Change auto model selection to best match from first match CORE-60.

Co-authored-by: Talmaj Marinc <talmaj@comfy.org>

.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

.github/workflows/tag-dispatch-cloud.yml

@@ -0,0 +1,45 @@
+name: Tag Dispatch to Cloud
+
+on:
+  push:
+    tags:
+      - 'v*'
+
+jobs:
+  dispatch-cloud:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Send repository dispatch to cloud
+        env:
+          DISPATCH_TOKEN: ${{ secrets.CLOUD_REPO_DISPATCH_TOKEN }}
+          RELEASE_TAG: ${{ github.ref_name }}
+        run: |
+          set -euo pipefail
+
+          if [ -z "${DISPATCH_TOKEN:-}" ]; then
+            echo "::error::CLOUD_REPO_DISPATCH_TOKEN is required but not set."
+            exit 1
+          fi
+
+          RELEASE_URL="https://github.com/${{ github.repository }}/releases/tag/${RELEASE_TAG}"
+
+          PAYLOAD="$(jq -n \
+            --arg release_tag "$RELEASE_TAG" \
+            --arg release_url "$RELEASE_URL" \
+            '{
+              event_type: "comfyui_tag_pushed",
+              client_payload: {
+                release_tag: $release_tag,
+                release_url: $release_url
+              }
+            }')"
+
+          curl -fsSL \
+            -X POST \
+            -H "Accept: application/vnd.github+json" \
+            -H "Content-Type: application/json" \
+            -H "Authorization: Bearer ${DISPATCH_TOKEN}" \
+            https://api.github.com/repos/Comfy-Org/cloud/dispatches \
+            -d "$PAYLOAD"
+
+          echo "✅ Dispatched ComfyUI tag ${RELEASE_TAG} to Comfy-Org/cloud"

.gitignore

@@ -21,6 +21,5 @@ venv*/
 *.log
 web_custom_versions/
 .DS_Store
-openapi.yaml
 filtered-openapi.yaml
 uv.lock

CODEOWNERS

@@ -1,2 +1,2 @@
 # Admins
-* @comfyanonymous @kosinkadink @guill
+* @comfyanonymous @kosinkadink @guill @alexisrolland @rattus128

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
 
@@ -194,7 +195,9 @@ The portable above currently comes with python 3.13 and pytorch cuda 13.0. Updat
 
 #### Alternative Downloads:
 
-[Experimental portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
+[Portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
+
+[Experimental portable for Intel GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z)
 
 [Portable with pytorch cuda 12.6 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z) (Supports Nvidia 10 series and older GPUs).
 
@@ -275,7 +278,7 @@ 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
 

api_server/routes/internal/internal_routes.py

@@ -67,7 +67,7 @@ class InternalRoutes:
                 (entry for entry in os.scandir(directory) if is_visible_file(entry)),
                 key=lambda entry: -entry.stat().st_mtime
             )
-            return web.json_response([entry.name for entry in sorted_files], status=200)
+            return web.json_response([f"{entry.name} [{directory_type}]" for entry in sorted_files], status=200)
 
 
     def get_app(self):

blueprints/.glsl/Glow_30.frag

@@ -2,7 +2,6 @@
 precision mediump float;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform int u_int0;      // Blend mode
 uniform int u_int1;      // Color tint
 uniform float u_float0;  // Intensity
@@ -75,7 +74,7 @@ void main() {
     float t0 = threshold - 0.15;
     float t1 = threshold + 0.15;
     
-    vec2 texelSize = 1.0 / u_resolution;
+    vec2 texelSize = 1.0 / vec2(textureSize(u_image0, 0));
     float radius2 = radius * radius;
     
     float sampleScale = clamp(radius * 0.75, 0.35, 1.0);

blueprints/.glsl/Image_Blur_1.frag

@@ -12,7 +12,6 @@ const int RADIAL_SAMPLES = 12;
 const float RADIAL_STRENGTH = 0.0003;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform int u_int0;      // Blur type (BLUR_GAUSSIAN, BLUR_BOX, BLUR_RADIAL)
 uniform float u_float0;  // Blur radius/amount
 uniform int u_pass;      // Pass index (0 = horizontal, 1 = vertical)
@@ -25,7 +24,7 @@ float gaussian(float x, float sigma) {
 }
 
 void main() {
-    vec2 texelSize = 1.0 / u_resolution;
+    vec2 texelSize = 1.0 / vec2(textureSize(u_image0, 0));
     float radius = max(u_float0, 0.0);
 
     // Radial (angular) blur - single pass, doesn't use separable

blueprints/.glsl/Sharpen_23.frag

@@ -2,14 +2,13 @@
 precision highp float;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0
 
 in vec2 v_texCoord;
 layout(location = 0) out vec4 fragColor0;
 
 void main() {
-    vec2 texel = 1.0 / u_resolution;
+    vec2 texel = 1.0 / vec2(textureSize(u_image0, 0));
     
     // Sample center and neighbors
     vec4 center = texture(u_image0, v_texCoord);

blueprints/.glsl/Unsharp_Mask_26.frag

@@ -2,7 +2,6 @@
 precision highp float;
 
 uniform sampler2D u_image0;
-uniform vec2 u_resolution;
 uniform float u_float0;  // amount    [0.0 - 3.0]  typical: 0.5-1.5
 uniform float u_float1;  // radius    [0.5 - 10.0] blur radius in pixels
 uniform float u_float2;  // threshold [0.0 - 0.1]  min difference to sharpen
@@ -19,7 +18,7 @@ float getLuminance(vec3 color) {
 }
 
 void main() {
-    vec2 texel = 1.0 / u_resolution;
+    vec2 texel = 1.0 / vec2(textureSize(u_image0, 0));
     float radius = max(u_float1, 0.5);
     float amount = u_float0;
     float threshold = u_float2;

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 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 / vec2(textureSize(u_image0, 0));\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/latent_formats.py

@@ -224,6 +224,7 @@ class Flux2(LatentFormat):
 
         self.latent_rgb_factors_bias = [-0.0329, -0.0718, -0.0851]
         self.latent_rgb_factors_reshape = lambda t: t.reshape(t.shape[0], 32, 2, 2, t.shape[-2], t.shape[-1]).permute(0, 1, 4, 2, 5, 3).reshape(t.shape[0], 32, t.shape[-2] * 2, t.shape[-1] * 2)
+        self.taesd_decoder_name = "taef2_decoder"
 
     def process_in(self, latent):
         return latent
@@ -783,3 +784,10 @@ class ZImagePixelSpace(ChromaRadiance):
     No VAE encoding/decoding — the model operates directly on RGB pixels.
     """
     pass
+
+class CogVideoX(LatentFormat):
+    latent_channels = 16
+    latent_dimensions = 3
+
+    def __init__(self):
+        self.scale_factor = 1.15258426

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/cogvideo/model.py

@@ -0,0 +1,573 @@
+# CogVideoX 3D Transformer - ported to ComfyUI native ops
+# Architecture reference: diffusers CogVideoXTransformer3DModel
+# Style reference: comfy/ldm/wan/model.py
+
+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.patcher_extension
+import comfy.ldm.common_dit
+
+
+def _get_1d_rotary_pos_embed(dim, pos, theta=10000.0):
+    """Returns (cos, sin) each with shape [seq_len, dim].
+
+    Frequencies are computed at dim//2 resolution then repeat_interleaved
+    to full dim, matching CogVideoX's interleaved (real, imag) pair format.
+    """
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float32, device=pos.device) / dim))
+    angles = torch.outer(pos.float(), freqs.float())
+    cos = angles.cos().repeat_interleave(2, dim=-1).float()
+    sin = angles.sin().repeat_interleave(2, dim=-1).float()
+    return (cos, sin)
+
+
+def apply_rotary_emb(x, freqs_cos_sin):
+    """Apply CogVideoX rotary embedding to query or key tensor.
+
+    x: [B, heads, seq_len, head_dim]
+    freqs_cos_sin: (cos, sin) each [seq_len, head_dim//2]
+
+    Uses interleaved pair rotation (same as diffusers CogVideoX/Flux).
+    head_dim is reshaped to (-1, 2) pairs, rotated, then flattened back.
+    """
+    cos, sin = freqs_cos_sin
+    cos = cos[None, None, :, :].to(x.device)
+    sin = sin[None, None, :, :].to(x.device)
+
+    # Interleaved pairs: [B, H, S, D] -> [B, H, S, D//2, 2] -> (real, imag)
+    x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)
+    x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
+
+    return (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
+
+
+def get_timestep_embedding(timesteps, dim, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1, max_period=10000):
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32, device=timesteps.device) / half)
+    args = timesteps[:, None].float() * freqs[None] * scale
+    embedding = torch.cat([torch.sin(args), torch.cos(args)], dim=-1)
+    if flip_sin_to_cos:
+        embedding = torch.cat([embedding[:, half:], embedding[:, :half]], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+def get_3d_sincos_pos_embed(embed_dim, spatial_size, temporal_size, spatial_interpolation_scale=1.0, temporal_interpolation_scale=1.0, device=None):
+    if isinstance(spatial_size, int):
+        spatial_size = (spatial_size, spatial_size)
+
+    grid_w = torch.arange(spatial_size[0], dtype=torch.float32, device=device) / spatial_interpolation_scale
+    grid_h = torch.arange(spatial_size[1], dtype=torch.float32, device=device) / spatial_interpolation_scale
+    grid_t = torch.arange(temporal_size, dtype=torch.float32, device=device) / temporal_interpolation_scale
+
+    grid_t, grid_h, grid_w = torch.meshgrid(grid_t, grid_h, grid_w, indexing="ij")
+
+    embed_dim_spatial = 2 * (embed_dim // 3)
+    embed_dim_temporal = embed_dim // 3
+
+    pos_embed_spatial = _get_2d_sincos_pos_embed(embed_dim_spatial, grid_h, grid_w, device=device)
+    pos_embed_temporal = _get_1d_sincos_pos_embed(embed_dim_temporal, grid_t[:, 0, 0], device=device)
+
+    T, H, W = grid_t.shape
+    pos_embed_temporal = pos_embed_temporal.unsqueeze(1).unsqueeze(1).expand(-1, H, W, -1)
+    pos_embed = torch.cat([pos_embed_temporal, pos_embed_spatial], dim=-1)
+
+    return pos_embed
+
+
+def _get_2d_sincos_pos_embed(embed_dim, grid_h, grid_w, device=None):
+    T, H, W = grid_h.shape
+    half_dim = embed_dim // 2
+    pos_h = _get_1d_sincos_pos_embed(half_dim, grid_h.reshape(-1), device=device).reshape(T, H, W, half_dim)
+    pos_w = _get_1d_sincos_pos_embed(half_dim, grid_w.reshape(-1), device=device).reshape(T, H, W, half_dim)
+    return torch.cat([pos_h, pos_w], dim=-1)
+
+
+def _get_1d_sincos_pos_embed(embed_dim, pos, device=None):
+    half = embed_dim // 2
+    freqs = torch.exp(-math.log(10000.0) * torch.arange(start=0, end=half, dtype=torch.float32, device=device) / half)
+    args = pos.float().reshape(-1)[:, None] * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if embed_dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+
+class CogVideoXPatchEmbed(nn.Module):
+    def __init__(self, patch_size=2, patch_size_t=None, in_channels=16, dim=1920,
+                 text_dim=4096, bias=True, sample_width=90, sample_height=60,
+                 sample_frames=49, temporal_compression_ratio=4,
+                 max_text_seq_length=226, spatial_interpolation_scale=1.875,
+                 temporal_interpolation_scale=1.0, use_positional_embeddings=True,
+                 use_learned_positional_embeddings=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.dim = dim
+        self.sample_height = sample_height
+        self.sample_width = sample_width
+        self.sample_frames = sample_frames
+        self.temporal_compression_ratio = temporal_compression_ratio
+        self.max_text_seq_length = max_text_seq_length
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.use_positional_embeddings = use_positional_embeddings
+        self.use_learned_positional_embeddings = use_learned_positional_embeddings
+
+        if patch_size_t is None:
+            self.proj = operations.Conv2d(in_channels, dim, kernel_size=patch_size, stride=patch_size, bias=bias, device=device, dtype=dtype)
+        else:
+            self.proj = operations.Linear(in_channels * patch_size * patch_size * patch_size_t, dim, device=device, dtype=dtype)
+
+        self.text_proj = operations.Linear(text_dim, dim, device=device, dtype=dtype)
+
+        if use_positional_embeddings or use_learned_positional_embeddings:
+            persistent = use_learned_positional_embeddings
+            pos_embedding = self._get_positional_embeddings(sample_height, sample_width, sample_frames)
+            self.register_buffer("pos_embedding", pos_embedding, persistent=persistent)
+
+    def _get_positional_embeddings(self, sample_height, sample_width, sample_frames, device=None):
+        post_patch_height = sample_height // self.patch_size
+        post_patch_width = sample_width // self.patch_size
+        post_time_compression_frames = (sample_frames - 1) // self.temporal_compression_ratio + 1
+        if self.patch_size_t is not None:
+            post_time_compression_frames = post_time_compression_frames // self.patch_size_t
+        num_patches = post_patch_height * post_patch_width * post_time_compression_frames
+
+        pos_embedding = get_3d_sincos_pos_embed(
+            self.dim,
+            (post_patch_width, post_patch_height),
+            post_time_compression_frames,
+            self.spatial_interpolation_scale,
+            self.temporal_interpolation_scale,
+            device=device,
+        )
+        pos_embedding = pos_embedding.reshape(-1, self.dim)
+        joint_pos_embedding = pos_embedding.new_zeros(
+            1, self.max_text_seq_length + num_patches, self.dim, requires_grad=False
+        )
+        joint_pos_embedding.data[:, self.max_text_seq_length:].copy_(pos_embedding)
+        return joint_pos_embedding
+
+    def forward(self, text_embeds, image_embeds):
+        input_dtype = text_embeds.dtype
+        text_embeds = self.text_proj(text_embeds.to(self.text_proj.weight.dtype)).to(input_dtype)
+        batch_size, num_frames, channels, height, width = image_embeds.shape
+
+        proj_dtype = self.proj.weight.dtype
+        if self.patch_size_t is None:
+            image_embeds = image_embeds.reshape(-1, channels, height, width)
+            image_embeds = self.proj(image_embeds.to(proj_dtype)).to(input_dtype)
+            image_embeds = image_embeds.view(batch_size, num_frames, *image_embeds.shape[1:])
+            image_embeds = image_embeds.flatten(3).transpose(2, 3)
+            image_embeds = image_embeds.flatten(1, 2)
+        else:
+            p = self.patch_size
+            p_t = self.patch_size_t
+            image_embeds = image_embeds.permute(0, 1, 3, 4, 2)
+            image_embeds = image_embeds.reshape(
+                batch_size, num_frames // p_t, p_t, height // p, p, width // p, p, channels
+            )
+            image_embeds = image_embeds.permute(0, 1, 3, 5, 7, 2, 4, 6).flatten(4, 7).flatten(1, 3)
+            image_embeds = self.proj(image_embeds.to(proj_dtype)).to(input_dtype)
+
+        embeds = torch.cat([text_embeds, image_embeds], dim=1).contiguous()
+
+        if self.use_positional_embeddings or self.use_learned_positional_embeddings:
+            text_seq_length = text_embeds.shape[1]
+            num_image_patches = image_embeds.shape[1]
+
+            if self.use_learned_positional_embeddings:
+                image_pos = self.pos_embedding[
+                    :, self.max_text_seq_length:self.max_text_seq_length + num_image_patches
+                ].to(device=embeds.device, dtype=embeds.dtype)
+            else:
+                image_pos = get_3d_sincos_pos_embed(
+                    self.dim,
+                    (width // self.patch_size, height // self.patch_size),
+                    num_image_patches // ((height // self.patch_size) * (width // self.patch_size)),
+                    self.spatial_interpolation_scale,
+                    self.temporal_interpolation_scale,
+                    device=embeds.device,
+                ).reshape(1, num_image_patches, self.dim).to(dtype=embeds.dtype)
+
+            # Build joint: zeros for text + sincos for image
+            joint_pos = torch.zeros(1, text_seq_length + num_image_patches, self.dim, device=embeds.device, dtype=embeds.dtype)
+            joint_pos[:, text_seq_length:] = image_pos
+            embeds = embeds + joint_pos
+
+        return embeds
+
+
+class CogVideoXLayerNormZero(nn.Module):
