=self.load_device, dtype=self.dtype))
+        return self.postprocess(outputs, orig_size)

comfy/model_base.py

@@ -52,6 +52,8 @@ import comfy.ldm.qwen_image.model
 import comfy.ldm.kandinsky5.model
 import comfy.ldm.anima.model
 import comfy.ldm.ace.ace_step15
+import comfy.ldm.rt_detr.rtdetr_v4
+import comfy.ldm.ernie.model
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -1957,3 +1959,18 @@ class Kandinsky5Image(Kandinsky5):
 
     def concat_cond(self, **kwargs):
         return None
+
+class RT_DETR_v4(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.rt_detr.rtdetr_v4.RTv4)
+
+class ErnieImage(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ernie.model.ErnieImageModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out

comfy/model_detection.py

@@ -696,6 +696,26 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
     if '{}encoder.lyric_encoder.layers.0.input_layernorm.weight'.format(key_prefix) in state_dict_keys:
         dit_config = {}
         dit_config["audio_model"] = "ace1.5"
+        head_dim = 128
+        dit_config["hidden_size"] = state_dict['{}decoder.layers.0.self_attn_norm.weight'.format(key_prefix)].shape[0]
+        dit_config["intermediate_size"] = state_dict['{}decoder.layers.0.mlp.gate_proj.weight'.format(key_prefix)].shape[0]
+        dit_config["num_heads"] = state_dict['{}decoder.layers.0.self_attn.q_proj.weight'.format(key_prefix)].shape[0] // head_dim
+
+        dit_config["encoder_hidden_size"] = state_dict['{}encoder.lyric_encoder.layers.0.input_layernorm.weight'.format(key_prefix)].shape[0]
+        dit_config["encoder_num_heads"] = state_dict['{}encoder.lyric_encoder.layers.0.self_attn.q_proj.weight'.format(key_prefix)].shape[0] // head_dim
+        dit_config["encoder_intermediate_size"] = state_dict['{}encoder.lyric_encoder.layers.0.mlp.gate_proj.weight'.format(key_prefix)].shape[0]
+        dit_config["num_dit_layers"] = count_blocks(state_dict_keys, '{}decoder.layers.'.format(key_prefix) + '{}.')
+        return dit_config
+
+    if '{}encoder.pan_blocks.1.cv4.conv.weight'.format(key_prefix) in state_dict_keys: # RT-DETR_v4
+        dit_config = {}
+        dit_config["image_model"] = "RT_DETR_v4"
+        dit_config["enc_h"] = state_dict['{}encoder.pan_blocks.1.cv4.conv.weight'.format(key_prefix)].shape[0]
+        return dit_config
+
+    if '{}layers.0.mlp.linear_fc2.weight'.format(key_prefix) in state_dict_keys: # Ernie Image
+        dit_config = {}
+        dit_config["image_model"] = "ernie"
         return dit_config
 
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:

comfy/model_management.py

@@ -1326,9 +1326,9 @@ MAX_PINNED_MEMORY = -1
 if not args.disable_pinned_memory:
     if is_nvidia() or is_amd():
         if WINDOWS:
-            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.45  # Windows limit is apparently 50%
+            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.40  # Windows limit is apparently 50%
         else:
-            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95
+            MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.90
         logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024)))
 
 PINNING_ALLOWED_TYPES = set(["Tensor", "Parameter", "QuantizedTensor"])
@@ -1403,8 +1403,6 @@ def unpin_memory(tensor):
 
     if torch.cuda.cudart().cudaHostUnregister(ptr) == 0:
         TOTAL_PINNED_MEMORY -= PINNED_MEMORY.pop(ptr)
-        if len(PINNED_MEMORY) == 0:
-            TOTAL_PINNED_MEMORY = 0
         return True
     else:
         logging.warning("Unpin error.")
@@ -1734,6 +1732,21 @@ def supports_mxfp8_compute(device=None):
 
     return True
 
+def supports_fp64(device=None):
+    if is_device_mps(device):
+        return False
+
+    if is_intel_xpu():
+        return False
+
+    if is_directml_enabled():
+        return False
+
+    if is_ixuca():
+        return False
+
+    return True
+
 def extended_fp16_support():
     # TODO: check why some models work with fp16 on newer torch versions but not on older
     if torch_version_numeric < (2, 7):

comfy/ops.py

@@ -1151,7 +1151,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                     if param is None:
                         continue
                     p = fn(param)
-                    if p.is_inference():
+                    if (not torch.is_inference_mode_enabled()) and p.is_inference():
                         p = p.clone()
                     self.register_parameter(key, torch.nn.Parameter(p, requires_grad=False))
                 for key, buf in self._buffers.items():

comfy/sd.py

@@ -62,6 +62,7 @@ import comfy.text_encoders.anima
 import comfy.text_encoders.ace15
 import comfy.text_encoders.longcat_image
 import comfy.text_encoders.qwen35
+import comfy.text_encoders.ernie
 
 import comfy.model_patcher
 import comfy.lora
@@ -556,12 +557,19 @@ class VAE:
                         old_memory_used_decode = self.memory_used_decode
                         self.memory_used_decode = lambda shape, dtype: old_memory_used_decode(shape, dtype) *  4.0
 
+                    decoder_ch = sd['decoder.conv_in.weight'].shape[0] // ddconfig['ch_mult'][-1]
+                    if decoder_ch != ddconfig['ch']:
+                        decoder_ddconfig = ddconfig.copy()
+                        decoder_ddconfig['ch'] = decoder_ch
+                    else:
+                        decoder_ddconfig = None
+
                     if 'post_quant_conv.weight' in sd:
-                        self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1])
+                        self.first_stage_model = AutoencoderKL(ddconfig=ddconfig, embed_dim=sd['post_quant_conv.weight'].shape[1], **({"decoder_ddconfig": decoder_ddconfig} if decoder_ddconfig is not None else {}))
                     else:
                         self.first_stage_model = AutoencodingEngine(regularizer_config={'target': "comfy.ldm.models.autoencoder.DiagonalGaussianRegularizer"},
                                                                     encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig},
-                                                                    decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig})
+                                                                    decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': decoder_ddconfig if decoder_ddconfig is not None else ddconfig})
             elif "decoder.layers.1.layers.0.beta" in sd:
                 config = {}
                 param_key = None
@@ -1228,6 +1236,7 @@ class TEModel(Enum):
     QWEN35_4B = 25
     QWEN35_9B = 26
     QWEN35_27B = 27
+    MINISTRAL_3_3B = 28
 
 
 def detect_te_model(sd):
@@ -1294,6 +1303,8 @@ def detect_te_model(sd):
                 return TEModel.MISTRAL3_24B
             else:
                 return TEModel.MISTRAL3_24B_PRUNED_FLUX2
+        if weight.shape[0] == 3072:
+            return TEModel.MINISTRAL_3_3B
 
         return TEModel.LLAMA3_8
     return None
@@ -1451,6 +1462,10 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif te_model == TEModel.QWEN3_06B:
             clip_target.clip = comfy.text_encoders.anima.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.anima.AnimaTokenizer
+        elif te_model == TEModel.MINISTRAL_3_3B:
+            clip_target.clip = comfy.text_encoders.ernie.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.ernie.ErnieTokenizer
+            tokenizer_data["tekken_model"] = clip_data[0].get("tekken_model", None)
         else:
             # clip_l
             if clip_type == CLIPType.SD3:
@@ -1736,15 +1751,18 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None, disable
     """
     dtype = model_options.get("dtype", None)
 
+    custom_operations = model_options.get("custom_operations", None)
+    if custom_operations is None:
+        sd, metadata = comfy.utils.convert_old_quants(sd, "", metadata=metadata)
+
     #Allow loading unets from checkpoint files
     diffusion_model_prefix = model_detection.unet_prefix_from_state_dict(sd)
     temp_sd = comfy.utils.state_dict_prefix_replace(sd, {diffusion_model_prefix: ""}, filter_keys=True)
     if len(temp_sd) > 0:
         sd = temp_sd
+        if custom_operations is None:
+            sd, metadata = comfy.utils.convert_old_quants(sd, "", metadata=metadata)
 
-    custom_operations = model_options.get("custom_operations", None)
-    if custom_operations is None:
-        sd, metadata = comfy.utils.convert_old_quants(sd, "", metadata=metadata)
     parameters = comfy.utils.calculate_parameters(sd)
     weight_dtype = comfy.utils.weight_dtype(sd)
 

comfy/supported_models.py

@@ -26,6 +26,7 @@ import comfy.text_encoders.z_image
 import comfy.text_encoders.anima
 import comfy.text_encoders.ace15
 import comfy.text_encoders.longcat_image
+import comfy.text_encoders.ernie
 
 from . import supported_models_base
 from . import latent_formats
@@ -1734,6 +1735,52 @@ class LongCatImage(supported_models_base.BASE):
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.longcat_image.LongCatImageTokenizer, comfy.text_encoders.longcat_image.te(**hunyuan_detect))
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, LongCatImage, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImagePixelSpace, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, WAN21_FlowRVS, WAN21_SCAIL, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
+
+class RT_DETR_v4(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "RT_DETR_v4",
+    }
+
+    supported_inference_dtypes = [torch.float16, torch.float32]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.RT_DETR_v4(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return None
+
+
+class ErnieImage(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "ernie",
+    }
+
+    sampling_settings = {
+        "multiplier": 1000.0,
+        "shift": 3.0,
+    }
+
+    memory_usage_factor = 10.0
+
+    unet_extra_config = {}
+    latent_format = latent_formats.Flux2
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.ErnieImage(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}ministral3_3b.transformer.".format(pref))
+        return supported_models_base.ClipTarget(comfy.text_encoders.ernie.ErnieTokenizer, comfy.text_encoders.ernie.te(**hunyuan_detect))
+
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, LongCatImage, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImagePixelSpace, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, WAN21_FlowRVS, WAN21_SCAIL, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima, RT_DETR_v4, ErnieImage]
 
 models += [SVD_img2vid]

