Merge branch 'master' into toolkit-nodes/aspect-ratio-blueprint

.coderabbit.yaml

@@ -0,0 +1,127 @@
+# yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json
+language: "en-US"
+early_access: false
+tone_instructions: "Only comment on issues introduced by this PR's changes. Do not flag pre-existing problems in moved, re-indented, or reformatted code."
+
+reviews:
+  profile: "chill"
+  request_changes_workflow: false
+  high_level_summary: false
+  poem: false
+  review_status: false
+  review_details: false
+  commit_status: true
+  collapse_walkthrough: true
+  changed_files_summary: false
+  sequence_diagrams: false
+  estimate_code_review_effort: false
+  assess_linked_issues: false
+  related_issues: false
+  related_prs: false
+  suggested_labels: false
+  auto_apply_labels: false
+  suggested_reviewers: false
+  auto_assign_reviewers: false
+  in_progress_fortune: false
+  enable_prompt_for_ai_agents: true
+
+  path_filters:
+    - "!comfy_api_nodes/apis/**"
+    - "!**/generated/*.pyi"
+    - "!.ci/**"
+    - "!script_examples/**"
+    - "!**/__pycache__/**"
+    - "!**/*.ipynb"
+    - "!**/*.png"
+    - "!**/*.bat"
+
+  path_instructions:
+    - path: "**"
+      instructions: |
+        IMPORTANT: Only comment on issues directly introduced by this PR's code changes.
+        Do NOT flag pre-existing issues in code that was merely moved, re-indented,
+        de-indented, or reformatted without logic changes. If code appears in the diff
+        only due to whitespace or structural reformatting (e.g., removing a `with:` block),
+        treat it as unchanged. Contributors should not feel obligated to address
+        pre-existing issues outside the scope of their contribution.
+    - path: "comfy/**"
+      instructions: |
+        Core ML/diffusion engine. Focus on:
+        - Backward compatibility (breaking changes affect all custom nodes)
+        - Memory management and GPU resource handling
+        - Performance implications in hot paths
+        - Thread safety for concurrent execution
+    - path: "comfy_api_nodes/**"
+      instructions: |
+        Third-party API integration nodes. Focus on:
+        - No hardcoded API keys or secrets
+        - Proper error handling for API failures (timeouts, rate limits, auth errors)
+        - Correct Pydantic model usage
+        - Security of user data passed to external APIs
+    - path: "comfy_extras/**"
+      instructions: |
+        Community-contributed extra nodes. Focus on:
+        - Consistency with node patterns (INPUT_TYPES, RETURN_TYPES, FUNCTION, CATEGORY)
+        - No breaking changes to existing node interfaces
+    - path: "comfy_execution/**"
+      instructions: |
+        Execution engine (graph execution, caching, jobs). Focus on:
+        - Caching correctness
+        - Concurrent execution safety
+        - Graph validation edge cases
+    - path: "nodes.py"
+      instructions: |
+        Core node definitions (2500+ lines). Focus on:
+        - Backward compatibility of NODE_CLASS_MAPPINGS
+        - Consistency of INPUT_TYPES return format
+    - path: "alembic_db/**"
+      instructions: |
+        Database migrations. Focus on:
+        - Migration safety and rollback support
+        - Data preservation during schema changes
+
+  auto_review:
+    enabled: true
+    auto_incremental_review: true
+    drafts: false
+    ignore_title_keywords:
+      - "WIP"
+      - "DO NOT REVIEW"
+      - "DO NOT MERGE"
+
+  finishing_touches:
+    docstrings:
+      enabled: false
+    unit_tests:
+      enabled: false
+
+  tools:
+    ruff:
+      enabled: false
+    pylint:
+      enabled: false
+    flake8:
+      enabled: false
+    gitleaks:
+      enabled: true
+    shellcheck:
+      enabled: false
+    markdownlint:
+      enabled: false
+    yamllint:
+      enabled: false
+    languagetool:
+      enabled: false
+    github-checks:
+      enabled: true
+      timeout_ms: 90000
+    ast-grep:
+      essential_rules: true
+
+chat:
+  auto_reply: true
+
+knowledge_base:
+  opt_out: false
+  learnings:
+    scope: "auto"

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -16,7 +16,7 @@ body:
 
         ## Very Important
 
-        Please make sure that you post ALL your ComfyUI logs in the bug report. A bug report without logs will likely be ignored.
+        Please make sure that you post ALL your ComfyUI logs in the bug report **even if there is no crash**. Just paste everything. The startup log (everything before "To see the GUI go to: ...") contains critical information to developers trying to help. For a performance issue or crash, paste everything from "got prompt" to the end, including the crash. More is better - always. A bug report without logs will likely be ignored.
   - type: checkboxes
     id: custom-nodes-test
     attributes:

.github/workflows/release-webhook.yml

@@ -7,6 +7,8 @@ on:
 jobs:
   send-webhook:
     runs-on: ubuntu-latest
+    env:
+      DESKTOP_REPO_DISPATCH_TOKEN: ${{ secrets.DESKTOP_REPO_DISPATCH_TOKEN }}
     steps:
       - name: Send release webhook
         env:
@@ -106,3 +108,37 @@ jobs:
             --fail --silent --show-error
           
           echo "✅ Release webhook sent successfully"
+
+      - name: Send repository dispatch to desktop
+        env:
+          DISPATCH_TOKEN: ${{ env.DESKTOP_REPO_DISPATCH_TOKEN }}
+          RELEASE_TAG: ${{ github.event.release.tag_name }}
+          RELEASE_URL: ${{ github.event.release.html_url }}
+        run: |
+          set -euo pipefail
+
+          if [ -z "${DISPATCH_TOKEN:-}" ]; then
+            echo "::error::DESKTOP_REPO_DISPATCH_TOKEN is required but not set."
+            exit 1
+          fi
+
+          PAYLOAD="$(jq -n \
+            --arg release_tag "$RELEASE_TAG" \
+            --arg release_url "$RELEASE_URL" \
+            '{
+              event_type: "comfyui_release_published",
+              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 ${DISPATCH_TOKEN}" \
+            https://api.github.com/repos/Comfy-Org/desktop/dispatches \
+            -d "$PAYLOAD"
+
+          echo "✅ Dispatched ComfyUI release ${RELEASE_TAG} to Comfy-Org/desktop"

.gitignore

@@ -11,7 +11,7 @@ extra_model_paths.yaml
 /.vs
 .vscode/
 .idea/
-venv/
+venv*/
 .venv/
 /web/extensions/*
 !/web/extensions/logging.js.example

README.md

@@ -189,8 +189,6 @@ The portable above currently comes with python 3.13 and pytorch cuda 13.0. Updat
 
 [Experimental portable for AMD GPUs](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z)
 
-[Portable with pytorch cuda 12.8 and python 3.12](https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu128.7z).
-
 [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).
 
 #### How do I share models between another UI and ComfyUI?
@@ -227,11 +225,11 @@ Put your VAE in: models/vae
 
 AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version:
 
-```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.4```
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.1```
 
-This is the command to install the nightly with ROCm 7.1 which might have some performance improvements:
+This is the command to install the nightly with ROCm 7.2 which might have some performance improvements:
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.1```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm7.2```
 
 
 ### AMD GPUs (Experimental: Windows and Linux), RDNA 3, 3.5 and 4 only.

app/node_replace_manager.py

@@ -0,0 +1,107 @@
+from __future__ import annotations
+
+from aiohttp import web
+
+from typing import TYPE_CHECKING, TypedDict
+if TYPE_CHECKING:
+    from comfy_api.latest._io_public import NodeReplace
+
+from comfy_execution.graph_utils import is_link
+import nodes
+
+class NodeStruct(TypedDict):
+    inputs: dict[str, str | int | float | bool | tuple[str, int]]
+    class_type: str
+    _meta: dict[str, str]
+
+def copy_node_struct(node_struct: NodeStruct, empty_inputs: bool = False) -> NodeStruct:
+    new_node_struct = node_struct.copy()
+    if empty_inputs:
+        new_node_struct["inputs"] = {}
+    else:
+        new_node_struct["inputs"] = node_struct["inputs"].copy()
+    new_node_struct["_meta"] = node_struct["_meta"].copy()
+    return new_node_struct
+
+
+class NodeReplaceManager:
+    """Manages node replacement registrations."""
+
+    def __init__(self):
+        self._replacements: dict[str, list[NodeReplace]] = {}
+
+    def register(self, node_replace: NodeReplace):
+        """Register a node replacement mapping."""
+        self._replacements.setdefault(node_replace.old_node_id, []).append(node_replace)
+
+    def get_replacement(self, old_node_id: str) -> list[NodeReplace] | None:
+        """Get replacements for an old node ID."""
+        return self._replacements.get(old_node_id)
+
+    def has_replacement(self, old_node_id: str) -> bool:
+        """Check if a replacement exists for an old node ID."""
+        return old_node_id in self._replacements
+
+    def apply_replacements(self, prompt: dict[str, NodeStruct]):
+        connections: dict[str, list[tuple[str, str, int]]] = {}
+        need_replacement: set[str] = set()
+        for node_number, node_struct in prompt.items():
+            if "class_type" not in node_struct or "inputs" not in node_struct:
+                continue
+            class_type = node_struct["class_type"]
+            # need replacement if not in NODE_CLASS_MAPPINGS and has replacement
+            if class_type not in nodes.NODE_CLASS_MAPPINGS.keys() and self.has_replacement(class_type):
+                need_replacement.add(node_number)
+            # keep track of connections
+            for input_id, input_value in node_struct["inputs"].items():
+                if is_link(input_value):
+                    conn_number = input_value[0]
+                    connections.setdefault(conn_number, []).append((node_number, input_id, input_value[1]))
+        for node_number in need_replacement:
+            node_struct = prompt[node_number]
+            class_type = node_struct["class_type"]
+            replacements = self.get_replacement(class_type)
+            if replacements is None:
+                continue
+            # just use the first replacement
+            replacement = replacements[0]
+            new_node_id = replacement.new_node_id
+            # if replacement is not a valid node, skip trying to replace it as will only cause confusion
+            if new_node_id not in nodes.NODE_CLASS_MAPPINGS.keys():
+                continue
+            # first, replace node id (class_type)
+            new_node_struct = copy_node_struct(node_struct, empty_inputs=True)
+            new_node_struct["class_type"] = new_node_id
+            # TODO: consider replacing display_name in _meta as well for error reporting purposes; would need to query node schema
+            # second, replace inputs
+            if replacement.input_mapping is not None:
+                for input_map in replacement.input_mapping:
+                    if "set_value" in input_map:
+                        new_node_struct["inputs"][input_map["new_id"]] = input_map["set_value"]
+                    elif "old_id" in input_map:
+                        new_node_struct["inputs"][input_map["new_id"]] = node_struct["inputs"][input_map["old_id"]]
+            # finalize input replacement
+            prompt[node_number] = new_node_struct
+            # third, replace outputs
+            if replacement.output_mapping is not None:
+                # re-mapping outputs requires changing the input values of nodes that receive connections from this one
+                if node_number in connections:
+                    for conns in connections[node_number]:
+                        conn_node_number, conn_input_id, old_output_idx = conns
+                        for output_map in replacement.output_mapping:
+                            if output_map["old_idx"] == old_output_idx:
+                                new_output_idx = output_map["new_idx"]
+                                previous_input = prompt[conn_node_number]["inputs"][conn_input_id]
+                                previous_input[1] = new_output_idx
+
+    def as_dict(self):
+        """Serialize all replacements to dict."""
+        return {
+            k: [v.as_dict() for v in v_list]
+            for k, v_list in self._replacements.items()
+        }
+
+    def add_routes(self, routes):
+        @routes.get("/node_replacements")
+        async def get_node_replacements(request):
+            return web.json_response(self.as_dict())

app/subgraph_manager.py

@@ -53,7 +53,7 @@ class SubgraphManager:
         return entry_id, entry
 
     async def load_entry_data(self, entry: SubgraphEntry):
-        with open(entry['path'], 'r') as f:
+        with open(entry['path'], 'r', encoding='utf-8') as f:
             entry['data'] = f.read()
         return entry
 

blueprints/.glsl/Brightness_and_Contrast_1.frag

@@ -0,0 +1,44 @@
+#version 300 es
+precision highp float;
+
+uniform sampler2D u_image0;
+uniform float u_float0; // Brightness slider -100..100
+uniform float u_float1; // Contrast slider -100..100
+
+in vec2 v_texCoord;
+out vec4 fragColor;
+
+const float MID_GRAY = 0.18;  // 18% reflectance
+
+// sRGB gamma 2.2 approximation
+vec3 srgbToLinear(vec3 c) {
+    return pow(max(c, 0.0), vec3(2.2));
+}
+
+vec3 linearToSrgb(vec3 c) {
+    return pow(max(c, 0.0), vec3(1.0/2.2));
+}
+
+float mapBrightness(float b) {
+    return clamp(b / 100.0, -1.0, 1.0);
+}
+
+float mapContrast(float c) {
+    return clamp(c / 100.0 + 1.0, 0.0, 2.0);
+}
+
+void main() {
+    vec4 orig = texture(u_image0, v_texCoord);
+
+    float brightness = mapBrightness(u_float0);
+    float contrast   = mapContrast(u_float1);
+
+    vec3 lin = srgbToLinear(orig.rgb);
+
+    lin = (lin - MID_GRAY) * contrast + brightness + MID_GRAY;
+
+    // Convert back to sRGB
+    vec3 result = linearToSrgb(clamp(lin, 0.0, 1.0));
+
+    fragColor = vec4(result, orig.a);
+}

blueprints/.glsl/Chromatic_Aberration_16.frag

@@ -0,0 +1,72 @@
+#version 300 es
+precision highp float;
+
+uniform sampler2D u_image0;
+uniform vec2 u_resolution;
+uniform int u_int0;      // Mode
+uniform float u_float0;  // Amount (0 to 100)
+
+in vec2 v_texCoord;
+out vec4 fragColor;
+
+const int MODE_LINEAR   = 0;
+const int MODE_RADIAL   = 1;
+const int MODE_BARREL   = 2;
+const int MODE_SWIRL    = 3;
+const int MODE_DIAGONAL = 4;
+
+const float AMOUNT_SCALE = 0.0005;
+const float RADIAL_MULT = 4.0;
+const float BARREL_MULT = 8.0;
+const float INV_SQRT2 = 0.70710678118;
+
+void main() {
+    vec2 uv = v_texCoord;
+    vec4 original = texture(u_image0, uv);
+
+    float amount = u_float0 * AMOUNT_SCALE;
+
+    if (amount < 0.000001) {
+        fragColor = original;
+        return;
+    }
+
+    // Aspect-corrected coordinates for circular effects
+    float aspect = u_resolution.x / u_resolution.y;
+    vec2 centered = uv - 0.5;
+    vec2 corrected = vec2(centered.x * aspect, centered.y);
+    float r = length(corrected);
+    vec2 dir = r > 0.0001 ? corrected / r : vec2(0.0);
+    vec2 offset = vec2(0.0);
+
+    if (u_int0 == MODE_LINEAR) {
+        // Horizontal shift (no aspect correction needed)
+        offset = vec2(amount, 0.0);
+    }
+    else if (u_int0 == MODE_RADIAL) {
+        // Outward from center, stronger at edges
+        offset = dir * r * amount * RADIAL_MULT;
+        offset.x /= aspect;  // Convert back to UV space
+    }
+    else if (u_int0 == MODE_BARREL) {
+        // Lens distortion simulation (r² falloff)
+        offset = dir * r * r * amount * BARREL_MULT;
+        offset.x /= aspect;  // Convert back to UV space
+    }
+    else if (u_int0 == MODE_SWIRL) {
+        // Perpendicular to radial (rotational aberration)
+        vec2 perp = vec2(-dir.y, dir.x);
+        offset = perp * r * amount * RADIAL_MULT;
+        offset.x /= aspect;  // Convert back to UV space
+    }
+    else if (u_int0 == MODE_DIAGONAL) {
+        // 45° offset (no aspect correction needed)
+        offset = vec2(amount, amount) * INV_SQRT2;
+    }
+    
+    float red = texture(u_image0, uv + offset).r;
+    float green = original.g;
+    float blue = texture(u_image0, uv - offset).b;
+    
+    fragColor = vec4(red, green, blue, original.a);
+}