+    def __init__(self, time_dim, dim, elementwise_affine=True, eps=1e-5, bias=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(time_dim, 6 * dim, bias=bias, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, encoder_hidden_states, temb):
+        shift, scale, gate, enc_shift, enc_scale, enc_gate = self.linear(self.silu(temb)).chunk(6, dim=1)
+        hidden_states = self.norm(hidden_states) * (1 + scale)[:, None, :] + shift[:, None, :]
+        encoder_hidden_states = self.norm(encoder_hidden_states) * (1 + enc_scale)[:, None, :] + enc_shift[:, None, :]
+        return hidden_states, encoder_hidden_states, gate[:, None, :], enc_gate[:, None, :]
+
+
+class CogVideoXAdaLayerNorm(nn.Module):
+    def __init__(self, time_dim, dim, elementwise_affine=True, eps=1e-5,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(time_dim, 2 * dim, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(dim, eps=eps, elementwise_affine=elementwise_affine, device=device, dtype=dtype)
+
+    def forward(self, x, temb):
+        temb = self.linear(self.silu(temb))
+        shift, scale = temb.chunk(2, dim=1)
+        x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :]
+        return x
+
+
+class CogVideoXBlock(nn.Module):
+    def __init__(self, dim, num_heads, head_dim, time_dim,
+                 eps=1e-5, ff_inner_dim=None, ff_bias=True,
+                 device=None, dtype=None, operations=None):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+
+        self.norm1 = CogVideoXLayerNormZero(time_dim, dim, eps=eps, device=device, dtype=dtype, operations=operations)
+
+        # Self-attention (joint text + latent)
+        self.q = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.k = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.v = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+        self.norm_q = operations.LayerNorm(head_dim, eps=1e-6, elementwise_affine=True, device=device, dtype=dtype)
+        self.norm_k = operations.LayerNorm(head_dim, eps=1e-6, elementwise_affine=True, device=device, dtype=dtype)
+        self.attn_out = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+
+        self.norm2 = CogVideoXLayerNormZero(time_dim, dim, eps=eps, device=device, dtype=dtype, operations=operations)
+
+        # Feed-forward (GELU approximate)
+        inner_dim = ff_inner_dim or dim * 4
+        self.ff_proj = operations.Linear(dim, inner_dim, bias=ff_bias, device=device, dtype=dtype)
+        self.ff_out = operations.Linear(inner_dim, dim, bias=ff_bias, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, encoder_hidden_states, temb, image_rotary_emb=None, transformer_options=None):
+        if transformer_options is None:
+            transformer_options = {}
+        text_seq_length = encoder_hidden_states.size(1)
+
+        # Norm & modulate
+        norm_hidden, norm_encoder, gate_msa, enc_gate_msa = self.norm1(hidden_states, encoder_hidden_states, temb)
+
+        # Joint self-attention
+        qkv_input = torch.cat([norm_encoder, norm_hidden], dim=1)
+        b, s, _ = qkv_input.shape
+        n, d = self.num_heads, self.head_dim
+
+        q = self.q(qkv_input).view(b, s, n, d)
+        k = self.k(qkv_input).view(b, s, n, d)
+        v = self.v(qkv_input)
+
+        q = self.norm_q(q).view(b, s, n, d)
+        k = self.norm_k(k).view(b, s, n, d)
+
+        # Apply rotary embeddings to image tokens only (diffusers format: [B, heads, seq, head_dim])
+        if image_rotary_emb is not None:
+            q_img = q[:, text_seq_length:].transpose(1, 2)  # [B, heads, img_seq, head_dim]
+            k_img = k[:, text_seq_length:].transpose(1, 2)
+            q_img = apply_rotary_emb(q_img, image_rotary_emb)
+            k_img = apply_rotary_emb(k_img, image_rotary_emb)
+            q = torch.cat([q[:, :text_seq_length], q_img.transpose(1, 2)], dim=1)
+            k = torch.cat([k[:, :text_seq_length], k_img.transpose(1, 2)], dim=1)
+
+        attn_out = optimized_attention(
+            q.reshape(b, s, n * d),
+            k.reshape(b, s, n * d),
+            v,
+            heads=self.num_heads,
+            transformer_options=transformer_options,
+        )
+
+        attn_out = self.attn_out(attn_out)
+
+        attn_encoder, attn_hidden = attn_out.split([text_seq_length, s - text_seq_length], dim=1)
+
+        hidden_states = hidden_states + gate_msa * attn_hidden
+        encoder_hidden_states = encoder_hidden_states + enc_gate_msa * attn_encoder
+
+        # Norm & modulate for FF
+        norm_hidden, norm_encoder, gate_ff, enc_gate_ff = self.norm2(hidden_states, encoder_hidden_states, temb)
+
+        # Feed-forward (GELU on concatenated text + latent)
+        ff_input = torch.cat([norm_encoder, norm_hidden], dim=1)
+        ff_output = self.ff_out(F.gelu(self.ff_proj(ff_input), approximate="tanh"))
+
+        hidden_states = hidden_states + gate_ff * ff_output[:, text_seq_length:]
+        encoder_hidden_states = encoder_hidden_states + enc_gate_ff * ff_output[:, :text_seq_length]
+
+        return hidden_states, encoder_hidden_states
+
+
+class CogVideoXTransformer3DModel(nn.Module):
+    def __init__(self,
+                 num_attention_heads=30,
+                 attention_head_dim=64,
+                 in_channels=16,
+                 out_channels=16,
+                 flip_sin_to_cos=True,
+                 freq_shift=0,
+                 time_embed_dim=512,
+                 ofs_embed_dim=None,
+                 text_embed_dim=4096,
+                 num_layers=30,
+                 dropout=0.0,
+                 attention_bias=True,
+                 sample_width=90,
+                 sample_height=60,
+                 sample_frames=49,
+                 patch_size=2,
+                 patch_size_t=None,
+                 temporal_compression_ratio=4,
+                 max_text_seq_length=226,
+                 spatial_interpolation_scale=1.875,
+                 temporal_interpolation_scale=1.0,
+                 use_rotary_positional_embeddings=False,
+                 use_learned_positional_embeddings=False,
+                 patch_bias=True,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+        super().__init__()
+        self.dtype = dtype
+        dim = num_attention_heads * attention_head_dim
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.patch_size = patch_size
+        self.patch_size_t = patch_size_t
+        self.max_text_seq_length = max_text_seq_length
+        self.use_rotary_positional_embeddings = use_rotary_positional_embeddings
+
+        # 1. Patch embedding
+        self.patch_embed = CogVideoXPatchEmbed(
+            patch_size=patch_size,
+            patch_size_t=patch_size_t,
+            in_channels=in_channels,
+            dim=dim,
+            text_dim=text_embed_dim,
+            bias=patch_bias,
+            sample_width=sample_width,
+            sample_height=sample_height,
+            sample_frames=sample_frames,
+            temporal_compression_ratio=temporal_compression_ratio,
+            max_text_seq_length=max_text_seq_length,
+            spatial_interpolation_scale=spatial_interpolation_scale,
+            temporal_interpolation_scale=temporal_interpolation_scale,
+            use_positional_embeddings=not use_rotary_positional_embeddings,
+            use_learned_positional_embeddings=use_learned_positional_embeddings,
+            device=device, dtype=torch.float32, operations=operations,
+        )
+
+        # 2. Time embedding
+        self.time_proj_dim = dim
+        self.time_proj_flip = flip_sin_to_cos
+        self.time_proj_shift = freq_shift
+        self.time_embedding_linear_1 = operations.Linear(dim, time_embed_dim, device=device, dtype=dtype)
+        self.time_embedding_act = nn.SiLU()
+        self.time_embedding_linear_2 = operations.Linear(time_embed_dim, time_embed_dim, device=device, dtype=dtype)
+
+        # Optional OFS embedding (CogVideoX 1.5 I2V)
+        self.ofs_proj_dim = ofs_embed_dim
+        if ofs_embed_dim:
+            self.ofs_embedding_linear_1 = operations.Linear(ofs_embed_dim, ofs_embed_dim, device=device, dtype=dtype)
+            self.ofs_embedding_act = nn.SiLU()
+            self.ofs_embedding_linear_2 = operations.Linear(ofs_embed_dim, ofs_embed_dim, device=device, dtype=dtype)
+        else:
+            self.ofs_embedding_linear_1 = None
+
+        # 3. Transformer blocks
+        self.blocks = nn.ModuleList([
+            CogVideoXBlock(
+                dim=dim,
+                num_heads=num_attention_heads,
+                head_dim=attention_head_dim,
+                time_dim=time_embed_dim,
+                eps=1e-5,
+                device=device, dtype=dtype, operations=operations,
+            )
+            for _ in range(num_layers)
+        ])
+
+        self.norm_final = operations.LayerNorm(dim, eps=1e-5, elementwise_affine=True, device=device, dtype=dtype)
+
+        # 4. Output
+        self.norm_out = CogVideoXAdaLayerNorm(
+            time_dim=time_embed_dim, dim=dim, eps=1e-5,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+        if patch_size_t is None:
+            output_dim = patch_size * patch_size * out_channels
+        else:
+            output_dim = patch_size * patch_size * patch_size_t * out_channels
+
+        self.proj_out = operations.Linear(dim, output_dim, device=device, dtype=dtype)
+
+        self.spatial_interpolation_scale = spatial_interpolation_scale
+        self.temporal_interpolation_scale = temporal_interpolation_scale
+        self.temporal_compression_ratio = temporal_compression_ratio
+
+    def forward(self, x, timestep, context, ofs=None, transformer_options=None, **kwargs):
+        if transformer_options is None:
+            transformer_options = {}
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, ofs, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, ofs=None, transformer_options=None, **kwargs):
+        if transformer_options is None:
+            transformer_options = {}
+        # ComfyUI passes [B, C, T, H, W]
+        batch_size, channels, t, h, w = x.shape
+
+        # Pad to patch size (temporal + spatial), same pattern as WAN
+        p_t = self.patch_size_t if self.patch_size_t is not None else 1
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, (p_t, self.patch_size, self.patch_size))
+
+        # CogVideoX expects [B, T, C, H, W]
+        x = x.permute(0, 2, 1, 3, 4)
+        batch_size, num_frames, channels, height, width = x.shape
+
+        # Time embedding
+        t_emb = get_timestep_embedding(timestep, self.time_proj_dim, self.time_proj_flip, self.time_proj_shift)
+        t_emb = t_emb.to(dtype=x.dtype)
+        emb = self.time_embedding_linear_2(self.time_embedding_act(self.time_embedding_linear_1(t_emb)))
+
+        if self.ofs_embedding_linear_1 is not None and ofs is not None:
+            ofs_emb = get_timestep_embedding(ofs, self.ofs_proj_dim, self.time_proj_flip, self.time_proj_shift)
+            ofs_emb = ofs_emb.to(dtype=x.dtype)
+            ofs_emb = self.ofs_embedding_linear_2(self.ofs_embedding_act(self.ofs_embedding_linear_1(ofs_emb)))
+            emb = emb + ofs_emb
+
+        # Patch embedding
+        hidden_states = self.patch_embed(context, x)
+
+        text_seq_length = context.shape[1]
+        encoder_hidden_states = hidden_states[:, :text_seq_length]
+        hidden_states = hidden_states[:, text_seq_length:]
+
+        # Rotary embeddings (if used)
+        image_rotary_emb = None
+        if self.use_rotary_positional_embeddings:
+            post_patch_height = height // self.patch_size
+            post_patch_width = width // self.patch_size
+            if self.patch_size_t is None:
+                post_time = num_frames
+            else:
+                post_time = num_frames // self.patch_size_t
+            image_rotary_emb = self._get_rotary_emb(post_patch_height, post_patch_width, post_time, device=x.device)
+
+        # Transformer blocks
+        for i, block in enumerate(self.blocks):
+            hidden_states, encoder_hidden_states = block(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                temb=emb,
+                image_rotary_emb=image_rotary_emb,
+                transformer_options=transformer_options,
+            )
+
+        hidden_states = self.norm_final(hidden_states)
+
+        # Output projection
+        hidden_states = self.norm_out(hidden_states, temb=emb)
+        hidden_states = self.proj_out(hidden_states)
+
+        # Unpatchify
+        p = self.patch_size
+        p_t = self.patch_size_t
+
+        if p_t is None:
+            output = hidden_states.reshape(batch_size, num_frames, height // p, width // p, -1, p, p)
+            output = output.permute(0, 1, 4, 2, 5, 3, 6).flatten(5, 6).flatten(3, 4)
+        else:
+            output = hidden_states.reshape(
+                batch_size, (num_frames + p_t - 1) // p_t, height // p, width // p, -1, p_t, p, p
+            )
+            output = output.permute(0, 1, 5, 4, 2, 6, 3, 7).flatten(6, 7).flatten(4, 5).flatten(1, 2)
+
+        # Back to ComfyUI format [B, C, T, H, W] and crop padding
+        output = output.permute(0, 2, 1, 3, 4)[:, :, :t, :h, :w]
+        return output
+
+    def _get_rotary_emb(self, h, w, t, device):
+        """Compute CogVideoX 3D rotary positional embeddings.
+
+        For CogVideoX 1.5 (patch_size_t != None): uses "slice" mode — grid positions
+        are integer arange computed at max_size, then sliced to actual size.
+        For CogVideoX 1.0 (patch_size_t == None): uses "linspace" mode with crop coords
+        scaled by spatial_interpolation_scale.
+        """
+        d = self.attention_head_dim
+        dim_t = d // 4
+        dim_h = d // 8 * 3
+        dim_w = d // 8 * 3
+
+        if self.patch_size_t is not None:
+            # CogVideoX 1.5: "slice" mode — positions are simple integer indices
+            # Compute at max(sample_size, actual_size) then slice to actual
+            base_h = self.patch_embed.sample_height // self.patch_size
+            base_w = self.patch_embed.sample_width // self.patch_size
+            max_h = max(base_h, h)
+            max_w = max(base_w, w)
+
+            grid_h = torch.arange(max_h, device=device, dtype=torch.float32)
+            grid_w = torch.arange(max_w, device=device, dtype=torch.float32)
+            grid_t = torch.arange(t, device=device, dtype=torch.float32)
+        else:
+            # CogVideoX 1.0: "linspace" mode with interpolation scale
+            grid_h = torch.linspace(0, h - 1, h, device=device, dtype=torch.float32) * self.spatial_interpolation_scale
+            grid_w = torch.linspace(0, w - 1, w, device=device, dtype=torch.float32) * self.spatial_interpolation_scale
+            grid_t = torch.arange(t, device=device, dtype=torch.float32)
+
+        freqs_t = _get_1d_rotary_pos_embed(dim_t, grid_t)
+        freqs_h = _get_1d_rotary_pos_embed(dim_h, grid_h)
+        freqs_w = _get_1d_rotary_pos_embed(dim_w, grid_w)
+
+        t_cos, t_sin = freqs_t
+        h_cos, h_sin = freqs_h
+        w_cos, w_sin = freqs_w
+
+        # Slice to actual size (for "slice" mode where grids may be larger)
+        t_cos, t_sin = t_cos[:t], t_sin[:t]
+        h_cos, h_sin = h_cos[:h], h_sin[:h]
+        w_cos, w_sin = w_cos[:w], w_sin[:w]
+
+        # Broadcast and concatenate into [T*H*W, head_dim]
+        t_cos = t_cos[:, None, None, :].expand(-1, h, w, -1)
+        t_sin = t_sin[:, None, None, :].expand(-1, h, w, -1)
+        h_cos = h_cos[None, :, None, :].expand(t, -1, w, -1)
+        h_sin = h_sin[None, :, None, :].expand(t, -1, w, -1)
+        w_cos = w_cos[None, None, :, :].expand(t, h, -1, -1)
+        w_sin = w_sin[None, None, :, :].expand(t, h, -1, -1)
+
+        cos = torch.cat([t_cos, h_cos, w_cos], dim=-1).reshape(t * h * w, -1)
+        sin = torch.cat([t_sin, h_sin, w_sin], dim=-1).reshape(t * h * w, -1)
+        return (cos, sin)