comfy/text_encoders/ernie.py

@@ -0,0 +1,38 @@
+from .flux import Mistral3Tokenizer
+from comfy import sd1_clip
+import comfy.text_encoders.llama
+
+class Ministral3_3BTokenizer(Mistral3Tokenizer):
+    def __init__(self, embedding_directory=None, embedding_size=5120, embedding_key='ministral3_3b', tokenizer_data={}):
+        return super().__init__(embedding_directory=embedding_directory, embedding_size=embedding_size, embedding_key=embedding_key, tokenizer_data=tokenizer_data)
+
+class ErnieTokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="ministral3_3b", tokenizer=Mistral3Tokenizer)
+
+    def tokenize_with_weights(self, text, return_word_ids=False, llama_template=None, **kwargs):
+        tokens = super().tokenize_with_weights(text, return_word_ids=return_word_ids, disable_weights=True, **kwargs)
+        return tokens
+
+
+class Ministral3_3BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="hidden", layer_idx=-2, dtype=None, attention_mask=True, model_options={}):
+        textmodel_json_config = {}
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config=textmodel_json_config, dtype=dtype, special_tokens={"start": 1, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Ministral3_3B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+
+class ErnieTEModel(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, name="ministral3_3b", clip_model=Ministral3_3BModel):
+        super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options)
+
+
+def te(dtype_llama=None, llama_quantization_metadata=None):
+    class ErnieTEModel_(ErnieTEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["quantization_metadata"] = llama_quantization_metadata
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return ErnieTEModel

comfy/text_encoders/flux.py

@@ -116,9 +116,9 @@ class MistralTokenizerClass:
         return LlamaTokenizerFast(**kwargs)
 
 class Mistral3Tokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
+    def __init__(self, embedding_directory=None, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_data={}):
         self.tekken_data = tokenizer_data.get("tekken_model", None)
-        super().__init__("", pad_with_end=False, embedding_directory=embedding_directory, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
+        super().__init__("", pad_with_end=False, embedding_directory=embedding_directory, embedding_size=embedding_size, embedding_key=embedding_key, tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, disable_weights=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"tekken_model": self.tekken_data}

comfy/text_encoders/llama.py

@@ -60,6 +60,30 @@ class Mistral3Small24BConfig:
     final_norm: bool = True
     lm_head: bool = False
 
+@dataclass
+class Ministral3_3BConfig:
+    vocab_size: int = 131072
+    hidden_size: int = 3072
+    intermediate_size: int = 9216
+    num_hidden_layers: int = 26
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 262144
+    rms_norm_eps: float = 1e-5
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = None
+    k_norm = None
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+    stop_tokens = [2]
+
 @dataclass
 class Qwen25_3BConfig:
     vocab_size: int = 151936
@@ -946,6 +970,15 @@ class Mistral3Small24B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+class Ministral3_3B(BaseLlama, BaseQwen3, BaseGenerate, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Ministral3_3BConfig(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
 class Qwen25_3B(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()

comfy_api_nodes/apis/bytedance.py

@@ -52,6 +52,26 @@ class TaskImageContent(BaseModel):
     role: Literal["first_frame", "last_frame", "reference_image"] | None = Field(None)
 
 
+class TaskVideoContentUrl(BaseModel):
+    url: str = Field(...)
+
+
+class TaskVideoContent(BaseModel):
+    type: str = Field("video_url")
+    video_url: TaskVideoContentUrl = Field(...)
+    role: str = Field("reference_video")
+
+
+class TaskAudioContentUrl(BaseModel):
+    url: str = Field(...)
+
+
+class TaskAudioContent(BaseModel):
+    type: str = Field("audio_url")
+    audio_url: TaskAudioContentUrl = Field(...)
+    role: str = Field("reference_audio")
+
+
 class Text2VideoTaskCreationRequest(BaseModel):
     model: str = Field(...)
     content: list[TaskTextContent] = Field(..., min_length=1)
@@ -64,6 +84,17 @@ class Image2VideoTaskCreationRequest(BaseModel):
     generate_audio: bool | None = Field(...)
 
 
+class Seedance2TaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    content: list[TaskTextContent | TaskImageContent | TaskVideoContent | TaskAudioContent] = Field(..., min_length=1)
+    generate_audio: bool | None = Field(None)
+    resolution: str | None = Field(None)
+    ratio: str | None = Field(None)
+    duration: int | None = Field(None, ge=4, le=15)
+    seed: int | None = Field(None, ge=0, le=2147483647)
+    watermark: bool | None = Field(None)
+
+
 class TaskCreationResponse(BaseModel):
     id: str = Field(...)
 
@@ -77,12 +108,27 @@ class TaskStatusResult(BaseModel):
     video_url: str = Field(...)
 
 
+class TaskStatusUsage(BaseModel):
+    completion_tokens: int = Field(0)
+    total_tokens: int = Field(0)
+
+
 class TaskStatusResponse(BaseModel):
     id: str = Field(...)
     model: str = Field(...)
     status: Literal["queued", "running", "cancelled", "succeeded", "failed"] = Field(...)
     error: TaskStatusError | None = Field(None)
     content: TaskStatusResult | None = Field(None)
+    usage: TaskStatusUsage | None = Field(None)
+
+
+# Dollars per 1K tokens, keyed by (model_id, has_video_input).
+SEEDANCE2_PRICE_PER_1K_TOKENS = {
+    ("dreamina-seedance-2-0-260128", False): 0.007,
+    ("dreamina-seedance-2-0-260128", True): 0.0043,
+    ("dreamina-seedance-2-0-fast-260128", False): 0.0056,
+    ("dreamina-seedance-2-0-fast-260128", True): 0.0033,
+}
 
 
 RECOMMENDED_PRESETS = [
@@ -112,6 +158,12 @@ RECOMMENDED_PRESETS_SEEDREAM_4 = [
     ("Custom", None, None),
 ]
 
+# Seedance 2.0 reference video pixel count limits per model.
+SEEDANCE2_REF_VIDEO_PIXEL_LIMITS = {
+    "dreamina-seedance-2-0-260128": {"min": 409_600, "max": 927_408},
+    "dreamina-seedance-2-0-fast-260128": {"min": 409_600, "max": 927_408},
+}
+
 # The time in this dictionary are given for 10 seconds duration.
 VIDEO_TASKS_EXECUTION_TIME = {
     "seedance-1-0-lite-t2v-250428": {

comfy_api_nodes/apis/wan.py

@@ -0,0 +1,226 @@
+from pydantic import BaseModel, Field
+
+
+class Text2ImageInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+
+
+class Image2ImageInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    images: list[str] = Field(..., min_length=1, max_length=2)
+
+
+class Text2VideoInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    audio_url: str | None = Field(None)
+
+
+class Image2VideoInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    img_url: str = Field(...)
+    audio_url: str | None = Field(None)
+
+
+class Reference2VideoInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    reference_video_urls: list[str] = Field(...)
+
+
+class Txt2ImageParametersField(BaseModel):
+    size: str = Field(...)
+    n: int = Field(1, description="Number of images to generate.")  # we support only value=1
+    seed: int = Field(..., ge=0, le=2147483647)
+    prompt_extend: bool = Field(True)
+    watermark: bool = Field(False)
+
+
+class Image2ImageParametersField(BaseModel):
+    size: str | None = Field(None)
+    n: int = Field(1, description="Number of images to generate.")  # we support only value=1
+    seed: int = Field(..., ge=0, le=2147483647)
+    watermark: bool = Field(False)
+
+
+class Text2VideoParametersField(BaseModel):
+    size: str = Field(...)
+    seed: int = Field(..., ge=0, le=2147483647)
+    duration: int = Field(5, ge=5, le=15)
+    prompt_extend: bool = Field(True)
+    watermark: bool = Field(False)
+    audio: bool = Field(False, description="Whether to generate audio automatically.")
+    shot_type: str = Field("single")
+
+
+class Image2VideoParametersField(BaseModel):
+    resolution: str = Field(...)
+    seed: int = Field(..., ge=0, le=2147483647)
+    duration: int = Field(5, ge=5, le=15)
+    prompt_extend: bool = Field(True)
+    watermark: bool = Field(False)
+    audio: bool = Field(False, description="Whether to generate audio automatically.")
+    shot_type: str = Field("single")
+
+
+class Reference2VideoParametersField(BaseModel):
+    size: str = Field(...)
+    duration: int = Field(5, ge=5, le=15)
+    shot_type: str = Field("single")
+    seed: int = Field(..., ge=0, le=2147483647)
+    watermark: bool = Field(False)
+
+
+class Text2ImageTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Text2ImageInputField = Field(...)
+    parameters: Txt2ImageParametersField = Field(...)
+
+
+class Image2ImageTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Image2ImageInputField = Field(...)
+    parameters: Image2ImageParametersField = Field(...)
+
+
+class Text2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Text2VideoInputField = Field(...)
+    parameters: Text2VideoParametersField = Field(...)
+
+
+class Image2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Image2VideoInputField = Field(...)
+    parameters: Image2VideoParametersField = Field(...)
+
+
+class Reference2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Reference2VideoInputField = Field(...)
+    parameters: Reference2VideoParametersField = Field(...)
+
+
+class Wan27MediaItem(BaseModel):
+    type: str = Field(...)
+    url: str = Field(...)
+
+
+class Wan27ReferenceVideoInputField(BaseModel):
+    prompt: str = Field(...)
+    negative_prompt: str | None = Field(None)
+    media: list[Wan27MediaItem] = Field(...)
+
+
+class Wan27ReferenceVideoParametersField(BaseModel):
+    resolution: str = Field(...)
+    ratio: str | None = Field(None)
+    duration: int = Field(5, ge=2, le=10)
+    watermark: bool = Field(False)
+    seed: int = Field(..., ge=0, le=2147483647)
+
+
+class Wan27ReferenceVideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Wan27ReferenceVideoInputField = Field(...)
+    parameters: Wan27ReferenceVideoParametersField = Field(...)
+
+
+class Wan27ImageToVideoInputField(BaseModel):
+    prompt: str | None = Field(None)
+    negative_prompt: str | None = Field(None)
+    media: list[Wan27MediaItem] = Field(...)
+
+
+class Wan27ImageToVideoParametersField(BaseModel):
+    resolution: str = Field(...)
+    duration: int = Field(5, ge=2, le=15)
+    prompt_extend: bool = Field(True)
+    watermark: bool = Field(False)
+    seed: int = Field(..., ge=0, le=2147483647)
+
+
+class Wan27ImageToVideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Wan27ImageToVideoInputField = Field(...)
+    parameters: Wan27ImageToVideoParametersField = Field(...)
+
+
+class Wan27VideoEditInputField(BaseModel):
+    prompt: str = Field(...)
+    media: list[Wan27MediaItem] = Field(...)
+
+
+class Wan27VideoEditParametersField(BaseModel):
+    resolution: str = Field(...)
+    ratio: str | None = Field(None)
+    duration: int = Field(0)
+    audio_setting: str = Field("auto")
+    watermark: bool = Field(False)
+    seed: int = Field(..., ge=0, le=2147483647)
+
+
+class Wan27VideoEditTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Wan27VideoEditInputField = Field(...)
+    parameters: Wan27VideoEditParametersField = Field(...)
+
+
+class Wan27Text2VideoParametersField(BaseModel):
+    resolution: str = Field(...)
+    ratio: str | None = Field(None)
+    duration: int = Field(5, ge=2, le=15)
+    prompt_extend: bool = Field(True)
+    watermark: bool = Field(False)
+    seed: int = Field(..., ge=0, le=2147483647)
+
+
+class Wan27Text2VideoTaskCreationRequest(BaseModel):
+    model: str = Field(...)
+    input: Text2VideoInputField = Field(...)
+    parameters: Wan27Text2VideoParametersField = Field(...)
+
+
+class TaskCreationOutputField(BaseModel):
+    task_id: str = Field(...)
+    task_status: str = Field(...)
+
+
+class TaskCreationResponse(BaseModel):
+    output: TaskCreationOutputField | None = Field(None)
+    request_id: str = Field(...)
+    code: str | None = Field(None, description="Error code for the failed request.")
+    message: str | None = Field(None, description="Details about the failed request.")
+
+
+class TaskResult(BaseModel):
+    url: str | None = Field(None)
+    code: str | None = Field(None)
+    message: str | None = Field(None)
+
+
+class ImageTaskStatusOutputField(TaskCreationOutputField):
+    task_id: str = Field(...)
+    task_status: str = Field(...)
+    results: list[TaskResult] | None = Field(None)
+
+
+class VideoTaskStatusOutputField(TaskCreationOutputField):
+    task_id: str = Field(...)
+    task_status: str = Field(...)
+    video_url: str | None = Field(None)
+    code: str | None = Field(None)
+    message: str | None = Field(None)
+
+
+class ImageTaskStatusResponse(BaseModel):
+    output: ImageTaskStatusOutputField | None = Field(None)
+    request_id: str = Field(...)
+
+
+class VideoTaskStatusResponse(BaseModel):
+    output: VideoTaskStatusOutputField | None = Field(None)
+    request_id: str = Field(...)