blueprints/.glsl/Color_Adjustment_15.frag

@@ -0,0 +1,78 @@
+#version 300 es
+precision highp float;
+
+uniform sampler2D u_image0;
+uniform float u_float0; // temperature (-100 to 100)
+uniform float u_float1; // tint (-100 to 100)
+uniform float u_float2; // vibrance (-100 to 100)
+uniform float u_float3; // saturation (-100 to 100)
+
+in vec2 v_texCoord;
+out vec4 fragColor;
+
+const float INPUT_SCALE = 0.01;
+const float TEMP_TINT_PRIMARY = 0.3;
+const float TEMP_TINT_SECONDARY = 0.15;
+const float VIBRANCE_BOOST = 2.0;
+const float SATURATION_BOOST = 2.0;
+const float SKIN_PROTECTION = 0.5;
+const float EPSILON = 0.001;
+const vec3 LUMA_WEIGHTS = vec3(0.299, 0.587, 0.114);
+
+void main() {
+    vec4 tex = texture(u_image0, v_texCoord);
+    vec3 color = tex.rgb;
+    
+    // Scale inputs: -100/100 → -1/1
+    float temperature = u_float0 * INPUT_SCALE;
+    float tint = u_float1 * INPUT_SCALE;
+    float vibrance = u_float2 * INPUT_SCALE;
+    float saturation = u_float3 * INPUT_SCALE;
+    
+    // Temperature (warm/cool): positive = warm, negative = cool
+    color.r += temperature * TEMP_TINT_PRIMARY;
+    color.b -= temperature * TEMP_TINT_PRIMARY;
+    
+    // Tint (green/magenta): positive = green, negative = magenta
+    color.g += tint * TEMP_TINT_PRIMARY;
+    color.r -= tint * TEMP_TINT_SECONDARY;
+    color.b -= tint * TEMP_TINT_SECONDARY;
+    
+    // Single clamp after temperature/tint
+    color = clamp(color, 0.0, 1.0);
+    
+    // Vibrance with skin protection
+    if (vibrance != 0.0) {
+        float maxC = max(color.r, max(color.g, color.b));
+        float minC = min(color.r, min(color.g, color.b));
+        float sat = maxC - minC;
+        float gray = dot(color, LUMA_WEIGHTS);
+        
+        if (vibrance < 0.0) {
+            // Desaturate: -100 → gray
+            color = mix(vec3(gray), color, 1.0 + vibrance);
+        } else {
+            // Boost less saturated colors more
+            float vibranceAmt = vibrance * (1.0 - sat);
+            
+            // Branchless skin tone protection
+            float isWarmTone = step(color.b, color.g) * step(color.g, color.r);
+            float warmth = (color.r - color.b) / max(maxC, EPSILON);
+            float skinTone = isWarmTone * warmth * sat * (1.0 - sat);
+            vibranceAmt *= (1.0 - skinTone * SKIN_PROTECTION);
+            
+            color = mix(vec3(gray), color, 1.0 + vibranceAmt * VIBRANCE_BOOST);
+        }
+    }
+    
+    // Saturation
+    if (saturation != 0.0) {
+        float gray = dot(color, LUMA_WEIGHTS);
+        float satMix = saturation < 0.0
+            ? 1.0 + saturation                      // -100 → gray
+            : 1.0 + saturation * SATURATION_BOOST;  // +100 → 3x boost
+        color = mix(vec3(gray), color, satMix);
+    }
+    
+    fragColor = vec4(clamp(color, 0.0, 1.0), tex.a);
+}

blueprints/.glsl/Edge-Preserving_Blur_128.frag

@@ -0,0 +1,94 @@
+#version 300 es
+precision highp float;
+
+uniform sampler2D u_image0;
+uniform float u_float0;   // Blur radius (0–20, default ~5)
+uniform float u_float1;   // Edge threshold (0–100, default ~30)
+uniform int u_int0;       // Step size (0/1 = every pixel, 2+ = skip pixels)
+
+in vec2 v_texCoord;
+out vec4 fragColor;
+
+const int MAX_RADIUS = 20;
+const float EPSILON = 0.0001;
+
+// Perceptual luminance
+float getLuminance(vec3 rgb) {
+    return dot(rgb, vec3(0.299, 0.587, 0.114));
+}
+
+vec4 bilateralFilter(vec2 uv, vec2 texelSize, int radius,
+                     float sigmaSpatial, float sigmaColor)
+{
+    vec4 center = texture(u_image0, uv);
+    vec3 centerRGB = center.rgb;
+
+    float invSpatial2 = -0.5 / (sigmaSpatial * sigmaSpatial);
+    float invColor2   = -0.5 / (sigmaColor * sigmaColor + EPSILON);
+
+    vec3 sumRGB = vec3(0.0);
+    float sumWeight = 0.0;
+
+    int step = max(u_int0, 1);
+    float radius2 = float(radius * radius);
+
+    for (int dy = -MAX_RADIUS; dy <= MAX_RADIUS; dy++) {
+        if (dy < -radius || dy > radius) continue;
+        if (abs(dy) % step != 0) continue;
+
+        for (int dx = -MAX_RADIUS; dx <= MAX_RADIUS; dx++) {
+            if (dx < -radius || dx > radius) continue;
+            if (abs(dx) % step != 0) continue;
+
+            vec2 offset = vec2(float(dx), float(dy));
+            float dist2 = dot(offset, offset);
+            if (dist2 > radius2) continue;
+
+            vec3 sampleRGB = texture(u_image0, uv + offset * texelSize).rgb;
+
+            // Spatial Gaussian
+            float spatialWeight = exp(dist2 * invSpatial2);
+
+            // Perceptual color distance (weighted RGB)
+            vec3 diff = sampleRGB - centerRGB;
+            float colorDist = dot(diff * diff, vec3(0.299, 0.587, 0.114));
+            float colorWeight = exp(colorDist * invColor2);
+
+            float w = spatialWeight * colorWeight;
+            sumRGB += sampleRGB * w;
+            sumWeight += w;
+        }
+    }
+
+    vec3 resultRGB = sumRGB / max(sumWeight, EPSILON);
+    return vec4(resultRGB, center.a); // preserve center alpha
+}
+
+void main() {
+    vec2 texelSize = 1.0 / vec2(textureSize(u_image0, 0));
+
+    float radiusF = clamp(u_float0, 0.0, float(MAX_RADIUS));
+    int radius = int(radiusF + 0.5);
+
+    if (radius == 0) {
+        fragColor = texture(u_image0, v_texCoord);
+        return;
+    }
+
+    // Edge threshold → color sigma
+    // Squared curve for better low-end control
+    float t = clamp(u_float1, 0.0, 100.0) / 100.0;
+    t *= t;
+    float sigmaColor = mix(0.01, 0.5, t);
+
+    // Spatial sigma tied to radius
+    float sigmaSpatial = max(radiusF * 0.75, 0.5);
+
+    fragColor = bilateralFilter(
+        v_texCoord,
+        texelSize,
+        radius,
+        sigmaSpatial,
+        sigmaColor
+    );
+}

blueprints/.glsl/Film_Grain_15.frag

@@ -0,0 +1,124 @@
+#version 300 es
+precision highp float;
+
+uniform sampler2D u_image0;
+uniform vec2 u_resolution;
+uniform float u_float0; // grain amount      [0.0 – 1.0]   typical: 0.2–0.8
+uniform float u_float1; // grain size        [0.3 – 3.0]   lower = finer grain
+uniform float u_float2; // color amount      [0.0 – 1.0]   0 = monochrome, 1 = RGB grain
+uniform float u_float3; // luminance bias    [0.0 – 1.0]   0 = uniform, 1 = shadows only
+uniform int   u_int0;   // noise mode        [0 or 1]      0 = smooth, 1 = grainy
+
+in vec2 v_texCoord;
+layout(location = 0) out vec4 fragColor0;
+
+// High-quality integer hash (pcg-like)
+uint pcg(uint v) {
+    uint state = v * 747796405u + 2891336453u;
+    uint word = ((state >> ((state >> 28u) + 4u)) ^ state) * 277803737u;
+    return (word >> 22u) ^ word;
+}
+
+// 2D -> 1D hash input
+uint hash2d(uvec2 p) {
+    return pcg(p.x + pcg(p.y));
+}
+
+// Hash to float [0, 1]
+float hashf(uvec2 p) {
+    return float(hash2d(p)) / float(0xffffffffu);
+}
+
+// Hash to float with offset (for RGB channels)
+float hashf(uvec2 p, uint offset) {
+    return float(pcg(hash2d(p) + offset)) / float(0xffffffffu);
+}
+
+// Convert uniform [0,1] to roughly Gaussian distribution
+// Using simple approximation: average of multiple samples
+float toGaussian(uvec2 p) {
+    float sum = hashf(p, 0u) + hashf(p, 1u) + hashf(p, 2u) + hashf(p, 3u);
+    return (sum - 2.0) * 0.7;  // Centered, scaled
+}
+
+float toGaussian(uvec2 p, uint offset) {
+    float sum = hashf(p, offset) + hashf(p, offset + 1u) 
+              + hashf(p, offset + 2u) + hashf(p, offset + 3u);
+    return (sum - 2.0) * 0.7;
+}
+
+// Smooth noise with better interpolation
+float smoothNoise(vec2 p) {
+    vec2 i = floor(p);
+    vec2 f = fract(p);
+    
+    // Quintic interpolation (less banding than cubic)
+    f = f * f * f * (f * (f * 6.0 - 15.0) + 10.0);
+    
+    uvec2 ui = uvec2(i);
+    float a = toGaussian(ui);
+    float b = toGaussian(ui + uvec2(1u, 0u));
+    float c = toGaussian(ui + uvec2(0u, 1u));
+    float d = toGaussian(ui + uvec2(1u, 1u));
+    
+    return mix(mix(a, b, f.x), mix(c, d, f.x), f.y);
+}
+
+float smoothNoise(vec2 p, uint offset) {
+    vec2 i = floor(p);
+    vec2 f = fract(p);
+    
+    f = f * f * f * (f * (f * 6.0 - 15.0) + 10.0);
+    
+    uvec2 ui = uvec2(i);
+    float a = toGaussian(ui, offset);
+    float b = toGaussian(ui + uvec2(1u, 0u), offset);
+    float c = toGaussian(ui + uvec2(0u, 1u), offset);
+    float d = toGaussian(ui + uvec2(1u, 1u), offset);
+    
+    return mix(mix(a, b, f.x), mix(c, d, f.x), f.y);
+}
+
+void main() {
+    vec4 color = texture(u_image0, v_texCoord);
+    
+    // Luminance (Rec.709)
+    float luma = dot(color.rgb, vec3(0.2126, 0.7152, 0.0722));
+    
+    // Grain UV (resolution-independent)
+    vec2 grainUV = v_texCoord * u_resolution / max(u_float1, 0.01);
+    uvec2 grainPixel = uvec2(grainUV);
+    
+    float g;
+    vec3 grainRGB;
+    
+    if (u_int0 == 1) {
+        // Grainy mode: pure hash noise (no interpolation = no banding)
+        g = toGaussian(grainPixel);
+        grainRGB = vec3(
+            toGaussian(grainPixel, 100u),
+            toGaussian(grainPixel, 200u),
+            toGaussian(grainPixel, 300u)
+        );
+    } else {
+        // Smooth mode: interpolated with quintic curve
+        g = smoothNoise(grainUV);
+        grainRGB = vec3(
+            smoothNoise(grainUV, 100u),
+            smoothNoise(grainUV, 200u),
+            smoothNoise(grainUV, 300u)
+        );
+    }
+    
+    // Luminance weighting (less grain in highlights)
+    float lumWeight = mix(1.0, 1.0 - luma, clamp(u_float3, 0.0, 1.0));
+    
+    // Strength
+    float strength = u_float0 * 0.15;
+    
+    // Color vs monochrome grain
+    vec3 grainColor = mix(vec3(g), grainRGB, clamp(u_float2, 0.0, 1.0));
+    
+    color.rgb += grainColor * strength * lumWeight;
+    fragColor0 = vec4(clamp(color.rgb, 0.0, 1.0), color.a);
+}

blueprints/.glsl/Glow_30.frag

@@ -0,0 +1,133 @@
+#version 300 es
+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
+uniform float u_float1;  // Radius
+uniform float u_float2;  // Threshold
+
+in vec2 v_texCoord;
+out vec4 fragColor;
+
+const int BLEND_ADD      = 0;
+const int BLEND_SCREEN   = 1;
+const int BLEND_SOFT     = 2;
+const int BLEND_OVERLAY  = 3;
+const int BLEND_LIGHTEN  = 4;
+
+const float GOLDEN_ANGLE = 2.39996323;
+const int MAX_SAMPLES = 48;
+const vec3 LUMA = vec3(0.299, 0.587, 0.114);
+
+float hash(vec2 p) {
+    p = fract(p * vec2(123.34, 456.21));
+    p += dot(p, p + 45.32);
+    return fract(p.x * p.y);
+}
+
+vec3 hexToRgb(int h) {
+    return vec3(
+        float((h >> 16) & 255),
+        float((h >> 8) & 255),
+        float(h & 255)
+    ) * (1.0 / 255.0);
+}
+
+vec3 blend(vec3 base, vec3 glow, int mode) {
+    if (mode == BLEND_SCREEN) {
+        return 1.0 - (1.0 - base) * (1.0 - glow);
+    }
+    if (mode == BLEND_SOFT) {
+        return mix(
+            base - (1.0 - 2.0 * glow) * base * (1.0 - base),
+            base + (2.0 * glow - 1.0) * (sqrt(base) - base),
+            step(0.5, glow)
+        );
+    }
+    if (mode == BLEND_OVERLAY) {
+        return mix(
+            2.0 * base * glow,
+            1.0 - 2.0 * (1.0 - base) * (1.0 - glow),
+            step(0.5, base)
+        );
+    }
+    if (mode == BLEND_LIGHTEN) {
+        return max(base, glow);
+    }
+    return base + glow;
+}
+
+void main() {
+    vec4 original = texture(u_image0, v_texCoord);
+    
+    float intensity = u_float0 * 0.05;
+    float radius = u_float1 * u_float1 * 0.012;
+    
+    if (intensity < 0.001 || radius < 0.1) {
+        fragColor = original;
+        return;
+    }
+    
+    float threshold = 1.0 - u_float2 * 0.01;
+    float t0 = threshold - 0.15;
+    float t1 = threshold + 0.15;
+    
+    vec2 texelSize = 1.0 / u_resolution;
+    float radius2 = radius * radius;
+    
+    float sampleScale = clamp(radius * 0.75, 0.35, 1.0);
+    int samples = int(float(MAX_SAMPLES) * sampleScale);
+    
+    float noise = hash(gl_FragCoord.xy);
+    float angleOffset = noise * GOLDEN_ANGLE;
+    float radiusJitter = 0.85 + noise * 0.3;
+    
+    float ca = cos(GOLDEN_ANGLE);
+    float sa = sin(GOLDEN_ANGLE);
+    vec2 dir = vec2(cos(angleOffset), sin(angleOffset));
+    
+    vec3 glow = vec3(0.0);
+    float totalWeight = 0.0;
+    
+    // Center tap
+    float centerMask = smoothstep(t0, t1, dot(original.rgb, LUMA));
+    glow += original.rgb * centerMask * 2.0;
+    totalWeight += 2.0;
+    
+    for (int i = 1; i < MAX_SAMPLES; i++) {
+        if (i >= samples) break;
+        
+        float fi = float(i);
+        float dist = sqrt(fi / float(samples)) * radius * radiusJitter;
+        
+        vec2 offset = dir * dist * texelSize;
+        vec3 c = texture(u_image0, v_texCoord + offset).rgb;
+        float mask = smoothstep(t0, t1, dot(c, LUMA));
+        
+        float w = 1.0 - (dist * dist) / (radius2 * 1.5);
+        w = max(w, 0.0);
+        w *= w;
+        
+        glow += c * mask * w;
+        totalWeight += w;
+        
+        dir = vec2(
+            dir.x * ca - dir.y * sa,
+            dir.x * sa + dir.y * ca
+        );
+    }
+    
+    glow *= intensity / max(totalWeight, 0.001);
+    
+    if (u_int1 > 0) {
+        glow *= hexToRgb(u_int1);
+    }
+    
+    vec3 result = blend(original.rgb, glow, u_int0);
+    result += (noise - 0.5) * (1.0 / 255.0);
+    
+    fragColor = vec4(clamp(result, 0.0, 1.0), original.a);
+}