comfy/ldm/cogvideo/vae.py

@@ -0,0 +1,566 @@
+# CogVideoX VAE - ported to ComfyUI native ops
+# Architecture reference: diffusers AutoencoderKLCogVideoX
+# Style reference: comfy/ldm/wan/vae.py
+
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+
+class CausalConv3d(nn.Module):
+    """Causal 3D convolution with temporal padding.
+
+    Uses comfy.ops.Conv3d with autopad='causal_zero' fast path: when input has
+    a single temporal frame and no cache, the 3D conv weight is sliced to act
+    as a 2D conv, avoiding computation on zero-padded temporal dimensions.
+    """
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, pad_mode="constant"):
+        super().__init__()
+        if isinstance(kernel_size, int):
+            kernel_size = (kernel_size,) * 3
+
+        time_kernel, height_kernel, width_kernel = kernel_size
+        self.time_kernel_size = time_kernel
+        self.pad_mode = pad_mode
+
+        height_pad = (height_kernel - 1) // 2
+        width_pad = (width_kernel - 1) // 2
+        self.time_causal_padding = (width_pad, width_pad, height_pad, height_pad, time_kernel - 1, 0)
+
+        stride = stride if isinstance(stride, tuple) else (stride, 1, 1)
+        dilation = (dilation, 1, 1)
+        self.conv = ops.Conv3d(
+            in_channels, out_channels, kernel_size,
+            stride=stride, dilation=dilation,
+            padding=(0, height_pad, width_pad),
+        )
+
+    def forward(self, x, conv_cache=None):
+        if self.pad_mode == "replicate":
+            x = F.pad(x, self.time_causal_padding, mode="replicate")
+            conv_cache = None
+        else:
+            kernel_t = self.time_kernel_size
+            if kernel_t > 1:
+                if conv_cache is None and x.shape[2] == 1:
+                    # Fast path: single frame, no cache. All temporal padding
+                    # frames are copies of the input (replicate-style), so the
+                    # 3D conv reduces to a 2D conv with summed temporal kernel.
+                    w = comfy.ops.cast_to_input(self.conv.weight, x)
+                    b = comfy.ops.cast_to_input(self.conv.bias, x) if self.conv.bias is not None else None
+                    w2d = w.sum(dim=2, keepdim=True)
+                    out = F.conv3d(x, w2d, b,
+                                   self.conv.stride, self.conv.padding,
+                                   self.conv.dilation, self.conv.groups)
+                    return out, None
+                cached = [conv_cache] if conv_cache is not None else [x[:, :, :1]] * (kernel_t - 1)
+                x = torch.cat(cached + [x], dim=2)
+            conv_cache = x[:, :, -self.time_kernel_size + 1:].clone() if self.time_kernel_size > 1 else None
+
+        out = self.conv(x)
+        return out, conv_cache
+
+
+def _interpolate_zq(zq, target_size):
+    """Interpolate latent z to target (T, H, W), matching CogVideoX's first-frame-special handling."""
+    t = target_size[0]
+    if t > 1 and t % 2 == 1:
+        z_first = F.interpolate(zq[:, :, :1], size=(1, target_size[1], target_size[2]))
+        z_rest = F.interpolate(zq[:, :, 1:], size=(t - 1, target_size[1], target_size[2]))
+        return torch.cat([z_first, z_rest], dim=2)
+    return F.interpolate(zq, size=target_size)
+
+
+class SpatialNorm3D(nn.Module):
+    """Spatially conditioned normalization."""
+    def __init__(self, f_channels, zq_channels, groups=32):
+        super().__init__()
+        self.norm_layer = ops.GroupNorm(num_channels=f_channels, num_groups=groups, eps=1e-6, affine=True)
+        self.conv_y = CausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+        self.conv_b = CausalConv3d(zq_channels, f_channels, kernel_size=1, stride=1)
+
+    def forward(self, f, zq, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        if zq.shape[-3:] != f.shape[-3:]:
+            zq = _interpolate_zq(zq, f.shape[-3:])
+
+        conv_y, new_cache["conv_y"] = self.conv_y(zq, conv_cache=conv_cache.get("conv_y"))
+        conv_b, new_cache["conv_b"] = self.conv_b(zq, conv_cache=conv_cache.get("conv_b"))
+
+        return self.norm_layer(f) * conv_y + conv_b, new_cache
+
+
+class ResnetBlock3D(nn.Module):
+    """3D ResNet block with optional spatial norm."""
+    def __init__(self, in_channels, out_channels=None, temb_channels=512, groups=32,
+                 eps=1e-6, act_fn="silu", spatial_norm_dim=None, pad_mode="first"):
+        super().__init__()
+        out_channels = out_channels or in_channels
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.spatial_norm_dim = spatial_norm_dim
+
+        if act_fn == "silu":
+            self.nonlinearity = nn.SiLU()
+        elif act_fn == "swish":
+            self.nonlinearity = nn.SiLU()
+        else:
+            self.nonlinearity = nn.SiLU()
+
+        if spatial_norm_dim is None:
+            self.norm1 = ops.GroupNorm(num_channels=in_channels, num_groups=groups, eps=eps)
+            self.norm2 = ops.GroupNorm(num_channels=out_channels, num_groups=groups, eps=eps)
+        else:
+            self.norm1 = SpatialNorm3D(in_channels, spatial_norm_dim, groups=groups)
+            self.norm2 = SpatialNorm3D(out_channels, spatial_norm_dim, groups=groups)
+
+        self.conv1 = CausalConv3d(in_channels, out_channels, kernel_size=3, pad_mode=pad_mode)
+
+        if temb_channels > 0:
+            self.temb_proj = ops.Linear(temb_channels, out_channels)
+
+        self.conv2 = CausalConv3d(out_channels, out_channels, kernel_size=3, pad_mode=pad_mode)
+
+        if in_channels != out_channels:
+            self.conv_shortcut = ops.Conv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+        else:
+            self.conv_shortcut = None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        residual = x
+
+        if zq is not None:
+            x, new_cache["norm1"] = self.norm1(x, zq, conv_cache=conv_cache.get("norm1"))
+        else:
+            x = self.norm1(x)
+
+        x = self.nonlinearity(x)
+        x, new_cache["conv1"] = self.conv1(x, conv_cache=conv_cache.get("conv1"))
+
+        if temb is not None and hasattr(self, "temb_proj"):
+            x = x + self.temb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if zq is not None:
+            x, new_cache["norm2"] = self.norm2(x, zq, conv_cache=conv_cache.get("norm2"))
+        else:
+            x = self.norm2(x)
+
+        x = self.nonlinearity(x)
+        x, new_cache["conv2"] = self.conv2(x, conv_cache=conv_cache.get("conv2"))
+
+        if self.conv_shortcut is not None:
+            residual = self.conv_shortcut(residual)
+
+        return x + residual, new_cache
+
+
+class Downsample3D(nn.Module):
+    """3D downsampling with optional temporal compression."""
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=2, padding=0, compress_time=False):
+        super().__init__()
+        self.conv = ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x):
+        if self.compress_time:
+            b, c, t, h, w = x.shape
+            x = x.permute(0, 3, 4, 1, 2).reshape(b * h * w, c, t)
+            if t % 2 == 1:
+                x_first, x_rest = x[..., 0], x[..., 1:]
+                if x_rest.shape[-1] > 0:
+                    x_rest = F.avg_pool1d(x_rest, kernel_size=2, stride=2)
+                x = torch.cat([x_first[..., None], x_rest], dim=-1)
+                x = x.reshape(b, h, w, c, x.shape[-1]).permute(0, 3, 4, 1, 2)
+            else:
+                x = F.avg_pool1d(x, kernel_size=2, stride=2)
+                x = x.reshape(b, h, w, c, x.shape[-1]).permute(0, 3, 4, 1, 2)
+
+        pad = (0, 1, 0, 1)
+        x = F.pad(x, pad, mode="constant", value=0)
+        b, c, t, h, w = x.shape
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.conv(x)
+        x = x.reshape(b, t, x.shape[1], x.shape[2], x.shape[3]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+class Upsample3D(nn.Module):
+    """3D upsampling with optional temporal decompression."""
+    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, padding=1, compress_time=False):
+        super().__init__()
+        self.conv = ops.Conv2d(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=padding)
+        self.compress_time = compress_time
+
+    def forward(self, x):
+        if self.compress_time:
+            if x.shape[2] > 1 and x.shape[2] % 2 == 1:
+                x_first, x_rest = x[:, :, 0], x[:, :, 1:]
+                x_first = F.interpolate(x_first, scale_factor=2.0)
+                x_rest = F.interpolate(x_rest, scale_factor=2.0)
+                x = torch.cat([x_first[:, :, None, :, :], x_rest], dim=2)
+            elif x.shape[2] > 1:
+                x = F.interpolate(x, scale_factor=2.0)
+            else:
+                x = x.squeeze(2)
+                x = F.interpolate(x, scale_factor=2.0)
+                x = x[:, :, None, :, :]
+        else:
+            b, c, t, h, w = x.shape
+            x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+            x = F.interpolate(x, scale_factor=2.0)
+            x = x.reshape(b, t, c, *x.shape[2:]).permute(0, 2, 1, 3, 4)
+
+        b, c, t, h, w = x.shape
+        x = x.permute(0, 2, 1, 3, 4).reshape(b * t, c, h, w)
+        x = self.conv(x)
+        x = x.reshape(b, t, *x.shape[1:]).permute(0, 2, 1, 3, 4)
+        return x
+
+
+class DownBlock3D(nn.Module):
+    def __init__(self, in_channels, out_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, add_downsample=True,
+                 compress_time=False, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels if i == 0 else out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels,
+                groups=groups, eps=eps, act_fn=act_fn, pad_mode=pad_mode,
+            )
+            for i in range(num_layers)
+        ])
+        self.downsamplers = nn.ModuleList([Downsample3D(out_channels, out_channels, compress_time=compress_time)]) if add_downsample else None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        if self.downsamplers is not None:
+            for ds in self.downsamplers:
+                x = ds(x)
+        return x, new_cache
+
+
+class MidBlock3D(nn.Module):
+    def __init__(self, in_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, spatial_norm_dim=None, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels, out_channels=in_channels,
+                temb_channels=temb_channels, groups=groups, eps=eps,
+                act_fn=act_fn, spatial_norm_dim=spatial_norm_dim, pad_mode=pad_mode,
+            )
+            for _ in range(num_layers)
+        ])
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        return x, new_cache
+
+
+class UpBlock3D(nn.Module):
+    def __init__(self, in_channels, out_channels, temb_channels=0, num_layers=1,
+                 eps=1e-6, act_fn="silu", groups=32, spatial_norm_dim=16,
+                 add_upsample=True, compress_time=False, pad_mode="first"):
+        super().__init__()
+        self.resnets = nn.ModuleList([
+            ResnetBlock3D(
+                in_channels=in_channels if i == 0 else out_channels,
+                out_channels=out_channels,
+                temb_channels=temb_channels, groups=groups, eps=eps,
+                act_fn=act_fn, spatial_norm_dim=spatial_norm_dim, pad_mode=pad_mode,
+            )
+            for i in range(num_layers)
+        ])
+        self.upsamplers = nn.ModuleList([Upsample3D(out_channels, out_channels, compress_time=compress_time)]) if add_upsample else None
+
+    def forward(self, x, temb=None, zq=None, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+        for i, resnet in enumerate(self.resnets):
+            x, new_cache[f"resnet_{i}"] = resnet(x, temb, zq, conv_cache=conv_cache.get(f"resnet_{i}"))
+        if self.upsamplers is not None:
+            for us in self.upsamplers:
+                x = us(x)
+        return x, new_cache
+
+
+class Encoder3D(nn.Module):
+    def __init__(self, in_channels=3, out_channels=16,
+                 block_out_channels=(128, 256, 256, 512),
+                 layers_per_block=3, act_fn="silu",
+                 eps=1e-6, groups=32, pad_mode="first",
+                 temporal_compression_ratio=4):
+        super().__init__()
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CausalConv3d(in_channels, block_out_channels[0], kernel_size=3, pad_mode=pad_mode)
+
+        self.down_blocks = nn.ModuleList()
+        output_channel = block_out_channels[0]
+        for i in range(len(block_out_channels)):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            self.down_blocks.append(DownBlock3D(
+                in_channels=input_channel, out_channels=output_channel,
+                temb_channels=0, num_layers=layers_per_block,
+                eps=eps, act_fn=act_fn, groups=groups,
+                add_downsample=not is_final, compress_time=compress_time,
+            ))
+
+        self.mid_block = MidBlock3D(
+            in_channels=block_out_channels[-1], temb_channels=0,
+            num_layers=2, eps=eps, act_fn=act_fn, groups=groups, pad_mode=pad_mode,
+        )
+
+        self.norm_out = ops.GroupNorm(groups, block_out_channels[-1], eps=1e-6)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(block_out_channels[-1], 2 * out_channels, kernel_size=3, pad_mode=pad_mode)
+
+    def forward(self, x, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        x, new_cache["conv_in"] = self.conv_in(x, conv_cache=conv_cache.get("conv_in"))
+
+        for i, block in enumerate(self.down_blocks):
+            key = f"down_block_{i}"
+            x, new_cache[key] = block(x, None, None, conv_cache.get(key))
+
+        x, new_cache["mid_block"] = self.mid_block(x, None, None, conv_cache=conv_cache.get("mid_block"))
+
+        x = self.norm_out(x)
+        x = self.conv_act(x)
+        x, new_cache["conv_out"] = self.conv_out(x, conv_cache=conv_cache.get("conv_out"))
+
+        return x, new_cache
+
+
+class Decoder3D(nn.Module):
+    def __init__(self, in_channels=16, out_channels=3,
+                 block_out_channels=(128, 256, 256, 512),
+                 layers_per_block=3, act_fn="silu",
+                 eps=1e-6, groups=32, pad_mode="first",
+                 temporal_compression_ratio=4):
+        super().__init__()
+        reversed_channels = list(reversed(block_out_channels))
+        temporal_compress_level = int(np.log2(temporal_compression_ratio))
+
+        self.conv_in = CausalConv3d(in_channels, reversed_channels[0], kernel_size=3, pad_mode=pad_mode)
+
+        self.mid_block = MidBlock3D(
+            in_channels=reversed_channels[0], temb_channels=0,
+            num_layers=2, eps=eps, act_fn=act_fn, groups=groups,
+            spatial_norm_dim=in_channels, pad_mode=pad_mode,
+        )
+
+        self.up_blocks = nn.ModuleList()
+        output_channel = reversed_channels[0]
+        for i in range(len(block_out_channels)):
+            prev_channel = output_channel
+            output_channel = reversed_channels[i]
+            is_final = i == len(block_out_channels) - 1
+            compress_time = i < temporal_compress_level
+
+            self.up_blocks.append(UpBlock3D(
+                in_channels=prev_channel, out_channels=output_channel,
+                temb_channels=0, num_layers=layers_per_block + 1,
+                eps=eps, act_fn=act_fn, groups=groups,
+                spatial_norm_dim=in_channels,
+                add_upsample=not is_final, compress_time=compress_time,
+            ))
+
+        self.norm_out = SpatialNorm3D(reversed_channels[-1], in_channels, groups=groups)
+        self.conv_act = nn.SiLU()
+        self.conv_out = CausalConv3d(reversed_channels[-1], out_channels, kernel_size=3, pad_mode=pad_mode)
+
+    def forward(self, sample, conv_cache=None):
+        new_cache = {}
+        conv_cache = conv_cache or {}
+
+        x, new_cache["conv_in"] = self.conv_in(sample, conv_cache=conv_cache.get("conv_in"))
+
+        x, new_cache["mid_block"] = self.mid_block(x, None, sample, conv_cache=conv_cache.get("mid_block"))
+
+        for i, block in enumerate(self.up_blocks):
+            key = f"up_block_{i}"
+            x, new_cache[key] = block(x, None, sample, conv_cache=conv_cache.get(key))
+
+        x, new_cache["norm_out"] = self.norm_out(x, sample, conv_cache=conv_cache.get("norm_out"))
+        x = self.conv_act(x)
+        x, new_cache["conv_out"] = self.conv_out(x, conv_cache=conv_cache.get("conv_out"))
+
+        return x, new_cache
+
+
+
+class AutoencoderKLCogVideoX(nn.Module):
+    """CogVideoX VAE. Spatial tiling/slicing handled by ComfyUI's VAE wrapper.
+
+    Uses rolling temporal decode: conv_in + mid_block + temporal up_blocks run
+    on the full (low-res) tensor, then the expensive spatial-only up_blocks +
+    norm_out + conv_out are processed in small temporal chunks with conv_cache
+    carrying causal state between chunks. This keeps peak VRAM proportional to
+    chunk_size rather than total frame count.
+    """
+
+    def __init__(self,
+                 in_channels=3, out_channels=3,
+                 block_out_channels=(128, 256, 256, 512),
+                 latent_channels=16, layers_per_block=3,
+                 act_fn="silu", eps=1e-6, groups=32,
+                 temporal_compression_ratio=4,
+                 ):
+        super().__init__()
+        self.latent_channels = latent_channels
+        self.temporal_compression_ratio = temporal_compression_ratio
+
+        self.encoder = Encoder3D(
+            in_channels=in_channels, out_channels=latent_channels,
+            block_out_channels=block_out_channels, layers_per_block=layers_per_block,
+            act_fn=act_fn, eps=eps, groups=groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+        self.decoder = Decoder3D(
+            in_channels=latent_channels, out_channels=out_channels,
+            block_out_channels=block_out_channels, layers_per_block=layers_per_block,
+            act_fn=act_fn, eps=eps, groups=groups,
+            temporal_compression_ratio=temporal_compression_ratio,
+        )
+
+        self.num_latent_frames_batch_size = 2
+        self.num_sample_frames_batch_size = 8
+
+    def encode(self, x):
+        t = x.shape[2]
+        frame_batch = self.num_sample_frames_batch_size
+        remainder = t % frame_batch
+        conv_cache = None
+        enc = []
+
+        # Process remainder frames first so only the first chunk can have an
+        # odd temporal dimension — where Downsample3D's first-frame-special
+        # handling in temporal compression is actually correct.
+        if remainder > 0:
+            chunk, conv_cache = self.encoder(x[:, :, :remainder], conv_cache=conv_cache)
+            enc.append(chunk.to(x.device))
+
+        for start in range(remainder, t, frame_batch):
+            chunk, conv_cache = self.encoder(x[:, :, start:start + frame_batch], conv_cache=conv_cache)
+            enc.append(chunk.to(x.device))
+
+        enc = torch.cat(enc, dim=2)
+        mean, _ = enc.chunk(2, dim=1)
+        return mean
+
+    def decode(self, z):
+        return self._decode_rolling(z)
+
+    def _decode_batched(self, z):
+        """Original batched decode - processes 2 latent frames through full decoder."""
+        t = z.shape[2]
+        frame_batch = self.num_latent_frames_batch_size
+        num_batches = max(t // frame_batch, 1)
+        conv_cache = None
+        dec = []
+        for i in range(num_batches):
+            remaining = t % frame_batch
+            start = frame_batch * i + (0 if i == 0 else remaining)
+            end = frame_batch * (i + 1) + remaining
+            chunk, conv_cache = self.decoder(z[:, :, start:end], conv_cache=conv_cache)
+            dec.append(chunk.cpu())
+        return torch.cat(dec, dim=2).to(z.device)
+
+    def _decode_rolling(self, z):
+        """Rolling decode - processes low-res layers on full tensor, then rolls
+        through expensive high-res layers in temporal chunks."""
+        decoder = self.decoder
+        device = z.device
+
+        # Determine which up_blocks have temporal upsample vs spatial-only.
+        # Temporal up_blocks are cheap (low res), spatial-only are expensive.
+        temporal_compress_level = int(np.log2(self.temporal_compression_ratio))
+        split_at = temporal_compress_level  # first N up_blocks do temporal upsample
+
+        # Phase 1: conv_in + mid_block + temporal up_blocks on full tensor (low/medium res)
+        x, _ = decoder.conv_in(z)
+        x, _ = decoder.mid_block(x, None, z)
+
+        for i in range(split_at):
+            x, _ = decoder.up_blocks[i](x, None, z)
+
+        # Phase 2: remaining spatial-only up_blocks + norm_out + conv_out in temporal chunks
+        remaining_blocks = list(range(split_at, len(decoder.up_blocks)))
+        chunk_size = 4  # pixel frames per chunk through high-res layers
+        t_expanded = x.shape[2]
+
+        if t_expanded <= chunk_size or len(remaining_blocks) == 0:
+            # Small enough to process in one go
+            for i in remaining_blocks:
+                x, _ = decoder.up_blocks[i](x, None, z)
+            x, _ = decoder.norm_out(x, z)
+            x = decoder.conv_act(x)
+            x, _ = decoder.conv_out(x)
+            return x
+
+        # Expand z temporally once to match Phase 2's time dimension.
+        # z stays at latent spatial resolution so this is small (~16 MB vs ~1.3 GB
+        # for the old approach of pre-interpolating to every pixel resolution).
+        z_time_expanded = _interpolate_zq(z, (t_expanded, z.shape[3], z.shape[4]))
+
+        # Process in temporal chunks, interpolating spatially per-chunk to avoid
+        # allocating full [B, C, t_expanded, H, W] tensors at each resolution.
+        dec_out = []
+        conv_caches = {}
+
+        for chunk_start in range(0, t_expanded, chunk_size):
+            chunk_end = min(chunk_start + chunk_size, t_expanded)
+            x_chunk = x[:, :, chunk_start:chunk_end]
+            z_t_chunk = z_time_expanded[:, :, chunk_start:chunk_end]
+            z_spatial_cache = {}
+
+            for i in remaining_blocks:
+                block = decoder.up_blocks[i]
+                cache_key = f"up_block_{i}"
+                hw_key = (x_chunk.shape[3], x_chunk.shape[4])
+                if hw_key not in z_spatial_cache:
+                    if z_t_chunk.shape[3] == hw_key[0] and z_t_chunk.shape[4] == hw_key[1]:
+                        z_spatial_cache[hw_key] = z_t_chunk
+                    else:
+                        z_spatial_cache[hw_key] = F.interpolate(z_t_chunk, size=(z_t_chunk.shape[2], hw_key[0], hw_key[1]))
+                x_chunk, new_cache = block(x_chunk, None, z_spatial_cache[hw_key], conv_cache=conv_caches.get(cache_key))
+                conv_caches[cache_key] = new_cache
+
+            hw_key = (x_chunk.shape[3], x_chunk.shape[4])
+            if hw_key not in z_spatial_cache:
+                z_spatial_cache[hw_key] = F.interpolate(z_t_chunk, size=(z_t_chunk.shape[2], hw_key[0], hw_key[1]))
+            x_chunk, new_cache = decoder.norm_out(x_chunk, z_spatial_cache[hw_key], conv_cache=conv_caches.get("norm_out"))
+            conv_caches["norm_out"] = new_cache
+            x_chunk = decoder.conv_act(x_chunk)
+            x_chunk, new_cache = decoder.conv_out(x_chunk, conv_cache=conv_caches.get("conv_out"))
+            conv_caches["conv_out"] = new_cache
+
+            dec_out.append(x_chunk.cpu())
+            del z_spatial_cache
+
+        del x, z_time_expanded
+        return torch.cat(dec_out, dim=2).to(device)

comfy/ldm/ernie/model.py

@@ -0,0 +1,301 @@
+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)
+
+        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 * (1 + scale_msa) + shift_msa
+
+        attn_out = self.self_attention(x_norm, attention_mask=attention_mask, image_rotary_emb=rotary_pos_emb)
+        x = residual + gate_msa * attn_out
+
+        residual = x
+        x_norm = self.adaLN_mlp_ln(x)
+        x_norm = x_norm * (1 + scale_mlp) + shift_mlp
+
+        return residual + gate_mlp * self.mlp(x_norm)
+
+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 = torch.addcmul(shift.unsqueeze(1), x, 1 + scale.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/lightricks/vae/audio_vae.py

@@ -4,9 +4,6 @@ import math
 import torch
 import torchaudio
 
-import comfy.model_management
-import comfy.model_patcher
-import comfy.utils as utils
 from comfy.ldm.mmaudio.vae.distributions import DiagonalGaussianDistribution
 from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
 from comfy.ldm.lightricks.vae.causal_audio_autoencoder import (
@@ -43,30 +40,6 @@ class AudioVAEComponentConfig:
 
         return cls(autoencoder=audio_config, vocoder=vocoder_config)
 
-
-class ModelDeviceManager:
-    """Manages device placement and GPU residency for the composed model."""
-
-    def __init__(self, module: torch.nn.Module):
-        load_device = comfy.model_management.get_torch_device()
-        offload_device = comfy.model_management.vae_offload_device()
-        self.patcher = comfy.model_patcher.ModelPatcher(module, load_device, offload_device)
-
-    def ensure_model_loaded(self) -> None:
-        comfy.model_management.free_memory(
-            self.patcher.model_size(),
-            self.patcher.load_device,
-        )
-        comfy.model_management.load_model_gpu(self.patcher)
-
-    def move_to_load_device(self, tensor: torch.Tensor) -> torch.Tensor:
-        return tensor.to(self.patcher.load_device)
-
-    @property
-    def load_device(self):
-        return self.patcher.load_device
-
-
 class AudioLatentNormalizer:
     """Applies per-channel statistics in patch space and restores original layout."""
 
@@ -132,23 +105,17 @@ class AudioPreprocessor:
 class AudioVAE(torch.nn.Module):
     """High-level Audio VAE wrapper exposing encode and decode entry points."""
 
-    def __init__(self, state_dict: dict, metadata: dict):
+    def __init__(self, metadata: dict):
         super().__init__()
 
         component_config = AudioVAEComponentConfig.from_metadata(metadata)
 
-        vae_sd = utils.state_dict_prefix_replace(state_dict, {"audio_vae.": ""}, filter_keys=True)
-        vocoder_sd = utils.state_dict_prefix_replace(state_dict, {"vocoder.": ""}, filter_keys=True)
-
         self.autoencoder = CausalAudioAutoencoder(config=component_config.autoencoder)
         if "bwe" in component_config.vocoder:
             self.vocoder = VocoderWithBWE(config=component_config.vocoder)
         else:
             self.vocoder = Vocoder(config=component_config.vocoder)
 
-        self.autoencoder.load_state_dict(vae_sd, strict=False)
-        self.vocoder.load_state_dict(vocoder_sd, strict=False)
-
         autoencoder_config = self.autoencoder.get_config()
         self.normalizer = AudioLatentNormalizer(
             AudioPatchifier(
@@ -168,18 +135,12 @@ class AudioVAE(torch.nn.Module):
             n_fft=autoencoder_config["n_fft"],
         )
 
-        self.device_manager = ModelDeviceManager(self)
-
-    def encode(self, audio: dict) -> torch.Tensor:
+    def encode(self, audio, sample_rate=44100) -> torch.Tensor:
         """Encode a waveform dictionary into normalized latent tensors."""
 
-        waveform = audio["waveform"]
-        waveform_sample_rate = audio["sample_rate"]
+        waveform = audio
+        waveform_sample_rate = sample_rate
         input_device = waveform.device
-        # Ensure that Audio VAE is loaded on the correct device.
-        self.device_manager.ensure_model_loaded()
-
-        waveform = self.device_manager.move_to_load_device(waveform)
         expected_channels = self.autoencoder.encoder.in_channels
         if waveform.shape[1] != expected_channels:
             if waveform.shape[1] == 1:
@@ -190,7 +151,7 @@ class AudioVAE(torch.nn.Module):
                 )
 
         mel_spec = self.preprocessor.waveform_to_mel(
-            waveform, waveform_sample_rate, device=self.device_manager.load_device
+            waveform, waveform_sample_rate, device=waveform.device
         )
 
         latents = self.autoencoder.encode(mel_spec)
@@ -204,17 +165,13 @@ class AudioVAE(torch.nn.Module):
         """Decode normalized latent tensors into an audio waveform."""
         original_shape = latents.shape
 
-        # Ensure that Audio VAE is loaded on the correct device.
-        self.device_manager.ensure_model_loaded()
-
-        latents = self.device_manager.move_to_load_device(latents)
         latents = self.normalizer.denormalize(latents)
 
         target_shape = self.target_shape_from_latents(original_shape)
         mel_spec = self.autoencoder.decode(latents, target_shape=target_shape)
 
         waveform = self.run_vocoder(mel_spec)
-        return self.device_manager.move_to_load_device(waveform)
+        return waveform
 
     def target_shape_from_latents(self, latents_shape):
         batch, _, time, _ = latents_shape

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/diffusionmodules/openaimodel.py

@@ -34,6 +34,16 @@ class TimestepBlock(nn.Module):
 #This is needed because accelerate makes a copy of transformer_options which breaks "transformer_index"
 def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, output_shape=None, time_context=None, num_video_frames=None, image_only_indicator=None):
     for layer in ts:
+        if "patches" in transformer_options and "forward_timestep_embed_patch" in transformer_options["patches"]:
+            found_patched = False
+            for class_type, handler in transformer_options["patches"]["forward_timestep_embed_patch"]:
+                if isinstance(layer, class_type):
+                    x = handler(layer, x, emb, context, transformer_options, output_shape, time_context, num_video_frames, image_only_indicator)
+                    found_patched = True
+                    break
+            if found_patched:
+                continue
+
         if isinstance(layer, VideoResBlock):
             x = layer(x, emb, num_video_frames, image_only_indicator)
         elif isinstance(layer, TimestepBlock):
@@ -49,15 +59,6 @@ def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, out
         elif isinstance(layer, Upsample):
             x = layer(x, output_shape=output_shape)
         else:
-            if "patches" in transformer_options and "forward_timestep_embed_patch" in transformer_options["patches"]:
-                found_patched = False
-                for class_type, handler in transformer_options["patches"]["forward_timestep_embed_patch"]:
-                    if isinstance(layer, class_type):
-                        x = handler(layer, x, emb, context, transformer_options, output_shape, time_context, num_video_frames, image_only_indicator)
-                        found_patched = True
-                        break
-                if found_patched:
-                    continue
             x = layer(x)
     return x
 
@@ -894,6 +895,12 @@ class UNetModel(nn.Module):
             h = forward_timestep_embed(self.middle_block, h, emb, context, transformer_options, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
         h = apply_control(h, control, 'middle')
 
+        if "middle_block_after_patch" in transformer_patches:
+            patch = transformer_patches["middle_block_after_patch"]
+            for p in patch:
+                out = p({"h": h, "x": x, "emb": emb, "context": context, "y": y,
+                         "timesteps": timesteps, "transformer_options": transformer_options})
+                h = out["h"]
 
         for id, module in enumerate(self.output_blocks):
             transformer_options["block"] = ("output", id)
@@ -905,8 +912,9 @@ class UNetModel(nn.Module):
                 for p in patch:
                     h, hsp = p(h, hsp, transformer_options)
 
-            h = th.cat([h, hsp], dim=1)
-            del hsp
+            if hsp is not None:
+                h = th.cat([h, hsp], dim=1)
+                del hsp
             if len(hs) > 0:
                 output_shape = hs[-1].shape
             else:

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/sam3/detector.py

@@ -0,0 +1,596 @@
+# SAM3 detector: transformer encoder-decoder, segmentation head, geometry encoder, scoring.
+
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.ops import roi_align
+
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.sam3.tracker import SAM3Tracker, SAM31Tracker
+from comfy.ldm.sam3.sam import SAM3VisionBackbone  # noqa: used in __init__
+from comfy.ldm.sam3.sam import MLP, PositionEmbeddingSine
+
+TRACKER_CLASSES = {"SAM3": SAM3Tracker, "SAM31": SAM31Tracker}
+from comfy.ops import cast_to_input
+
+
+def box_cxcywh_to_xyxy(x):
+    cx, cy, w, h = x.unbind(-1)
+    return torch.stack([cx - 0.5 * w, cy - 0.5 * h, cx + 0.5 * w, cy + 0.5 * h], dim=-1)
+
+
+def gen_sineembed_for_position(pos_tensor, num_feats=256):
+    """Per-coordinate sinusoidal embedding: (..., N) -> (..., N * num_feats)."""
+    assert num_feats % 2 == 0
+    hdim = num_feats // 2
+    freqs = 10000.0 ** (2 * (torch.arange(hdim, dtype=torch.float32, device=pos_tensor.device) // 2) / hdim)
+    embeds = []
+    for c in range(pos_tensor.shape[-1]):
+        raw = (pos_tensor[..., c].float() * 2 * math.pi).unsqueeze(-1) / freqs
+        embeds.append(torch.stack([raw[..., 0::2].sin(), raw[..., 1::2].cos()], dim=-1).flatten(-2))
+    return torch.cat(embeds, dim=-1).to(pos_tensor.dtype)
+
+
+class SplitMHA(nn.Module):
+    """Multi-head attention with separate Q/K/V projections (split from fused in_proj_weight)."""
+    def __init__(self, d_model, num_heads=8, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.num_heads = num_heads
+        self.q_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.k_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.v_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.out_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+
+    def forward(self, q_input, k_input=None, v_input=None, mask=None):
+        q = self.q_proj(q_input)
+        if k_input is None:
+            k = self.k_proj(q_input)
+            v = self.v_proj(q_input)
+        else:
+            k = self.k_proj(k_input)
+            v = self.v_proj(v_input if v_input is not None else k_input)
+        if mask is not None and mask.ndim == 2:
+            mask = mask[:, None, None, :]  # [B, T] -> [B, 1, 1, T] for SDPA broadcast
+        dtype = q.dtype  # manual_cast may produce mixed dtypes
+        out = optimized_attention(q, k.to(dtype), v.to(dtype), self.num_heads, mask=mask, low_precision_attention=False)
+        return self.out_proj(out)
+
+
+class MLPWithNorm(nn.Module):
+    """MLP with residual connection and output LayerNorm."""
+    def __init__(self, input_dim, hidden_dim, output_dim, num_layers, residual=True, device=None, dtype=None, operations=None):
+        super().__init__()
+        dims = [input_dim] + [hidden_dim] * (num_layers - 1) + [output_dim]
+        self.layers = nn.ModuleList([
+            operations.Linear(dims[i], dims[i + 1], device=device, dtype=dtype)
+            for i in range(num_layers)
+        ])
+        self.out_norm = operations.LayerNorm(output_dim, device=device, dtype=dtype)
+        self.residual = residual and (input_dim == output_dim)
+
+    def forward(self, x):
+        orig = x
+        for i, layer in enumerate(self.layers):
+            x = layer(x)
+            if i < len(self.layers) - 1:
+                x = F.relu(x)
+        if self.residual:
+            x = x + orig
+        return self.out_norm(x)
+
+
+class EncoderLayer(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, dim_ff=2048, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.self_attn = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.cross_attn_image = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.linear1 = operations.Linear(d_model, dim_ff, device=device, dtype=dtype)
+        self.linear2 = operations.Linear(dim_ff, 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.norm3 = operations.LayerNorm(d_model, device=device, dtype=dtype)
+
+    def forward(self, x, pos, text_memory=None, text_mask=None):
+        normed = self.norm1(x)
+        q_k = normed + pos
+        x = x + self.self_attn(q_k, q_k, normed)
+        if text_memory is not None:
+            normed = self.norm2(x)
+            x = x + self.cross_attn_image(normed, text_memory, text_memory, mask=text_mask)
+        normed = self.norm3(x)
+        x = x + self.linear2(F.relu(self.linear1(normed)))
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    """Checkpoint: transformer.encoder.layers.N.*"""
+    def __init__(self, d_model=256, num_heads=8, dim_ff=2048, num_layers=6, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.layers = nn.ModuleList([
+            EncoderLayer(d_model, num_heads, dim_ff, device=device, dtype=dtype, operations=operations)
+            for _ in range(num_layers)
+        ])
+
+    def forward(self, x, pos, text_memory=None, text_mask=None):
+        for layer in self.layers:
+            x = layer(x, pos, text_memory, text_mask)
+        return x
+
+
+class DecoderLayer(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, dim_ff=2048, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.self_attn = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.cross_attn = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.ca_text = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.norm1 = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.norm2 = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.norm3 = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.catext_norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.linear1 = operations.Linear(d_model, dim_ff, device=device, dtype=dtype)
+        self.linear2 = operations.Linear(dim_ff, d_model, device=device, dtype=dtype)
+
+    def forward(self, x, memory, x_pos, memory_pos, text_memory=None, text_mask=None, cross_attn_bias=None):
+        q_k = x + x_pos
+        x = self.norm2(x + self.self_attn(q_k, q_k, x))
+        if text_memory is not None:
+            x = self.catext_norm(x + self.ca_text(x + x_pos, text_memory, text_memory, mask=text_mask))
+        x = self.norm1(x + self.cross_attn(x + x_pos, memory + memory_pos, memory, mask=cross_attn_bias))
+        x = self.norm3(x + self.linear2(F.relu(self.linear1(x))))
+        return x
+
+
+class TransformerDecoder(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, dim_ff=2048, num_layers=6,
+                 num_queries=200, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.d_model = d_model
+        self.num_queries = num_queries
+
+        self.layers = nn.ModuleList([
+            DecoderLayer(d_model, num_heads, dim_ff, device=device, dtype=dtype, operations=operations)
+            for _ in range(num_layers)
+        ])
+        self.norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.query_embed = operations.Embedding(num_queries, d_model, device=device, dtype=dtype)
+        self.reference_points = operations.Embedding(num_queries, 4, device=device, dtype=dtype) # Reference points: Embedding(num_queries, 4) — learned anchor boxes
+        self.ref_point_head = MLP(d_model * 2, d_model, d_model, 2, device=device, dtype=dtype, operations=operations) # ref_point_head input: 512 (4 coords * 128 sine features each)
+        self.bbox_embed = MLP(d_model, d_model, 4, 3, device=device, dtype=dtype, operations=operations)
+
+        self.boxRPB_embed_x = MLP(2, d_model, num_heads, 2, device=device, dtype=dtype, operations=operations)
+        self.boxRPB_embed_y = MLP(2, d_model, num_heads, 2, device=device, dtype=dtype, operations=operations)
+
+        self.presence_token = operations.Embedding(1, d_model, device=device, dtype=dtype)
+        self.presence_token_head = MLP(d_model, d_model, 1, 3, device=device, dtype=dtype, operations=operations)
+        self.presence_token_out_norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+
+    @staticmethod
+    def _inverse_sigmoid(x):
+        return torch.log(x / (1 - x + 1e-6) + 1e-6)
+
+    def _compute_box_rpb(self, ref_points, H, W):
+        """Box rotary position bias: (B, Q, 4) cxcywh -> (B, n_heads, Q+1, H*W) bias."""
+        boxes_xyxy = box_cxcywh_to_xyxy(ref_points)
+        B, Q, _ = boxes_xyxy.shape
+        coords_h = torch.arange(H, device=ref_points.device, dtype=torch.float32) / H
+        coords_w = torch.arange(W, device=ref_points.device, dtype=torch.float32) / W
+        deltas_x = coords_w.view(1, 1, -1, 1) - boxes_xyxy[:, :, None, 0:3:2]
+        deltas_y = coords_h.view(1, 1, -1, 1) - boxes_xyxy[:, :, None, 1:4:2]
+
+        log2_8 = float(math.log2(8))
+        def log_scale(d):
+            return torch.sign(d * 8) * torch.log2(torch.abs(d * 8) + 1.0) / log2_8
+
+        rpb_x = self.boxRPB_embed_x(log_scale(deltas_x).to(ref_points.dtype))
+        rpb_y = self.boxRPB_embed_y(log_scale(deltas_y).to(ref_points.dtype))
+
+        bias = (rpb_y.unsqueeze(3) + rpb_x.unsqueeze(2)).flatten(2, 3).permute(0, 3, 1, 2)
+        pres_bias = torch.zeros(B, bias.shape[1], 1, bias.shape[3], device=bias.device, dtype=bias.dtype)
+        return torch.cat([pres_bias, bias], dim=2)
+
+    def forward(self, memory, memory_pos, text_memory=None, text_mask=None, H=72, W=72):
+        B = memory.shape[0]
+        tgt = cast_to_input(self.query_embed.weight, memory).unsqueeze(0).expand(B, -1, -1)
+        presence_out = cast_to_input(self.presence_token.weight, memory)[None].expand(B, -1, -1)
+        ref_points = cast_to_input(self.reference_points.weight, memory).unsqueeze(0).expand(B, -1, -1).sigmoid()
+
+        for layer_idx, layer in enumerate(self.layers):
+            query_pos = self.ref_point_head(gen_sineembed_for_position(ref_points, self.d_model))
+            tgt_with_pres = torch.cat([presence_out, tgt], dim=1)
+            pos_with_pres = torch.cat([torch.zeros_like(presence_out), query_pos], dim=1)
+            tgt_with_pres = layer(tgt_with_pres, memory, pos_with_pres, memory_pos,
+                                  text_memory, text_mask, self._compute_box_rpb(ref_points, H, W))
+            presence_out, tgt = tgt_with_pres[:, :1], tgt_with_pres[:, 1:]
+            if layer_idx < len(self.layers) - 1:
+                ref_inv = self._inverse_sigmoid(ref_points)
+                ref_points = (ref_inv + self.bbox_embed(self.norm(tgt))).sigmoid().detach()
+
+        query_out = self.norm(tgt)
+        ref_inv = self._inverse_sigmoid(ref_points)
+        boxes = (ref_inv + self.bbox_embed(query_out)).sigmoid()
+        presence = self.presence_token_head(self.presence_token_out_norm(presence_out)).squeeze(-1)
+        return {"decoder_output": query_out, "pred_boxes": boxes, "presence": presence}
+
+
+class Transformer(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, dim_ff=2048, enc_layers=6, dec_layers=6,
+                 num_queries=200, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.encoder = TransformerEncoder(d_model, num_heads, dim_ff, enc_layers, device=device, dtype=dtype, operations=operations)
+        self.decoder = TransformerDecoder(d_model, num_heads, dim_ff, dec_layers, num_queries, device=device, dtype=dtype, operations=operations)
+
+
+class GeometryEncoder(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, num_layers=3, roi_size=7, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.d_model = d_model
+        self.roi_size = roi_size
+        self.pos_enc = PositionEmbeddingSine(num_pos_feats=d_model, normalize=True)
+        self.points_direct_project = operations.Linear(2, d_model, device=device, dtype=dtype)
+        self.points_pool_project = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.points_pos_enc_project = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.boxes_direct_project = operations.Linear(4, d_model, device=device, dtype=dtype)
+        self.boxes_pool_project = operations.Conv2d(d_model, d_model, kernel_size=roi_size, device=device, dtype=dtype)
+        self.boxes_pos_enc_project = operations.Linear(d_model + 2, d_model, device=device, dtype=dtype)
+        self.label_embed = operations.Embedding(2, d_model, device=device, dtype=dtype)
+        self.cls_embed = operations.Embedding(1, d_model, device=device, dtype=dtype)
+        self.norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.img_pre_norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.encode = nn.ModuleList([
+            EncoderLayer(d_model, num_heads, 2048, device=device, dtype=dtype, operations=operations)
+            for _ in range(num_layers)
+        ])
+        self.encode_norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.final_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+
+    def _encode_points(self, coords, labels, img_feat_2d):
+        """Encode point prompts: direct + pool + pos_enc + label. coords: [B, N, 2] normalized."""
+        B, N, _ = coords.shape
+        embed = self.points_direct_project(coords)
+        # Pool features from backbone at point locations via grid_sample
+        grid = (coords * 2 - 1).unsqueeze(2)  # [B, N, 1, 2] in [-1, 1]
+        sampled = F.grid_sample(img_feat_2d, grid, align_corners=False)  # [B, C, N, 1]
+        embed = embed + self.points_pool_project(sampled.squeeze(-1).permute(0, 2, 1))  # [B, N, C]
+        # Positional encoding of coordinates
+        x, y = coords[:, :, 0], coords[:, :, 1]  # [B, N]
+        pos_x, pos_y = self.pos_enc._encode_xy(x.flatten(), y.flatten())
+        enc = torch.cat([pos_x, pos_y], dim=-1).view(B, N, -1)
+        embed = embed + self.points_pos_enc_project(cast_to_input(enc, embed))
+        embed = embed + cast_to_input(self.label_embed(labels.long()), embed)
+        return embed
+
+    def _encode_boxes(self, boxes, labels, img_feat_2d):
+        """Encode box prompts: direct + pool + pos_enc + label. boxes: [B, N, 4] normalized cxcywh."""
+        B, N, _ = boxes.shape
+        embed = self.boxes_direct_project(boxes)
+        # ROI align from backbone at box regions
+        H, W = img_feat_2d.shape[-2:]
+        boxes_xyxy = box_cxcywh_to_xyxy(boxes)
+        scale = torch.tensor([W, H, W, H], dtype=boxes_xyxy.dtype, device=boxes_xyxy.device)
+        boxes_scaled = boxes_xyxy * scale
+        sampled = roi_align(img_feat_2d, boxes_scaled.view(-1, 4).split(N), self.roi_size)
+        proj = self.boxes_pool_project(sampled).view(B, N, -1)  # Conv2d(roi_size) -> [B*N, C, 1, 1] -> [B, N, C]
+        embed = embed + proj
+        # Positional encoding of box center + size
+        cx, cy, w, h = boxes[:, :, 0], boxes[:, :, 1], boxes[:, :, 2], boxes[:, :, 3]
+        enc = self.pos_enc.encode_boxes(cx.flatten(), cy.flatten(), w.flatten(), h.flatten())
+        enc = enc.view(B, N, -1)
+        embed = embed + self.boxes_pos_enc_project(cast_to_input(enc, embed))
+        embed = embed + cast_to_input(self.label_embed(labels.long()), embed)
+        return embed
+
+    def forward(self, points=None, boxes=None, image_features=None):
+        """Encode geometry prompts. image_features: [B, HW, C] flattened backbone features."""
+        # Prepare 2D image features for pooling
+        img_feat_2d = None
+        if image_features is not None:
+            B = image_features.shape[0]
+            HW, C = image_features.shape[1], image_features.shape[2]
+            hw = int(math.sqrt(HW))
+            img_normed = self.img_pre_norm(image_features)
+            img_feat_2d = img_normed.permute(0, 2, 1).view(B, C, hw, hw)
+
+        embeddings = []
+        if points is not None:
+            coords, labels = points
+            embeddings.append(self._encode_points(coords, labels, img_feat_2d))
+        if boxes is not None:
+            B = boxes.shape[0]
+            box_labels = torch.ones(B, boxes.shape[1], dtype=torch.long, device=boxes.device)
+            embeddings.append(self._encode_boxes(boxes, box_labels, img_feat_2d))
+        if not embeddings:
+            return None
+        geo = torch.cat(embeddings, dim=1)
+        geo = self.norm(geo)
+        if image_features is not None:
+            for layer in self.encode:
+                geo = layer(geo, torch.zeros_like(geo), image_features)
+        geo = self.encode_norm(geo)
+        return self.final_proj(geo)
+
+
+class PixelDecoder(nn.Module):
+    """Top-down FPN pixel decoder with GroupNorm + ReLU + nearest interpolation."""
+    def __init__(self, d_model=256, num_stages=3, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.conv_layers = nn.ModuleList([operations.Conv2d(d_model, d_model, kernel_size=3, padding=1, device=device, dtype=dtype) for _ in range(num_stages)])
+        self.norms = nn.ModuleList([operations.GroupNorm(8, d_model, device=device, dtype=dtype) for _ in range(num_stages)])
+
+    def forward(self, backbone_features):
+        prev = backbone_features[-1]
+        for i, feat in enumerate(backbone_features[:-1][::-1]):
+            prev = F.relu(self.norms[i](self.conv_layers[i](feat + F.interpolate(prev, size=feat.shape[-2:], mode="nearest"))))
+        return prev
+
+
+class MaskPredictor(nn.Module):
+    def __init__(self, d_model=256, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.mask_embed = MLP(d_model, d_model, d_model, 3, device=device, dtype=dtype, operations=operations)
+
+    def forward(self, query_embeddings, pixel_features):
+        mask_embed = self.mask_embed(query_embeddings)
+        return torch.einsum("bqc,bchw->bqhw", mask_embed, pixel_features)
+
+
+class SegmentationHead(nn.Module):
+    def __init__(self, d_model=256, num_heads=8, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.d_model = d_model
+        self.pixel_decoder = PixelDecoder(d_model, 3, device=device, dtype=dtype, operations=operations)
+        self.mask_predictor = MaskPredictor(d_model, device=device, dtype=dtype, operations=operations)
+        self.cross_attend_prompt = SplitMHA(d_model, num_heads, device=device, dtype=dtype, operations=operations)
+        self.cross_attn_norm = operations.LayerNorm(d_model, device=device, dtype=dtype)
+        self.instance_seg_head = operations.Conv2d(d_model, d_model, kernel_size=1, device=device, dtype=dtype)
+        self.semantic_seg_head = operations.Conv2d(d_model, 1, kernel_size=1, device=device, dtype=dtype)
+
+    def forward(self, query_embeddings, backbone_features, encoder_hidden_states=None, prompt=None, prompt_mask=None):
+        if encoder_hidden_states is not None and prompt is not None:
+            enc_normed = self.cross_attn_norm(encoder_hidden_states)
+            enc_cross = self.cross_attend_prompt(enc_normed, prompt, prompt, mask=prompt_mask)
+            encoder_hidden_states = enc_cross + encoder_hidden_states
+
+        if encoder_hidden_states is not None:
+            B, H, W = encoder_hidden_states.shape[0], backbone_features[-1].shape[-2], backbone_features[-1].shape[-1]
+            encoder_visual = encoder_hidden_states[:, :H * W].permute(0, 2, 1).view(B, self.d_model, H, W)
+            backbone_features = list(backbone_features)
+            backbone_features[-1] = encoder_visual
+
+        pixel_features = self.pixel_decoder(backbone_features)
+        instance_features = self.instance_seg_head(pixel_features)
+        masks = self.mask_predictor(query_embeddings, instance_features)
+        return masks
+
+
+class DotProductScoring(nn.Module):
+    def __init__(self, d_model=256, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.hs_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.prompt_proj = operations.Linear(d_model, d_model, device=device, dtype=dtype)
+        self.prompt_mlp = MLPWithNorm(d_model, 2048, d_model, 2, device=device, dtype=dtype, operations=operations)
+        self.scale = 1.0 / (d_model ** 0.5)
+
+    def forward(self, query_embeddings, prompt_embeddings, prompt_mask=None):
+        prompt = self.prompt_mlp(prompt_embeddings)
+        if prompt_mask is not None:
+            weight = prompt_mask.unsqueeze(-1).to(dtype=prompt.dtype)
+            pooled = (prompt * weight).sum(dim=1) / weight.sum(dim=1).clamp(min=1)
+        else:
+            pooled = prompt.mean(dim=1)
+        hs = self.hs_proj(query_embeddings)
+        pp = self.prompt_proj(pooled).unsqueeze(-1).to(hs.dtype)
+        scores = torch.matmul(hs, pp)
+        return (scores * self.scale).clamp(-12.0, 12.0).squeeze(-1)
+
+
+class SAM3Detector(nn.Module):
+    def __init__(self, d_model=256, embed_dim=1024, num_queries=200, device=None, dtype=None, operations=None, **kwargs):
+        super().__init__()
+        image_model = kwargs.pop("image_model", "SAM3")
+        for k in ("num_heads", "num_head_channels"):
+            kwargs.pop(k, None)
+        multiplex = image_model == "SAM31"
+        # SAM3: 4 FPN levels, drop last (scalp=1); SAM3.1: 3 levels, use all (scalp=0)
+        self.scalp = 0 if multiplex else 1
+        self.backbone = nn.ModuleDict({
+            "vision_backbone": SAM3VisionBackbone(embed_dim=embed_dim, d_model=d_model, multiplex=multiplex, device=device, dtype=dtype, operations=operations, **kwargs),
+            "language_backbone": nn.ModuleDict({"resizer": operations.Linear(embed_dim, d_model, device=device, dtype=dtype)}),
+        })
+        self.transformer = Transformer(d_model=d_model, num_queries=num_queries, device=device, dtype=dtype, operations=operations)
+        self.segmentation_head = SegmentationHead(d_model=d_model, device=device, dtype=dtype, operations=operations)
+        self.geometry_encoder = GeometryEncoder(d_model=d_model, device=device, dtype=dtype, operations=operations)
+        self.dot_prod_scoring = DotProductScoring(d_model=d_model, device=device, dtype=dtype, operations=operations)
+
+    def _get_backbone_features(self, images):
+        """Run backbone and return (detector_features, detector_positions, tracker_features, tracker_positions)."""
+        bb = self.backbone["vision_backbone"]
+        if bb.multiplex:
+            all_f, all_p, tf, tp = bb(images, tracker_mode="propagation")
+        else:
+            all_f, all_p, tf, tp = bb(images, need_tracker=True)
+        return all_f, all_p, tf, tp
+
+    @staticmethod
+    def _run_geo_layer(layer, x, memory, memory_pos):
+        x = x + layer.self_attn(layer.norm1(x))
+        x = x + layer.cross_attn_image(layer.norm2(x), memory + memory_pos, memory)
+        x = x + layer.linear2(F.relu(layer.linear1(layer.norm3(x))))
+        return x
+
+    def _detect(self, features, positions, text_embeddings=None, text_mask=None,
+                points=None, boxes=None):
+        """Shared detection: geometry encoding, transformer, scoring, segmentation."""
+        B = features[0].shape[0]
+        # Scalp for encoder (use top-level feature), but keep all levels for segmentation head
+        seg_features = features
+        if self.scalp > 0:
+            features = features[:-self.scalp]
+            positions = positions[:-self.scalp]
+        enc_feat, enc_pos = features[-1], positions[-1]
+        _, _, H, W = enc_feat.shape
+        img_flat = enc_feat.flatten(2).permute(0, 2, 1)
+        pos_flat = enc_pos.flatten(2).permute(0, 2, 1)
+
+        has_prompts = text_embeddings is not None or points is not None or boxes is not None
+        if has_prompts:
+            geo_enc = self.geometry_encoder
+            geo_prompts = geo_enc(points=points, boxes=boxes, image_features=img_flat)
+            geo_cls = geo_enc.norm(geo_enc.final_proj(cast_to_input(geo_enc.cls_embed.weight, img_flat).view(1, 1, -1).expand(B, -1, -1)))
+            for layer in geo_enc.encode:
+                geo_cls = self._run_geo_layer(layer, geo_cls, img_flat, pos_flat)
+            geo_cls = geo_enc.encode_norm(geo_cls)
+            if text_embeddings is not None and text_embeddings.shape[0] != B:
+                text_embeddings = text_embeddings.expand(B, -1, -1)
+            if text_mask is not None and text_mask.shape[0] != B:
+                text_mask = text_mask.expand(B, -1)
+            parts = [t for t in [text_embeddings, geo_prompts, geo_cls] if t is not None]
+            text_embeddings = torch.cat(parts, dim=1)
+            n_new = text_embeddings.shape[1] - (text_mask.shape[1] if text_mask is not None else 0)
+            if text_mask is not None:
+                text_mask = torch.cat([text_mask, torch.ones(B, n_new, dtype=torch.bool, device=text_mask.device)], dim=1)
+            else:
+                text_mask = torch.ones(B, text_embeddings.shape[1], dtype=torch.bool, device=text_embeddings.device)
+
+        memory = self.transformer.encoder(img_flat, pos_flat, text_embeddings, text_mask)
+        dec_out = self.transformer.decoder(memory, pos_flat, text_embeddings, text_mask, H, W)
+        query_out, pred_boxes = dec_out["decoder_output"], dec_out["pred_boxes"]
+
+        if text_embeddings is not None:
+            scores = self.dot_prod_scoring(query_out, text_embeddings, text_mask)
+        else:
+            scores = torch.zeros(B, query_out.shape[1], device=query_out.device)
+
+        masks = self.segmentation_head(query_out, seg_features, encoder_hidden_states=memory, prompt=text_embeddings, prompt_mask=text_mask)
+        return box_cxcywh_to_xyxy(pred_boxes), scores, masks, dec_out
+
+    def forward(self, images, text_embeddings=None, text_mask=None, points=None, boxes=None, threshold=0.3, orig_size=None):
+        features, positions, _, _ = self._get_backbone_features(images)
+
+        if text_embeddings is not None:
+            text_embeddings = self.backbone["language_backbone"]["resizer"](text_embeddings)
+            if text_mask is not None:
+                text_mask = text_mask.bool()
+
+        boxes_xyxy, scores, masks, dec_out = self._detect(
+            features, positions, text_embeddings, text_mask, points, boxes)
+
+        if orig_size is not None:
+            oh, ow = orig_size
+            boxes_xyxy = boxes_xyxy * torch.tensor([ow, oh, ow, oh], device=boxes_xyxy.device, dtype=boxes_xyxy.dtype)
+            masks = F.interpolate(masks, size=orig_size, mode="bilinear", align_corners=False)
+
+        return {
+            "boxes": boxes_xyxy,
+            "scores": scores,
+            "masks": masks,
+            "presence": dec_out.get("presence"),
+        }
+
+    def forward_from_trunk(self, trunk_out, text_embeddings, text_mask):
+        """Run detection using a pre-computed ViTDet trunk output.
+
+        text_embeddings must already be resized through language_backbone.resizer.
+        Returns dict with boxes (normalized xyxy), scores, masks at detector resolution.
+        """
+        bb = self.backbone["vision_backbone"]
+        features = [conv(trunk_out) for conv in bb.convs]
+        positions = [cast_to_input(bb.position_encoding(f), f) for f in features]
+
+        if text_mask is not None:
+            text_mask = text_mask.bool()
+
+        boxes_xyxy, scores, masks, _ = self._detect(features, positions, text_embeddings, text_mask)
+        return {"boxes": boxes_xyxy, "scores": scores, "masks": masks}
+
+
+class SAM3Model(nn.Module):
+    def __init__(self, device=None, dtype=None, operations=None, **kwargs):
+        super().__init__()
+        self.dtype = dtype
+        image_model = kwargs.get("image_model", "SAM3")
+        tracker_cls = TRACKER_CLASSES[image_model]
+        self.detector = SAM3Detector(device=device, dtype=dtype, operations=operations, **kwargs)
+        self.tracker = tracker_cls(device=device, dtype=dtype, operations=operations, **kwargs)
+
+    def forward(self, images, **kwargs):
+        return self.detector(images, **kwargs)
+
+    def forward_segment(self, images, point_inputs=None, box_inputs=None, mask_inputs=None):
+        """Interactive segmentation using SAM decoder with point/box/mask prompts.
+
+        Args:
+            images: [B, 3, 1008, 1008] preprocessed images
+            point_inputs: {"point_coords": [B, N, 2], "point_labels": [B, N]} in 1008x1008 pixel space
+            box_inputs: [B, 2, 2] box corners (top-left, bottom-right) in 1008x1008 pixel space
+            mask_inputs: [B, 1, H, W] coarse mask logits to refine
+        Returns:
+            [B, 1, image_size, image_size] high-res mask logits
+        """
+        bb = self.detector.backbone["vision_backbone"]
+        if bb.multiplex:
+            _, _, tracker_features, tracker_positions = bb(images, tracker_mode="interactive")
+        else:
+            _, _, tracker_features, tracker_positions = bb(images, need_tracker=True)
+            if self.detector.scalp > 0:
+                tracker_features = tracker_features[:-self.detector.scalp]
+                tracker_positions = tracker_positions[:-self.detector.scalp]
+
+        high_res = list(tracker_features[:-1])
+        backbone_feat = tracker_features[-1]
+        B, C, H, W = backbone_feat.shape
+        # Add no-memory embedding (init frame path)
+        no_mem = getattr(self.tracker, 'interactivity_no_mem_embed', None)
+        if no_mem is None:
+            no_mem = getattr(self.tracker, 'no_mem_embed', None)
+        if no_mem is not None:
+            feat_flat = backbone_feat.flatten(2).permute(0, 2, 1)
+            feat_flat = feat_flat + cast_to_input(no_mem, feat_flat)
+            backbone_feat = feat_flat.view(B, H, W, C).permute(0, 3, 1, 2)
+
+        num_pts = 0 if point_inputs is None else point_inputs["point_labels"].size(1)
+        _, high_res_masks, _, _ = self.tracker._forward_sam_heads(
+            backbone_features=backbone_feat,
+            point_inputs=point_inputs,
+            mask_inputs=mask_inputs,
+            box_inputs=box_inputs,
+            high_res_features=high_res,
+            multimask_output=(0 < num_pts <= 1),
+        )
+        return high_res_masks
+
+    def forward_video(self, images, initial_masks, pbar=None, text_prompts=None,
+                       new_det_thresh=0.5, max_objects=0, detect_interval=1):
+        """Track video with optional per-frame text-prompted detection."""
+        bb = self.detector.backbone["vision_backbone"]
+
+        def backbone_fn(frame, frame_idx=None):
+            trunk_out = bb.trunk(frame)
+            if bb.multiplex:
+                _, _, tf, tp = bb(frame, tracker_mode="propagation", cached_trunk=trunk_out, tracker_only=True)
+            else:
+                _, _, tf, tp = bb(frame, need_tracker=True, cached_trunk=trunk_out, tracker_only=True)
+            return tf, tp, trunk_out
+
+        detect_fn = None
+        if text_prompts:
+            resizer = self.detector.backbone["language_backbone"]["resizer"]
+            resized = [(resizer(emb), m.bool() if m is not None else None) for emb, m in text_prompts]
+            def detect_fn(trunk_out):
+                all_scores, all_masks = [], []
+                for emb, mask in resized:
+                    det = self.detector.forward_from_trunk(trunk_out, emb, mask)
+                    all_scores.append(det["scores"])
+                    all_masks.append(det["masks"])
+                return {"scores": torch.cat(all_scores, dim=1), "masks": torch.cat(all_masks, dim=1)}
+
+        if hasattr(self.tracker, 'track_video_with_detection'):
+            return self.tracker.track_video_with_detection(
+                backbone_fn, images, initial_masks, detect_fn,
+                new_det_thresh=new_det_thresh, max_objects=max_objects,
+                detect_interval=detect_interval, backbone_obj=bb, pbar=pbar)
+        # SAM3 (non-multiplex) — no detection support, requires initial masks
+        if initial_masks is None:
+            raise ValueError("SAM3 (non-multiplex) requires initial_mask for video tracking")
+        return self.tracker.track_video(backbone_fn, images, initial_masks, pbar=pbar, backbone_obj=bb)