comfy_api_nodes/nodes_bytedance.py

@@ -8,16 +8,23 @@ from comfy_api.latest import IO, ComfyExtension, Input
 from comfy_api_nodes.apis.bytedance import (
     RECOMMENDED_PRESETS,
     RECOMMENDED_PRESETS_SEEDREAM_4,
+    SEEDANCE2_PRICE_PER_1K_TOKENS,
+    SEEDANCE2_REF_VIDEO_PIXEL_LIMITS,
     VIDEO_TASKS_EXECUTION_TIME,
     Image2VideoTaskCreationRequest,
     ImageTaskCreationResponse,
+    Seedance2TaskCreationRequest,
     Seedream4Options,
     Seedream4TaskCreationRequest,
+    TaskAudioContent,
+    TaskAudioContentUrl,
     TaskCreationResponse,
     TaskImageContent,
     TaskImageContentUrl,
     TaskStatusResponse,
     TaskTextContent,
+    TaskVideoContent,
+    TaskVideoContentUrl,
     Text2ImageTaskCreationRequest,
     Text2VideoTaskCreationRequest,
 )
@@ -29,7 +36,10 @@ from comfy_api_nodes.util import (
     image_tensor_pair_to_batch,
     poll_op,
     sync_op,
+    upload_audio_to_comfyapi,
+    upload_image_to_comfyapi,
     upload_images_to_comfyapi,
+    upload_video_to_comfyapi,
     validate_image_aspect_ratio,
     validate_image_dimensions,
     validate_string,
@@ -46,12 +56,56 @@ SEEDREAM_MODELS = {
 # Long-running tasks endpoints(e.g., video)
 BYTEPLUS_TASK_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks"
 BYTEPLUS_TASK_STATUS_ENDPOINT = "/proxy/byteplus/api/v3/contents/generations/tasks"  # + /{task_id}
+BYTEPLUS_SEEDANCE2_TASK_STATUS_ENDPOINT = "/proxy/byteplus-seedance2/api/v3/contents/generations/tasks"  # + /{task_id}
+
+SEEDANCE_MODELS = {
+    "Seedance 2.0": "dreamina-seedance-2-0-260128",
+    "Seedance 2.0 Fast": "dreamina-seedance-2-0-fast-260128",
+}
 
 DEPRECATED_MODELS = {"seedance-1-0-lite-t2v-250428", "seedance-1-0-lite-i2v-250428"}
 
+
 logger = logging.getLogger(__name__)
 
 
+def _validate_ref_video_pixels(video: Input.Video, model_id: str, index: int) -> None:
+    """Validate reference video pixel count against Seedance 2.0 model limits."""
+    limits = SEEDANCE2_REF_VIDEO_PIXEL_LIMITS.get(model_id)
+    if not limits:
+        return
+    try:
+        w, h = video.get_dimensions()
+    except Exception:
+        return
+    pixels = w * h
+    min_px = limits.get("min")
+    max_px = limits.get("max")
+    if min_px and pixels < min_px:
+        raise ValueError(
+            f"Reference video {index} is too small: {w}x{h} = {pixels:,}px. " f"Minimum is {min_px:,}px for this model."
+        )
+    if max_px and pixels > max_px:
+        raise ValueError(
+            f"Reference video {index} is too large: {w}x{h} = {pixels:,}px. "
+            f"Maximum is {max_px:,}px for this model. Try downscaling the video."
+        )
+
+
+def _seedance2_price_extractor(model_id: str, has_video_input: bool):
+    """Returns a price_extractor closure for Seedance 2.0 poll_op."""
+    rate = SEEDANCE2_PRICE_PER_1K_TOKENS.get((model_id, has_video_input))
+    if rate is None:
+        return None
+
+    def extractor(response: TaskStatusResponse) -> float | None:
+        if response.usage is None:
+            return None
+        return response.usage.total_tokens * 1.43 * rate / 1_000.0
+
+    return extractor
+
+
 def get_image_url_from_response(response: ImageTaskCreationResponse) -> str:
     if response.error:
         error_msg = f"ByteDance request failed. Code: {response.error['code']}, message: {response.error['message']}"
@@ -335,8 +389,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
         mp_provided = out_num_pixels / 1_000_000.0
         if ("seedream-4-5" in model or "seedream-5-0" in model) and out_num_pixels < 3686400:
             raise ValueError(
-                f"Minimum image resolution for the selected model is 3.68MP, "
-                f"but {mp_provided:.2f}MP provided."
+                f"Minimum image resolution for the selected model is 3.68MP, " f"but {mp_provided:.2f}MP provided."
             )
         if "seedream-4-0" in model and out_num_pixels < 921600:
             raise ValueError(
@@ -952,33 +1005,6 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
         )
 
 
-async def process_video_task(
-    cls: type[IO.ComfyNode],
-    payload: Text2VideoTaskCreationRequest | Image2VideoTaskCreationRequest,
-    estimated_duration: int | None,
-) -> IO.NodeOutput:
-    if payload.model in DEPRECATED_MODELS:
-        logger.warning(
-            "Model '%s' is deprecated and will be deactivated on May 13, 2026. "
-            "Please switch to a newer model. Recommended: seedance-1-0-pro-fast-251015.",
-            payload.model,
-        )
-    initial_response = await sync_op(
-        cls,
-        ApiEndpoint(path=BYTEPLUS_TASK_ENDPOINT, method="POST"),
-        data=payload,
-        response_model=TaskCreationResponse,
-    )
-    response = await poll_op(
-        cls,
-        ApiEndpoint(path=f"{BYTEPLUS_TASK_STATUS_ENDPOINT}/{initial_response.id}"),
-        status_extractor=lambda r: r.status,
-        estimated_duration=estimated_duration,
-        response_model=TaskStatusResponse,
-    )
-    return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
-
-
 def raise_if_text_params(prompt: str, text_params: list[str]) -> None:
     for i in text_params:
         if f"--{i} " in prompt:
@@ -1040,6 +1066,530 @@ PRICE_BADGE_VIDEO = IO.PriceBadge(
 )
 