blueprints/.glsl/Hue_and_Saturation_1.frag

@@ -0,0 +1,222 @@
+#version 300 es
+precision highp float;
+
+uniform sampler2D u_image0;
+uniform int u_int0;      // Mode: 0=Master, 1=Reds, 2=Yellows, 3=Greens, 4=Cyans, 5=Blues, 6=Magentas, 7=Colorize
+uniform int u_int1;      // Color Space: 0=HSL, 1=HSB/HSV
+uniform float u_float0;  // Hue (-180 to 180)
+uniform float u_float1;  // Saturation (-100 to 100)
+uniform float u_float2;  // Lightness/Brightness (-100 to 100)
+uniform float u_float3;  // Overlap (0 to 100) - feathering between adjacent color ranges
+
+in vec2 v_texCoord;
+out vec4 fragColor;
+
+// Color range modes
+const int MODE_MASTER   = 0;
+const int MODE_RED      = 1;
+const int MODE_YELLOW   = 2;
+const int MODE_GREEN    = 3;
+const int MODE_CYAN     = 4;
+const int MODE_BLUE     = 5;
+const int MODE_MAGENTA  = 6;
+const int MODE_COLORIZE = 7;
+
+// Color space modes
+const int COLORSPACE_HSL = 0;
+const int COLORSPACE_HSB = 1;
+
+const float EPSILON = 0.0001;
+
+//=============================================================================
+// RGB <-> HSL Conversions
+//=============================================================================
+
+vec3 rgb2hsl(vec3 c) {
+    float maxC = max(max(c.r, c.g), c.b);
+    float minC = min(min(c.r, c.g), c.b);
+    float delta = maxC - minC;
+
+    float h = 0.0;
+    float s = 0.0;
+    float l = (maxC + minC) * 0.5;
+
+    if (delta > EPSILON) {
+        s = l < 0.5
+            ? delta / (maxC + minC)
+            : delta / (2.0 - maxC - minC);
+
+        if (maxC == c.r) {
+            h = (c.g - c.b) / delta + (c.g < c.b ? 6.0 : 0.0);
+        } else if (maxC == c.g) {
+            h = (c.b - c.r) / delta + 2.0;
+        } else {
+            h = (c.r - c.g) / delta + 4.0;
+        }
+        h /= 6.0;
+    }
+
+    return vec3(h, s, l);
+}
+
+float hue2rgb(float p, float q, float t) {
+    t = fract(t);
+    if (t < 1.0/6.0) return p + (q - p) * 6.0 * t;
+    if (t < 0.5)       return q;
+    if (t < 2.0/3.0)   return p + (q - p) * (2.0/3.0 - t) * 6.0;
+    return p;
+}
+
+vec3 hsl2rgb(vec3 hsl) {
+    if (hsl.y < EPSILON) return vec3(hsl.z);
+
+    float q = hsl.z < 0.5
+        ? hsl.z * (1.0 + hsl.y)
+        : hsl.z + hsl.y - hsl.z * hsl.y;
+    float p = 2.0 * hsl.z - q;
+
+    return vec3(
+        hue2rgb(p, q, hsl.x + 1.0/3.0),
+        hue2rgb(p, q, hsl.x),
+        hue2rgb(p, q, hsl.x - 1.0/3.0)
+    );
+}
+
+vec3 rgb2hsb(vec3 c) {
+    float maxC = max(max(c.r, c.g), c.b);
+    float minC = min(min(c.r, c.g), c.b);
+    float delta = maxC - minC;
+
+    float h = 0.0;
+    float s = (maxC > EPSILON) ? delta / maxC : 0.0;
+    float b = maxC;
+
+    if (delta > EPSILON) {
+        if (maxC == c.r) {
+            h = (c.g - c.b) / delta + (c.g < c.b ? 6.0 : 0.0);
+        } else if (maxC == c.g) {
+            h = (c.b - c.r) / delta + 2.0;
+        } else {
+            h = (c.r - c.g) / delta + 4.0;
+        }
+        h /= 6.0;
+    }
+
+    return vec3(h, s, b);
+}
+
+vec3 hsb2rgb(vec3 hsb) {
+    vec3 rgb = clamp(abs(mod(hsb.x * 6.0 + vec3(0.0, 4.0, 2.0), 6.0) - 3.0) - 1.0, 0.0, 1.0);
+    return hsb.z * mix(vec3(1.0), rgb, hsb.y);
+}
+
+//=============================================================================
+// Color Range Weight Calculation
+//=============================================================================
+
+float hueDistance(float a, float b) {
+    float d = abs(a - b);
+    return min(d, 1.0 - d);
+}
+
+float getHueWeight(float hue, float center, float overlap) {
+    float baseWidth = 1.0 / 6.0;
+    float feather = baseWidth * overlap;
+
+    float d = hueDistance(hue, center);
+
+    float inner = baseWidth * 0.5;
+    float outer = inner + feather;
+
+    return 1.0 - smoothstep(inner, outer, d);
+}
+
+float getModeWeight(float hue, int mode, float overlap) {
+    if (mode == MODE_MASTER || mode == MODE_COLORIZE) return 1.0;
+
+    if (mode == MODE_RED) {
+        return max(
+            getHueWeight(hue, 0.0, overlap),
+            getHueWeight(hue, 1.0, overlap)
+        );
+    }
+
+    float center = float(mode - 1) / 6.0;
+    return getHueWeight(hue, center, overlap);
+}
+
+//=============================================================================
+// Adjustment Functions
+//=============================================================================
+
+float adjustLightness(float l, float amount) {
+    return amount > 0.0
+        ? l + (1.0 - l) * amount
+        : l + l * amount;
+}
+
+float adjustBrightness(float b, float amount) {
+    return clamp(b + amount, 0.0, 1.0);
+}
+
+float adjustSaturation(float s, float amount) {
+    return amount > 0.0
+        ? s + (1.0 - s) * amount
+        : s + s * amount;
+}
+
+vec3 colorize(vec3 rgb, float hue, float sat, float light) {
+    float lum = dot(rgb, vec3(0.299, 0.587, 0.114));
+    float l = adjustLightness(lum, light);
+
+    vec3 hsl = vec3(fract(hue), clamp(sat, 0.0, 1.0), clamp(l, 0.0, 1.0));
+    return hsl2rgb(hsl);
+}
+
+//=============================================================================
+// Main
+//=============================================================================
+
+void main() {
+    vec4 original = texture(u_image0, v_texCoord);
+
+    float hueShift   = u_float0 / 360.0;   // -180..180 -> -0.5..0.5
+    float satAmount  = u_float1 / 100.0;   // -100..100 -> -1..1
+    float lightAmount= u_float2 / 100.0;   // -100..100 -> -1..1
+    float overlap    = u_float3 / 100.0;   // 0..100 -> 0..1
+
+    vec3 result;
+
+    if (u_int0 == MODE_COLORIZE) {
+        result = colorize(original.rgb, hueShift, satAmount, lightAmount);
+        fragColor = vec4(result, original.a);
+        return;
+    }
+
+    vec3 hsx = (u_int1 == COLORSPACE_HSL)
+        ? rgb2hsl(original.rgb)
+        : rgb2hsb(original.rgb);
+
+    float weight = getModeWeight(hsx.x, u_int0, overlap);
+
+    if (u_int0 != MODE_MASTER && hsx.y < EPSILON) {
+        weight = 0.0;
+    }
+
+    if (weight > EPSILON) {
+        float h = fract(hsx.x + hueShift * weight);
+        float s = clamp(adjustSaturation(hsx.y, satAmount * weight), 0.0, 1.0);
+        float v = (u_int1 == COLORSPACE_HSL)
+            ? clamp(adjustLightness(hsx.z, lightAmount * weight), 0.0, 1.0)
+            : clamp(adjustBrightness(hsx.z, lightAmount * weight), 0.0, 1.0);
+
+        vec3 adjusted = vec3(h, s, v);
+        result = (u_int1 == COLORSPACE_HSL)
+            ? hsl2rgb(adjusted)
+            : hsb2rgb(adjusted);
+    } else {
+        result = original.rgb;
+    }
+
+    fragColor = vec4(result, original.a);
+}

blueprints/.glsl/Image_Blur_1.frag

@@ -0,0 +1,111 @@
+#version 300 es
+#pragma passes 2
+precision highp float;
+
+// Blur type constants
+const int BLUR_GAUSSIAN = 0;
+const int BLUR_BOX = 1;
+const int BLUR_RADIAL = 2;
+
+// Radial blur config
+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)
+
+in vec2 v_texCoord;
+layout(location = 0) out vec4 fragColor0;
+
+float gaussian(float x, float sigma) {
+    return exp(-(x * x) / (2.0 * sigma * sigma));
+}
+
+void main() {
+    vec2 texelSize = 1.0 / u_resolution;
+    float radius = max(u_float0, 0.0);
+
+    // Radial (angular) blur - single pass, doesn't use separable
+    if (u_int0 == BLUR_RADIAL) {
+        // Only execute on first pass
+        if (u_pass > 0) {
+            fragColor0 = texture(u_image0, v_texCoord);
+            return;
+        }
+
+        vec2 center = vec2(0.5);
+        vec2 dir = v_texCoord - center;
+        float dist = length(dir);
+
+        if (dist < 1e-4) {
+            fragColor0 = texture(u_image0, v_texCoord);
+            return;
+        }
+
+        vec4 sum = vec4(0.0);
+        float totalWeight = 0.0;
+        float angleStep = radius * RADIAL_STRENGTH;
+
+        dir /= dist;
+
+        float cosStep = cos(angleStep);
+        float sinStep = sin(angleStep);
+
+        float negAngle = -float(RADIAL_SAMPLES) * angleStep;
+        vec2 rotDir = vec2(
+            dir.x * cos(negAngle) - dir.y * sin(negAngle),
+            dir.x * sin(negAngle) + dir.y * cos(negAngle)
+        );
+
+        for (int i = -RADIAL_SAMPLES; i <= RADIAL_SAMPLES; i++) {
+            vec2 uv = center + rotDir * dist;
+            float w = 1.0 - abs(float(i)) / float(RADIAL_SAMPLES);
+            sum += texture(u_image0, uv) * w;
+            totalWeight += w;
+
+            rotDir = vec2(
+                rotDir.x * cosStep - rotDir.y * sinStep,
+                rotDir.x * sinStep + rotDir.y * cosStep
+            );
+        }
+
+        fragColor0 = sum / max(totalWeight, 0.001);
+        return;
+    }
+
+    // Separable Gaussian / Box blur
+    int samples = int(ceil(radius));
+
+    if (samples == 0) {
+        fragColor0 = texture(u_image0, v_texCoord);
+        return;
+    }
+
+    // Direction: pass 0 = horizontal, pass 1 = vertical
+    vec2 dir = (u_pass == 0) ? vec2(1.0, 0.0) : vec2(0.0, 1.0);
+
+    vec4 color = vec4(0.0);
+    float totalWeight = 0.0;
+    float sigma = radius / 2.0;
+
+    for (int i = -samples; i <= samples; i++) {
+        vec2 offset = dir * float(i) * texelSize;
+        vec4 sample_color = texture(u_image0, v_texCoord + offset);
+
+        float weight;
+        if (u_int0 == BLUR_GAUSSIAN) {
+            weight = gaussian(float(i), sigma);
+        } else {
+            // BLUR_BOX
+            weight = 1.0;
+        }
+
+        color += sample_color * weight;
+        totalWeight += weight;
+    }
+
+    fragColor0 = color / totalWeight;
+}

blueprints/.glsl/Image_Channels_23.frag

@@ -0,0 +1,19 @@
+#version 300 es
+precision highp float;
+
+uniform sampler2D u_image0;
+
+in vec2 v_texCoord;
+layout(location = 0) out vec4 fragColor0;
+layout(location = 1) out vec4 fragColor1;
+layout(location = 2) out vec4 fragColor2;
+layout(location = 3) out vec4 fragColor3;
+
+void main() {
+  vec4 color = texture(u_image0, v_texCoord);
+  // Output each channel as grayscale to separate render targets
+  fragColor0 = vec4(vec3(color.r), 1.0);  // Red channel
+  fragColor1 = vec4(vec3(color.g), 1.0);  // Green channel
+  fragColor2 = vec4(vec3(color.b), 1.0);  // Blue channel
+  fragColor3 = vec4(vec3(color.a), 1.0);  // Alpha channel
+}

blueprints/.glsl/Image_Levels_1.frag

@@ -0,0 +1,71 @@
+#version 300 es
+precision highp float;
+
+// Levels Adjustment
+// u_int0:   channel      (0=RGB, 1=R, 2=G, 3=B)         default: 0
+// u_float0: input black  (0-255)                        default: 0
+// u_float1: input white  (0-255)                        default: 255
+// u_float2: gamma        (0.01-9.99)                    default: 1.0
+// u_float3: output black (0-255)                        default: 0
+// u_float4: output white (0-255)                        default: 255
+
+uniform sampler2D u_image0;
+uniform int u_int0;
+uniform float u_float0;
+uniform float u_float1;
+uniform float u_float2;
+uniform float u_float3;
+uniform float u_float4;
+
+in vec2 v_texCoord;
+out vec4 fragColor;
+
+vec3 applyLevels(vec3 color, float inBlack, float inWhite, float gamma, float outBlack, float outWhite) {
+    float inRange = max(inWhite - inBlack, 0.0001);
+    vec3 result = clamp((color - inBlack) / inRange, 0.0, 1.0);
+    result = pow(result, vec3(1.0 / gamma));
+    result = mix(vec3(outBlack), vec3(outWhite), result);
+    return result;
+}
+
+float applySingleChannel(float value, float inBlack, float inWhite, float gamma, float outBlack, float outWhite) {
+    float inRange = max(inWhite - inBlack, 0.0001);
+    float result = clamp((value - inBlack) / inRange, 0.0, 1.0);
+    result = pow(result, 1.0 / gamma);
+    result = mix(outBlack, outWhite, result);
+    return result;
+}
+
+void main() {
+    vec4 texColor = texture(u_image0, v_texCoord);
+    vec3 color = texColor.rgb;
+    
+    float inBlack = u_float0 / 255.0;
+    float inWhite = u_float1 / 255.0;
+    float gamma = u_float2;
+    float outBlack = u_float3 / 255.0;
+    float outWhite = u_float4 / 255.0;
+    
+    vec3 result;
+    
+    if (u_int0 == 0) {
+        result = applyLevels(color, inBlack, inWhite, gamma, outBlack, outWhite);
+    }
+    else if (u_int0 == 1) {
+        result = color;
+        result.r = applySingleChannel(color.r, inBlack, inWhite, gamma, outBlack, outWhite);
+    }
+    else if (u_int0 == 2) {
+        result = color;
+        result.g = applySingleChannel(color.g, inBlack, inWhite, gamma, outBlack, outWhite);
+    }
+    else if (u_int0 == 3) {
+        result = color;
+        result.b = applySingleChannel(color.b, inBlack, inWhite, gamma, outBlack, outWhite);
+    }
+    else {
+        result = color;
+    }
+    
+    fragColor = vec4(result, texColor.a);
+}