comfy/ldm/sam3/sam.py

@@ -0,0 +1,425 @@
+# SAM3 shared components: primitives, ViTDet backbone, FPN neck, position encodings.
+
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.math import apply_rope
+from comfy.ldm.flux.layers import EmbedND
+from comfy.ops import cast_to_input
+
+
+class MLP(nn.Module):
+    def __init__(self, input_dim, hidden_dim, output_dim, num_layers, sigmoid_output=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        dims = [input_dim] + [hidden_dim] * (num_layers - 1) + [output_dim]
+        self.layers = nn.ModuleList([operations.Linear(dims[i], dims[i + 1], device=device, dtype=dtype) for i in range(num_layers)])
+        self.sigmoid_output = sigmoid_output
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = F.relu(layer(x)) if i < len(self.layers) - 1 else layer(x)
+        return torch.sigmoid(x) if self.sigmoid_output else x
+
+
+class SAMAttention(nn.Module):
+    def __init__(self, embedding_dim, num_heads, downsample_rate=1, kv_in_dim=None, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.num_heads = num_heads
+        internal_dim = embedding_dim // downsample_rate
+        kv_dim = kv_in_dim if kv_in_dim is not None else embedding_dim
+        self.q_proj = operations.Linear(embedding_dim, internal_dim, device=device, dtype=dtype)
+        self.k_proj = operations.Linear(kv_dim, internal_dim, device=device, dtype=dtype)
+        self.v_proj = operations.Linear(kv_dim, internal_dim, device=device, dtype=dtype)
+        self.out_proj = operations.Linear(internal_dim, embedding_dim, device=device, dtype=dtype)
+
+    def forward(self, q, k, v):
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+        return self.out_proj(optimized_attention(q, k, v, self.num_heads, low_precision_attention=False))
+
+
+class TwoWayAttentionBlock(nn.Module):
+    def __init__(self, embedding_dim, num_heads, mlp_dim=2048, attention_downsample_rate=2, skip_first_layer_pe=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.skip_first_layer_pe = skip_first_layer_pe
+        self.self_attn = SAMAttention(embedding_dim, num_heads, device=device, dtype=dtype, operations=operations)
+        self.cross_attn_token_to_image = SAMAttention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate, device=device, dtype=dtype, operations=operations)
+        self.cross_attn_image_to_token = SAMAttention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate, device=device, dtype=dtype, operations=operations)
+        self.mlp = nn.Sequential(operations.Linear(embedding_dim, mlp_dim, device=device, dtype=dtype), nn.ReLU(), operations.Linear(mlp_dim, embedding_dim, device=device, dtype=dtype))
+        self.norm1 = operations.LayerNorm(embedding_dim, device=device, dtype=dtype)
+        self.norm2 = operations.LayerNorm(embedding_dim, device=device, dtype=dtype)
+        self.norm3 = operations.LayerNorm(embedding_dim, device=device, dtype=dtype)
+        self.norm4 = operations.LayerNorm(embedding_dim, device=device, dtype=dtype)
+
+    def forward(self, queries, keys, query_pe, key_pe):
+        if self.skip_first_layer_pe:
+            queries = self.norm1(self.self_attn(queries, queries, queries))
+        else:
+            q = queries + query_pe
+            queries = self.norm1(queries + self.self_attn(q, q, queries))
+        q, k = queries + query_pe, keys + key_pe
+        queries = self.norm2(queries + self.cross_attn_token_to_image(q, k, keys))
+        queries = self.norm3(queries + self.mlp(queries))
+        q, k = queries + query_pe, keys + key_pe
+        keys = self.norm4(keys + self.cross_attn_image_to_token(k, q, queries))
+        return queries, keys
+
+
+class TwoWayTransformer(nn.Module):
+    def __init__(self, depth=2, embedding_dim=256, num_heads=8, mlp_dim=2048, attention_downsample_rate=2, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.layers = nn.ModuleList([
+            TwoWayAttentionBlock(embedding_dim, num_heads, mlp_dim, attention_downsample_rate,
+                                 skip_first_layer_pe=(i == 0), device=device, dtype=dtype, operations=operations)
+            for i in range(depth)
+        ])
+        self.final_attn_token_to_image = SAMAttention(embedding_dim, num_heads, downsample_rate=attention_downsample_rate, device=device, dtype=dtype, operations=operations)
+        self.norm_final = operations.LayerNorm(embedding_dim, device=device, dtype=dtype)
+
+    def forward(self, image_embedding, image_pe, point_embedding):
+        queries, keys = point_embedding, image_embedding
+        for layer in self.layers:
+            queries, keys = layer(queries, keys, point_embedding, image_pe)
+        q, k = queries + point_embedding, keys + image_pe
+        queries = self.norm_final(queries + self.final_attn_token_to_image(q, k, keys))
+        return queries, keys
+
+
+class PositionEmbeddingRandom(nn.Module):
+    """Fourier feature positional encoding with random gaussian projection."""
+    def __init__(self, num_pos_feats=64, scale=None):
+        super().__init__()
+        self.register_buffer("positional_encoding_gaussian_matrix", (scale or 1.0) * torch.randn(2, num_pos_feats))
+
+    def _encode(self, normalized_coords):
+        """Map normalized [0,1] coordinates to fourier features via random projection. Computes in fp32."""
+        orig_dtype = normalized_coords.dtype
+        proj_matrix = self.positional_encoding_gaussian_matrix.to(device=normalized_coords.device, dtype=torch.float32)
+        projected = 2 * math.pi * (2 * normalized_coords.float() - 1) @ proj_matrix
+        return torch.cat([projected.sin(), projected.cos()], dim=-1).to(orig_dtype)
+
+    def forward(self, size, device=None):
+        h, w = size
+        dev = device if device is not None else self.positional_encoding_gaussian_matrix.device
+        ones = torch.ones((h, w), device=dev, dtype=torch.float32)
+        norm_xy = torch.stack([(ones.cumsum(1) - 0.5) / w, (ones.cumsum(0) - 0.5) / h], dim=-1)
+        return self._encode(norm_xy).permute(2, 0, 1).unsqueeze(0)
+
+    def forward_with_coords(self, pixel_coords, image_size):
+        norm = pixel_coords.clone()
+        norm[:, :, 0] /= image_size[1]
+        norm[:, :, 1] /= image_size[0]
+        return self._encode(norm)
+
+
+# ViTDet backbone + FPN neck
+
+def window_partition(x: torch.Tensor, window_size: int):
+    B, H, W, C = x.shape
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(windows: torch.Tensor, window_size: int, pad_hw, hw):
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+def rope_2d(end_x: int, end_y: int, dim: int, theta: float = 10000.0, scale_pos: float = 1.0):
+    """Generate 2D axial RoPE using flux EmbedND. Returns [1, 1, HW, dim//2, 2, 2]."""
+    t = torch.arange(end_x * end_y, dtype=torch.float32)
+    ids = torch.stack([(t % end_x) * scale_pos,
+                       torch.div(t, end_x, rounding_mode="floor") * scale_pos], dim=-1)
+    return EmbedND(dim=dim, theta=theta, axes_dim=[dim // 2, dim // 2])(ids.unsqueeze(0))
+
+
+class _ViTMLP(nn.Module):
+    def __init__(self, dim, mlp_ratio=4.0, device=None, dtype=None, operations=None):
+        super().__init__()
+        hidden = int(dim * mlp_ratio)
+        self.fc1 = operations.Linear(dim, hidden, device=device, dtype=dtype)
+        self.act = nn.GELU()
+        self.fc2 = operations.Linear(hidden, dim, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.fc2(self.act(self.fc1(x)))
+
+
+class Attention(nn.Module):
+    """ViTDet multi-head attention with fused QKV projection."""
+
+    def __init__(self, dim, num_heads=8, qkv_bias=True, use_rope=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.use_rope = use_rope
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, device=device, dtype=dtype)
+        self.proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+
+    def forward(self, x, freqs_cis=None):
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, self.head_dim)
+        q, k, v = qkv.permute(2, 0, 3, 1, 4).unbind(dim=0)
+        if self.use_rope and freqs_cis is not None:
+            q, k = apply_rope(q, k, freqs_cis)
+        return self.proj(optimized_attention(q, k, v, self.num_heads, skip_reshape=True, low_precision_attention=False))
+
+
+class Block(nn.Module):
+    def __init__(self, dim, num_heads, mlp_ratio=4.0, qkv_bias=True, window_size=0, use_rope=False, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.window_size = window_size
+        self.norm1 = operations.LayerNorm(dim, device=device, dtype=dtype)
+        self.attn = Attention(dim, num_heads, qkv_bias, use_rope, device=device, dtype=dtype, operations=operations)
+        self.norm2 = operations.LayerNorm(dim, device=device, dtype=dtype)
+        self.mlp = _ViTMLP(dim, mlp_ratio, device=device, dtype=dtype, operations=operations)
+
+    def forward(self, x, freqs_cis=None):
+        shortcut = x
+        x = self.norm1(x)
+        if self.window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, self.window_size)
+            x = x.view(x.shape[0], self.window_size * self.window_size, -1)
+            x = self.attn(x, freqs_cis=freqs_cis)
+            x = x.view(-1, self.window_size, self.window_size, x.shape[-1])
+            x = window_unpartition(x, self.window_size, pad_hw, (H, W))
+        else:
+            B, H, W, C = x.shape
+            x = x.view(B, H * W, C)
+            x = self.attn(x, freqs_cis=freqs_cis)
+            x = x.view(B, H, W, C)
+        x = shortcut + x
+        x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    def __init__(self, patch_size=14, in_chans=3, embed_dim=1024, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.proj = operations.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size, bias=False, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.proj(x)
+
+
+class ViTDet(nn.Module):
+    def __init__(self, img_size=1008, patch_size=14, embed_dim=1024, depth=32, num_heads=16, mlp_ratio=4.625, qkv_bias=True, window_size=24,
+                 global_att_blocks=(7, 15, 23, 31), use_rope=True, pretrain_img_size=336, device=None, dtype=None, operations=None, **kwargs):
+        super().__init__()
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.global_att_blocks = set(global_att_blocks)
+
+        self.patch_embed = PatchEmbed(patch_size, 3, embed_dim, device=device, dtype=dtype, operations=operations)
+
+        num_patches = (pretrain_img_size // patch_size) ** 2 + 1  # +1 for cls token
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim, device=device, dtype=dtype))
+
+        self.ln_pre = operations.LayerNorm(embed_dim, device=device, dtype=dtype)
+
+        grid_size = img_size // patch_size
+        pretrain_grid = pretrain_img_size // patch_size
+
+        self.blocks = nn.ModuleList()
+        for i in range(depth):
+            is_global = i in self.global_att_blocks
+            self.blocks.append(Block(
+                embed_dim, num_heads, mlp_ratio, qkv_bias,
+                window_size=0 if is_global else window_size,
+                use_rope=use_rope,
+                device=device, dtype=dtype, operations=operations,
+            ))
+
+        if use_rope:
+            rope_scale = pretrain_grid / grid_size
+            self.register_buffer("freqs_cis", rope_2d(grid_size, grid_size, embed_dim // num_heads, scale_pos=rope_scale), persistent=False)
+            self.register_buffer("freqs_cis_window", rope_2d(window_size, window_size, embed_dim // num_heads), persistent=False)
+        else:
+            self.freqs_cis = None
+            self.freqs_cis_window = None
+
+    def _get_pos_embed(self, num_tokens):
+        pos = self.pos_embed
+        if pos.shape[1] == num_tokens:
+            return pos
+        cls_pos = pos[:, :1]
+        spatial_pos = pos[:, 1:]
+        old_size = int(math.sqrt(spatial_pos.shape[1]))
+        new_size = int(math.sqrt(num_tokens - 1)) if num_tokens > 1 else old_size
+        spatial_2d = spatial_pos.reshape(1, old_size, old_size, -1).permute(0, 3, 1, 2)
+        tiles_h = new_size // old_size + 1
+        tiles_w = new_size // old_size + 1
+        tiled = spatial_2d.tile([1, 1, tiles_h, tiles_w])[:, :, :new_size, :new_size]
+        tiled = tiled.permute(0, 2, 3, 1).reshape(1, new_size * new_size, -1)
+        return torch.cat([cls_pos, tiled], dim=1)
+
+    def forward(self, x):
+        x = self.patch_embed(x)
+        B, C, Hp, Wp = x.shape
+        x = x.permute(0, 2, 3, 1).reshape(B, Hp * Wp, C)
+
+        pos = cast_to_input(self._get_pos_embed(Hp * Wp + 1), x)
+        x = x + pos[:, 1:Hp * Wp + 1]
+
+        x = x.view(B, Hp, Wp, C)
+        x = self.ln_pre(x)
+
+        freqs_cis_global = self.freqs_cis
+        freqs_cis_win = self.freqs_cis_window
+        if freqs_cis_global is not None:
+            freqs_cis_global = cast_to_input(freqs_cis_global, x)
+        if freqs_cis_win is not None:
+            freqs_cis_win = cast_to_input(freqs_cis_win, x)
+
+        for block in self.blocks:
+            fc = freqs_cis_win if block.window_size > 0 else freqs_cis_global
+            x = block(x, freqs_cis=fc)
+
+        return x.permute(0, 3, 1, 2)
+
+
+class FPNScaleConv(nn.Module):
+    def __init__(self, in_dim, out_dim, scale, device=None, dtype=None, operations=None):
+        super().__init__()
+        if scale == 4.0:
+            self.dconv_2x2_0 = operations.ConvTranspose2d(in_dim, in_dim // 2, kernel_size=2, stride=2, device=device, dtype=dtype)
+            self.dconv_2x2_1 = operations.ConvTranspose2d(in_dim // 2, in_dim // 4, kernel_size=2, stride=2, device=device, dtype=dtype)
+            proj_in = in_dim // 4
+        elif scale == 2.0:
+            self.dconv_2x2 = operations.ConvTranspose2d(in_dim, in_dim // 2, kernel_size=2, stride=2, device=device, dtype=dtype)
+            proj_in = in_dim // 2
+        elif scale == 1.0:
+            proj_in = in_dim
+        elif scale == 0.5:
+            self.pool = nn.MaxPool2d(kernel_size=2, stride=2)
+            proj_in = in_dim
+        self.scale = scale
+        self.conv_1x1 = operations.Conv2d(proj_in, out_dim, kernel_size=1, device=device, dtype=dtype)
+        self.conv_3x3 = operations.Conv2d(out_dim, out_dim, kernel_size=3, padding=1, device=device, dtype=dtype)
+
+    def forward(self, x):
+        if self.scale == 4.0:
+            x = F.gelu(self.dconv_2x2_0(x))
+            x = self.dconv_2x2_1(x)
+        elif self.scale == 2.0:
+            x = self.dconv_2x2(x)
+        elif self.scale == 0.5:
+            x = self.pool(x)
+        x = self.conv_1x1(x)
+        x = self.conv_3x3(x)
+        return x
+
+
+class PositionEmbeddingSine(nn.Module):
+    """2D sinusoidal position encoding (DETR-style) with result caching."""
+    def __init__(self, num_pos_feats=256, temperature=10000.0, normalize=True, scale=None):
+        super().__init__()
+        assert num_pos_feats % 2 == 0
+        self.half_dim = num_pos_feats // 2
+        self.temperature = temperature
+        self.normalize = normalize
+        self.scale = scale if scale is not None else 2 * math.pi
+        self._cache = {}
+
+    def _sincos(self, vals):
+        """Encode 1D values to interleaved sin/cos features."""
+        freqs = self.temperature ** (2 * (torch.arange(self.half_dim, dtype=torch.float32, device=vals.device) // 2) / self.half_dim)
+        raw = vals[..., None] * self.scale / freqs
+        return torch.stack((raw[..., 0::2].sin(), raw[..., 1::2].cos()), dim=-1).flatten(-2)
+
+    def _encode_xy(self, x, y):
+        """Encode normalized x, y coordinates to sinusoidal features. Returns (pos_x, pos_y) each [N, half_dim]."""
+        dim_t = self.temperature ** (2 * (torch.arange(self.half_dim, dtype=torch.float32, device=x.device) // 2) / self.half_dim)
+        pos_x = x[:, None] * self.scale / dim_t
+        pos_y = y[:, None] * self.scale / dim_t
+        pos_x = torch.stack((pos_x[:, 0::2].sin(), pos_x[:, 1::2].cos()), dim=2).flatten(1)
+        pos_y = torch.stack((pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2).flatten(1)
+        return pos_x, pos_y
+
+    def encode_boxes(self, cx, cy, w, h):
+        """Encode box center + size to [N, d_model+2] features."""
+        pos_x, pos_y = self._encode_xy(cx, cy)
+        return torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        key = (H, W, x.device)
+        if key not in self._cache:
+            gy = torch.arange(H, dtype=torch.float32, device=x.device)
+            gx = torch.arange(W, dtype=torch.float32, device=x.device)
+            if self.normalize:
+                gy, gx = gy / (H - 1 + 1e-6), gx / (W - 1 + 1e-6)
+            yy, xx = torch.meshgrid(gy, gx, indexing="ij")
+            self._cache[key] = torch.cat((self._sincos(yy), self._sincos(xx)), dim=-1).permute(2, 0, 1).unsqueeze(0)
+        return self._cache[key].expand(B, -1, -1, -1)
+
+
+class SAM3VisionBackbone(nn.Module):
+    def __init__(self, embed_dim=1024, d_model=256, multiplex=False, device=None, dtype=None, operations=None, **kwargs):
+        super().__init__()
+        self.trunk = ViTDet(embed_dim=embed_dim, device=device, dtype=dtype, operations=operations, **kwargs)
+        self.position_encoding = PositionEmbeddingSine(num_pos_feats=d_model, normalize=True)
+        self.multiplex = multiplex
+
+        fpn_args = dict(device=device, dtype=dtype, operations=operations)
+        if multiplex:
+            scales = [4.0, 2.0, 1.0]
+            self.convs = nn.ModuleList([FPNScaleConv(embed_dim, d_model, s, **fpn_args) for s in scales])
+            self.propagation_convs = nn.ModuleList([FPNScaleConv(embed_dim, d_model, s, **fpn_args) for s in scales])
+            self.interactive_convs = nn.ModuleList([FPNScaleConv(embed_dim, d_model, s, **fpn_args) for s in scales])
+        else:
+            scales = [4.0, 2.0, 1.0, 0.5]
+            self.convs = nn.ModuleList([FPNScaleConv(embed_dim, d_model, s, **fpn_args) for s in scales])
+            self.sam2_convs = nn.ModuleList([FPNScaleConv(embed_dim, d_model, s, **fpn_args) for s in scales])
+
+    def forward(self, images, need_tracker=False, tracker_mode=None, cached_trunk=None, tracker_only=False):
+        backbone_out = cached_trunk if cached_trunk is not None else self.trunk(images)
+
+        if tracker_only:
+            # Skip detector FPN when only tracker features are needed (video tracking)
+            if self.multiplex:
+                tracker_convs = self.propagation_convs if tracker_mode == "propagation" else self.interactive_convs
+            else:
+                tracker_convs = self.sam2_convs
+            tracker_features = [conv(backbone_out) for conv in tracker_convs]
+            tracker_positions = [cast_to_input(self.position_encoding(f), f) for f in tracker_features]
+            return None, None, tracker_features, tracker_positions
+
+        features = [conv(backbone_out) for conv in self.convs]
+        positions = [cast_to_input(self.position_encoding(f), f) for f in features]
+
+        if self.multiplex:
+            if tracker_mode == "propagation":
+                tracker_convs = self.propagation_convs
+            elif tracker_mode == "interactive":
+                tracker_convs = self.interactive_convs
+            else:
+                return features, positions, None, None
+        elif need_tracker:
+            tracker_convs = self.sam2_convs
+        else:
+            return features, positions, None, None
+
+        tracker_features = [conv(backbone_out) for conv in tracker_convs]
+        tracker_positions = [cast_to_input(self.position_encoding(f), f) for f in tracker_features]
+        return features, positions, tracker_features, tracker_positions