 
+def _seedance2_text_inputs():
+    return [
+        IO.String.Input(
+            "prompt",
+            multiline=True,
+            default="",
+            tooltip="Text prompt for video generation.",
+        ),
+        IO.Combo.Input(
+            "resolution",
+            options=["480p", "720p"],
+            tooltip="Resolution of the output video.",
+        ),
+        IO.Combo.Input(
+            "ratio",
+            options=["16:9", "4:3", "1:1", "3:4", "9:16", "21:9", "adaptive"],
+            tooltip="Aspect ratio of the output video.",
+        ),
+        IO.Int.Input(
+            "duration",
+            default=7,
+            min=4,
+            max=15,
+            step=1,
+            tooltip="Duration of the output video in seconds (4-15).",
+            display_mode=IO.NumberDisplay.slider,
+        ),
+        IO.Boolean.Input(
+            "generate_audio",
+            default=True,
+            tooltip="Enable audio generation for the output video.",
+        ),
+    ]
+
+
+class ByteDance2TextToVideoNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="ByteDance2TextToVideoNode",
+            display_name="ByteDance Seedance 2.0 Text to Video",
+            category="api node/video/ByteDance",
+            description="Generate video using Seedance 2.0 models based on a text prompt.",
+            inputs=[
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option("Seedance 2.0", _seedance2_text_inputs()),
+                        IO.DynamicCombo.Option("Seedance 2.0 Fast", _seedance2_text_inputs()),
+                    ],
+                    tooltip="Seedance 2.0 for maximum quality; Seedance 2.0 Fast for speed optimization.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+                IO.Boolean.Input(
+                    "watermark",
+                    default=False,
+                    tooltip="Whether to add a watermark to the video.",
+                    advanced=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=["model", "model.resolution", "model.duration"]),
+                expr="""
+                (
+                  $rate480 := 10044;
+                  $rate720 := 21600;
+                  $m := widgets.model;
+                  $pricePer1K := $contains($m, "fast") ? 0.008008 : 0.01001;
+                  $res := $lookup(widgets, "model.resolution");
+                  $dur := $lookup(widgets, "model.duration");
+                  $rate := $res = "720p" ? $rate720 : $rate480;
+                  $cost := $dur * $rate * $pricePer1K / 1000;
+                  {"type": "usd", "usd": $cost, "format": {"approximate": true}}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: dict,
+        seed: int,
+        watermark: bool,
+    ) -> IO.NodeOutput:
+        validate_string(model["prompt"], strip_whitespace=True, min_length=1)
+        model_id = SEEDANCE_MODELS[model["model"]]
+        initial_response = await sync_op(
+            cls,
+            ApiEndpoint(path=BYTEPLUS_TASK_ENDPOINT, method="POST"),
+            data=Seedance2TaskCreationRequest(
+                model=model_id,
+                content=[TaskTextContent(text=model["prompt"])],
+                generate_audio=model["generate_audio"],
+                resolution=model["resolution"],
+                ratio=model["ratio"],
+                duration=model["duration"],
+                seed=seed,
+                watermark=watermark,
+            ),
+            response_model=TaskCreationResponse,
+        )
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"{BYTEPLUS_SEEDANCE2_TASK_STATUS_ENDPOINT}/{initial_response.id}"),
+            response_model=TaskStatusResponse,
+            status_extractor=lambda r: r.status,
+            price_extractor=_seedance2_price_extractor(model_id, has_video_input=False),
+            poll_interval=9,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
+
+
+class ByteDance2FirstLastFrameNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="ByteDance2FirstLastFrameNode",
+            display_name="ByteDance Seedance 2.0 First-Last-Frame to Video",
+            category="api node/video/ByteDance",
+            description="Generate video using Seedance 2.0 from a first frame image and optional last frame image.",
+            inputs=[
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option("Seedance 2.0", _seedance2_text_inputs()),
+                        IO.DynamicCombo.Option("Seedance 2.0 Fast", _seedance2_text_inputs()),
+                    ],
+                    tooltip="Seedance 2.0 for maximum quality; Seedance 2.0 Fast for speed optimization.",
+                ),
+                IO.Image.Input(
+                    "first_frame",
+                    tooltip="First frame image for the video.",
+                ),
+                IO.Image.Input(
+                    "last_frame",
+                    tooltip="Last frame image for the video.",
+                    optional=True,
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+                IO.Boolean.Input(
+                    "watermark",
+                    default=False,
+                    tooltip="Whether to add a watermark to the video.",
+                    advanced=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=["model", "model.resolution", "model.duration"]),
+                expr="""
+                (
+                  $rate480 := 10044;
+                  $rate720 := 21600;
+                  $m := widgets.model;
+                  $pricePer1K := $contains($m, "fast") ? 0.008008 : 0.01001;
+                  $res := $lookup(widgets, "model.resolution");
+                  $dur := $lookup(widgets, "model.duration");
+                  $rate := $res = "720p" ? $rate720 : $rate480;
+                  $cost := $dur * $rate * $pricePer1K / 1000;
+                  {"type": "usd", "usd": $cost, "format": {"approximate": true}}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: dict,
+        first_frame: Input.Image,
+        seed: int,
+        watermark: bool,
+        last_frame: Input.Image | None = None,
+    ) -> IO.NodeOutput:
+        validate_string(model["prompt"], strip_whitespace=True, min_length=1)
+        model_id = SEEDANCE_MODELS[model["model"]]
+
+        content: list[TaskTextContent | TaskImageContent] = [
+            TaskTextContent(text=model["prompt"]),
+            TaskImageContent(
+                image_url=TaskImageContentUrl(
+                    url=await upload_image_to_comfyapi(cls, first_frame, wait_label="Uploading first frame.")
+                ),
+                role="first_frame",
+            ),
+        ]
+        if last_frame is not None:
+            content.append(
+                TaskImageContent(
+                    image_url=TaskImageContentUrl(
+                        url=await upload_image_to_comfyapi(cls, last_frame, wait_label="Uploading last frame.")
+                    ),
+                    role="last_frame",
+                ),
+            )
+
+        initial_response = await sync_op(
+            cls,
+            ApiEndpoint(path=BYTEPLUS_TASK_ENDPOINT, method="POST"),
+            data=Seedance2TaskCreationRequest(
+                model=model_id,
+                content=content,
+                generate_audio=model["generate_audio"],
+                resolution=model["resolution"],
+                ratio=model["ratio"],
+                duration=model["duration"],
+                seed=seed,
+                watermark=watermark,
+            ),
+            response_model=TaskCreationResponse,
+        )
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"{BYTEPLUS_SEEDANCE2_TASK_STATUS_ENDPOINT}/{initial_response.id}"),
+            response_model=TaskStatusResponse,
+            status_extractor=lambda r: r.status,
+            price_extractor=_seedance2_price_extractor(model_id, has_video_input=False),
+            poll_interval=9,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
+
+
+def _seedance2_reference_inputs():
+    return [
+        *_seedance2_text_inputs(),
+        IO.Autogrow.Input(
+            "reference_images",
+            template=IO.Autogrow.TemplateNames(
+                IO.Image.Input("reference_image"),
+                names=[
+                    "image_1",
+                    "image_2",
+                    "image_3",
+                    "image_4",
+                    "image_5",
+                    "image_6",
+                    "image_7",
+                    "image_8",
+                    "image_9",
+                ],
+                min=0,
+            ),
+        ),
+        IO.Autogrow.Input(
+            "reference_videos",
+            template=IO.Autogrow.TemplateNames(
+                IO.Video.Input("reference_video"),
+                names=["video_1", "video_2", "video_3"],
+                min=0,
+            ),
+        ),
+        IO.Autogrow.Input(
+            "reference_audios",
+            template=IO.Autogrow.TemplateNames(
+                IO.Audio.Input("reference_audio"),
+                names=["audio_1", "audio_2", "audio_3"],
+                min=0,
+            ),
+        ),
+    ]
+
+
+class ByteDance2ReferenceNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="ByteDance2ReferenceNode",
+            display_name="ByteDance Seedance 2.0 Reference to Video",
+            category="api node/video/ByteDance",
+            description="Generate, edit, or extend video using Seedance 2.0 with reference images, "
+            "videos, and audio. Supports multimodal reference, video editing, and video extension.",
+            inputs=[
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option("Seedance 2.0", _seedance2_reference_inputs()),
+                        IO.DynamicCombo.Option("Seedance 2.0 Fast", _seedance2_reference_inputs()),
+                    ],
+                    tooltip="Seedance 2.0 for maximum quality; Seedance 2.0 Fast for speed optimization.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+                IO.Boolean.Input(
+                    "watermark",
+                    default=False,
+                    tooltip="Whether to add a watermark to the video.",
+                    advanced=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=["model", "model.resolution", "model.duration"],
+                    input_groups=["model.reference_videos"],
+                ),
+                expr="""
+                (
+                  $rate480 := 10044;
+                  $rate720 := 21600;
+                  $m := widgets.model;
+                  $hasVideo := $lookup(inputGroups, "model.reference_videos") > 0;
+                  $noVideoPricePer1K := $contains($m, "fast") ? 0.008008 : 0.01001;
+                  $videoPricePer1K := $contains($m, "fast") ? 0.004719 : 0.006149;
+                  $res := $lookup(widgets, "model.resolution");
+                  $dur := $lookup(widgets, "model.duration");
+                  $rate := $res = "720p" ? $rate720 : $rate480;
+                  $noVideoCost := $dur * $rate * $noVideoPricePer1K / 1000;
+                  $minVideoFactor := $ceil($dur * 5 / 3);
+                  $minVideoCost := $minVideoFactor * $rate * $videoPricePer1K / 1000;
+                  $maxVideoCost := (15 + $dur) * $rate * $videoPricePer1K / 1000;
+                  $hasVideo
+                    ? {
+                        "type": "range_usd",
+                        "min_usd": $minVideoCost,
+                        "max_usd": $maxVideoCost,
+                        "format": {"approximate": true}
+                      }
+                    : {
+                        "type": "usd",
+                        "usd": $noVideoCost,
+                        "format": {"approximate": true}
+                      }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: dict,
+        seed: int,
+        watermark: bool,
+    ) -> IO.NodeOutput:
+        validate_string(model["prompt"], strip_whitespace=True, min_length=1)
+
+        reference_images = model.get("reference_images", {})
+        reference_videos = model.get("reference_videos", {})
+        reference_audios = model.get("reference_audios", {})
+
+        if not reference_images and not reference_videos:
+            raise ValueError("At least one reference image or video is required.")
+
+        model_id = SEEDANCE_MODELS[model["model"]]
+        has_video_input = len(reference_videos) > 0
+        total_video_duration = 0.0
+        for i, key in enumerate(reference_videos, 1):
+            video = reference_videos[key]
+            _validate_ref_video_pixels(video, model_id, i)
+            try:
+                dur = video.get_duration()
+                if dur < 1.8:
+                    raise ValueError(f"Reference video {i} is too short: {dur:.1f}s. Minimum duration is 1.8 seconds.")
+                total_video_duration += dur
+            except ValueError:
+                raise
+            except Exception:
+                pass
+        if total_video_duration > 15.1:
+            raise ValueError(f"Total reference video duration is {total_video_duration:.1f}s. Maximum is 15.1 seconds.")
+
+        total_audio_duration = 0.0
+        for i, key in enumerate(reference_audios, 1):
+            audio = reference_audios[key]
+            dur = int(audio["waveform"].shape[-1]) / int(audio["sample_rate"])
+            if dur < 1.8:
+                raise ValueError(f"Reference audio {i} is too short: {dur:.1f}s. Minimum duration is 1.8 seconds.")
+            total_audio_duration += dur
+        if total_audio_duration > 15.1:
+            raise ValueError(f"Total reference audio duration is {total_audio_duration:.1f}s. Maximum is 15.1 seconds.")
+
+        content: list[TaskTextContent | TaskImageContent | TaskVideoContent | TaskAudioContent] = [
+            TaskTextContent(text=model["prompt"]),
+        ]
+        for i, key in enumerate(reference_images, 1):
+            content.append(
+                TaskImageContent(
+                    image_url=TaskImageContentUrl(
+                        url=await upload_image_to_comfyapi(
+                            cls,
+                            image=reference_images[key],
+                            wait_label=f"Uploading image {i}",
+                        ),
+                    ),
+                    role="reference_image",
+                ),
+            )
+        for i, key in enumerate(reference_videos, 1):
+            content.append(
+                TaskVideoContent(
+                    video_url=TaskVideoContentUrl(
+                        url=await upload_video_to_comfyapi(
+                            cls,
+                            reference_videos[key],
+                            wait_label=f"Uploading video {i}",
+                        ),
+                    ),
+                ),
+            )
+        for key in reference_audios:
+            content.append(
+                TaskAudioContent(
+                    audio_url=TaskAudioContentUrl(
+                        url=await upload_audio_to_comfyapi(
+                            cls,
+                            reference_audios[key],
+                            container_format="mp3",
+                            codec_name="libmp3lame",
+                            mime_type="audio/mpeg",
+                        ),
+                    ),
+                ),
+            )
+        initial_response = await sync_op(
+            cls,
+            ApiEndpoint(path=BYTEPLUS_TASK_ENDPOINT, method="POST"),
+            data=Seedance2TaskCreationRequest(
+                model=model_id,
+                content=content,
+                generate_audio=model["generate_audio"],
+                resolution=model["resolution"],
+                ratio=model["ratio"],
+                duration=model["duration"],
+                seed=seed,
+                watermark=watermark,
+            ),
+            response_model=TaskCreationResponse,
+        )
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"{BYTEPLUS_SEEDANCE2_TASK_STATUS_ENDPOINT}/{initial_response.id}"),
+            response_model=TaskStatusResponse,
+            status_extractor=lambda r: r.status,
+            price_extractor=_seedance2_price_extractor(model_id, has_video_input=has_video_input),
+            poll_interval=9,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
+
+
+async def process_video_task(
+    cls: type[IO.ComfyNode],
+    payload: Text2VideoTaskCreationRequest | Image2VideoTaskCreationRequest,
+    estimated_duration: int | None,
+) -> IO.NodeOutput:
+    if payload.model in DEPRECATED_MODELS:
+        logger.warning(
+            "Model '%s' is deprecated and will be deactivated on May 13, 2026. "
+            "Please switch to a newer model. Recommended: seedance-1-0-pro-fast-251015.",
+            payload.model,
+        )
+    initial_response = await sync_op(
+        cls,
+        ApiEndpoint(path=BYTEPLUS_TASK_ENDPOINT, method="POST"),
+        data=payload,
+        response_model=TaskCreationResponse,
+    )
+    response = await poll_op(
+        cls,
+        ApiEndpoint(path=f"{BYTEPLUS_TASK_STATUS_ENDPOINT}/{initial_response.id}"),
+        status_extractor=lambda r: r.status,
+        estimated_duration=estimated_duration,
+        response_model=TaskStatusResponse,
+    )
+    return IO.NodeOutput(await download_url_to_video_output(response.content.video_url))
+
+
 class ByteDanceExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -1050,6 +1600,9 @@ class ByteDanceExtension(ComfyExtension):
             ByteDanceImageToVideoNode,
             ByteDanceFirstLastFrameNode,
             ByteDanceImageReferenceNode,
+            ByteDance2TextToVideoNode,
+            ByteDance2FirstLastFrameNode,
+            ByteDance2ReferenceNode,
         ]
 