blueprints/.glsl/README.md

@@ -0,0 +1,28 @@
+# GLSL Shader Sources
+
+This folder contains the GLSL fragment shaders extracted from blueprint JSON files for easier editing and version control.
+
+## File Naming Convention
+
+`{Blueprint_Name}_{node_id}.frag`
+
+- **Blueprint_Name**: The JSON filename with spaces/special chars replaced by underscores
+- **node_id**: The GLSLShader node ID within the subgraph
+
+## Usage
+
+```bash
+# Extract shaders from blueprint JSONs to this folder
+python update_blueprints.py extract
+
+# Patch edited shaders back into blueprint JSONs
+python update_blueprints.py patch
+```
+
+## Workflow
+
+1. Run `extract` to pull current shaders from JSONs
+2. Edit `.frag` files
+3. Run `patch` to update the blueprint JSONs
+4. Test
+5. Commit both `.frag` files and updated JSONs

blueprints/.glsl/Sharpen_23.frag

@@ -0,0 +1,28 @@
+#version 300 es
+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;
+    
+    // Sample center and neighbors
+    vec4 center = texture(u_image0, v_texCoord);
+    vec4 top    = texture(u_image0, v_texCoord + vec2( 0.0, -texel.y));
+    vec4 bottom = texture(u_image0, v_texCoord + vec2( 0.0,  texel.y));
+    vec4 left   = texture(u_image0, v_texCoord + vec2(-texel.x,  0.0));
+    vec4 right  = texture(u_image0, v_texCoord + vec2( texel.x,  0.0));
+    
+    // Edge enhancement (Laplacian)
+    vec4 edges = center * 4.0 - top - bottom - left - right;
+    
+    // Add edges back scaled by strength
+    vec4 sharpened = center + edges * u_float0;
+    
+    fragColor0 = vec4(clamp(sharpened.rgb, 0.0, 1.0), center.a);
+}

blueprints/.glsl/Unsharp_Mask_26.frag

@@ -0,0 +1,61 @@
+#version 300 es
+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
+
+in vec2 v_texCoord;
+layout(location = 0) out vec4 fragColor0;
+
+float gaussian(float x, float sigma) {
+    return exp(-(x * x) / (2.0 * sigma * sigma));
+}
+
+float getLuminance(vec3 color) {
+    return dot(color, vec3(0.2126, 0.7152, 0.0722));
+}
+
+void main() {
+    vec2 texel = 1.0 / u_resolution;
+    float radius = max(u_float1, 0.5);
+    float amount = u_float0;
+    float threshold = u_float2;
+
+    vec4 original = texture(u_image0, v_texCoord);
+
+    // Gaussian blur for the "unsharp" mask
+    int samples = int(ceil(radius));
+    float sigma = radius / 2.0;
+
+    vec4 blurred = vec4(0.0);
+    float totalWeight = 0.0;
+
+    for (int x = -samples; x <= samples; x++) {
+        for (int y = -samples; y <= samples; y++) {
+            vec2 offset = vec2(float(x), float(y)) * texel;
+            vec4 sample_color = texture(u_image0, v_texCoord + offset);
+
+            float dist = length(vec2(float(x), float(y)));
+            float weight = gaussian(dist, sigma);
+            blurred += sample_color * weight;
+            totalWeight += weight;
+        }
+    }
+    blurred /= totalWeight;
+
+    // Unsharp mask = original - blurred
+    vec3 mask = original.rgb - blurred.rgb;
+
+    // Luminance-based threshold with smooth falloff
+    float lumaDelta = abs(getLuminance(original.rgb) - getLuminance(blurred.rgb));
+    float thresholdScale = smoothstep(0.0, threshold, lumaDelta);
+    mask *= thresholdScale;
+
+    // Sharpen: original + mask * amount
+    vec3 sharpened = original.rgb + mask * amount;
+
+    fragColor0 = vec4(clamp(sharpened, 0.0, 1.0), original.a);
+}

blueprints/.glsl/update_blueprints.py

@@ -0,0 +1,159 @@
+#!/usr/bin/env python3
+"""
+Shader Blueprint Updater
+
+Syncs GLSL shader files between this folder and blueprint JSON files.
+
+File naming convention:
+    {Blueprint Name}_{node_id}.frag
+
+Usage:
+    python update_blueprints.py extract   # Extract shaders from JSONs to here
+    python update_blueprints.py patch     # Patch shaders back into JSONs
+    python update_blueprints.py           # Same as patch (default)
+"""
+
+import json
+import logging
+import sys
+import re
+from pathlib import Path
+
+logging.basicConfig(level=logging.INFO, format='%(message)s')
+logger = logging.getLogger(__name__)
+
+GLSL_DIR = Path(__file__).parent
+BLUEPRINTS_DIR = GLSL_DIR.parent
+
+
+def get_blueprint_files():
+    """Get all blueprint JSON files."""
+    return sorted(BLUEPRINTS_DIR.glob("*.json"))
+
+
+def sanitize_filename(name):
+    """Convert blueprint name to safe filename."""
+    return re.sub(r'[^\w\-]', '_', name)
+
+
+def extract_shaders():
+    """Extract all shaders from blueprint JSONs to this folder."""
+    extracted = 0
+    for json_path in get_blueprint_files():
+        blueprint_name = json_path.stem
+
+        try:
+            with open(json_path, 'r') as f:
+                data = json.load(f)
+        except (json.JSONDecodeError, IOError) as e:
+            logger.warning("Skipping %s: %s", json_path.name, e)
+            continue
+
+        # Find GLSLShader nodes in subgraphs
+        for subgraph in data.get('definitions', {}).get('subgraphs', []):
+            for node in subgraph.get('nodes', []):
+                if node.get('type') == 'GLSLShader':
+                    node_id = node.get('id')
+                    widgets = node.get('widgets_values', [])
+
+                    # Find shader code (first string that looks like GLSL)
+                    for widget in widgets:
+                        if isinstance(widget, str) and widget.startswith('#version'):
+                            safe_name = sanitize_filename(blueprint_name)
+                            frag_name = f"{safe_name}_{node_id}.frag"
+                            frag_path = GLSL_DIR / frag_name
+
+                            with open(frag_path, 'w') as f:
+                                f.write(widget)
+
+                            logger.info("  Extracted: %s", frag_name)
+                            extracted += 1
+                            break
+
+    logger.info("\nExtracted %d shader(s)", extracted)
+
+
+def patch_shaders():
+    """Patch shaders from this folder back into blueprint JSONs."""
+    # Build lookup: blueprint_name -> [(node_id, shader_code), ...]
+    shader_updates = {}
+
+    for frag_path in sorted(GLSL_DIR.glob("*.frag")):
+        # Parse filename: {blueprint_name}_{node_id}.frag
+        parts = frag_path.stem.rsplit('_', 1)
+        if len(parts) != 2:
+            logger.warning("Skipping %s: invalid filename format", frag_path.name)
+            continue
+
+        blueprint_name, node_id_str = parts
+
+        try:
+            node_id = int(node_id_str)
+        except ValueError:
+            logger.warning("Skipping %s: invalid node_id", frag_path.name)
+            continue
+
+        with open(frag_path, 'r') as f:
+            shader_code = f.read()
+
+        if blueprint_name not in shader_updates:
+            shader_updates[blueprint_name] = []
+        shader_updates[blueprint_name].append((node_id, shader_code))
+
+    # Apply updates to JSON files
+    patched = 0
+    for json_path in get_blueprint_files():
+        blueprint_name = sanitize_filename(json_path.stem)
+
+        if blueprint_name not in shader_updates:
+            continue
+
+        try:
+            with open(json_path, 'r') as f:
+                data = json.load(f)
+        except (json.JSONDecodeError, IOError) as e:
+            logger.error("Error reading %s: %s", json_path.name, e)
+            continue
+
+        modified = False
+        for node_id, shader_code in shader_updates[blueprint_name]:
+            # Find the node and update
+            for subgraph in data.get('definitions', {}).get('subgraphs', []):
+                for node in subgraph.get('nodes', []):
+                    if node.get('id') == node_id and node.get('type') == 'GLSLShader':
+                        widgets = node.get('widgets_values', [])
+                        if len(widgets) > 0 and widgets[0] != shader_code:
+                            widgets[0] = shader_code
+                            modified = True
+                            logger.info("  Patched: %s (node %d)", json_path.name, node_id)
+                            patched += 1
+
+        if modified:
+            with open(json_path, 'w') as f:
+                json.dump(data, f)
+
+    if patched == 0:
+        logger.info("No changes to apply.")
+    else:
+        logger.info("\nPatched %d shader(s)", patched)
+
+
+def main():
+    if len(sys.argv) < 2:
+        command = "patch"
+    else:
+        command = sys.argv[1].lower()
+
+    if command == "extract":
+        logger.info("Extracting shaders from blueprints...")
+        extract_shaders()
+    elif command in ("patch", "update", "apply"):
+        logger.info("Patching shaders into blueprints...")
+        patch_shaders()
+    else:
+        logger.info(__doc__)
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

blueprints/Image Channels.json

@@ -0,0 +1 @@
+{"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

@@ -0,0 +1 @@
+{"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

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

blueprints/Video Upscale(GAN x4).json

@@ -0,0 +1 @@
+{"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/checkpoint_pickle.py

@@ -1,13 +0,0 @@
-import pickle
-
-load = pickle.load
-
-class Empty:
-    pass
-
-class Unpickler(pickle.Unpickler):
-    def find_class(self, module, name):
-        #TODO: safe unpickle
-        if module.startswith("pytorch_lightning"):
-            return Empty
-        return super().find_class(module, name)

comfy/comfy_types/node_typing.py

@@ -176,6 +176,8 @@ class InputTypeOptions(TypedDict):
     """COMBO type only. Specifies the configuration for a multi-select widget.
     Available after ComfyUI frontend v1.13.4
     https://github.com/Comfy-Org/ComfyUI_frontend/pull/2987"""
+    gradient_stops: NotRequired[list[list[float]]]
+    """Gradient color stops for gradientslider display mode. Each stop is [offset, r, g, b] (``FLOAT``)."""
 
 
 class HiddenInputTypeDict(TypedDict):

comfy/conds.py

@@ -4,6 +4,25 @@ import comfy.utils
 import logging
 
 
+def is_equal(x, y):
+    if torch.is_tensor(x) and torch.is_tensor(y):
+        return torch.equal(x, y)
+    elif isinstance(x, dict) and isinstance(y, dict):
+        if x.keys() != y.keys():
+            return False
+        return all(is_equal(x[k], y[k]) for k in x)
+    elif isinstance(x, (list, tuple)) and isinstance(y, (list, tuple)):
+        if type(x) is not type(y) or len(x) != len(y):
+            return False
+        return all(is_equal(a, b) for a, b in zip(x, y))
+    else:
+        try:
+            return x == y
+        except Exception:
+            logging.warning("comparison issue with COND")
+            return False
+
+
 class CONDRegular:
     def __init__(self, cond):
         self.cond = cond
@@ -84,7 +103,7 @@ class CONDConstant(CONDRegular):
         return self._copy_with(self.cond)
 
     def can_concat(self, other):
-        if self.cond != other.cond:
+        if not is_equal(self.cond, other.cond):
             return False
         return True
 

comfy/controlnet.py

@@ -297,6 +297,30 @@ class ControlNet(ControlBase):
         self.model_sampling_current = None
         super().cleanup()
 
+
+class QwenFunControlNet(ControlNet):
+    def get_control(self, x_noisy, t, cond, batched_number, transformer_options):
+        # Fun checkpoints are more sensitive to high strengths in the generic
+        # ControlNet merge path. Use a soft response curve so strength=1.0 stays
+        # unchanged while >1 grows more gently.
+        original_strength = self.strength
+        self.strength = math.sqrt(max(self.strength, 0.0))
+        try:
+            return super().get_control(x_noisy, t, cond, batched_number, transformer_options)
+        finally:
+            self.strength = original_strength
+
+    def pre_run(self, model, percent_to_timestep_function):
+        super().pre_run(model, percent_to_timestep_function)
+        self.set_extra_arg("base_model", model.diffusion_model)
+
+    def copy(self):
+        c = QwenFunControlNet(None, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype)
+        c.control_model = self.control_model
+        c.control_model_wrapped = self.control_model_wrapped
+        self.copy_to(c)
+        return c
+
 class ControlLoraOps:
     class Linear(torch.nn.Module, comfy.ops.CastWeightBiasOp):
         def __init__(self, in_features: int, out_features: int, bias: bool = True,
@@ -560,6 +584,7 @@ def load_controlnet_hunyuandit(controlnet_data, model_options={}):
 def load_controlnet_flux_xlabs_mistoline(sd, mistoline=False, model_options={}):
     model_config, operations, load_device, unet_dtype, manual_cast_dtype, offload_device = controlnet_config(sd, model_options=model_options)
     control_model = comfy.ldm.flux.controlnet.ControlNetFlux(mistoline=mistoline, operations=operations, device=offload_device, dtype=unet_dtype, **model_config.unet_config)
+    sd = model_config.process_unet_state_dict(sd)
     control_model = controlnet_load_state_dict(control_model, sd)
     extra_conds = ['y', 'guidance']
     control = ControlNet(control_model, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
@@ -605,6 +630,53 @@ def load_controlnet_qwen_instantx(sd, model_options={}):
     control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, concat_mask=concat_mask, load_device=load_device, manual_cast_dtype=manual_cast_dtype, extra_conds=extra_conds)
     return control
 
+
+def load_controlnet_qwen_fun(sd, model_options={}):
+    load_device = comfy.model_management.get_torch_device()
+    weight_dtype = comfy.utils.weight_dtype(sd)
+    unet_dtype = model_options.get("dtype", weight_dtype)
+    manual_cast_dtype = comfy.model_management.unet_manual_cast(unet_dtype, load_device)
+
+    operations = model_options.get("custom_operations", None)
+    if operations is None:
+        operations = comfy.ops.pick_operations(unet_dtype, manual_cast_dtype, disable_fast_fp8=True)
+
+    in_features = sd["control_img_in.weight"].shape[1]
+    inner_dim = sd["control_img_in.weight"].shape[0]
+
+    block_weight = sd["control_blocks.0.attn.to_q.weight"]
+    attention_head_dim = sd["control_blocks.0.attn.norm_q.weight"].shape[0]
+    num_attention_heads = max(1, block_weight.shape[0] // max(1, attention_head_dim))
+
+    model = comfy.ldm.qwen_image.controlnet.QwenImageFunControlNetModel(
+        control_in_features=in_features,
+        inner_dim=inner_dim,
+        num_attention_heads=num_attention_heads,
+        attention_head_dim=attention_head_dim,
+        num_control_blocks=5,
+        main_model_double=60,
+        injection_layers=(0, 12, 24, 36, 48),
+        operations=operations,
+        device=comfy.model_management.unet_offload_device(),
+        dtype=unet_dtype,
+    )
+    model = controlnet_load_state_dict(model, sd)
+
+    latent_format = comfy.latent_formats.Wan21()
+    control = QwenFunControlNet(
+        model,
+        compression_ratio=1,
+        latent_format=latent_format,
+        # Fun checkpoints already expect their own 33-channel context handling.
+        # Enabling generic concat_mask injects an extra mask channel at apply-time
+        # and breaks the intended fallback packing path.
+        concat_mask=False,
+        load_device=load_device,
+        manual_cast_dtype=manual_cast_dtype,
+        extra_conds=[],
+    )
+    return control
+
 def convert_mistoline(sd):
     return comfy.utils.state_dict_prefix_replace(sd, {"single_controlnet_blocks.": "controlnet_single_blocks."})
 