comfy/ldm/supir/supir_modules.py

@@ -0,0 +1,226 @@
+import torch
+import torch.nn as nn
+
+from comfy.ldm.modules.diffusionmodules.util import timestep_embedding
+from comfy.ldm.modules.diffusionmodules.openaimodel import Downsample, TimestepEmbedSequential, ResBlock, SpatialTransformer
+from comfy.ldm.modules.attention import optimized_attention
+
+
+class ZeroSFT(nn.Module):
+    def __init__(self, label_nc, norm_nc, concat_channels=0, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        ks = 3
+        pw = ks // 2
+
+        self.param_free_norm = operations.GroupNorm(32, norm_nc + concat_channels, dtype=dtype, device=device)
+
+        nhidden = 128
+
+        self.mlp_shared = nn.Sequential(
+            operations.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw, dtype=dtype, device=device),
+            nn.SiLU()
+        )
+        self.zero_mul = operations.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw, dtype=dtype, device=device)
+        self.zero_add = operations.Conv2d(nhidden, norm_nc + concat_channels, kernel_size=ks, padding=pw, dtype=dtype, device=device)
+
+        self.zero_conv = operations.Conv2d(label_nc, norm_nc, 1, 1, 0, dtype=dtype, device=device)
+        self.pre_concat = bool(concat_channels != 0)
+
+    def forward(self, c, h, h_ori=None, control_scale=1):
+        if h_ori is not None and self.pre_concat:
+            h_raw = torch.cat([h_ori, h], dim=1)
+        else:
+            h_raw = h
+
+        h = h + self.zero_conv(c)
+        if h_ori is not None and self.pre_concat:
+            h = torch.cat([h_ori, h], dim=1)
+        actv = self.mlp_shared(c)
+        gamma = self.zero_mul(actv)
+        beta = self.zero_add(actv)
+        h = self.param_free_norm(h)
+        h = torch.addcmul(h + beta, h, gamma)
+        if h_ori is not None and not self.pre_concat:
+            h = torch.cat([h_ori, h], dim=1)
+        return torch.lerp(h_raw, h, control_scale)
+
+
+class _CrossAttnInner(nn.Module):
+    """Inner cross-attention module matching the state_dict layout of the original CrossAttention."""
+    def __init__(self, query_dim, context_dim, heads, dim_head, dtype=None, device=None, operations=None):
+        super().__init__()
+        inner_dim = dim_head * heads
+        self.heads = heads
+        self.to_q = operations.Linear(query_dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.to_k = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.to_v = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.to_out = nn.Sequential(
+            operations.Linear(inner_dim, query_dim, dtype=dtype, device=device),
+        )
+
+    def forward(self, x, context):
+        q = self.to_q(x)
+        k = self.to_k(context)
+        v = self.to_v(context)
+        return self.to_out(optimized_attention(q, k, v, self.heads))
+
+
+class ZeroCrossAttn(nn.Module):
+    def __init__(self, context_dim, query_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        heads = query_dim // 64
+        dim_head = 64
+        self.attn = _CrossAttnInner(query_dim, context_dim, heads, dim_head, dtype=dtype, device=device, operations=operations)
+        self.norm1 = operations.GroupNorm(32, query_dim, dtype=dtype, device=device)
+        self.norm2 = operations.GroupNorm(32, context_dim, dtype=dtype, device=device)
+
+    def forward(self, context, x, control_scale=1):
+        b, c, h, w = x.shape
+        x_in = x
+
+        x = self.attn(
+            self.norm1(x).flatten(2).transpose(1, 2),
+            self.norm2(context).flatten(2).transpose(1, 2),
+        ).transpose(1, 2).unflatten(2, (h, w))
+
+        return x_in + x * control_scale
+
+
+class GLVControl(nn.Module):
+    """SUPIR's Guided Latent Vector control encoder. Truncated UNet (input + middle blocks only)."""
+    def __init__(
+        self,
+        in_channels=4,
+        model_channels=320,
+        num_res_blocks=2,
+        attention_resolutions=(4, 2),
+        channel_mult=(1, 2, 4),
+        num_head_channels=64,
+        transformer_depth=(1, 2, 10),
+        context_dim=2048,
+        adm_in_channels=2816,
+        use_linear_in_transformer=True,
+        use_checkpoint=False,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs,
+    ):
+        super().__init__()
+        self.model_channels = model_channels
+        time_embed_dim = model_channels * 4
+
+        self.time_embed = nn.Sequential(
+            operations.Linear(model_channels, time_embed_dim, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(time_embed_dim, time_embed_dim, dtype=dtype, device=device),
+        )
+
+        self.label_emb = nn.Sequential(
+            nn.Sequential(
+                operations.Linear(adm_in_channels, time_embed_dim, dtype=dtype, device=device),
+                nn.SiLU(),
+                operations.Linear(time_embed_dim, time_embed_dim, dtype=dtype, device=device),
+            )
+        )
+
+        self.input_blocks = nn.ModuleList([
+            TimestepEmbedSequential(
+                operations.Conv2d(in_channels, model_channels, 3, padding=1, dtype=dtype, device=device)
+            )
+        ])
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(num_res_blocks):
+                layers = [
+                    ResBlock(ch, time_embed_dim, 0, out_channels=mult * model_channels,
+                             dtype=dtype, device=device, operations=operations)
+                ]
+                ch = mult * model_channels
+                if ds in attention_resolutions:
+                    num_heads = ch // num_head_channels
+                    layers.append(
+                        SpatialTransformer(ch, num_heads, num_head_channels,
+                                           depth=transformer_depth[level], context_dim=context_dim,
+                                           use_linear=use_linear_in_transformer,
+                                           use_checkpoint=use_checkpoint,
+                                           dtype=dtype, device=device, operations=operations)
+                    )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+            if level != len(channel_mult) - 1:
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        Downsample(ch, True, out_channels=ch, dtype=dtype, device=device, operations=operations)
+                    )
+                )
+                ds *= 2
+
+        num_heads = ch // num_head_channels
+        self.middle_block = TimestepEmbedSequential(
+            ResBlock(ch, time_embed_dim, 0, dtype=dtype, device=device, operations=operations),
+            SpatialTransformer(ch, num_heads, num_head_channels,
+                               depth=transformer_depth[-1], context_dim=context_dim,
+                               use_linear=use_linear_in_transformer,
+                               use_checkpoint=use_checkpoint,
+                               dtype=dtype, device=device, operations=operations),
+            ResBlock(ch, time_embed_dim, 0, dtype=dtype, device=device, operations=operations),
+        )
+
+        self.input_hint_block = TimestepEmbedSequential(
+            operations.Conv2d(in_channels, model_channels, 3, padding=1, dtype=dtype, device=device)
+        )
+
+    def forward(self, x, timesteps, xt, context=None, y=None, **kwargs):
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
+        emb = self.time_embed(t_emb) + self.label_emb(y)
+
+        guided_hint = self.input_hint_block(x, emb, context)
+
+        hs = []
+        h = xt
+        for module in self.input_blocks:
+            if guided_hint is not None:
+                h = module(h, emb, context)
+                h += guided_hint
+                guided_hint = None
+            else:
+                h = module(h, emb, context)
+            hs.append(h)
+        h = self.middle_block(h, emb, context)
+        hs.append(h)
+        return hs
+
+
+class SUPIR(nn.Module):
+    """
+    SUPIR model containing GLVControl (control encoder) and project_modules (adapters).
+    State dict keys match the original SUPIR checkpoint layout:
+      control_model.*           -> GLVControl
+      project_modules.*         -> nn.ModuleList of ZeroSFT/ZeroCrossAttn
+    """
+    def __init__(self, device=None, dtype=None, operations=None):
+        super().__init__()
+
+        self.control_model = GLVControl(dtype=dtype, device=device, operations=operations)
+
+        project_channel_scale = 2
+        cond_output_channels = [320] * 4 + [640] * 3 + [1280] * 3
+        project_channels = [int(c * project_channel_scale) for c in [160] * 4 + [320] * 3 + [640] * 3]
+        concat_channels = [320] * 2 + [640] * 3 + [1280] * 4 + [0]
+        cross_attn_insert_idx = [6, 3]
+
+        self.project_modules = nn.ModuleList()
+        for i in range(len(cond_output_channels)):
+            self.project_modules.append(ZeroSFT(
+                project_channels[i], cond_output_channels[i],
+                concat_channels=concat_channels[i],
+                dtype=dtype, device=device, operations=operations,
+            ))
+
+        for i in cross_attn_insert_idx:
+            self.project_modules.insert(i, ZeroCrossAttn(
+                cond_output_channels[i], concat_channels[i],
+                dtype=dtype, device=device, operations=operations,
+            ))

comfy/ldm/supir/supir_patch.py

@@ -0,0 +1,103 @@
+import torch
+from comfy.ldm.modules.diffusionmodules.openaimodel import Upsample
+
+
+class SUPIRPatch:
+    """
+    Holds GLVControl (control encoder) + project_modules (ZeroSFT/ZeroCrossAttn adapters).
+    Runs GLVControl lazily on first patch invocation per step, applies adapters through
+    middle_block_after_patch, output_block_merge_patch, and forward_timestep_embed_patch.
+    """
+    SIGMA_MAX = 14.6146
+
+    def __init__(self, model_patch, project_modules, hint_latent, strength_start, strength_end):
+        self.model_patch = model_patch           # CoreModelPatcher wrapping GLVControl
+        self.project_modules = project_modules   # nn.ModuleList of ZeroSFT/ZeroCrossAttn
+        self.hint_latent = hint_latent           # encoded LQ image latent
+        self.strength_start = strength_start
+        self.strength_end = strength_end
+        self.cached_features = None
+        self.adapter_idx = 0
+        self.control_idx = 0
+        self.current_control_idx = 0
+        self.active = True
+
+    def _ensure_features(self, kwargs):
+        """Run GLVControl on first call per step, cache results."""
+        if self.cached_features is not None:
+            return
+        x = kwargs["x"]
+        b = x.shape[0]
+        hint = self.hint_latent.to(device=x.device, dtype=x.dtype)
+        if hint.shape[0] != b:
+            hint = hint.expand(b, -1, -1, -1) if hint.shape[0] == 1 else hint.repeat((b + hint.shape[0] - 1) // hint.shape[0], 1, 1, 1)[:b]
+        self.cached_features = self.model_patch.model.control_model(
+            hint, kwargs["timesteps"], x,
+            kwargs["context"], kwargs["y"]
+        )
+        self.adapter_idx = len(self.project_modules) - 1
+        self.control_idx = len(self.cached_features) - 1
+
+    def _get_control_scale(self, kwargs):
+        if self.strength_start == self.strength_end:
+            return self.strength_end
+        sigma = kwargs["transformer_options"].get("sigmas")
+        if sigma is None:
+            return self.strength_end
+        s = sigma[0].item() if sigma.dim() > 0 else sigma.item()
+        t = min(s / self.SIGMA_MAX, 1.0)
+        return t * (self.strength_start - self.strength_end) + self.strength_end
+
+    def middle_after(self, kwargs):
+        """middle_block_after_patch: run GLVControl lazily, apply last adapter after middle block."""
+        self.cached_features = None  # reset from previous step
+        self.current_scale = self._get_control_scale(kwargs)
+        self.active = self.current_scale > 0
+        if not self.active:
+            return {"h": kwargs["h"]}
+        self._ensure_features(kwargs)
+        h = kwargs["h"]
+        h = self.project_modules[self.adapter_idx](
+            self.cached_features[self.control_idx], h, control_scale=self.current_scale
+        )
+        self.adapter_idx -= 1
+        self.control_idx -= 1
+        return {"h": h}
+
+    def output_block(self, h, hsp, transformer_options):
+        """output_block_patch: ZeroSFT adapter fusion replaces cat([h, hsp]). Returns (h, None) to skip cat."""
+        if not self.active:
+            return h, hsp
+        self.current_control_idx = self.control_idx
+        h = self.project_modules[self.adapter_idx](
+            self.cached_features[self.control_idx], hsp, h, control_scale=self.current_scale
+        )
+        self.adapter_idx -= 1
+        self.control_idx -= 1
+        return h, None
+
+    def pre_upsample(self, layer, x, emb, context, transformer_options, output_shape, *args, **kw):
+        """forward_timestep_embed_patch for Upsample: extra cross-attn adapter before upsample."""
+        block_type, _ = transformer_options["block"]
+        if block_type == "output" and self.active and self.cached_features is not None:
+            x = self.project_modules[self.adapter_idx](
+                self.cached_features[self.current_control_idx], x, control_scale=self.current_scale
+            )
+            self.adapter_idx -= 1
+        return layer(x, output_shape=output_shape)
+
+    def to(self, device_or_dtype):
+        if isinstance(device_or_dtype, torch.device):
+            self.cached_features = None
+            if self.hint_latent is not None:
+                self.hint_latent = self.hint_latent.to(device_or_dtype)
+        return self
+
+    def models(self):
+        return [self.model_patch]
+
+    def register(self, model_patcher):
+        """Register all patches on a cloned model patcher."""
+        model_patcher.set_model_patch(self.middle_after, "middle_block_after_patch")
+        model_patcher.set_model_output_block_patch(self.output_block)
+        model_patcher.set_model_patch((Upsample, self.pre_upsample), "forward_timestep_embed_patch")

comfy/lora.py

@@ -342,6 +342,12 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["base_model.model.{}".format(key_lora)] = k  # Official base model loras
                 key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k  # LyCORIS/LoKR format
 
+    if isinstance(model, comfy.model_base.ErnieImage):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"):
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                key_map["transformer.{}".format(key_lora)] = k
+
     return key_map
 