 

comfy_api_nodes/nodes_grok.py

@@ -558,7 +558,7 @@ class GrokVideoReferenceNode(IO.ComfyNode):
                 (
                   $res := $lookup(widgets, "model.resolution");
                   $dur := $lookup(widgets, "model.duration");
-                  $refs := inputGroups["model.reference_images"];
+                  $refs := $lookup(inputGroups, "model.reference_images");
                   $rate := $res = "720p" ? 0.07 : 0.05;
                   $price := ($rate * $dur + 0.002 * $refs) * 1.43;
                   {"type":"usd","usd": $price}

comfy_api_nodes/nodes_hunyuan3d.py

@@ -132,7 +132,7 @@ class TencentTextToModelNode(IO.ComfyNode):
                     tooltip="The LowPoly option is unavailable for the `3.1` model.",
                 ),
                 IO.String.Input("prompt", multiline=True, default="", tooltip="Supports up to 1024 characters."),
-                IO.Int.Input("face_count", default=500000, min=40000, max=1500000),
+                IO.Int.Input("face_count", default=500000, min=3000, max=1500000),
                 IO.DynamicCombo.Input(
                     "generate_type",
                     options=[
@@ -251,7 +251,7 @@ class TencentImageToModelNode(IO.ComfyNode):
                 IO.Image.Input("image_left", optional=True),
                 IO.Image.Input("image_right", optional=True),
                 IO.Image.Input("image_back", optional=True),
-                IO.Int.Input("face_count", default=500000, min=40000, max=1500000),
+                IO.Int.Input("face_count", default=500000, min=3000, max=1500000),
                 IO.DynamicCombo.Input(
                     "generate_type",
                     options=[
@@ -422,6 +422,7 @@ class TencentModelTo3DUVNode(IO.ComfyNode):
             outputs=[
                 IO.File3DOBJ.Output(display_name="OBJ"),
                 IO.File3DFBX.Output(display_name="FBX"),
+                IO.Image.Output(display_name="uv_image"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -468,9 +469,16 @@ class TencentModelTo3DUVNode(IO.ComfyNode):
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
+        uv_image_file = get_file_from_response(result.ResultFile3Ds, "uv_image", raise_if_not_found=False)
+        uv_image = (
+            await download_url_to_image_tensor(uv_image_file.Url)
+            if uv_image_file is not None
+            else torch.zeros(1, 1, 1, 3)
+        )
         return IO.NodeOutput(
             await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj"),
             await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"),
+            uv_image,
         )
 
 

comfy_api_nodes/nodes_sonilo.py

@@ -0,0 +1,287 @@
+import base64
+import json
+import logging
+import time
+from urllib.parse import urljoin
+
+import aiohttp
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    audio_bytes_to_audio_input,
+    upload_video_to_comfyapi,
+    validate_string,
+)
+from comfy_api_nodes.util._helpers import (
+    default_base_url,
+    get_auth_header,
+    get_node_id,
+    is_processing_interrupted,
+)
+from comfy_api_nodes.util.common_exceptions import ProcessingInterrupted
+from server import PromptServer
+
+logger = logging.getLogger(__name__)
+
+
+class SoniloVideoToMusic(IO.ComfyNode):
+    """Generate music from video using Sonilo's AI model."""
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="SoniloVideoToMusic",
+            display_name="Sonilo Video to Music",
+            category="api node/audio/Sonilo",
+            description="Generate music from video content using Sonilo's AI model. "
+            "Analyzes the video and creates matching music.",
+            inputs=[
+                IO.Video.Input(
+                    "video",
+                    tooltip="Input video to generate music from. Maximum duration: 6 minutes.",
+                ),
+                IO.String.Input(
+                    "prompt",
+                    default="",
+                    multiline=True,
+                    tooltip="Optional text prompt to guide music generation. "
+                    "Leave empty for best quality - the model will fully analyze the video content.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=0xFFFFFFFFFFFFFFFF,
+                    control_after_generate=True,
+                    tooltip="Seed for reproducibility. Currently ignored by the Sonilo "
+                    "service but kept for graph consistency.",
+                ),
+            ],
+            outputs=[IO.Audio.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(
+                expr='{"type":"usd","usd":0.009,"format":{"suffix":"/second"}}',
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        video: Input.Video,
+        prompt: str = "",
+        seed: int = 0,
+    ) -> IO.NodeOutput:
+        video_url = await upload_video_to_comfyapi(cls, video, max_duration=360)
+        form = aiohttp.FormData()
+        form.add_field("video_url", video_url)
+        if prompt.strip():
+            form.add_field("prompt", prompt.strip())
+        audio_bytes = await _stream_sonilo_music(
+            cls,
+            ApiEndpoint(path="/proxy/sonilo/v2m/generate", method="POST"),
+            form,
+        )
+        return IO.NodeOutput(audio_bytes_to_audio_input(audio_bytes))
+
+
+class SoniloTextToMusic(IO.ComfyNode):
+    """Generate music from a text prompt using Sonilo's AI model."""
+
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="SoniloTextToMusic",
+            display_name="Sonilo Text to Music",
+            category="api node/audio/Sonilo",
+            description="Generate music from a text prompt using Sonilo's AI model. "
+            "Leave duration at 0 to let the model infer it from the prompt.",
+            inputs=[
+                IO.String.Input(
+                    "prompt",
+                    default="",
+                    multiline=True,
+                    tooltip="Text prompt describing the music to generate.",
+                ),
+                IO.Int.Input(
+                    "duration",
+                    default=0,
+                    min=0,
+                    max=360,
+                    tooltip="Target duration in seconds. Set to 0 to let the model "
+                    "infer the duration from the prompt. Maximum: 6 minutes.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=0xFFFFFFFFFFFFFFFF,
+                    control_after_generate=True,
+                    tooltip="Seed for reproducibility. Currently ignored by the Sonilo "
+                    "service but kept for graph consistency.",
+                ),
+            ],
+            outputs=[IO.Audio.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=["duration"]),
+                expr="""
+                (
+                  widgets.duration > 0
+                    ? {"type":"usd","usd": 0.005 * widgets.duration}
+                    : {"type":"usd","usd": 0.005, "format":{"suffix":"/second"}}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        prompt: str,
+        duration: int = 0,
+        seed: int = 0,
+    ) -> IO.NodeOutput:
+        validate_string(prompt, strip_whitespace=True, min_length=1)
+        form = aiohttp.FormData()
+        form.add_field("prompt", prompt)
+        if duration > 0:
+            form.add_field("duration", str(duration))
+        audio_bytes = await _stream_sonilo_music(
+            cls,
+            ApiEndpoint(path="/proxy/sonilo/t2m/generate", method="POST"),
+            form,
+        )
+        return IO.NodeOutput(audio_bytes_to_audio_input(audio_bytes))
+
+
+async def _stream_sonilo_music(
+    cls: type[IO.ComfyNode],
+    endpoint: ApiEndpoint,
+    form: aiohttp.FormData,
+) -> bytes:
+    """POST ``form`` to Sonilo, read the NDJSON stream, and return the first stream's audio bytes."""
+    url = urljoin(default_base_url().rstrip("/") + "/", endpoint.path.lstrip("/"))
+
+    headers: dict[str, str] = {}
+    headers.update(get_auth_header(cls))
+    headers.update(endpoint.headers)
+
+    node_id = get_node_id(cls)
+    start_ts = time.monotonic()
+    last_chunk_status_ts = 0.0
+    audio_streams: dict[int, list[bytes]] = {}
+    title: str | None = None
+
+    timeout = aiohttp.ClientTimeout(total=1200.0, sock_read=300.0)
+    async with aiohttp.ClientSession(timeout=timeout) as session:
+        PromptServer.instance.send_progress_text("Status: Queued", node_id)
+        async with session.post(url, data=form, headers=headers) as resp:
+            if resp.status >= 400:
+                msg = await _extract_error_message(resp)
+                raise Exception(f"Sonilo API error ({resp.status}): {msg}")
+
+            while True:
+                if is_processing_interrupted():
+                    raise ProcessingInterrupted("Task cancelled")
+
+                raw_line = await resp.content.readline()
+                if not raw_line:
+                    break
+
+                line = raw_line.decode("utf-8").strip()
+                if not line:
+                    continue
+
+                try:
+                    evt = json.loads(line)
+                except json.JSONDecodeError:
+                    logger.warning("Sonilo: skipping malformed NDJSON line")
+                    continue
+
+                evt_type = evt.get("type")
+                if evt_type == "error":
+                    code = evt.get("code", "UNKNOWN")
+                    message = evt.get("message", "Unknown error")
+                    raise Exception(f"Sonilo generation error ({code}): {message}")
+                if evt_type == "duration":
+                    duration_sec = evt.get("duration_sec")
+                    if duration_sec is not None:
+                        PromptServer.instance.send_progress_text(
+                            f"Status: Generating\nVideo duration: {duration_sec:.1f}s",
+                            node_id,
+                        )
+                elif evt_type in ("titles", "title"):
+                    # v2m sends a "titles" list, t2m sends a scalar "title"
+                    if evt_type == "titles":
+                        titles = evt.get("titles", [])
+                        if titles:
+                            title = titles[0]
+                    else:
+                        title = evt.get("title") or title
+                    if title:
+                        PromptServer.instance.send_progress_text(
+                            f"Status: Generating\nTitle: {title}",
+                            node_id,
+                        )
+                elif evt_type == "audio_chunk":
+                    stream_idx = evt.get("stream_index", 0)
+                    chunk_data = base64.b64decode(evt["data"])
+
+                    if stream_idx not in audio_streams:
+                        audio_streams[stream_idx] = []
+                    audio_streams[stream_idx].append(chunk_data)
+
+                    now = time.monotonic()
+                    if now - last_chunk_status_ts >= 1.0:
+                        total_chunks = sum(len(chunks) for chunks in audio_streams.values())
+                        elapsed = int(now - start_ts)
+                        status_lines = ["Status: Receiving audio"]
+                        if title:
+                            status_lines.append(f"Title: {title}")
+                        status_lines.append(f"Chunks received: {total_chunks}")
+                        status_lines.append(f"Time elapsed: {elapsed}s")
+                        PromptServer.instance.send_progress_text("\n".join(status_lines), node_id)
+                        last_chunk_status_ts = now
+                elif evt_type == "complete":
+                    break
+
+    if not audio_streams:
+        raise Exception("Sonilo API returned no audio data.")
+
+    PromptServer.instance.send_progress_text("Status: Completed", node_id)
+    selected_stream = 0 if 0 in audio_streams else min(audio_streams)
+    return b"".join(audio_streams[selected_stream])
+
+
+async def _extract_error_message(resp: aiohttp.ClientResponse) -> str:
+    """Extract a human-readable error message from an HTTP error response."""
+    try:
+        error_body = await resp.json()
+        detail = error_body.get("detail", {})
+        if isinstance(detail, dict):
+            return detail.get("message", str(detail))
+        return str(detail)
+    except Exception:
+        return await resp.text()
+
+
+class SoniloExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [SoniloVideoToMusic, SoniloTextToMusic]
+
+
+async def comfy_entrypoint() -> SoniloExtension:
+    return SoniloExtension()