@@ -682,6 +754,8 @@ def load_controlnet_state_dict(state_dict, model=None, model_options={}):
             return load_controlnet_qwen_instantx(controlnet_data, model_options=model_options)
         elif "controlnet_x_embedder.weight" in controlnet_data:
             return load_controlnet_flux_instantx(controlnet_data, model_options=model_options)
+    elif "control_blocks.0.after_proj.weight" in controlnet_data and "control_img_in.weight" in controlnet_data:
+        return load_controlnet_qwen_fun(controlnet_data, model_options=model_options)
 
     elif "controlnet_blocks.0.linear.weight" in controlnet_data: #mistoline flux
         return load_controlnet_flux_xlabs_mistoline(convert_mistoline(controlnet_data), mistoline=True, model_options=model_options)

comfy/k_diffusion/sampling.py

@@ -1,12 +1,11 @@
 import math
-import time
 from functools import partial
 
 from scipy import integrate
 import torch
 from torch import nn
 import torchsde
-from tqdm.auto import trange as trange_, tqdm
+from tqdm.auto import tqdm
 
 from . import utils
 from . import deis
@@ -15,34 +14,7 @@ import comfy.model_patcher
 import comfy.model_sampling
 
 import comfy.memory_management
-
-
-def trange(*args, **kwargs):
-    if comfy.memory_management.aimdo_allocator is None:
-        return trange_(*args, **kwargs)
-
-    pbar = trange_(*args, **kwargs, smoothing=1.0)
-    pbar._i = 0
-    pbar.set_postfix_str("  Model Initializing ...  ")
-
-    _update = pbar.update
-
-    def warmup_update(n=1):
-        pbar._i += 1
-        if pbar._i == 1:
-            pbar.i1_time = time.time()
-            pbar.set_postfix_str(" Model Initialization complete!  ")
-        elif pbar._i == 2:
-            #bring forward the effective start time based the the diff between first and second iteration
-            #to attempt to remove load overhead from the final step rate estimate.
-            pbar.start_t = pbar.i1_time - (time.time() - pbar.i1_time)
-            pbar.set_postfix_str("")
-
-        _update(n)
-
-    pbar.update = warmup_update
-    return pbar
-
+from comfy.utils import model_trange as trange
 
 def append_zero(x):
     return torch.cat([x, x.new_zeros([1])])

comfy/latent_formats.py

@@ -755,6 +755,10 @@ class ACEAudio(LatentFormat):
     latent_channels = 8
     latent_dimensions = 2
 
+class ACEAudio15(LatentFormat):
+    latent_channels = 64
+    latent_dimensions = 1
+
 class ChromaRadiance(LatentFormat):
     latent_channels = 3
     spacial_downscale_ratio = 1

comfy/ldm/anima/model.py

@@ -179,8 +179,8 @@ class LLMAdapter(nn.Module):
             if source_attention_mask.ndim == 2:
                 source_attention_mask = source_attention_mask.unsqueeze(1).unsqueeze(1)
 
-        x = self.in_proj(self.embed(target_input_ids))
         context = source_hidden_states
+        x = self.in_proj(self.embed(target_input_ids, out_dtype=context.dtype))
         position_ids = torch.arange(x.shape[1], device=x.device).unsqueeze(0)
         position_ids_context = torch.arange(context.shape[1], device=x.device).unsqueeze(0)
         position_embeddings = self.rotary_emb(x, position_ids)
@@ -195,8 +195,20 @@ class Anima(MiniTrainDIT):
         super().__init__(*args, **kwargs)
         self.llm_adapter = LLMAdapter(device=kwargs.get("device"), dtype=kwargs.get("dtype"), operations=kwargs.get("operations"))
 
-    def preprocess_text_embeds(self, text_embeds, text_ids):
+    def preprocess_text_embeds(self, text_embeds, text_ids, t5xxl_weights=None):
         if text_ids is not None:
-            return self.llm_adapter(text_embeds, text_ids)
+            out = self.llm_adapter(text_embeds, text_ids)
+            if t5xxl_weights is not None:
+                out = out * t5xxl_weights
+
+            if out.shape[1] < 512:
+                out = torch.nn.functional.pad(out, (0, 0, 0, 512 - out.shape[1]))
+            return out
         else:
             return text_embeds
+
+    def forward(self, x, timesteps, context, **kwargs):
+        t5xxl_ids = kwargs.pop("t5xxl_ids", None)
+        if t5xxl_ids is not None:
+            context = self.preprocess_text_embeds(context, t5xxl_ids, t5xxl_weights=kwargs.pop("t5xxl_weights", None))
+        return super().forward(x, timesteps, context, **kwargs)

comfy/ldm/chroma/layers.py

@@ -3,7 +3,6 @@ from torch import Tensor, nn
 
 from comfy.ldm.flux.layers import (
     MLPEmbedder,
-    RMSNorm,
     ModulationOut,
 )
 
@@ -29,7 +28,7 @@ class Approximator(nn.Module):
         super().__init__()
         self.in_proj = operations.Linear(in_dim, hidden_dim, bias=True, dtype=dtype, device=device)
         self.layers = nn.ModuleList([MLPEmbedder(hidden_dim, hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
-        self.norms = nn.ModuleList([RMSNorm(hidden_dim, dtype=dtype, device=device, operations=operations) for x in range( n_layers)])
+        self.norms = nn.ModuleList([operations.RMSNorm(hidden_dim, dtype=dtype, device=device) for x in range( n_layers)])
         self.out_proj = operations.Linear(hidden_dim, out_dim, dtype=dtype, device=device)
 
     @property

comfy/ldm/chroma/model.py

@@ -152,6 +152,7 @@ class Chroma(nn.Module):
         transformer_options={},
         attn_mask: Tensor = None,
     ) -> Tensor:
+        transformer_options = transformer_options.copy()
         patches_replace = transformer_options.get("patches_replace", {})
 
         # running on sequences img
@@ -228,6 +229,7 @@ class Chroma(nn.Module):
 
         transformer_options["total_blocks"] = len(self.single_blocks)
         transformer_options["block_type"] = "single"
+        transformer_options["img_slice"] = [txt.shape[1], img.shape[1]]
         for i, block in enumerate(self.single_blocks):
             transformer_options["block_index"] = i
             if i not in self.skip_dit:

comfy/ldm/chroma_radiance/layers.py

@@ -4,8 +4,6 @@ from functools import lru_cache
 import torch
 from torch import nn
 
-from comfy.ldm.flux.layers import RMSNorm
-
 
 class NerfEmbedder(nn.Module):
     """
@@ -145,7 +143,7 @@ class NerfGLUBlock(nn.Module):
         # We now need to generate parameters for 3 matrices.
         total_params = 3 * hidden_size_x**2 * mlp_ratio
         self.param_generator = operations.Linear(hidden_size_s, total_params, dtype=dtype, device=device)
-        self.norm = RMSNorm(hidden_size_x, dtype=dtype, device=device, operations=operations)
+        self.norm = operations.RMSNorm(hidden_size_x, dtype=dtype, device=device)
         self.mlp_ratio = mlp_ratio
 
 
@@ -178,7 +176,7 @@ class NerfGLUBlock(nn.Module):
 class NerfFinalLayer(nn.Module):
     def __init__(self, hidden_size, out_channels, dtype=None, device=None, operations=None):
         super().__init__()
-        self.norm = RMSNorm(hidden_size, dtype=dtype, device=device, operations=operations)
+        self.norm = operations.RMSNorm(hidden_size, dtype=dtype, device=device)
         self.linear = operations.Linear(hidden_size, out_channels, dtype=dtype, device=device)
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
@@ -190,7 +188,7 @@ class NerfFinalLayer(nn.Module):
 class NerfFinalLayerConv(nn.Module):
     def __init__(self, hidden_size: int, out_channels: int, dtype=None, device=None, operations=None):
         super().__init__()
-        self.norm = RMSNorm(hidden_size, dtype=dtype, device=device, operations=operations)
+        self.norm = operations.RMSNorm(hidden_size, dtype=dtype, device=device)
         self.conv = operations.Conv2d(
             in_channels=hidden_size,
             out_channels=out_channels,

comfy/ldm/cosmos/predict2.py

@@ -335,7 +335,7 @@ class FinalLayer(nn.Module):
         device=None, dtype=None, operations=None
     ):
         super().__init__()
-        self.layer_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.layer_norm = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
         self.linear = operations.Linear(
             hidden_size, spatial_patch_size * spatial_patch_size * temporal_patch_size * out_channels, bias=False, device=device, dtype=dtype
         )
@@ -463,6 +463,8 @@ class Block(nn.Module):
         extra_per_block_pos_emb: Optional[torch.Tensor] = None,
         transformer_options: Optional[dict] = {},
     ) -> torch.Tensor:
+        residual_dtype = x_B_T_H_W_D.dtype
+        compute_dtype = emb_B_T_D.dtype
         if extra_per_block_pos_emb is not None:
             x_B_T_H_W_D = x_B_T_H_W_D + extra_per_block_pos_emb
 
@@ -512,7 +514,7 @@ class Block(nn.Module):
         result_B_T_H_W_D = rearrange(
             self.self_attn(
                 # normalized_x_B_T_HW_D,
-                rearrange(normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                rearrange(normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"),
                 None,
                 rope_emb=rope_emb_L_1_1_D,
                 transformer_options=transformer_options,
@@ -522,7 +524,7 @@ class Block(nn.Module):
             h=H,
             w=W,
         )
-        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D * result_B_T_H_W_D
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_self_attn_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype)
 
         def _x_fn(
             _x_B_T_H_W_D: torch.Tensor,
@@ -536,7 +538,7 @@ class Block(nn.Module):
             )
             _result_B_T_H_W_D = rearrange(
                 self.cross_attn(
-                    rearrange(_normalized_x_B_T_H_W_D, "b t h w d -> b (t h w) d"),
+                    rearrange(_normalized_x_B_T_H_W_D.to(compute_dtype), "b t h w d -> b (t h w) d"),
                     crossattn_emb,
                     rope_emb=rope_emb_L_1_1_D,
                     transformer_options=transformer_options,
@@ -555,7 +557,7 @@ class Block(nn.Module):
             shift_cross_attn_B_T_1_1_D,
             transformer_options=transformer_options,
         )
-        x_B_T_H_W_D = result_B_T_H_W_D * gate_cross_attn_B_T_1_1_D + x_B_T_H_W_D
+        x_B_T_H_W_D = result_B_T_H_W_D.to(residual_dtype) * gate_cross_attn_B_T_1_1_D.to(residual_dtype) + x_B_T_H_W_D
 
         normalized_x_B_T_H_W_D = _fn(
             x_B_T_H_W_D,
@@ -563,8 +565,8 @@ class Block(nn.Module):
             scale_mlp_B_T_1_1_D,
             shift_mlp_B_T_1_1_D,
         )
-        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D)
-        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D * result_B_T_H_W_D
+        result_B_T_H_W_D = self.mlp(normalized_x_B_T_H_W_D.to(compute_dtype))
+        x_B_T_H_W_D = x_B_T_H_W_D + gate_mlp_B_T_1_1_D.to(residual_dtype) * result_B_T_H_W_D.to(residual_dtype)
         return x_B_T_H_W_D
 
 
@@ -876,6 +878,14 @@ class MiniTrainDIT(nn.Module):
             "extra_per_block_pos_emb": extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
             "transformer_options": kwargs.get("transformer_options", {}),
         }
+
+        # The residual stream for this model has large values. To make fp16 compute_dtype work, we keep the residual stream
+        # in fp32, but run attention and MLP modules in fp16.
+        # An alternate method that clamps fp16 values "works" in the sense that it makes coherent images, but there is noticeable
+        # quality degradation and visual artifacts.
+        if x_B_T_H_W_D.dtype == torch.float16:
+            x_B_T_H_W_D = x_B_T_H_W_D.float()
+
         for block in self.blocks:
             x_B_T_H_W_D = block(
                 x_B_T_H_W_D,
@@ -884,6 +894,6 @@ class MiniTrainDIT(nn.Module):
                 **block_kwargs,
             )
 
-        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D, t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
+        x_B_T_H_W_O = self.final_layer(x_B_T_H_W_D.to(crossattn_emb.dtype), t_embedding_B_T_D, adaln_lora_B_T_3D=adaln_lora_B_T_3D)
         x_B_C_Tt_Hp_Wp = self.unpatchify(x_B_T_H_W_O)[:, :, :orig_shape[-3], :orig_shape[-2], :orig_shape[-1]]
         return x_B_C_Tt_Hp_Wp

comfy/ldm/flux/layers.py

@@ -5,9 +5,9 @@ import torch
 from torch import Tensor, nn
 
 from .math import attention, rope
-import comfy.ops
-import comfy.ldm.common_dit
 
+# Fix import for some custom nodes, TODO: delete eventually.
+RMSNorm = None
 
 class EmbedND(nn.Module):
     def __init__(self, dim: int, theta: int, axes_dim: list):
@@ -87,20 +87,12 @@ def build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=False, yak_mlp=False, dt
             operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
         )
 
-class RMSNorm(torch.nn.Module):
-    def __init__(self, dim: int, dtype=None, device=None, operations=None):
-        super().__init__()
-        self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
-
-    def forward(self, x: Tensor):
-        return comfy.ldm.common_dit.rms_norm(x, self.scale, 1e-6)
-
 
 class QKNorm(torch.nn.Module):
     def __init__(self, dim: int, dtype=None, device=None, operations=None):
         super().__init__()
-        self.query_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
-        self.key_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
+        self.query_norm = operations.RMSNorm(dim, dtype=dtype, device=device)
+        self.key_norm = operations.RMSNorm(dim, dtype=dtype, device=device)
 
     def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple:
         q = self.query_norm(q)
@@ -169,7 +161,7 @@ class SiLUActivation(nn.Module):
 
 
 class DoubleStreamBlock(nn.Module):
-    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None):
         super().__init__()
 
         mlp_hidden_dim = int(hidden_size * mlp_ratio)
@@ -197,8 +189,6 @@ class DoubleStreamBlock(nn.Module):
 
         self.txt_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
 
-        self.flipped_img_txt = flipped_img_txt
-
     def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}):
         if self.modulation:
             img_mod1, img_mod2 = self.img_mod(vec)
@@ -206,6 +196,9 @@ class DoubleStreamBlock(nn.Module):
         else:
             (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
 
+        transformer_patches = transformer_options.get("patches", {})
+        extra_options = transformer_options.copy()
+
         # prepare image for attention
         img_modulated = self.img_norm1(img)
         img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img)
@@ -224,32 +217,23 @@ class DoubleStreamBlock(nn.Module):
         del txt_qkv
         txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
 