 

comfy/model_base.py

@@ -52,7 +52,10 @@ 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.cogvideo.model
 import comfy.ldm.rt_detr.rtdetr_v4
+import comfy.ldm.ernie.model
+import comfy.ldm.sam3.detector
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -79,6 +82,7 @@ class ModelType(Enum):
     IMG_TO_IMG = 9
     FLOW_COSMOS = 10
     IMG_TO_IMG_FLOW = 11
+    V_PREDICTION_DDPM = 12
 
 
 def model_sampling(model_config, model_type):
@@ -113,6 +117,8 @@ def model_sampling(model_config, model_type):
         s = comfy.model_sampling.ModelSamplingCosmosRFlow
     elif model_type == ModelType.IMG_TO_IMG_FLOW:
         c = comfy.model_sampling.IMG_TO_IMG_FLOW
+    elif model_type == ModelType.V_PREDICTION_DDPM:
+        c = comfy.model_sampling.V_PREDICTION_DDPM
 
     class ModelSampling(s, c):
         pass
@@ -577,8 +583,8 @@ class Stable_Zero123(BaseModel):
     def __init__(self, model_config, model_type=ModelType.EPS, device=None, cc_projection_weight=None, cc_projection_bias=None):
         super().__init__(model_config, model_type, device=device)
         self.cc_projection = comfy.ops.manual_cast.Linear(cc_projection_weight.shape[1], cc_projection_weight.shape[0], dtype=self.get_dtype(), device=device)
-        self.cc_projection.weight.copy_(cc_projection_weight)
-        self.cc_projection.bias.copy_(cc_projection_bias)
+        self.cc_projection.weight = torch.nn.Parameter(cc_projection_weight.clone())
+        self.cc_projection.bias = torch.nn.Parameter(cc_projection_bias.clone())
 
     def extra_conds(self, **kwargs):
         out = {}
@@ -1962,3 +1968,74 @@ class Kandinsky5Image(Kandinsky5):
 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
+
+class SAM3(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.sam3.detector.SAM3Model)
+
+class CogVideoX(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.V_PREDICTION_DDPM, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.cogvideo.model.CogVideoXTransformer3DModel)
+        self.image_to_video = image_to_video
+
+    def concat_cond(self, **kwargs):
+        noise = kwargs.get("noise", None)
+        # Detect extra channels needed (e.g. 32 - 16 = 16 for ref latent)
+        extra_channels = self.diffusion_model.in_channels - noise.shape[1]
+        if extra_channels == 0:
+            return None
+
+        image = kwargs.get("concat_latent_image", None)
+        device = kwargs["device"]
+
+        if image is None:
+            shape = list(noise.shape)
+            shape[1] = extra_channels
+            return torch.zeros(shape, dtype=noise.dtype, layout=noise.layout, device=noise.device)
+
+        latent_dim = self.latent_format.latent_channels
+        image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+
+        if noise.ndim == 5 and image.ndim == 5:
+            if image.shape[-3] < noise.shape[-3]:
+                image = torch.nn.functional.pad(image, (0, 0, 0, 0, 0, noise.shape[-3] - image.shape[-3]), "constant", 0)
+            elif image.shape[-3] > noise.shape[-3]:
+                image = image[:, :, :noise.shape[-3]]
+
+        for i in range(0, image.shape[1], latent_dim):
+            image[:, i:i + latent_dim] = self.process_latent_in(image[:, i:i + latent_dim])
+        image = utils.resize_to_batch_size(image, noise.shape[0])
+
+        if image.shape[1] > extra_channels:
+            image = image[:, :extra_channels]
+        elif image.shape[1] < extra_channels:
+            repeats = extra_channels // image.shape[1]
+            remainder = extra_channels % image.shape[1]
+            parts = [image] * repeats
+            if remainder > 0:
+                parts.append(image[:, :remainder])
+            image = torch.cat(parts, dim=1)
+
+        return image
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        # OFS embedding (CogVideoX 1.5 I2V), default 2.0 as used by SparkVSR
+        if self.diffusion_model.ofs_proj_dim is not None:
+            ofs = kwargs.get("ofs", None)
+            if ofs is None:
+                noise = kwargs.get("noise", None)
+                ofs = torch.full((noise.shape[0],), 2.0, device=noise.device, dtype=noise.dtype)
+            out['ofs'] = comfy.conds.CONDRegular(ofs)
+        return out

comfy/model_detection.py

@@ -490,6 +490,54 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
         return dit_config
 
+    if '{}blocks.0.norm1.linear.weight'.format(key_prefix) in state_dict_keys:  # CogVideoX
+        dit_config = {}
+        dit_config["image_model"] = "cogvideox"
+
+        # Extract config from weight shapes
+        norm1_weight = state_dict['{}blocks.0.norm1.linear.weight'.format(key_prefix)]
+        time_embed_dim = norm1_weight.shape[1]
+        dim = norm1_weight.shape[0] // 6
+
+        dit_config["num_attention_heads"] = dim // 64
+        dit_config["attention_head_dim"] = 64
+        dit_config["time_embed_dim"] = time_embed_dim
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')
+
+        # Detect in_channels from patch_embed
+        patch_proj_key = '{}patch_embed.proj.weight'.format(key_prefix)
+        if patch_proj_key in state_dict_keys:
+            w = state_dict[patch_proj_key]
+            if w.ndim == 4:
+                # Conv2d: [out, in, kh, kw] — CogVideoX 1.0
+                dit_config["in_channels"] = w.shape[1]
+                dit_config["patch_size"] = w.shape[2]
+            elif w.ndim == 2:
+                # Linear: [out, in_channels * patch_size * patch_size * patch_size_t] — CogVideoX 1.5
+                dit_config["patch_size"] = 2
+                dit_config["patch_size_t"] = 2
+                dit_config["in_channels"] = w.shape[1] // (2 * 2 * 2)  # 256 // 8 = 32
+
+        text_proj_key = '{}patch_embed.text_proj.weight'.format(key_prefix)
+        if text_proj_key in state_dict_keys:
+            dit_config["text_embed_dim"] = state_dict[text_proj_key].shape[1]
+
+        # Detect OFS embedding
+        ofs_key = '{}ofs_embedding_linear_1.weight'.format(key_prefix)
+        if ofs_key in state_dict_keys:
+            dit_config["ofs_embed_dim"] = state_dict[ofs_key].shape[1]
+
+        # Detect positional embedding type
+        pos_key = '{}patch_embed.pos_embedding'.format(key_prefix)
+        if pos_key in state_dict_keys:
+            dit_config["use_learned_positional_embeddings"] = True
+            dit_config["use_rotary_positional_embeddings"] = False
+        else:
+            dit_config["use_learned_positional_embeddings"] = False
+            dit_config["use_rotary_positional_embeddings"] = True
+
+        return dit_config
+
     if '{}head.modulation'.format(key_prefix) in state_dict_keys:  # Wan 2.1
         dit_config = {}
         dit_config["image_model"] = "wan2.1"
@@ -696,6 +744,15 @@ 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
@@ -704,6 +761,19 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         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 'detector.backbone.vision_backbone.trunk.blocks.0.attn.qkv.weight' in state_dict_keys: # SAM3 / SAM3.1
+        if 'detector.transformer.decoder.query_embed.weight' in state_dict_keys:
+            dit_config = {}
+            dit_config["image_model"] = "SAM3"
+            if 'detector.backbone.vision_backbone.propagation_convs.0.conv_1x1.weight' in state_dict_keys:
+                dit_config["image_model"] = "SAM31"
+            return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
@@ -835,9 +905,16 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
     return unet_config
 
 def model_config_from_unet_config(unet_config, state_dict=None):
+    best_match = None
+    best_score = -1
     for model_config in comfy.supported_models.models:
-        if model_config.matches(unet_config, state_dict):
-            return model_config(unet_config)
+        score = model_config.match_score(unet_config, state_dict)
+        if score > best_score:
+            best_score = score
+            best_match = model_config
+
+    if best_match is not None:
+        return best_match(unet_config)
 
     logging.error("no match {}".format(unet_config))
     return None
@@ -859,6 +936,10 @@ def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=Fal
     return model_config
 
 def unet_prefix_from_state_dict(state_dict):
+    # SAM3: detector.* and tracker.* at top level, no common prefix
+    if any(k.startswith("detector.") for k in state_dict) and any(k.startswith("tracker.") for k in state_dict):
+        return ""
+
     candidates = ["model.diffusion_model.", #ldm/sgm models
                   "model.model.", #audio models
                   "net.", #cosmos

comfy/model_management.py

@@ -663,6 +663,7 @@ def minimum_inference_memory():
 
 def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, pins_required=0, ram_required=0):
     cleanup_models_gc()
+    comfy.memory_management.extra_ram_release(max(pins_required, ram_required))
     unloaded_model = []
     can_unload = []
     unloaded_models = []
@@ -1732,6 +1733,21 @@ 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):
@@ -1786,7 +1802,7 @@ def debug_memory_summary():
         return torch.cuda.memory.memory_summary()
     return ""
 
-class InterruptProcessingException(Exception):
+class InterruptProcessingException(BaseException):
     pass
 
 interrupt_processing_mutex = threading.RLock()

comfy/model_patcher.py

@@ -31,6 +31,7 @@ import comfy.float
 import comfy.hooks
 import comfy.lora
 import comfy.model_management
+import comfy.ops
 import comfy.patcher_extension
 import comfy.utils
 from comfy.comfy_types import UnetWrapperFunction
@@ -506,6 +507,10 @@ class ModelPatcher:
     def set_model_noise_refiner_patch(self, patch):
         self.set_model_patch(patch, "noise_refiner")
 
+    def set_model_middle_block_after_patch(self, patch):
+        self.set_model_patch(patch, "middle_block_after_patch")
+
+
     def set_model_rope_options(self, scale_x, shift_x, scale_y, shift_y, scale_t, shift_t, **kwargs):
         rope_options = self.model_options["transformer_options"].get("rope_options", {})
         rope_options["scale_x"] = scale_x
@@ -681,9 +686,9 @@ class ModelPatcher:
                         sd.pop(k)
             return sd
 
-    def patch_weight_to_device(self, key, device_to=None, inplace_update=False, return_weight=False):
+    def patch_weight_to_device(self, key, device_to=None, inplace_update=False, return_weight=False, force_cast=False):
         weight, set_func, convert_func = get_key_weight(self.model, key)
-        if key not in self.patches:
+        if key not in self.patches and not force_cast:
             return weight
 
         inplace_update = self.weight_inplace_update or inplace_update
@@ -691,7 +696,7 @@ class ModelPatcher:
         if key not in self.backup and not return_weight:
             self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
 
-        temp_dtype = comfy.model_management.lora_compute_dtype(device_to)
+        temp_dtype = comfy.model_management.lora_compute_dtype(device_to) if key in self.patches else None
         if device_to is not None:
             temp_weight = comfy.model_management.cast_to_device(weight, device_to, temp_dtype, copy=True)
         else:
@@ -699,9 +704,10 @@ class ModelPatcher:
         if convert_func is not None:
             temp_weight = convert_func(temp_weight, inplace=True)
 
-        out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key)
+        out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key) if key in self.patches else temp_weight
         if set_func is None:
-            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key))
+            if key in self.patches:
+                out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key))
             if return_weight:
                 return out_weight
             elif inplace_update:
@@ -851,7 +857,9 @@ class ModelPatcher:
                     if m.comfy_patched_weights == True:
                         continue
 
-                for param in params:
+                for param, param_value in params.items():
+                    if hasattr(m, "comfy_cast_weights") and getattr(param_value, "is_meta", False):
+                        comfy.ops.disable_weight_init._zero_init_parameter(m, param)
                     key = key_param_name_to_key(n, param)
                     self.unpin_weight(key)
                     self.patch_weight_to_device(key, device_to=device_to)
@@ -1580,7 +1588,7 @@ class ModelPatcherDynamic(ModelPatcher):
                     key = key_param_name_to_key(n, param_key)
                     if key in self.backup:
                         comfy.utils.set_attr_param(self.model, key, self.backup[key].weight)
-                    self.patch_weight_to_device(key, device_to=device_to)
+                    self.patch_weight_to_device(key, device_to=device_to, force_cast=True)
                     weight, _, _ = get_key_weight(self.model, key)
                     if weight is not None:
                         self.model.model_loaded_weight_memory += weight.numel() * weight.element_size()
@@ -1605,6 +1613,10 @@ class ModelPatcherDynamic(ModelPatcher):
                             m._v = vbar.alloc(v_weight_size)
                         allocated_size += v_weight_size
 
+                    for param in params:
+                        if param not in ("weight", "bias"):
+                            force_load_param(self, param, device_to)
+
                 else:
                     for param in params:
                         key = key_param_name_to_key(n, param)

comfy/model_sampling.py

@@ -54,6 +54,30 @@ class V_PREDICTION(EPS):
         sigma = reshape_sigma(sigma, model_output.ndim)
         return model_input * self.sigma_data ** 2 / (sigma ** 2 + self.sigma_data ** 2) - model_output * sigma * self.sigma_data / (sigma ** 2 + self.sigma_data ** 2) ** 0.5
 
+class V_PREDICTION_DDPM:
+    """CogVideoX v-prediction: model receives raw x_t (unscaled), predicts velocity v.
+    x_0 = sqrt(alpha) * x_t - sqrt(1-alpha) * v
+        = x_t / sqrt(sigma^2 + 1) - v * sigma / sqrt(sigma^2 + 1)
+    """
+    def calculate_input(self, sigma, noise):
+        return noise
+
+    def calculate_denoised(self, sigma, model_output, model_input):
+        sigma = reshape_sigma(sigma, model_output.ndim)
+        return model_input / (sigma ** 2 + 1.0) ** 0.5 - model_output * sigma / (sigma ** 2 + 1.0) ** 0.5
+
+    def noise_scaling(self, sigma, noise, latent_image, max_denoise=False):
+        sigma = reshape_sigma(sigma, noise.ndim)
+        if max_denoise:
+            noise = noise * torch.sqrt(1.0 + sigma ** 2.0)
+        else:
+            noise = noise * sigma
+        noise += latent_image
+        return noise
+
+    def inverse_noise_scaling(self, sigma, latent):
+        return latent
+
 class EDM(V_PREDICTION):
     def calculate_denoised(self, sigma, model_output, model_input):
         sigma = reshape_sigma(sigma, model_output.ndim)

comfy/ops.py

@@ -79,14 +79,21 @@ def cast_to_input(weight, input, non_blocking=False, copy=True):
     return comfy.model_management.cast_to(weight, input.dtype, input.device, non_blocking=non_blocking, copy=copy)
 
 
-def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant):
+def materialize_meta_param(s, param_keys):
+    for param_key in param_keys:
+        param = getattr(s, param_key, None)
+        if param is not None and getattr(param, "is_meta", False):
+            setattr(s, param_key, torch.nn.Parameter(torch.zeros(param.shape, dtype=param.dtype), requires_grad=param.requires_grad))
 
+
+def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant):
     #vbar doesn't support CPU weights, but some custom nodes have weird paths
     #that might switch the layer to the CPU and expect it to work. We have to take
     #a clone conservatively as we are mmapped and some SFT files are packed misaligned
     #If you are a custom node author reading this, please move your layer to the GPU
     #or declare your ModelPatcher as CPU in the first place.
     if comfy.model_management.is_device_cpu(device):
+        materialize_meta_param(s, ["weight", "bias"])
         weight = s.weight.to(dtype=dtype, copy=True)
         if isinstance(weight, QuantizedTensor):
             weight = weight.dequantize()
@@ -108,6 +115,7 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
 
     if not resident:
+        materialize_meta_param(s, ["weight", "bias"])
         cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
         cast_dest = None
 
@@ -306,6 +314,12 @@ class CastWeightBiasOp:
     bias_function = []
 
 class disable_weight_init:
+    @staticmethod
+    def _zero_init_parameter(module, name):
+        param = getattr(module, name)
+        device = None if getattr(param, "is_meta", False) else param.device
+        setattr(module, name, torch.nn.Parameter(torch.zeros(param.shape, device=device, dtype=param.dtype), requires_grad=False))
+
     @staticmethod
     def _lazy_load_from_state_dict(module, state_dict, prefix, local_metadata,
                                    missing_keys, unexpected_keys, weight_shape,
@@ -1151,7 +1165,7 @@ 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/pinned_memory.py

@@ -2,7 +2,6 @@ import comfy.model_management
 import comfy.memory_management
 import comfy_aimdo.host_buffer
 import comfy_aimdo.torch
-import psutil
 
 from comfy.cli_args import args
 
@@ -12,11 +11,6 @@ def get_pin(module):
 def pin_memory(module):
     if module.pin_failed or args.disable_pinned_memory or get_pin(module) is not None:
         return
-    #FIXME: This is a RAM cache trigger event
-    ram_headroom = comfy.memory_management.RAM_CACHE_HEADROOM
-    #we split the difference and assume half the RAM cache headroom is for us
-    if ram_headroom > 0 and psutil.virtual_memory().available < (ram_headroom * 0.5):
-        comfy.memory_management.extra_ram_release(ram_headroom)
 
     size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ])
 

comfy/sd.py

@@ -12,11 +12,13 @@ from .ldm.cascade.stage_c_coder import StageC_coder
 from .ldm.audio.autoencoder import AudioOobleckVAE
 import comfy.ldm.genmo.vae.model
 import comfy.ldm.lightricks.vae.causal_video_autoencoder
+import comfy.ldm.lightricks.vae.audio_vae
 import comfy.ldm.cosmos.vae
 import comfy.ldm.wan.vae
 import comfy.ldm.wan.vae2_2
 import comfy.ldm.hunyuan3d.vae
 import comfy.ldm.ace.vae.music_dcae_pipeline
+import comfy.ldm.cogvideo.vae
 import comfy.ldm.hunyuan_video.vae
 import comfy.ldm.mmaudio.vae.autoencoder
 import comfy.pixel_space_convert
@@ -62,6 +64,7 @@ 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
@@ -476,7 +479,10 @@ class VAE:
                                                             encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': encoder_config},
                                                             decoder_config={'target': "comfy.ldm.modules.temporal_ae.VideoDecoder", 'params': decoder_config})
             elif "taesd_decoder.1.weight" in sd:
-                self.latent_channels = sd["taesd_decoder.1.weight"].shape[1]
+                if isinstance(metadata, dict) and "tae_latent_channels" in metadata:
+                    self.latent_channels = metadata["tae_latent_channels"]
+                else:
+                    self.latent_channels = sd["taesd_decoder.1.weight"].shape[1]
                 self.first_stage_model = comfy.taesd.taesd.TAESD(latent_channels=self.latent_channels)
             elif "vquantizer.codebook.weight" in sd: #VQGan: stage a of stable cascade
                 self.first_stage_model = StageA()
@@ -556,12 +562,19 @@ 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
@@ -643,6 +656,17 @@ class VAE:
 
                 self.memory_used_encode = lambda shape, dtype: (1400 * 9 * shape[-2] * shape[-1]) * model_management.dtype_size(dtype)
                 self.memory_used_decode = lambda shape, dtype: (3600 * 4 * shape[-2] * shape[-1] * 16 * 16) * model_management.dtype_size(dtype)
+            elif "decoder.conv_in.conv.weight" in sd and "decoder.mid_block.resnets.0.norm1.norm_layer.weight" in sd:  # CogVideoX VAE
+                self.upscale_ratio = (lambda a: max(0, a * 4 - 3), 8, 8)
+                self.upscale_index_formula = (4, 8, 8)
+                self.downscale_ratio = (lambda a: max(0, math.floor((a + 3) / 4)), 8, 8)
+                self.downscale_index_formula = (4, 8, 8)
+                self.latent_dim = 3
+                self.latent_channels = sd["encoder.conv_out.conv.weight"].shape[0] // 2
+                self.first_stage_model = comfy.ldm.cogvideo.vae.AutoencoderKLCogVideoX(latent_channels=self.latent_channels)
+                self.memory_used_decode = lambda shape, dtype: (2800 * max(2, ((shape[2] - 1) * 4) + 1) * shape[3] * shape[4] * (8 * 8)) * model_management.dtype_size(dtype)
+                self.memory_used_encode = lambda shape, dtype: (1400 * max(1, shape[2]) * shape[3] * shape[4]) * model_management.dtype_size(dtype)
+                self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
             elif "decoder.conv_in.conv.weight" in sd:
                 ddconfig = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
                 ddconfig["conv3d"] = True
@@ -797,6 +821,24 @@ class VAE:
                     self.downscale_index_formula = (4, 8, 8)
                     self.memory_used_encode = lambda shape, dtype: (700 * (max(1, (shape[-3] ** 0.66 * 0.11)) * shape[-2] * shape[-1]) * model_management.dtype_size(dtype))
                     self.memory_used_decode = lambda shape, dtype: (50 * (max(1, (shape[-3] ** 0.65 * 0.26)) * shape[-2] * shape[-1] * 32 * 32) * model_management.dtype_size(dtype))
+            elif "vocoder.resblocks.0.convs1.0.weight" in sd or "vocoder.vocoder.resblocks.0.convs1.0.weight" in sd: # LTX Audio
+                sd = comfy.utils.state_dict_prefix_replace(sd, {"audio_vae.": "autoencoder."})
+                self.first_stage_model = comfy.ldm.lightricks.vae.audio_vae.AudioVAE(metadata=metadata)
+                self.memory_used_encode = lambda shape, dtype: (shape[2] * 330) * model_management.dtype_size(dtype)
+                self.memory_used_decode = lambda shape, dtype: (shape[2] * shape[3] * 87000) * model_management.dtype_size(dtype)
+                self.latent_channels = self.first_stage_model.latent_channels
+                self.audio_sample_rate_output = self.first_stage_model.output_sample_rate
+                self.autoencoder = self.first_stage_model.autoencoder  # TODO: remove hack for ltxv custom nodes
+                self.output_channels = 2
+                self.pad_channel_value = "replicate"
+                self.upscale_ratio = 4096
+                self.downscale_ratio = 4096
+                self.latent_dim = 2
+                self.process_output = lambda audio: audio
+                self.process_input = lambda audio: audio
+                self.working_dtypes = [torch.float32]
+                self.disable_offload = True
+                self.extra_1d_channel = 16
             else:
                 logging.warning("WARNING: No VAE weights detected, VAE not initalized.")
                 self.first_stage_model = None
@@ -1228,6 +1270,7 @@ class TEModel(Enum):
     QWEN35_4B = 25
     QWEN35_9B = 26
     QWEN35_27B = 27
+    MINISTRAL_3_3B = 28
 
 
 def detect_te_model(sd):
@@ -1294,6 +1337,8 @@ 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 +1496,10 @@ 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:
@@ -1745,6 +1794,8 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None, disable
     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)
 
     parameters = comfy.utils.calculate_parameters(sd)
     weight_dtype = comfy.utils.weight_dtype(sd)

comfy/supported_models.py

@@ -26,6 +26,8 @@ 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
+import comfy.text_encoders.cogvideo
 
 from . import supported_models_base
 from . import latent_formats
@@ -1749,6 +1751,214 @@ class RT_DETR_v4(supported_models_base.BASE):
     def clip_target(self, state_dict={}):
         return None
 