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,
@@ -617,6 +619,7 @@ 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["X-Comfy-Env"] = get_deploy_environment()
         if cfg.endpoint.headers:
             payload_headers.update(cfg.endpoint.headers)
 

comfy_extras/nodes_ace.py

@@ -80,7 +80,7 @@ class EmptyAceStepLatentAudio(io.ComfyNode):
     @classmethod
     def execute(cls, seconds, batch_size) -> io.NodeOutput:
         length = int(seconds * 44100 / 512 / 8)
-        latent = torch.zeros([batch_size, 8, 16, length], device=comfy.model_management.intermediate_device())
+        latent = torch.zeros([batch_size, 8, 16, length], device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype())
         return io.NodeOutput({"samples": latent, "type": "audio"})
 
 
@@ -103,7 +103,7 @@ class EmptyAceStep15LatentAudio(io.ComfyNode):
     @classmethod
     def execute(cls, seconds, batch_size) -> io.NodeOutput:
         length = round((seconds * 48000 / 1920))
-        latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device())
+        latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device(), dtype=comfy.model_management.intermediate_dtype())
         return io.NodeOutput({"samples": latent, "type": "audio"})
 
 class ReferenceAudio(io.ComfyNode):

comfy_extras/nodes_curve.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 
+import numpy as np
+
 from comfy_api.latest import ComfyExtension, io
 from comfy_api.input import CurveInput
 from typing_extensions import override
@@ -32,10 +34,58 @@ class CurveEditor(io.ComfyNode):
         return io.NodeOutput(result, ui=ui) if ui else io.NodeOutput(result)
 
 
+class ImageHistogram(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="ImageHistogram",
+            display_name="Image Histogram",
+            category="utils",
+            inputs=[
+                io.Image.Input("image"),
+            ],
+            outputs=[
+                io.Histogram.Output("rgb"),
+                io.Histogram.Output("luminance"),
+                io.Histogram.Output("red"),
+                io.Histogram.Output("green"),
+                io.Histogram.Output("blue"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, image) -> io.NodeOutput:
+        img = image[0].cpu().numpy()
+        img_uint8 = np.clip(img * 255, 0, 255).astype(np.uint8)
+
+        def bincount(data):
+            return np.bincount(data.ravel(), minlength=256)[:256]
+
+        hist_r = bincount(img_uint8[:, :, 0])
+        hist_g = bincount(img_uint8[:, :, 1])
+        hist_b = bincount(img_uint8[:, :, 2])
+
+        # Average of R, G, B histograms (same as Photoshop's RGB composite)
+        rgb = ((hist_r + hist_g + hist_b) // 3).tolist()
+
+        # ITU-R BT.709-6, Item 3.2 (p.6) — Derivation of luminance signal
+        # https://www.itu.int/rec/R-REC-BT.709-6-201506-I/en
+        lum = 0.2126 * img[:, :, 0] + 0.7152 * img[:, :, 1] + 0.0722 * img[:, :, 2]
+        luminance = bincount(np.clip(lum * 255, 0, 255).astype(np.uint8)).tolist()
+
+        return io.NodeOutput(
+            rgb,
+            luminance,
+            hist_r.tolist(),
+            hist_g.tolist(),
+            hist_b.tolist(),
+        )
+
+
 class CurveExtension(ComfyExtension):
     @override
     async def get_node_list(self):
-        return [CurveEditor]
+        return [CurveEditor, ImageHistogram]
 
 
 async def comfy_entrypoint():

comfy_extras/nodes_preview_any.py

@@ -11,7 +11,7 @@ class PreviewAny():
             "required": {"source": (IO.ANY, {})},
         }
 
-    RETURN_TYPES = ()
+    RETURN_TYPES = (IO.STRING,)
     FUNCTION = "main"
     OUTPUT_NODE = True
 
@@ -33,7 +33,7 @@ class PreviewAny():
                 except Exception:
                     value = 'source exists, but could not be serialized.'
 
-        return {"ui": {"text": (value,)}}
+        return {"ui": {"text": (value,)}, "result": (value,)}
 