-        if self.flipped_img_txt:
-            q = torch.cat((img_q, txt_q), dim=2)
-            del img_q, txt_q
-            k = torch.cat((img_k, txt_k), dim=2)
-            del img_k, txt_k
-            v = torch.cat((img_v, txt_v), dim=2)
-            del img_v, txt_v
-            # run actual attention
-            attn = attention(q, k, v,
-                             pe=pe, mask=attn_mask, transformer_options=transformer_options)
-            del q, k, v
+        q = torch.cat((txt_q, img_q), dim=2)
+        del txt_q, img_q
+        k = torch.cat((txt_k, img_k), dim=2)
+        del txt_k, img_k
+        v = torch.cat((txt_v, img_v), dim=2)
+        del txt_v, img_v
+        # run actual attention
+        attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
+        del q, k, v
 
-            img_attn, txt_attn = attn[:, : img.shape[1]], attn[:, img.shape[1]:]
-        else:
-            q = torch.cat((txt_q, img_q), dim=2)
-            del txt_q, img_q
-            k = torch.cat((txt_k, img_k), dim=2)
-            del txt_k, img_k
-            v = torch.cat((txt_v, img_v), dim=2)
-            del txt_v, img_v
-            # run actual attention
-            attn = attention(q, k, v,
-                             pe=pe, mask=attn_mask, transformer_options=transformer_options)
-            del q, k, v
+        if "attn1_output_patch" in transformer_patches:
+            extra_options["img_slice"] = [txt.shape[1], attn.shape[1]]
+            patch = transformer_patches["attn1_output_patch"]
+            for p in patch:
+                attn = p(attn, extra_options)
 
-            txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
+        txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
         # calculate the img bloks
         img += apply_mod(self.img_attn.proj(img_attn), img_mod1.gate, None, modulation_dims_img)
@@ -328,6 +312,9 @@ class SingleStreamBlock(nn.Module):
         else:
             mod = vec
 
+        transformer_patches = transformer_options.get("patches", {})
+        extra_options = transformer_options.copy()
+
         qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim_first], dim=-1)
 
         q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
@@ -337,6 +324,12 @@ class SingleStreamBlock(nn.Module):
         # compute attention
         attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
         del q, k, v
+
+        if "attn1_output_patch" in transformer_patches:
+            patch = transformer_patches["attn1_output_patch"]
+            for p in patch:
+                attn = p(attn, extra_options)
+
         # compute activation in mlp stream, cat again and run second linear layer
         if self.yak_mlp:
             mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2]

comfy/ldm/flux/math.py

@@ -29,19 +29,34 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     return out.to(dtype=torch.float32, device=pos.device)
 
 
+def _apply_rope1(x: Tensor, freqs_cis: Tensor):
+    x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
+
+    x_out = freqs_cis[..., 0] * x_[..., 0]
+    x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
+
+    return x_out.reshape(*x.shape).type_as(x)
+
+
+def _apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
+    return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+
+
 try:
     import comfy.quant_ops
-    apply_rope = comfy.quant_ops.ck.apply_rope
-    apply_rope1 = comfy.quant_ops.ck.apply_rope1
+    q_apply_rope = comfy.quant_ops.ck.apply_rope
+    q_apply_rope1 = comfy.quant_ops.ck.apply_rope1
+    def apply_rope(xq, xk, freqs_cis):
+        if comfy.model_management.in_training:
+            return _apply_rope(xq, xk, freqs_cis)
+        else:
+            return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+    def apply_rope1(x, freqs_cis):
+        if comfy.model_management.in_training:
+            return _apply_rope1(x, freqs_cis)
+        else:
+            return q_apply_rope1(x, freqs_cis)
 except:
     logging.warning("No comfy kitchen, using old apply_rope functions.")
-    def apply_rope1(x: Tensor, freqs_cis: Tensor):
-        x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
-
-        x_out = freqs_cis[..., 0] * x_[..., 0]
-        x_out.addcmul_(freqs_cis[..., 1], x_[..., 1])
-
-        return x_out.reshape(*x.shape).type_as(x)
-
-    def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
-        return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)
+    apply_rope = _apply_rope
+    apply_rope1 = _apply_rope1

comfy/ldm/flux/model.py

@@ -16,7 +16,6 @@ from .layers import (
     SingleStreamBlock,
     timestep_embedding,
     Modulation,
-    RMSNorm
 )
 
 @dataclass
@@ -81,7 +80,7 @@ class Flux(nn.Module):
         self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device)
 
         if params.txt_norm:
-            self.txt_norm = RMSNorm(params.context_in_dim, dtype=dtype, device=device, operations=operations)
+            self.txt_norm = operations.RMSNorm(params.context_in_dim, dtype=dtype, device=device)
         else:
             self.txt_norm = None
 
@@ -143,6 +142,7 @@ class Flux(nn.Module):
         attn_mask: Tensor = None,
     ) -> Tensor:
 
+        transformer_options = transformer_options.copy()
         patches = transformer_options.get("patches", {})
         patches_replace = transformer_options.get("patches_replace", {})
         if img.ndim != 3 or txt.ndim != 3:
@@ -232,6 +232,7 @@ class Flux(nn.Module):
 
         transformer_options["total_blocks"] = len(self.single_blocks)
         transformer_options["block_type"] = "single"
+        transformer_options["img_slice"] = [txt.shape[1], img.shape[1]]
         for i, block in enumerate(self.single_blocks):
             transformer_options["block_index"] = i
             if ("single_block", i) in blocks_replace:

comfy/ldm/hunyuan_video/model.py

@@ -241,7 +241,6 @@ class HunyuanVideo(nn.Module):
                     self.num_heads,
                     mlp_ratio=params.mlp_ratio,
                     qkv_bias=params.qkv_bias,
-                    flipped_img_txt=True,
                     dtype=dtype, device=device, operations=operations
                 )
                 for _ in range(params.depth)
@@ -305,6 +304,7 @@ class HunyuanVideo(nn.Module):
         control=None,
         transformer_options={},
     ) -> Tensor:
+        transformer_options = transformer_options.copy()
         patches_replace = transformer_options.get("patches_replace", {})
 
         initial_shape = list(img.shape)
@@ -378,14 +378,14 @@ class HunyuanVideo(nn.Module):
             extra_txt_ids = torch.zeros((txt_ids.shape[0], txt_vision_states.shape[1], txt_ids.shape[-1]), device=txt_ids.device, dtype=txt_ids.dtype)
             txt_ids = torch.cat((txt_ids, extra_txt_ids), dim=1)
 
-        ids = torch.cat((img_ids, txt_ids), dim=1)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
         pe = self.pe_embedder(ids)
 
         img_len = img.shape[1]
         if txt_mask is not None:
             attn_mask_len = img_len + txt.shape[1]
             attn_mask = torch.zeros((1, 1, attn_mask_len), dtype=img.dtype, device=img.device)
-            attn_mask[:, 0, img_len:] = txt_mask
+            attn_mask[:, 0, :txt.shape[1]] = txt_mask
         else:
             attn_mask = None
 
@@ -413,10 +413,11 @@ class HunyuanVideo(nn.Module):
                     if add is not None:
                         img += add
 
-        img = torch.cat((img, txt), 1)
+        img = torch.cat((txt, img), 1)
 
         transformer_options["total_blocks"] = len(self.single_blocks)
         transformer_options["block_type"] = "single"
+        transformer_options["img_slice"] = [txt.shape[1], img.shape[1]]
         for i, block in enumerate(self.single_blocks):
             transformer_options["block_index"] = i
             if ("single_block", i) in blocks_replace:
@@ -435,9 +436,9 @@ class HunyuanVideo(nn.Module):
                 if i < len(control_o):
                     add = control_o[i]
                     if add is not None:
-                        img[:, : img_len] += add
+                        img[:, txt.shape[1]: img_len + txt.shape[1]] += add
 
-        img = img[:, : img_len]
+        img = img[:, txt.shape[1]: img_len + txt.shape[1]]
         if ref_latent is not None:
             img = img[:, ref_latent.shape[1]:]
 

comfy/ldm/lightricks/av_model.py

@@ -9,6 +9,7 @@ from comfy.ldm.lightricks.model import (
     LTXVModel,
 )
 from comfy.ldm.lightricks.symmetric_patchifier import AudioPatchifier
+from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
 import comfy.ldm.common_dit
 
 class CompressedTimestep:
@@ -217,7 +218,7 @@ class BasicAVTransformerBlock(nn.Module):
     def forward(
         self, x: Tuple[torch.Tensor, torch.Tensor], v_context=None, a_context=None, attention_mask=None, v_timestep=None, a_timestep=None,
         v_pe=None, a_pe=None, v_cross_pe=None, a_cross_pe=None, v_cross_scale_shift_timestep=None, a_cross_scale_shift_timestep=None,
-        v_cross_gate_timestep=None, a_cross_gate_timestep=None, transformer_options=None,
+        v_cross_gate_timestep=None, a_cross_gate_timestep=None, transformer_options=None, self_attention_mask=None,
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         run_vx = transformer_options.get("run_vx", True)
         run_ax = transformer_options.get("run_ax", True)
@@ -233,7 +234,7 @@ class BasicAVTransformerBlock(nn.Module):
             vshift_msa, vscale_msa = (self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(0, 2)))
             norm_vx = comfy.ldm.common_dit.rms_norm(vx) * (1 + vscale_msa) + vshift_msa
             del vshift_msa, vscale_msa
-            attn1_out = self.attn1(norm_vx, pe=v_pe, transformer_options=transformer_options)
+            attn1_out = self.attn1(norm_vx, pe=v_pe, mask=self_attention_mask, transformer_options=transformer_options)
             del norm_vx
             # video cross-attention
             vgate_msa = self.get_ada_values(self.scale_shift_table, vx.shape[0], v_timestep, slice(2, 3))[0]
@@ -450,6 +451,29 @@ class LTXAVModel(LTXVModel):
             operations=self.operations,
         )
 
+        self.audio_embeddings_connector = Embeddings1DConnector(
+            split_rope=True,
+            double_precision_rope=True,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+
+        self.video_embeddings_connector = Embeddings1DConnector(
+            split_rope=True,
+            double_precision_rope=True,
+            dtype=dtype,
+            device=device,
+            operations=self.operations,
+        )
+
+    def preprocess_text_embeds(self, context):
+        if context.shape[-1] == self.caption_channels * 2:
+            return context
+        out_vid = self.video_embeddings_connector(context)[0]
+        out_audio = self.audio_embeddings_connector(context)[0]
+        return torch.concat((out_vid, out_audio), dim=-1)
+
     def _init_transformer_blocks(self, device, dtype, **kwargs):
         """Initialize transformer blocks for LTXAV."""
         self.transformer_blocks = nn.ModuleList(
@@ -702,7 +726,7 @@ class LTXAVModel(LTXVModel):
         return [(v_pe, av_cross_video_freq_cis), (a_pe, av_cross_audio_freq_cis)]
 
     def _process_transformer_blocks(
-        self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs
+        self, x, context, attention_mask, timestep, pe, transformer_options={}, self_attention_mask=None, **kwargs
     ):
         vx = x[0]
         ax = x[1]
@@ -746,6 +770,7 @@ class LTXAVModel(LTXVModel):
                         v_cross_gate_timestep=args["v_cross_gate_timestep"],
                         a_cross_gate_timestep=args["a_cross_gate_timestep"],
                         transformer_options=args["transformer_options"],
+                        self_attention_mask=args.get("self_attention_mask"),
                     )
                     return out
 
@@ -766,6 +791,7 @@ class LTXAVModel(LTXVModel):
                         "v_cross_gate_timestep": av_ca_a2v_gate_noise_timestep,
                         "a_cross_gate_timestep": av_ca_v2a_gate_noise_timestep,
                         "transformer_options": transformer_options,
+                        "self_attention_mask": self_attention_mask,
                     },
                     {"original_block": block_wrap},
                 )
@@ -787,6 +813,7 @@ class LTXAVModel(LTXVModel):
                     v_cross_gate_timestep=av_ca_a2v_gate_noise_timestep,
                     a_cross_gate_timestep=av_ca_v2a_gate_noise_timestep,
                     transformer_options=transformer_options,
+                    self_attention_mask=self_attention_mask,
                 )
 
         return [vx, ax]

comfy/ldm/lightricks/embeddings_connector.py

@@ -157,11 +157,9 @@ class Embeddings1DConnector(nn.Module):
         self.num_learnable_registers = num_learnable_registers
         if self.num_learnable_registers:
             self.learnable_registers = nn.Parameter(
-                torch.rand(
+                torch.empty(
                     self.num_learnable_registers, inner_dim, dtype=dtype, device=device
                 )
-                * 2.0
-                - 1.0
             )
 
     def get_fractional_positions(self, indices_grid):
@@ -234,7 +232,7 @@ class Embeddings1DConnector(nn.Module):
 
         return indices
 
-    def precompute_freqs_cis(self, indices_grid, spacing="exp"):
+    def precompute_freqs_cis(self, indices_grid, spacing="exp", out_dtype=None):
         dim = self.inner_dim
         n_elem = 2  # 2 because of cos and sin
         freqs = self.precompute_freqs(indices_grid, spacing)
@@ -247,7 +245,7 @@ class Embeddings1DConnector(nn.Module):
             )
         else:
             cos_freq, sin_freq = interleaved_freqs_cis(freqs, dim % n_elem)
-        return cos_freq.to(self.dtype), sin_freq.to(self.dtype), self.split_rope
+        return cos_freq.to(dtype=out_dtype), sin_freq.to(dtype=out_dtype), self.split_rope
 
     def forward(
         self,
@@ -288,7 +286,7 @@ class Embeddings1DConnector(nn.Module):
             hidden_states.shape[1], dtype=torch.float32, device=hidden_states.device
         )
         indices_grid = indices_grid[None, None, :]
-        freqs_cis = self.precompute_freqs_cis(indices_grid)
+        freqs_cis = self.precompute_freqs_cis(indices_grid, out_dtype=hidden_states.dtype)
 
         # 2. Blocks
         for block_idx, block in enumerate(self.transformer_1d_blocks):

comfy/ldm/lightricks/model.py

@@ -1,6 +1,7 @@
 from abc import ABC, abstractmethod
 from enum import Enum
 import functools
+import logging
 import math
 from typing import Dict, Optional, Tuple
 
@@ -14,6 +15,8 @@ import comfy.ldm.common_dit
 
 from .symmetric_patchifier import SymmetricPatchifier, latent_to_pixel_coords
 
+logger = logging.getLogger(__name__)
+
 def _log_base(x, base):
     return np.log(x) / np.log(base)
 
@@ -415,12 +418,12 @@ class BasicTransformerBlock(nn.Module):
 
         self.scale_shift_table = nn.Parameter(torch.empty(6, dim, device=device, dtype=dtype))
 
-    def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None, transformer_options={}):
+    def forward(self, x, context=None, attention_mask=None, timestep=None, pe=None, transformer_options={}, self_attention_mask=None):
         shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None, None].to(device=x.device, dtype=x.dtype) + timestep.reshape(x.shape[0], timestep.shape[1], self.scale_shift_table.shape[0], -1)).unbind(dim=2)
 
         attn1_input = comfy.ldm.common_dit.rms_norm(x)
         attn1_input = torch.addcmul(attn1_input, attn1_input, scale_msa).add_(shift_msa)
-        attn1_input = self.attn1(attn1_input, pe=pe, transformer_options=transformer_options)
+        attn1_input = self.attn1(attn1_input, pe=pe, mask=self_attention_mask, transformer_options=transformer_options)
         x.addcmul_(attn1_input, gate_msa)
         del attn1_input
 
@@ -638,8 +641,16 @@ class LTXBaseModel(torch.nn.Module, ABC):
         """Process input data. Must be implemented by subclasses."""
         pass
 