-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]
 
-models += [SVD_img2vid]
+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))
+
+
+class SAM3(supported_models_base.BASE):
+    unet_config = {"image_model": "SAM3"}
+    supported_inference_dtypes = [torch.float16, torch.bfloat16, torch.float32]
+    text_encoder_key_prefix = ["detector.backbone.language_backbone."]
+    unet_extra_prefix = ""
+
+    def process_clip_state_dict(self, state_dict):
+        clip_keys = getattr(self, "_clip_stash", {})
+        clip_keys = utils.state_dict_prefix_replace(clip_keys, {"detector.backbone.language_backbone.": "", "backbone.language_backbone.": ""}, filter_keys=True)
+        clip_keys = utils.clip_text_transformers_convert(clip_keys, "encoder.", "sam3_clip.transformer.")
+        return {k: v for k, v in clip_keys.items() if not k.startswith("resizer.")}
+
+    def process_unet_state_dict(self, state_dict):
+        self._clip_stash = {k: state_dict.pop(k) for k in list(state_dict.keys()) if "language_backbone" in k and "resizer" not in k}
+        # SAM3.1: remap tracker.model.* -> tracker.*
+        for k in list(state_dict.keys()):
+            if k.startswith("tracker.model."):
+                state_dict["tracker." + k[len("tracker.model."):]] = state_dict.pop(k)
+        # SAM3.1: remove per-block freqs_cis buffers (computed dynamically)
+        for k in [k for k in list(state_dict.keys()) if ".attn.freqs_cis" in k]:
+            state_dict.pop(k)
+        # Split fused QKV projections
+        for k in [k for k in list(state_dict.keys()) if k.endswith((".in_proj_weight", ".in_proj_bias"))]:
+            t = state_dict.pop(k)
+            base, suffix = k.rsplit(".in_proj_", 1)
+            s = ".weight" if suffix == "weight" else ".bias"
+            d = t.shape[0] // 3
+            state_dict[base + ".q_proj" + s] = t[:d]
+            state_dict[base + ".k_proj" + s] = t[d:2*d]
+            state_dict[base + ".v_proj" + s] = t[2*d:]
+        # Remap tracker SAM decoder transformer key names to match sam.py TwoWayTransformer
+        for k in list(state_dict.keys()):
+            if "sam_mask_decoder.transformer." not in k:
+                continue
+            new_k = k.replace(".mlp.lin1.", ".mlp.0.").replace(".mlp.lin2.", ".mlp.2.").replace(".norm_final_attn.", ".norm_final.")
+            if new_k != k:
+                state_dict[new_k] = state_dict.pop(k)
+        return state_dict
+
+    def get_model(self, state_dict, prefix="", device=None):
+        return model_base.SAM3(self, device=device)
+
+    def clip_target(self, state_dict={}):
+        import comfy.text_encoders.sam3_clip
+        return supported_models_base.ClipTarget(comfy.text_encoders.sam3_clip.SAM3TokenizerWrapper, comfy.text_encoders.sam3_clip.SAM3ClipModelWrapper)
+
+
+class SAM31(SAM3):
+    unet_config = {"image_model": "SAM31"}
+
+
+class CogVideoX_T2V(supported_models_base.BASE):
+    unet_config = {
+        "image_model": "cogvideox",
+    }
+
+    sampling_settings = {
+        "linear_start": 0.00085,
+        "linear_end": 0.012,
+        "beta_schedule": "linear",
+        "zsnr": True,
+    }
+
+    unet_extra_config = {}
+    latent_format = latent_formats.CogVideoX
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        # CogVideoX 1.5 (patch_size_t=2) has different training base dimensions for RoPE
+        if self.unet_config.get("patch_size_t") is not None:
+            self.unet_config.setdefault("sample_height", 96)
+            self.unet_config.setdefault("sample_width", 170)
+            self.unet_config.setdefault("sample_frames", 81)
+        out = model_base.CogVideoX(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        return supported_models_base.ClipTarget(comfy.text_encoders.cogvideo.CogVideoXT5Tokenizer, comfy.text_encoders.sd3_clip.T5XXLModel)
+
+class CogVideoX_I2V(CogVideoX_T2V):
+    unet_config = {
+        "image_model": "cogvideox",
+        "in_channels": 32,
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        if self.unet_config.get("patch_size_t") is not None:
+            self.unet_config.setdefault("sample_height", 96)
+            self.unet_config.setdefault("sample_width", 170)
+            self.unet_config.setdefault("sample_frames", 81)
+        out = model_base.CogVideoX(self, image_to_video=True, device=device)
+        return out
+
+
+models = [
+    LotusD,
+    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,
+    SAM3,
+    SAM31,
+    CogVideoX_I2V,
+    CogVideoX_T2V,
+    SVD_img2vid,
+]

comfy/supported_models_base.py

@@ -54,15 +54,31 @@ class BASE:
     optimizations = {"fp8": False}
 
     @classmethod
-    def matches(s, unet_config, state_dict=None):
+    def match_score(s, unet_config, state_dict=None):
+        """Return a non-negative specificity score if this model matches the given
+        ``unet_config``/``state_dict``, otherwise ``-1``.
+
+        The score is the total number of ``unet_config`` keys (plus ``required_keys``
+        when ``state_dict`` is provided) that this model declares and that match the
+        input. Higher scores indicate a more specific match, allowing
+        :func:`model_config_from_unet_config` to prefer the most specific model when
+        several configs would otherwise match by subset.
+        """
+        score = 0
         for k in s.unet_config:
             if k not in unet_config or s.unet_config[k] != unet_config[k]:
-                return False
+                return -1
+            score += 1
         if state_dict is not None:
             for k in s.required_keys:
                 if k not in state_dict:
-                    return False
-        return True
+                    return -1
+                score += 1
+        return score
+
+    @classmethod
+    def matches(s, unet_config, state_dict=None):
+        return s.match_score(unet_config, state_dict) >= 0
 
     def model_type(self, state_dict, prefix=""):
         return model_base.ModelType.EPS

comfy/taesd/taehv.py

@@ -7,6 +7,7 @@ from tqdm.auto import tqdm
 from collections import namedtuple, deque
 
 import comfy.ops
+import comfy.model_management
 operations=comfy.ops.disable_weight_init
 
 DecoderResult = namedtuple("DecoderResult", ("frame", "memory"))
@@ -47,11 +48,14 @@ class TGrow(nn.Module):
         x = self.conv(x)
         return x.reshape(-1, C, H, W)
 
-def apply_model_with_memblocks(model, x, parallel, show_progress_bar):
+def apply_model_with_memblocks(model, x, parallel, show_progress_bar, output_device=None,
+                               patch_size=1, decode=False):
 
     B, T, C, H, W = x.shape
     if parallel:
         x = x.reshape(B*T, C, H, W)
+        if not decode and patch_size > 1:
+            x = F.pixel_unshuffle(x, patch_size)
         # parallel over input timesteps, iterate over blocks
         for b in tqdm(model, disable=not show_progress_bar):
             if isinstance(b, MemBlock):
@@ -62,20 +66,27 @@ def apply_model_with_memblocks(model, x, parallel, show_progress_bar):
                 x = b(x, mem)
             else:
                 x = b(x)
-        BT, C, H, W = x.shape
-        T = BT // B
-        x = x.view(B, T, C, H, W)
+        if decode and patch_size > 1:
+            x = F.pixel_shuffle(x, patch_size)
+        x = x.view(B, x.shape[0] // B, *x.shape[1:])
+        x = x.to(output_device)
     else:
         out = []
-        work_queue = deque([TWorkItem(xt, 0) for t, xt in enumerate(x.reshape(B, T * C, H, W).chunk(T, dim=1))])
+        # Chunk along the time dim directly (chunks are [B,1,C,H,W] views, squeeze to [B,C,H,W] views).
+        # Avoids forcing a contiguous copy when x is non-contiguous (e.g. after movedim in encode/decode).
+        work_queue = deque([TWorkItem(xt.squeeze(1), 0) for xt in x.chunk(T, dim=1)])
         progress_bar = tqdm(range(T), disable=not show_progress_bar)
         mem = [None] * len(model)
         while work_queue:
             xt, i = work_queue.popleft()
             if i == 0:
                 progress_bar.update(1)
+                if not decode and patch_size > 1:
+                    xt = F.pixel_unshuffle(xt, patch_size)
             if i == len(model):
-                out.append(xt)
+                if decode and patch_size > 1:
+                    xt = F.pixel_shuffle(xt, patch_size)
+                out.append(xt.to(output_device))
                 del xt
             else:
                 b = model[i]
@@ -165,24 +176,20 @@ class TAEHV(nn.Module):
 
     def encode(self, x, **kwargs):
         x = x.movedim(2, 1)  # [B, C, T, H, W] -> [B, T, C, H, W]
-        if self.patch_size > 1:
-            B, T, C, H, W = x.shape
-            x = x.reshape(B * T, C, H, W)
-            x = F.pixel_unshuffle(x, self.patch_size)
-            x = x.reshape(B, T, C * self.patch_size ** 2, H // self.patch_size, W // self.patch_size)
         if x.shape[1] % self.t_downscale != 0:
             # pad at end to multiple of t_downscale
             n_pad = self.t_downscale - x.shape[1] % self.t_downscale
             padding = x[:, -1:].repeat_interleave(n_pad, dim=1)
             x = torch.cat([x, padding], 1)
-        x = apply_model_with_memblocks(self.encoder, x, self.parallel, self.show_progress_bar).movedim(2, 1)
+        x = apply_model_with_memblocks(self.encoder, x, self.parallel, self.show_progress_bar,
+                                        patch_size=self.patch_size).movedim(2, 1)
         return self.process_out(x)
 
     def decode(self, x, **kwargs):
         x = x.unsqueeze(0) if x.ndim == 4 else x  # [T, C, H, W] -> [1, T, C, H, W]
         x = x.movedim(1, 2) if x.shape[1] != self.latent_channels else x  # [B, T, C, H, W] or [B, C, T, H, W]
         x = self.process_in(x).movedim(2, 1)  # [B, C, T, H, W] -> [B, T, C, H, W]
-        x = apply_model_with_memblocks(self.decoder, x, self.parallel, self.show_progress_bar)
-        if self.patch_size > 1:
-            x = F.pixel_shuffle(x, self.patch_size)
+        x = apply_model_with_memblocks(self.decoder, x, self.parallel, self.show_progress_bar,
+                                        output_device=comfy.model_management.intermediate_device(),
+                                        patch_size=self.patch_size, decode=True)
         return x[:, self.frames_to_trim:].movedim(2, 1)

comfy/taesd/taesd.py

@@ -17,32 +17,79 @@ class Clamp(nn.Module):
         return torch.tanh(x / 3) * 3
 
 class Block(nn.Module):
-    def __init__(self, n_in, n_out):
+    def __init__(self, n_in: int, n_out: int, use_midblock_gn: bool = False):
         super().__init__()
         self.conv = nn.Sequential(conv(n_in, n_out), nn.ReLU(), conv(n_out, n_out), nn.ReLU(), conv(n_out, n_out))
         self.skip = comfy.ops.disable_weight_init.Conv2d(n_in, n_out, 1, bias=False) if n_in != n_out else nn.Identity()
         self.fuse = nn.ReLU()
-    def forward(self, x):
+        if not use_midblock_gn:
+            self.pool = None
+            return
+        n_gn = n_in * 4
+        self.pool = nn.Sequential(
+            comfy.ops.disable_weight_init.Conv2d(n_in, n_gn, 1, bias=False),
+            comfy.ops.disable_weight_init.GroupNorm(4, n_gn),
+            nn.ReLU(inplace=True),
+            comfy.ops.disable_weight_init.Conv2d(n_gn, n_in, 1, bias=False),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.pool is not None:
+            x = x + self.pool(x)
         return self.fuse(self.conv(x) + self.skip(x))
 
-def Encoder(latent_channels=4):
-    return nn.Sequential(
-        conv(3, 64), Block(64, 64),
-        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
-        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
-        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
-        conv(64, latent_channels),
-    )
+class Encoder(nn.Sequential):
+    def __init__(self, latent_channels: int = 4, use_gn: bool = False):
+        super().__init__(
+            conv(3, 64), Block(64, 64),
+            conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
+            conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
+            conv(64, 64, stride=2, bias=False), Block(64, 64, use_gn), Block(64, 64, use_gn), Block(64, 64, use_gn),
+            conv(64, latent_channels),
+        )
 
+class Decoder(nn.Sequential):
+    def __init__(self, latent_channels: int = 4, use_gn: bool = False):
+        super().__init__(
+            Clamp(), conv(latent_channels, 64), nn.ReLU(),
+            Block(64, 64, use_gn), Block(64, 64, use_gn), Block(64, 64, use_gn), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
+            Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
+            Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
+            Block(64, 64), conv(64, 3),
+        )
+
+class DecoderFlux2(Decoder):
+    def __init__(self, latent_channels: int = 128, use_gn: bool = True):
+        if latent_channels != 128 or not use_gn:
+            raise ValueError("Unexpected parameters for Flux2 TAE module")
+        super().__init__(latent_channels=32, use_gn=True)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, C, H, W = x.shape
+        x = (
+            x
+            .reshape(B, 32, 2, 2, H, W)
+            .permute(0, 1, 4, 2, 5, 3)
+            .reshape(B, 32, H * 2, W * 2)
+        )
+        return super().forward(x)
+
+class EncoderFlux2(Encoder):
+    def __init__(self, latent_channels: int = 128, use_gn: bool = True):
+        if latent_channels != 128 or not use_gn:
+            raise ValueError("Unexpected parameters for Flux2 TAE module")
+        super().__init__(latent_channels=32, use_gn=True)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        result = super().forward(x)
+        B, C, H, W = result.shape
+        return (
+            result
+            .reshape(B, C, H // 2, 2, W // 2, 2)
+            .permute(0, 1, 3, 5, 2, 4)
+            .reshape(B, 128, H // 2, W // 2)
+        )
 
-def Decoder(latent_channels=4):
-    return nn.Sequential(
-        Clamp(), conv(latent_channels, 64), nn.ReLU(),
-        Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
-        Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
-        Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
-        Block(64, 64), conv(64, 3),
-    )
 
 class TAESD(nn.Module):
     latent_magnitude = 3
@@ -51,8 +98,15 @@ class TAESD(nn.Module):
     def __init__(self, encoder_path=None, decoder_path=None, latent_channels=4):
         """Initialize pretrained TAESD on the given device from the given checkpoints."""
         super().__init__()
-        self.taesd_encoder = Encoder(latent_channels=latent_channels)
-        self.taesd_decoder = Decoder(latent_channels=latent_channels)
+        if latent_channels == 128:
+            encoder_class = EncoderFlux2
+            decoder_class = DecoderFlux2
+        else:
+            encoder_class = Encoder
+            decoder_class = Decoder
+        self.taesd_encoder = encoder_class(latent_channels=latent_channels)
+        self.taesd_decoder = decoder_class(latent_channels=latent_channels)
+
         self.vae_scale = torch.nn.Parameter(torch.tensor(1.0))
         self.vae_shift = torch.nn.Parameter(torch.tensor(0.0))
         if encoder_path is not None:
@@ -61,19 +115,19 @@ class TAESD(nn.Module):
             self.taesd_decoder.load_state_dict(comfy.utils.load_torch_file(decoder_path, safe_load=True))
 
     @staticmethod
-    def scale_latents(x):
+    def scale_latents(x: torch.Tensor) -> torch.Tensor:
         """raw latents -> [0, 1]"""
         return x.div(2 * TAESD.latent_magnitude).add(TAESD.latent_shift).clamp(0, 1)
 
     @staticmethod
-    def unscale_latents(x):
+    def unscale_latents(x: torch.Tensor) -> torch.Tensor:
         """[0, 1] -> raw latents"""
         return x.sub(TAESD.latent_shift).mul(2 * TAESD.latent_magnitude)
 
-    def decode(self, x):
+    def decode(self, x: torch.Tensor) -> torch.Tensor:
         x_sample = self.taesd_decoder((x - self.vae_shift) * self.vae_scale)
         x_sample = x_sample.sub(0.5).mul(2)
         return x_sample
 
-    def encode(self, x):
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
         return (self.taesd_encoder(x * 0.5 + 0.5) / self.vae_scale) + self.vae_shift

comfy/text_encoders/cogvideo.py

@@ -0,0 +1,6 @@
+import comfy.text_encoders.sd3_clip
+
+
+class CogVideoXT5Tokenizer(comfy.text_encoders.sd3_clip.T5XXLTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, min_length=226)

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/text_encoders/sam3_clip.py

@@ -0,0 +1,97 @@
+import re
+from comfy import sd1_clip
+
+SAM3_CLIP_CONFIG = {
+    "architectures": ["CLIPTextModel"],
+    "hidden_act": "quick_gelu",
+    "hidden_size": 1024,
+    "intermediate_size": 4096,
+    "num_attention_heads": 16,
+    "num_hidden_layers": 24,
+    "max_position_embeddings": 32,
+    "projection_dim": 512,
+    "vocab_size": 49408,
+    "layer_norm_eps": 1e-5,
+    "eos_token_id": 49407,
+}
+
+
+class SAM3ClipModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, max_length=32, layer="last", textmodel_json_config=SAM3_CLIP_CONFIG, special_tokens={"start": 49406, "end": 49407, "pad": 0}, return_projected_pooled=False, return_attention_masks=True, enable_attention_masks=True, model_options=model_options)
+
+
+class SAM3Tokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(max_length=32, pad_with_end=False, pad_token=0, embedding_directory=embedding_directory, embedding_size=1024, embedding_key="sam3_clip", tokenizer_data=tokenizer_data)
+        self.disable_weights = True
+
+
+def _parse_prompts(text):
+    """Split comma-separated prompts with optional :N max detections per category"""
+    text = text.replace("(", "").replace(")", "")
+    parts = [p.strip() for p in text.split(",") if p.strip()]
+    result = []
+    for part in parts:
+        m = re.match(r'^(.+?)\s*:\s*([\d.]+)\s*$', part)
+        if m:
+            text_part = m.group(1).strip()
+            val = m.group(2)
+            max_det = max(1, round(float(val)))
+            result.append((text_part, max_det))
+        else:
+            result.append((part, 1))
+    return result
+
+
+class SAM3TokenizerWrapper(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, clip_name="l", tokenizer=SAM3Tokenizer, name="sam3_clip")
+
+    def tokenize_with_weights(self, text: str, return_word_ids=False, **kwargs):
+        parsed = _parse_prompts(text)
+        if len(parsed) <= 1 and (not parsed or parsed[0][1] == 1):
+            return super().tokenize_with_weights(text, return_word_ids, **kwargs)
+        # Tokenize each prompt part separately, store per-part batches and metadata
+        inner = getattr(self, self.clip)
+        per_prompt = []
+        for prompt_text, max_det in parsed:
+            batches = inner.tokenize_with_weights(prompt_text, return_word_ids, **kwargs)
+            per_prompt.append((batches, max_det))
+        # Main output uses first prompt's tokens (for compatibility)
+        out = {self.clip_name: per_prompt[0][0], "sam3_per_prompt": per_prompt}
+        return out
+
+
+class SAM3ClipModelWrapper(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}, **kwargs):
+        super().__init__(device=device, dtype=dtype, model_options=model_options, clip_name="l", clip_model=SAM3ClipModel, name="sam3_clip")
+
+    def encode_token_weights(self, token_weight_pairs):
+        per_prompt = token_weight_pairs.pop("sam3_per_prompt", None)
+        if per_prompt is None:
+            return super().encode_token_weights(token_weight_pairs)
+
+        # Encode each prompt separately, pack into extra dict
+        inner = getattr(self, self.clip)
+        multi_cond = []
+        first_pooled = None
+        for batches, max_det in per_prompt:
+            out = inner.encode_token_weights(batches)
+            cond, pooled = out[0], out[1]
+            extra = out[2] if len(out) > 2 else {}
+            if first_pooled is None:
+                first_pooled = pooled
+            multi_cond.append({
+                "cond": cond,
+                "attention_mask": extra.get("attention_mask"),
+                "max_detections": max_det,
+            })
+
+        # Return first prompt as main (for non-SAM3 consumers), all prompts in metadata
+        main = multi_cond[0]
+        main_extra = {}
+        if main["attention_mask"] is not None:
+            main_extra["attention_mask"] = main["attention_mask"]
+        main_extra["sam3_multi_cond"] = multi_cond
+        return (main["cond"], first_pooled, main_extra)