 NODE_CLASS_MAPPINGS = {
     "PreviewAny": PreviewAny,

comfy_extras/nodes_rtdetr.py

@@ -0,0 +1,156 @@
+from typing_extensions import override
+
+import torch
+from comfy.ldm.rt_detr.rtdetr_v4 import COCO_CLASSES
+import comfy.model_management
+import comfy.utils
+from comfy_api.latest import ComfyExtension, io
+from torchvision.transforms import ToPILImage, ToTensor
+from PIL import ImageDraw, ImageFont
+
+
+class RTDETR_detect(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="RTDETR_detect",
+            display_name="RT-DETR Detect",
+            category="detection/",
+            search_aliases=["bbox", "bounding box", "object detection", "coco"],
+            inputs=[
+                io.Model.Input("model", display_name="model"),
+                io.Image.Input("image", display_name="image"),
+                io.Float.Input("threshold", display_name="threshold", default=0.5),
+                io.Combo.Input("class_name", options=["all"] + COCO_CLASSES, default="all", tooltip="Filter detections by class. Set to 'all' to disable filtering."),
+                io.Int.Input("max_detections", display_name="max_detections", default=100, tooltip="Maximum number of detections to return per image. In order of descending confidence score."),
+            ],
+            outputs=[
+                io.BoundingBox.Output("bboxes")],
+        )
+
+    @classmethod
+    def execute(cls, model, image, threshold, class_name, max_detections) -> io.NodeOutput:
+        B, H, W, C = image.shape
+
+        comfy.model_management.load_model_gpu(model)
+        results = []
+        for i in range(0, B, 32):
+            batch = image[i:i + 32]
+            image_in = comfy.utils.common_upscale(batch.movedim(-1, 1), 640, 640, "bilinear", crop="disabled")
+            results.extend(model.model.diffusion_model(image_in, (W, H)))
+
+        all_bbox_dicts = []
+
+        for det in results:
+            keep   = det['scores'] > threshold
+            boxes  = det['boxes'][keep].cpu()
+            labels = det['labels'][keep].cpu()
+            scores = det['scores'][keep].cpu()
+
+            bbox_dicts = [
+                {
+                    "x": float(box[0]),
+                    "y": float(box[1]),
+                    "width": float(box[2] - box[0]),
+                    "height": float(box[3] - box[1]),
+                    "label": COCO_CLASSES[int(label)],
+                    "score": float(score)
+                }
+                for box, label, score in zip(boxes, labels, scores)
+                if class_name == "all" or COCO_CLASSES[int(label)] == class_name
+            ]
+            bbox_dicts.sort(key=lambda d: d["score"], reverse=True)
+            all_bbox_dicts.append(bbox_dicts[:max_detections])
+
+        return io.NodeOutput(all_bbox_dicts)
+
+
+class DrawBBoxes(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="DrawBBoxes",
+            display_name="Draw BBoxes",
+            category="detection/",
+            search_aliases=["bbox", "bounding box", "object detection", "rt_detr", "visualize detections", "coco"],
+            inputs=[
+                io.Image.Input("image", optional=True),
+                io.BoundingBox.Input("bboxes", force_input=True),
+            ],
+            outputs=[
+                io.Image.Output("out_image"),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, bboxes, image=None) -> io.NodeOutput:
+        # Normalise to list[list[dict]], then fit to batch size B.
+        B = image.shape[0] if image is not None else 1
+        if isinstance(bboxes, dict):
+            bboxes = [[bboxes]]
+        elif not isinstance(bboxes, list) or not bboxes:
+            bboxes = [[]]
+        elif isinstance(bboxes[0], dict):
+            bboxes = [bboxes]  # flat list → same detections for every image
+
+        if len(bboxes) == 1:
+            bboxes = bboxes * B
+        bboxes = (bboxes + [[]] * B)[:B]
+
+        if image is None:
+            B = len(bboxes)
+            max_w = max((int(d["x"] + d["width"])  for frame in bboxes for d in frame), default=640)
+            max_h = max((int(d["y"] + d["height"]) for frame in bboxes for d in frame), default=640)
+            image = torch.zeros((B, max_h, max_w, 3), dtype=torch.float32)
+
+        all_out_images = []
+        for i in range(B):
+            detections = bboxes[i]
+            if detections:
+                boxes  = torch.tensor([[d["x"], d["y"], d["x"] + d["width"], d["y"] + d["height"]] for d in detections])
+                labels = [d.get("label") if d.get("label") in COCO_CLASSES else None for d in detections]
+                scores = torch.tensor([d.get("score", 1.0) for d in detections])
+            else:
+                boxes  = torch.zeros((0, 4))
+                labels = []
+                scores = torch.zeros((0,))
+
+            pil_image = image[i].movedim(-1, 0)
+            img = ToPILImage()(pil_image)
+            if detections:
+                img = cls.draw_detections(img, boxes, labels, scores)
+            all_out_images.append(ToTensor()(img).unsqueeze(0).movedim(1, -1))
+
+        out_images = torch.cat(all_out_images, dim=0).to(comfy.model_management.intermediate_device())
+        return io.NodeOutput(out_images)
+
+    @classmethod
+    def draw_detections(cls, img, boxes, labels, scores):
+        draw = ImageDraw.Draw(img)
+        try:
+            font = ImageFont.truetype('arial.ttf', 16)
+        except Exception:
+            font = ImageFont.load_default()
+        colors = [(255,0,0),(0,200,0),(0,0,255),(255,165,0),(128,0,128),
+                (0,255,255),(255,20,147),(100,149,237)]
+        for box, label, score in sorted(zip(boxes, labels, scores), key=lambda x: x[2].item()):
+            x1, y1, x2, y2 = box.tolist()
+            color_idx = COCO_CLASSES.index(label) if label is not None else 0
+            c = colors[color_idx % len(colors)]
+            draw.rectangle([x1, y1, x2, y2], outline=c, width=3)
+            if label is not None:
+                draw.text((x1 + 2, y1 + 2), f'{label} {score:.2f}', fill=c, font=font)
+        return img
+
+
+class RTDETRExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [
+            RTDETR_detect,
+            DrawBBoxes,
+        ]
+
+
+async def comfy_entrypoint() -> RTDETRExtension:
+    return RTDETRExtension()

comfy_extras/nodes_sdpose.py

@@ -1,5 +1,6 @@
 import torch
 import comfy.utils
+import comfy.model_management
 import numpy as np
 import math
 import colorsys
@@ -410,7 +411,9 @@ class SDPoseDrawKeypoints(io.ComfyNode):
             pose_outputs.append(canvas)
 
         pose_outputs_np = np.stack(pose_outputs) if len(pose_outputs) > 1 else np.expand_dims(pose_outputs[0], 0)
-        final_pose_output = torch.from_numpy(pose_outputs_np).float() / 255.0
+        final_pose_output = torch.from_numpy(pose_outputs_np).to(
+            device=comfy.model_management.intermediate_device(),
+            dtype=comfy.model_management.intermediate_dtype()) / 255.0
         return io.NodeOutput(final_pose_output)
 
 class SDPoseKeypointExtractor(io.ComfyNode):
@@ -459,6 +462,27 @@ class SDPoseKeypointExtractor(io.ComfyNode):
         model_h = int(head.heatmap_size[0]) * 4   # e.g. 192 * 4 = 768
         model_w = int(head.heatmap_size[1]) * 4   # e.g. 256 * 4 = 1024
 
+        def _resize_to_model(imgs):
+            """Aspect-preserving resize + zero-pad BHWC images to (model_h, model_w). Returns (resized_bhwc, scale, pad_top, pad_left)."""
+            h, w = imgs.shape[-3], imgs.shape[-2]
+            scale = min(model_h / h, model_w / w)
+            sh, sw = int(round(h * scale)), int(round(w * scale))
+            pt, pl = (model_h - sh) // 2, (model_w - sw) // 2
+            chw = imgs.permute(0, 3, 1, 2).float()
+            scaled = comfy.utils.common_upscale(chw, sw, sh, upscale_method="bilinear", crop="disabled")
+            padded = torch.zeros(scaled.shape[0], scaled.shape[1], model_h, model_w, dtype=scaled.dtype, device=scaled.device)
+            padded[:, :, pt:pt + sh, pl:pl + sw] = scaled
+            return padded.permute(0, 2, 3, 1), scale, pt, pl
+
+        def _remap_keypoints(kp, scale, pad_top, pad_left, offset_x=0, offset_y=0):
+            """Remap keypoints from model space back to original image space."""
+            kp = kp.copy() if isinstance(kp, np.ndarray) else np.array(kp, dtype=np.float32)
+            invalid = kp[..., 0] < 0
+            kp[..., 0] = (kp[..., 0] - pad_left) / scale + offset_x
+            kp[..., 1] = (kp[..., 1] - pad_top)  / scale + offset_y
+            kp[invalid] = -1
+            return kp
+
         def _run_on_latent(latent_batch):
             """Run one forward pass and return (keypoints_list, scores_list) for the batch."""
             nonlocal captured_feat
@@ -504,36 +528,19 @@ class SDPoseKeypointExtractor(io.ComfyNode):
                         if x2 <= x1 or y2 <= y1:
                             continue
 
-                        crop_h_px, crop_w_px = y2 - y1, x2 - x1
                         crop = img[:, y1:y2, x1:x2, :]  # (1, crop_h, crop_w, C)
-
-                        # scale to fit inside (model_h, model_w) while preserving aspect ratio, then pad to exact model size.
-                        scale = min(model_h / crop_h_px, model_w / crop_w_px)
-                        scaled_h, scaled_w = int(round(crop_h_px * scale)), int(round(crop_w_px * scale))
-                        pad_top, pad_left  = (model_h - scaled_h) // 2, (model_w - scaled_w) // 2
-
-                        crop_chw = crop.permute(0, 3, 1, 2).float()  # BHWC → BCHW
-                        scaled = comfy.utils.common_upscale(crop_chw, scaled_w, scaled_h, upscale_method="bilinear", crop="disabled")
-                        padded = torch.zeros(1, scaled.shape[1], model_h, model_w, dtype=scaled.dtype, device=scaled.device)
-                        padded[:, :, pad_top:pad_top + scaled_h, pad_left:pad_left + scaled_w] = scaled
-                        crop_resized = padded.permute(0, 2, 3, 1)  # BCHW → BHWC
+                        crop_resized, scale, pad_top, pad_left = _resize_to_model(crop)
 
                         latent_crop = vae.encode(crop_resized)
                         kp_batch, sc_batch = _run_on_latent(latent_crop)
-                        kp, sc = kp_batch[0], sc_batch[0]  # (K, 2), coords in model pixel space
-
-                        # remove padding offset, undo scale, offset to full-image coordinates.
-                        kp = kp.copy() if isinstance(kp, np.ndarray) else np.array(kp, dtype=np.float32)
-                        kp[..., 0] = (kp[..., 0] - pad_left) / scale + x1
-                        kp[..., 1] = (kp[..., 1] - pad_top)  / scale + y1
-
+                        kp = _remap_keypoints(kp_batch[0], scale, pad_top, pad_left, x1, y1)
                         img_keypoints.append(kp)
-                        img_scores.append(sc)
+                        img_scores.append(sc_batch[0])
                 else:
-                    # No bboxes for this image – run on the full image
-                    latent_img = vae.encode(img)
+                    img_resized, scale, pad_top, pad_left = _resize_to_model(img)
+                    latent_img = vae.encode(img_resized)
                     kp_batch, sc_batch = _run_on_latent(latent_img)
-                    img_keypoints.append(kp_batch[0])
+                    img_keypoints.append(_remap_keypoints(kp_batch[0], scale, pad_top, pad_left))
                     img_scores.append(sc_batch[0])
 
                 all_keypoints.append(img_keypoints)
@@ -541,19 +548,16 @@ class SDPoseKeypointExtractor(io.ComfyNode):
                 pbar.update(1)
 
         else: # full-image mode, batched
-            tqdm_pbar = tqdm(total=total_images, desc="Extracting keypoints")
-            for batch_start in range(0, total_images, batch_size):
-                batch_end = min(batch_start + batch_size, total_images)
-                latent_batch = vae.encode(image[batch_start:batch_end])
-
+            for batch_start in tqdm(range(0, total_images, batch_size), desc="Extracting keypoints"):
+                batch_resized, scale, pad_top, pad_left = _resize_to_model(image[batch_start:batch_start + batch_size])
+                latent_batch = vae.encode(batch_resized)
                 kp_batch, sc_batch = _run_on_latent(latent_batch)
 
                 for kp, sc in zip(kp_batch, sc_batch):
-                    all_keypoints.append([kp])
+                    all_keypoints.append([_remap_keypoints(kp, scale, pad_top, pad_left)])
                     all_scores.append([sc])
-                    tqdm_pbar.update(1)
 