+    def _build_guide_self_attention_mask(self, x, transformer_options, merged_args):
+        """Build self-attention mask for per-guide attention attenuation.
+
+        Base implementation returns None (no attenuation). Subclasses that
+        support guide-based attention control should override this.
+        """
+        return None
+
     @abstractmethod
-    def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, **kwargs):
+    def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, self_attention_mask=None, **kwargs):
         """Process transformer blocks. Must be implemented by subclasses."""
         pass
 
@@ -788,9 +799,17 @@ class LTXBaseModel(torch.nn.Module, ABC):
         attention_mask = self._prepare_attention_mask(attention_mask, input_dtype)
         pe = self._prepare_positional_embeddings(pixel_coords, frame_rate, input_dtype)
 
+        # Build self-attention mask for per-guide attenuation
+        self_attention_mask = self._build_guide_self_attention_mask(
+            x, transformer_options, merged_args
+        )
+
         # Process transformer blocks
         x = self._process_transformer_blocks(
-            x, context, attention_mask, timestep, pe, transformer_options=transformer_options, **merged_args
+            x, context, attention_mask, timestep, pe,
+            transformer_options=transformer_options,
+            self_attention_mask=self_attention_mask,
+            **merged_args,
         )
 
         # Process output
@@ -890,13 +909,243 @@ class LTXVModel(LTXBaseModel):
             pixel_coords = pixel_coords[:, :, grid_mask, ...]
 
             kf_grid_mask = grid_mask[-keyframe_idxs.shape[2]:]
+
+            # Compute per-guide surviving token counts from guide_attention_entries.
+            # Each entry tracks one guide reference; they are appended in order and
+            # their pre_filter_counts partition the kf_grid_mask.
+            guide_entries = kwargs.get("guide_attention_entries", None)
+            if guide_entries:
+                total_pfc = sum(e["pre_filter_count"] for e in guide_entries)
+                if total_pfc != len(kf_grid_mask):
+                    raise ValueError(
+                        f"guide pre_filter_counts ({total_pfc}) != "
+                        f"keyframe grid mask length ({len(kf_grid_mask)})"
+                    )
+                resolved_entries = []
+                offset = 0
+                for entry in guide_entries:
+                    pfc = entry["pre_filter_count"]
+                    entry_mask = kf_grid_mask[offset:offset + pfc]
+                    surviving = int(entry_mask.sum().item())
+                    resolved_entries.append({
+                        **entry,
+                        "surviving_count": surviving,
+                    })
+                    offset += pfc
+                additional_args["resolved_guide_entries"] = resolved_entries
+
             keyframe_idxs = keyframe_idxs[..., kf_grid_mask, :]
             pixel_coords[:, :, -keyframe_idxs.shape[2]:, :] = keyframe_idxs
 
+            # Total surviving guide tokens (all guides)
+            additional_args["num_guide_tokens"] = keyframe_idxs.shape[2]
+
         x = self.patchify_proj(x)
         return x, pixel_coords, additional_args
 
-    def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, transformer_options={}, **kwargs):
+    def _build_guide_self_attention_mask(self, x, transformer_options, merged_args):
+        """Build self-attention mask for per-guide attention attenuation.
+
+        Reads resolved_guide_entries from merged_args (computed in _process_input)
+        to build a log-space additive bias mask that attenuates noisy ↔ guide
+        attention for each guide reference independently.
+
+        Returns None if no attenuation is needed (all strengths == 1.0 and no
+        spatial masks, or no guide tokens).
+        """
+        if isinstance(x, list):
+            # AV model: x = [vx, ax]; use vx for token count and device
+            total_tokens = x[0].shape[1]
+            device = x[0].device
+            dtype = x[0].dtype
+        else:
+            total_tokens = x.shape[1]
+            device = x.device
+            dtype = x.dtype
+
+        num_guide_tokens = merged_args.get("num_guide_tokens", 0)
+        if num_guide_tokens == 0:
+            return None
+
+        resolved_entries = merged_args.get("resolved_guide_entries", None)
+        if not resolved_entries:
+            return None
+
+        # Check if any attenuation is actually needed
+        needs_attenuation = any(
+            e["strength"] < 1.0 or e.get("pixel_mask") is not None
+            for e in resolved_entries
+        )
+        if not needs_attenuation:
+            return None
+
+        # Build per-guide-token weights for all tracked guide tokens.
+        # Guides are appended in order at the end of the sequence.
+        guide_start = total_tokens - num_guide_tokens
+        all_weights = []
+        total_tracked = 0
+
+        for entry in resolved_entries:
+            surviving = entry["surviving_count"]
+            if surviving == 0:
+                continue
+
+            strength = entry["strength"]
+            pixel_mask = entry.get("pixel_mask")
+            latent_shape = entry.get("latent_shape")
+
+            if pixel_mask is not None and latent_shape is not None:
+                f_lat, h_lat, w_lat = latent_shape
+                per_token = self._downsample_mask_to_latent(
+                    pixel_mask.to(device=device, dtype=dtype),
+                    f_lat, h_lat, w_lat,
+                )
+                # per_token shape: (B, f_lat*h_lat*w_lat).
+                # Collapse batch dim — the mask is assumed identical across the
+                # batch; validate and take the first element to get (1, tokens).
+                if per_token.shape[0] > 1:
+                    ref = per_token[0]
+                    for bi in range(1, per_token.shape[0]):
+                        if not torch.equal(ref, per_token[bi]):
+                            logger.warning(
+                                "pixel_mask differs across batch elements; "
+                                "using first element only."
+                            )
+                            break
+                    per_token = per_token[:1]
+                # `surviving` is the post-grid_mask token count.
+                # Clamp to surviving to handle any mismatch safely.
+                n_weights = min(per_token.shape[1], surviving)
+                weights = per_token[:, :n_weights] * strength  # (1, n_weights)
+            else:
+                weights = torch.full(
+                    (1, surviving), strength, device=device, dtype=dtype
+                )
+
+            all_weights.append(weights)
+            total_tracked += weights.shape[1]
+
+        if not all_weights:
+            return None
+
+        # Concatenate per-token weights for all tracked guides
+        tracked_weights = torch.cat(all_weights, dim=1)  # (1, total_tracked)
+
+        # Check if any weight is actually < 1.0 (otherwise no attenuation needed)
+        if (tracked_weights >= 1.0).all():
+            return None
+
+        # Build the mask: guide tokens are at the end of the sequence.
+        # Tracked guides come first (in order), untracked follow.
+        return self._build_self_attention_mask(
+            total_tokens, num_guide_tokens, total_tracked,
+            tracked_weights, guide_start, device, dtype,
+        )
+
+    @staticmethod
+    def _downsample_mask_to_latent(mask, f_lat, h_lat, w_lat):
+        """Downsample a pixel-space mask to per-token latent weights.
+
+        Args:
+            mask: (B, 1, F_pix, H_pix, W_pix) pixel-space mask with values in [0, 1].
+            f_lat: Number of latent frames (pre-dilation original count).
+            h_lat: Latent height (pre-dilation original height).
+            w_lat: Latent width (pre-dilation original width).
+
+        Returns:
+            (B, F_lat * H_lat * W_lat) flattened per-token weights.
+        """
+        b = mask.shape[0]
+        f_pix = mask.shape[2]
+
+        # Spatial downsampling: area interpolation per frame
+        spatial_down = torch.nn.functional.interpolate(
+            rearrange(mask, "b 1 f h w -> (b f) 1 h w"),
+            size=(h_lat, w_lat),
+            mode="area",
+        )
+        spatial_down = rearrange(spatial_down, "(b f) 1 h w -> b 1 f h w", b=b)
+
+        # Temporal downsampling: first pixel frame maps to first latent frame,
+        # remaining pixel frames are averaged in groups for causal temporal structure.
+        first_frame = spatial_down[:, :, :1, :, :]
+        if f_pix > 1 and f_lat > 1:
+            remaining_pix = f_pix - 1
+            remaining_lat = f_lat - 1
+            t = remaining_pix // remaining_lat
+            if t < 1:
+                # Fewer pixel frames than latent frames — upsample by repeating
+                # the available pixel frames via nearest interpolation.
+                rest_flat = rearrange(
+                    spatial_down[:, :, 1:, :, :],
+                    "b 1 f h w -> (b h w) 1 f",
+                )
+                rest_up = torch.nn.functional.interpolate(
+                    rest_flat, size=remaining_lat, mode="nearest",
+                )
+                rest = rearrange(
+                    rest_up, "(b h w) 1 f -> b 1 f h w",
+                    b=b, h=h_lat, w=w_lat,
+                )
+            else:
+                # Trim trailing pixel frames that don't fill a complete group
+                usable = remaining_lat * t
+                rest = rearrange(
+                    spatial_down[:, :, 1:1 + usable, :, :],
+                    "b 1 (f t) h w -> b 1 f t h w",
+                    t=t,
+                )
+                rest = rest.mean(dim=3)
+            latent_mask = torch.cat([first_frame, rest], dim=2)
+        elif f_lat > 1:
+            # Single pixel frame but multiple latent frames — repeat the
+            # single frame across all latent frames.
+            latent_mask = first_frame.expand(-1, -1, f_lat, -1, -1)
+        else:
+            latent_mask = first_frame
+
+        return rearrange(latent_mask, "b 1 f h w -> b (f h w)")
+
+    @staticmethod
+    def _build_self_attention_mask(total_tokens, num_guide_tokens, tracked_count,
+                                    tracked_weights, guide_start, device, dtype):
+        """Build a log-space additive self-attention bias mask.
+
+        Attenuates attention between noisy tokens and tracked guide tokens.
+        Untracked guide tokens (at the end of the guide portion) keep full attention.
+
+        Args:
+            total_tokens: Total sequence length.
+            num_guide_tokens: Total guide tokens (all guides) at end of sequence.
+            tracked_count: Number of tracked guide tokens (first in the guide portion).
+            tracked_weights: (1, tracked_count) tensor, values in [0, 1].
+            guide_start: Index where guide tokens begin in the sequence.
+            device: Target device.
+            dtype: Target dtype.
+
+        Returns:
+            (1, 1, total_tokens, total_tokens) additive bias mask.
+            0.0 = full attention, negative = attenuated, finfo.min = effectively fully masked.
+        """
+        finfo = torch.finfo(dtype)
+        mask = torch.zeros((1, 1, total_tokens, total_tokens), device=device, dtype=dtype)
+        tracked_end = guide_start + tracked_count
+
+        # Convert weights to log-space bias
+        w = tracked_weights.to(device=device, dtype=dtype)  # (1, tracked_count)
+        log_w = torch.full_like(w, finfo.min)
+        positive_mask = w > 0
+        if positive_mask.any():
+            log_w[positive_mask] = torch.log(w[positive_mask].clamp(min=finfo.tiny))
+
+        # noisy → tracked guides: each noisy row gets the same per-guide weight
+        mask[:, :, :guide_start, guide_start:tracked_end] = log_w.view(1, 1, 1, -1)
+        # tracked guides → noisy: each guide row broadcasts its weight across noisy cols
+        mask[:, :, guide_start:tracked_end, :guide_start] = log_w.view(1, 1, -1, 1)
+
+        return mask
+
+    def _process_transformer_blocks(self, x, context, attention_mask, timestep, pe, transformer_options={}, self_attention_mask=None, **kwargs):
         """Process transformer blocks for LTXV."""
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
@@ -906,10 +1155,10 @@ class LTXVModel(LTXBaseModel):
 
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"], transformer_options=args["transformer_options"])
+                    out["img"] = block(args["img"], context=args["txt"], attention_mask=args["attention_mask"], timestep=args["vec"], pe=args["pe"], transformer_options=args["transformer_options"], self_attention_mask=args.get("self_attention_mask"))
                     return out
 
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "attention_mask": attention_mask, "vec": timestep, "pe": pe, "transformer_options": transformer_options}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "attention_mask": attention_mask, "vec": timestep, "pe": pe, "transformer_options": transformer_options, "self_attention_mask": self_attention_mask}, {"original_block": block_wrap})
                 x = out["img"]
             else:
                 x = block(
@@ -919,6 +1168,7 @@ class LTXVModel(LTXBaseModel):
                     timestep=timestep,
                     pe=pe,
                     transformer_options=transformer_options,
+                    self_attention_mask=self_attention_mask,
                 )
 
         return x

comfy/ldm/modules/attention.py

@@ -524,6 +524,9 @@ def attention_pytorch(q, k, v, heads, mask=None, attn_precision=None, skip_resha
 
 @wrap_attn
 def attention_sage(q, k, v, heads, mask=None, attn_precision=None, skip_reshape=False, skip_output_reshape=False, **kwargs):
+    if kwargs.get("low_precision_attention", True) is False:
+        return attention_pytorch(q, k, v, heads, mask=mask, skip_reshape=skip_reshape, skip_output_reshape=skip_output_reshape, **kwargs)
+
     exception_fallback = False
     if skip_reshape:
         b, _, _, dim_head = q.shape

comfy/ldm/modules/diffusionmodules/model.py

@@ -102,19 +102,7 @@ class VideoConv3d(nn.Module):
         return self.conv(x)
 
 def interpolate_up(x, scale_factor):
-    try:
-        return torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode="nearest")
-    except: #operation not implemented for bf16
-        orig_shape = list(x.shape)
-        out_shape = orig_shape[:2]
-        for i in range(len(orig_shape) - 2):
-            out_shape.append(round(orig_shape[i + 2] * scale_factor[i]))
-        out = torch.empty(out_shape, dtype=x.dtype, layout=x.layout, device=x.device)
-        split = 8
-        l = out.shape[1] // split
-        for i in range(0, out.shape[1], l):
-            out[:,i:i+l] = torch.nn.functional.interpolate(x[:,i:i+l].to(torch.float32), scale_factor=scale_factor, mode="nearest").to(x.dtype)
-        return out
+    return torch.nn.functional.interpolate(x, scale_factor=scale_factor, mode="nearest")
 
 class Upsample(nn.Module):
     def __init__(self, in_channels, with_conv, conv_op=ops.Conv2d, scale_factor=2.0):

comfy/ldm/modules/diffusionmodules/openaimodel.py

@@ -18,6 +18,8 @@ import comfy.patcher_extension
 import comfy.ops
 ops = comfy.ops.disable_weight_init
 
+from ..sdpose import HeatmapHead
+
 class TimestepBlock(nn.Module):
     """
     Any module where forward() takes timestep embeddings as a second argument.
@@ -441,6 +443,7 @@ class UNetModel(nn.Module):
         disable_temporal_crossattention=False,
         max_ddpm_temb_period=10000,
         attn_precision=None,
+        heatmap_head=False,
         device=None,
         operations=ops,
     ):
@@ -827,6 +830,9 @@ class UNetModel(nn.Module):
             #nn.LogSoftmax(dim=1)  # change to cross_entropy and produce non-normalized logits
         )
 