-                pbar.update(batch_end - batch_start)
+                pbar.update(len(kp_batch))
 
         openpose_frames = _to_openpose_frames(all_keypoints, all_scores, height, width)
         return io.NodeOutput(openpose_frames)
@@ -661,6 +665,7 @@ class CropByBBoxes(io.ComfyNode):
                 io.Int.Input("output_width",  default=512, min=64, max=4096, step=8, tooltip="Width each crop is resized to."),
                 io.Int.Input("output_height", default=512, min=64, max=4096, step=8, tooltip="Height each crop is resized to."),
                 io.Int.Input("padding", default=0, min=0, max=1024, step=1, tooltip="Extra padding in pixels added on each side of the bbox before cropping."),
+                io.Combo.Input("keep_aspect", options=["stretch", "pad"], default="stretch", tooltip="Whether to stretch the crop to fit the output size, or pad with black pixels to preserve aspect ratio."),
             ],
             outputs=[
                 io.Image.Output(tooltip="All crops stacked into a single image batch."),
@@ -668,7 +673,7 @@ class CropByBBoxes(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, image, bboxes, output_width, output_height, padding) -> io.NodeOutput:
+    def execute(cls, image, bboxes, output_width, output_height, padding, keep_aspect="stretch") -> io.NodeOutput:
         total_frames = image.shape[0]
         img_h = image.shape[1]
         img_w = image.shape[2]
@@ -716,7 +721,19 @@ class CropByBBoxes(io.ComfyNode):
                 x1, y1, x2, y2 = fb_x1, fb_y1, fb_x2, fb_y2
 
             crop_chw = frame_chw[:, :, y1:y2, x1:x2]  # (1, C, crop_h, crop_w)
-            resized = comfy.utils.common_upscale(crop_chw, output_width, output_height, upscale_method="bilinear", crop="disabled")
+
+            if keep_aspect == "pad":
+                crop_h, crop_w = y2 - y1, x2 - x1
+                scale = min(output_width / crop_w, output_height / crop_h)
+                scaled_w = int(round(crop_w * scale))
+                scaled_h = int(round(crop_h * scale))
+                scaled = comfy.utils.common_upscale(crop_chw, scaled_w, scaled_h, upscale_method="bilinear", crop="disabled")
+                pad_left = (output_width  - scaled_w) // 2
+                pad_top  = (output_height - scaled_h) // 2
+                resized = torch.zeros(1, num_ch, output_height, output_width, dtype=image.dtype, device=image.device)
+                resized[:, :, pad_top:pad_top + scaled_h, pad_left:pad_left + scaled_w] = scaled
+            else:  # "stretch"
+                resized = comfy.utils.common_upscale(crop_chw, output_width, output_height, upscale_method="bilinear", crop="disabled")
             crops.append(resized)
 
         if not crops:

comfy_extras/nodes_string.py

@@ -9,9 +9,9 @@ class StringConcatenate(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringConcatenate",
-            display_name="Concatenate",
+            display_name="Text Concatenate",
             category="utils/string",
-            search_aliases=["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", "concat", "concatenate", "append text", "combine text", "string"],
             inputs=[
                 io.String.Input("string_a", multiline=True),
                 io.String.Input("string_b", multiline=True),
@@ -32,8 +32,8 @@ class StringSubstring(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringSubstring",
-            search_aliases=["extract text", "text portion"],
-            display_name="Substring",
+            search_aliases=["Substring", "extract text", "text portion"],
+            display_name="Text Substring",
             category="utils/string",
             inputs=[
                 io.String.Input("string", multiline=True),
@@ -55,8 +55,8 @@ class StringLength(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringLength",
-            search_aliases=["character count", "text size"],
-            display_name="Length",
+            search_aliases=["character count", "text size", "string length"],
+            display_name="Text Length",
             category="utils/string",
             inputs=[
                 io.String.Input("string", multiline=True),
@@ -76,8 +76,8 @@ class CaseConverter(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="CaseConverter",
-            search_aliases=["text case", "uppercase", "lowercase", "capitalize"],
-            display_name="Case Converter",
+            search_aliases=["Case Converter", "text case", "uppercase", "lowercase", "capitalize"],
+            display_name="Text Case Converter",
             category="utils/string",
             inputs=[
                 io.String.Input("string", multiline=True),
@@ -109,8 +109,8 @@ class StringTrim(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringTrim",
-            search_aliases=["clean whitespace", "remove whitespace"],
-            display_name="Trim",
+            search_aliases=["Trim", "clean whitespace", "remove whitespace", "strip"],
+            display_name="Text Trim",
             category="utils/string",
             inputs=[
                 io.String.Input("string", multiline=True),
@@ -140,8 +140,8 @@ class StringReplace(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringReplace",
-            search_aliases=["find and replace", "substitute", "swap text"],
-            display_name="Replace",
+            search_aliases=["Replace", "find and replace", "substitute", "swap text"],
+            display_name="Text Replace",
             category="utils/string",
             inputs=[
                 io.String.Input("string", multiline=True),
@@ -163,8 +163,8 @@ class StringContains(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringContains",
-            search_aliases=["text includes", "string includes"],
-            display_name="Contains",
+            search_aliases=["Contains", "text includes", "string includes"],
+            display_name="Text Contains",
             category="utils/string",
             inputs=[
                 io.String.Input("string", multiline=True),
@@ -191,8 +191,8 @@ class StringCompare(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="StringCompare",
-            search_aliases=["text match", "string equals", "starts with", "ends with"],
-            display_name="Compare",
+            search_aliases=["Compare", "text match", "string equals", "starts with", "ends with"],
+            display_name="Text Compare",
             category="utils/string",
             inputs=[
                 io.String.Input("string_a", multiline=True),
@@ -227,8 +227,8 @@ class RegexMatch(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="RegexMatch",
-            search_aliases=["pattern match", "text contains", "string match"],
-            display_name="Regex Match",
+            search_aliases=["Regex Match", "regex", "pattern match", "text contains", "string match"],
+            display_name="Text Match",
             category="utils/string",
             inputs=[
                 io.String.Input("string", multiline=True),
@@ -268,8 +268,8 @@ class RegexExtract(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="RegexExtract",
-            search_aliases=["pattern extract", "text parser", "parse text"],
-            display_name="Regex Extract",
+            search_aliases=["Regex Extract", "regex", "pattern extract", "text parser", "parse text"],
+            display_name="Text Extract Substring",
             category="utils/string",
             inputs=[
                 io.String.Input("string", multiline=True),
@@ -343,8 +343,8 @@ class RegexReplace(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="RegexReplace",
-            search_aliases=["pattern replace", "find and replace", "substitution"],
-            display_name="Regex Replace",
+            search_aliases=["Regex Replace", "regex", "pattern replace", "regex replace", "substitution"],
+            display_name="Text Replace (Regex)",
             category="utils/string",
             description="Find and replace text using regex patterns.",
             inputs=[

comfy_extras/nodes_textgen.py

@@ -35,6 +35,7 @@ class TextGenerate(io.ComfyNode):
                 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."),
+                io.Boolean.Input("use_default_template", optional=True, default=True, tooltip="Use the built in system prompt/template if the model has one.", advanced=True),
             ],
             outputs=[
                 io.String.Output(display_name="generated_text"),
@@ -42,9 +43,9 @@ class TextGenerate(io.ComfyNode):
         )
 
     @classmethod
-    def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False) -> io.NodeOutput:
+    def execute(cls, clip, prompt, max_length, sampling_mode, image=None, thinking=False, use_default_template=True) -> io.NodeOutput:
 
-        tokens = clip.tokenize(prompt, image=image, skip_template=False, min_length=1, thinking=thinking)
+        tokens = clip.tokenize(prompt, image=image, skip_template=not use_default_template, min_length=1, thinking=thinking)
 
         # Get sampling parameters from dynamic combo
         do_sample = sampling_mode.get("sampling_mode") == "on"

comfy_extras/nodes_upscale_model.py

@@ -6,6 +6,7 @@ import comfy.utils
 import folder_paths
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, io
+import comfy.model_management
 
 try:
     from spandrel_extra_arches import EXTRA_REGISTRY
@@ -78,13 +79,15 @@ class ImageUpscaleWithModel(io.ComfyNode):
         tile = 512
         overlap = 32
 
+        output_device = comfy.model_management.intermediate_device()
+
         oom = True
         try:
             while oom:
                 try:
                     steps = in_img.shape[0] * comfy.utils.get_tiled_scale_steps(in_img.shape[3], in_img.shape[2], tile_x=tile, tile_y=tile, overlap=overlap)
                     pbar = comfy.utils.ProgressBar(steps)
-                    s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar)
+                    s = comfy.utils.tiled_scale(in_img, lambda a: upscale_model(a.float()), tile_x=tile, tile_y=tile, overlap=overlap, upscale_amount=upscale_model.scale, pbar=pbar, output_device=output_device)
                     oom = False
                 except Exception as e:
                     model_management.raise_non_oom(e)
@@ -94,7 +97,7 @@ class ImageUpscaleWithModel(io.ComfyNode):
         finally:
             upscale_model.to("cpu")
 
-        s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0)
+        s = torch.clamp(s.movedim(-3,-1), min=0, max=1.0).to(comfy.model_management.intermediate_dtype())
         return io.NodeOutput(s)
 
     upscale = execute  # TODO: remove

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.18.1"
+__version__ = "0.19.1"

nodes.py

@@ -2457,6 +2457,7 @@ async def init_builtin_extra_nodes():
         "nodes_number_convert.py",
         "nodes_painter.py",
         "nodes_curve.py",
+        "nodes_rtdetr.py"
     ]
 
     import_failed = []

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.18.1"
+version = "0.19.1"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

requirements.txt

@@ -1,5 +1,5 @@
-comfyui-frontend-package==1.42.8
-comfyui-workflow-templates==0.9.39
+comfyui-frontend-package==1.42.11
+comfyui-workflow-templates==0.9.54
 comfyui-embedded-docs==0.4.3
 torch
 torchsde

server.py

@@ -146,6 +146,10 @@ def is_loopback(host):
 def create_origin_only_middleware():
     @web.middleware
     async def origin_only_middleware(request: web.Request, handler):
+        if 'Sec-Fetch-Site' in request.headers:
+            sec_fetch_site = request.headers['Sec-Fetch-Site']
+            if sec_fetch_site == 'cross-site':
+                return web.Response(status=403)
         #this code is used to prevent the case where a random website can queue comfy workflows by making a POST to 127.0.0.1 which browsers don't prevent for some dumb reason.
         #in that case the Host and Origin hostnames won't match
         #I know the proper fix would be to add a cookie but this should take care of the problem in the meantime