+        if heatmap_head:
+            self.heatmap_head = HeatmapHead(device=device, dtype=self.dtype, operations=operations)
+
     def forward(self, x, timesteps=None, context=None, y=None, control=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
             self._forward,

comfy/ldm/modules/sdpose.py

@@ -0,0 +1,130 @@
+import torch
+import numpy as np
+from scipy.ndimage import gaussian_filter
+
+class HeatmapHead(torch.nn.Module):
+    def __init__(
+            self,
+            in_channels=640,
+            out_channels=133,
+            input_size=(768, 1024),
+            heatmap_scale=4,
+            deconv_out_channels=(640,),
+            deconv_kernel_sizes=(4,),
+            conv_out_channels=(640,),
+            conv_kernel_sizes=(1,),
+            final_layer_kernel_size=1,
+            device=None, dtype=None, operations=None
+        ):
+        super().__init__()
+
+        self.heatmap_size = (input_size[0] // heatmap_scale, input_size[1] // heatmap_scale)
+        self.scale_factor = ((np.array(input_size) - 1) / (np.array(self.heatmap_size) - 1)).astype(np.float32)
+
+        # Deconv layers
+        if deconv_out_channels:
+            deconv_layers = []
+            for out_ch, kernel_size in zip(deconv_out_channels, deconv_kernel_sizes):
+                if kernel_size == 4:
+                    padding, output_padding = 1, 0
+                elif kernel_size == 3:
+                    padding, output_padding = 1, 1
+                elif kernel_size == 2:
+                    padding, output_padding = 0, 0
+                else:
+                    raise ValueError(f'Unsupported kernel size {kernel_size}')
+
+                deconv_layers.extend([
+                    operations.ConvTranspose2d(in_channels, out_ch, kernel_size,
+                                     stride=2, padding=padding, output_padding=output_padding, bias=False, device=device, dtype=dtype),
+                    torch.nn.InstanceNorm2d(out_ch, device=device, dtype=dtype),
+                    torch.nn.SiLU(inplace=True)
+                ])
+                in_channels = out_ch
+            self.deconv_layers = torch.nn.Sequential(*deconv_layers)
+        else:
+            self.deconv_layers = torch.nn.Identity()
+
+        # Conv layers
+        if conv_out_channels:
+            conv_layers = []
+            for out_ch, kernel_size in zip(conv_out_channels, conv_kernel_sizes):
+                padding = (kernel_size - 1) // 2
+                conv_layers.extend([
+                    operations.Conv2d(in_channels, out_ch, kernel_size,
+                            stride=1, padding=padding, device=device, dtype=dtype),
+                    torch.nn.InstanceNorm2d(out_ch, device=device, dtype=dtype),
+                    torch.nn.SiLU(inplace=True)
+                ])
+                in_channels = out_ch
+            self.conv_layers = torch.nn.Sequential(*conv_layers)
+        else:
+            self.conv_layers = torch.nn.Identity()
+
+        self.final_layer = operations.Conv2d(in_channels, out_channels, kernel_size=final_layer_kernel_size, padding=final_layer_kernel_size // 2, device=device, dtype=dtype)
+
+    def forward(self, x): # Decode heatmaps to keypoints
+        heatmaps = self.final_layer(self.conv_layers(self.deconv_layers(x)))
+        heatmaps_np = heatmaps.float().cpu().numpy()  # (B, K, H, W)
+        B, K, H, W = heatmaps_np.shape
+
+        batch_keypoints = []
+        batch_scores = []
+
+        for b in range(B):
+            hm = heatmaps_np[b].copy()  # (K, H, W)
+
+            # --- vectorised argmax ---
+            flat = hm.reshape(K, -1)
+            idx = np.argmax(flat, axis=1)
+            scores = flat[np.arange(K), idx].copy()
+            y_locs, x_locs = np.unravel_index(idx, (H, W))
+            keypoints = np.stack([x_locs, y_locs], axis=-1).astype(np.float32)  # (K, 2) in heatmap space
+            invalid = scores <= 0.
+            keypoints[invalid] = -1
+
+            # --- DARK sub-pixel refinement (UDP) ---
+            # 1. Gaussian blur with max-preserving normalisation
+            border = 5  # (kernel-1)//2 for kernel=11
+            for k in range(K):
+                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)
+                hm[k] = dr[border:-border, border:-border].copy()
+                cur_max = np.max(hm[k])
+                if cur_max > 0:
+                    hm[k] *= origin_max / cur_max
+            # 2. Log-space for Taylor expansion
+            np.clip(hm, 1e-3, 50., hm)
+            np.log(hm, hm)
+            # 3. Hessian-based Newton step
+            hm_pad = np.pad(hm, ((0, 0), (1, 1), (1, 1)), mode='edge').flatten()
+            index = keypoints[:, 0] + 1 + (keypoints[:, 1] + 1) * (W + 2)
+            index += (W + 2) * (H + 2) * np.arange(0, K)
+            index = index.astype(int).reshape(-1, 1)
+            i_       = hm_pad[index]
+            ix1      = hm_pad[index + 1]
+            iy1      = hm_pad[index + W + 2]
+            ix1y1    = hm_pad[index + W + 3]
+            ix1_y1_  = hm_pad[index - W - 3]
+            ix1_     = hm_pad[index - 1]
+            iy1_     = hm_pad[index - 2 - W]
+            dx = 0.5 * (ix1 - ix1_)
+            dy = 0.5 * (iy1 - iy1_)
+            derivative = np.concatenate([dx, dy], axis=1).reshape(K, 2, 1)
+            dxx = ix1  - 2 * i_ + ix1_
+            dyy = iy1  - 2 * i_ + iy1_
+            dxy = 0.5 * (ix1y1 - ix1 - iy1 + i_ + i_ - ix1_ - iy1_ + ix1_y1_)
+            hessian = np.concatenate([dxx, dxy, dxy, dyy], axis=1).reshape(K, 2, 2)
+            hessian = np.linalg.inv(hessian + np.finfo(np.float32).eps * np.eye(2))
+            keypoints -= np.einsum('imn,ink->imk', hessian, derivative).squeeze(axis=-1)
+
+            # --- restore to input image space ---
+            keypoints = keypoints * self.scale_factor
+            keypoints[invalid] = -1
+
+            batch_keypoints.append(keypoints)
+            batch_scores.append(scores)
+
+        return batch_keypoints, batch_scores

comfy/ldm/qwen_image/controlnet.py

@@ -2,6 +2,196 @@ import torch
 import math
 
 from .model import QwenImageTransformer2DModel
+from .model import QwenImageTransformerBlock
+
+
+class QwenImageFunControlBlock(QwenImageTransformerBlock):
+    def __init__(self, dim, num_attention_heads, attention_head_dim, has_before_proj=False, dtype=None, device=None, operations=None):
+        super().__init__(
+            dim=dim,
+            num_attention_heads=num_attention_heads,
+            attention_head_dim=attention_head_dim,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+        self.has_before_proj = has_before_proj
+        if has_before_proj:
+            self.before_proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+        self.after_proj = operations.Linear(dim, dim, device=device, dtype=dtype)
+
+
+class QwenImageFunControlNetModel(torch.nn.Module):
+    def __init__(
+        self,
+        control_in_features=132,
+        inner_dim=3072,
+        num_attention_heads=24,
+        attention_head_dim=128,
+        num_control_blocks=5,
+        main_model_double=60,
+        injection_layers=(0, 12, 24, 36, 48),
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.main_model_double = main_model_double
+        self.injection_layers = tuple(injection_layers)
+        # Keep base hint scaling at 1.0 so user-facing strength behaves similarly
+        # to the reference Gen2/VideoX implementation around strength=1.
+        self.hint_scale = 1.0
+        self.control_img_in = operations.Linear(control_in_features, inner_dim, device=device, dtype=dtype)
+
+        self.control_blocks = torch.nn.ModuleList([])
+        for i in range(num_control_blocks):
+            self.control_blocks.append(
+                QwenImageFunControlBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    has_before_proj=(i == 0),
+                    dtype=dtype,
+                    device=device,
+                    operations=operations,
+                )
+            )
+
+    def _process_hint_tokens(self, hint):
+        if hint is None:
+            return None
+        if hint.ndim == 4:
+            hint = hint.unsqueeze(2)
+
+        # Fun checkpoints are trained with 33 latent channels before 2x2 packing:
+        # [control_latent(16), mask(1), inpaint_latent(16)] -> 132 features.
+        # Default behavior (no inpaint input in stock Apply ControlNet) should use
+        # zeros for mask/inpaint branches, matching VideoX fallback semantics.
+        expected_c = self.control_img_in.weight.shape[1] // 4
+        if hint.shape[1] == 16 and expected_c == 33:
+            zeros_mask = torch.zeros_like(hint[:, :1])
+            zeros_inpaint = torch.zeros_like(hint)
+            hint = torch.cat([hint, zeros_mask, zeros_inpaint], dim=1)
+
+        bs, c, t, h, w = hint.shape
+        hidden_states = torch.nn.functional.pad(hint, (0, w % 2, 0, h % 2))
+        orig_shape = hidden_states.shape
+        hidden_states = hidden_states.view(
+            orig_shape[0],
+            orig_shape[1],
+            orig_shape[-3],
+            orig_shape[-2] // 2,
+            2,
+            orig_shape[-1] // 2,
+            2,
+        )
+        hidden_states = hidden_states.permute(0, 2, 3, 5, 1, 4, 6)
+        hidden_states = hidden_states.reshape(
+            bs,
+            t * ((h + 1) // 2) * ((w + 1) // 2),
+            c * 4,
+        )
+
+        expected_in = self.control_img_in.weight.shape[1]
+        cur_in = hidden_states.shape[-1]
+        if cur_in < expected_in:
+            pad = torch.zeros(
+                (hidden_states.shape[0], hidden_states.shape[1], expected_in - cur_in),
+                device=hidden_states.device,
+                dtype=hidden_states.dtype,
+            )
+            hidden_states = torch.cat([hidden_states, pad], dim=-1)
+        elif cur_in > expected_in:
+            hidden_states = hidden_states[:, :, :expected_in]
+
+        return hidden_states
+
+    def forward(
+        self,
+        x,
+        timesteps,
+        context,
+        attention_mask=None,
+        guidance: torch.Tensor = None,
+        hint=None,
+        transformer_options={},
+        base_model=None,
+        **kwargs,
+    ):
+        if base_model is None:
+            raise RuntimeError("Qwen Fun ControlNet requires a QwenImage base model at runtime.")
+
+        encoder_hidden_states_mask = attention_mask
+        # Keep attention mask disabled inside Fun control blocks to mirror
+        # VideoX behavior (they rely on seq lengths for RoPE, not masked attention).
+        encoder_hidden_states_mask = None
+
+        hidden_states, img_ids, _ = base_model.process_img(x)
+        hint_tokens = self._process_hint_tokens(hint)
+        if hint_tokens is None:
+            raise RuntimeError("Qwen Fun ControlNet requires a control hint image.")
+
+        if hint_tokens.shape[1] != hidden_states.shape[1]:
+            max_tokens = min(hint_tokens.shape[1], hidden_states.shape[1])
+            hint_tokens = hint_tokens[:, :max_tokens]
+            hidden_states = hidden_states[:, :max_tokens]
+            img_ids = img_ids[:, :max_tokens]
+
+        txt_start = round(
+            max(
+                ((x.shape[-1] + (base_model.patch_size // 2)) // base_model.patch_size) // 2,
+                ((x.shape[-2] + (base_model.patch_size // 2)) // base_model.patch_size) // 2,
+            )
+        )
+        txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        image_rotary_emb = base_model.pe_embedder(ids).to(x.dtype).contiguous()
+
+        hidden_states = base_model.img_in(hidden_states)
+        encoder_hidden_states = base_model.txt_norm(context)
+        encoder_hidden_states = base_model.txt_in(encoder_hidden_states)
+
+        if guidance is not None:
+            guidance = guidance * 1000
+
+        temb = (
+            base_model.time_text_embed(timesteps, hidden_states)
+            if guidance is None
+            else base_model.time_text_embed(timesteps, guidance, hidden_states)
+        )
+
+        c = self.control_img_in(hint_tokens)
+
+        for i, block in enumerate(self.control_blocks):
+            if i == 0:
+                c_in = block.before_proj(c) + hidden_states
+                all_c = []
+            else:
+                all_c = list(torch.unbind(c, dim=0))
+                c_in = all_c.pop(-1)
+
+            encoder_hidden_states, c_out = block(
+                hidden_states=c_in,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_hidden_states_mask=encoder_hidden_states_mask,
+                temb=temb,
+                image_rotary_emb=image_rotary_emb,
+                transformer_options=transformer_options,
+            )
+
+            c_skip = block.after_proj(c_out) * self.hint_scale
+            all_c += [c_skip, c_out]
+            c = torch.stack(all_c, dim=0)
+
+        hints = torch.unbind(c, dim=0)[:-1]
+
+        controlnet_block_samples = [None] * self.main_model_double
+        for local_idx, base_idx in enumerate(self.injection_layers):
+            if local_idx < len(hints) and base_idx < len(controlnet_block_samples):
+                controlnet_block_samples[base_idx] = hints[local_idx]
+
+        return {"input": controlnet_block_samples}
 
 
 class QwenImageControlNetModel(QwenImageTransformer2DModel):

comfy/ldm/wan/vae.py

@@ -459,6 +459,7 @@ class WanVAE(nn.Module):
                  attn_scales=[],
                  temperal_downsample=[True, True, False],
                  image_channels=3,
+                 conv_out_channels=3,
                  dropout=0.0):
         super().__init__()
         self.dim = dim
@@ -474,7 +475,7 @@ class WanVAE(nn.Module):
                                  attn_scales, self.temperal_downsample, dropout)
         self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
         self.conv2 = CausalConv3d(z_dim, z_dim, 1)
-        self.decoder = Decoder3d(dim, z_dim, image_channels, dim_mult, num_res_blocks,
+        self.decoder = Decoder3d(dim, z_dim, conv_out_channels, dim_mult, num_res_blocks,
                                  attn_scales, self.temperal_upsample, dropout)
 
     def encode(self, x):

comfy/lora.py

@@ -332,6 +332,13 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["{}".format(key_lora)] = k
                 key_map["transformer.{}".format(key_lora)] = k
 
+    if isinstance(model, comfy.model_base.ACEStep15):
+        for k in sdk:
+            if k.startswith("diffusion_model.decoder.") and k.endswith(".weight"):
+                key_lora = k[len("diffusion_model.decoder."):-len(".weight")]
+                key_map["base_model.model.{}".format(key_lora)] = k  # Official base model loras
+                key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k  # LyCORIS/LoKR format
+
     return key_map
 
 
@@ -368,6 +375,31 @@ def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Ten
 
     return padded_tensor
 
+def calculate_shape(patches, weight, key, original_weights=None):
+    current_shape = weight.shape
+
+    for p in patches:
+        v = p[1]
+        offset = p[3]
+
+        # Offsets restore the old shape; lists force a diff without metadata
+        if offset is not None or isinstance(v, list):
+            continue
+
+        if isinstance(v, weight_adapter.WeightAdapterBase):
+            adapter_shape = v.calculate_shape(key)
+            if adapter_shape is not None:
+                current_shape = adapter_shape
+            continue
+
+        # Standard diff logic with padding
+        if len(v) == 2:
+            patch_type, patch_data = v[0], v[1]
+            if patch_type == "diff" and len(patch_data) > 1 and patch_data[1]['pad_weight']:
+                current_shape = patch_data[0].shape
+
+    return current_shape
+
 def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, original_weights=None):
     for p in patches:
         strength = p[0]

comfy/lora_convert.py

@@ -5,7 +5,7 @@ import comfy.utils
 def convert_lora_bfl_control(sd): #BFL loras for Flux
     sd_out = {}
     for k in sd:
-        k_to = "diffusion_model.{}".format(k.replace(".lora_B.bias", ".diff_b").replace("_norm.scale", "_norm.scale.set_weight"))
+        k_to = "diffusion_model.{}".format(k.replace(".lora_B.bias", ".diff_b").replace("_norm.scale", "_norm.set_weight"))
         sd_out[k_to] = sd[k]
 
     sd_out["diffusion_model.img_in.reshape_weight"] = torch.tensor([sd["img_in.lora_B.weight"].shape[0], sd["img_in.lora_A.weight"].shape[1]])

comfy/memory_management.py

@@ -78,4 +78,4 @@ def interpret_gathered_like(tensors, gathered):
 
     return dest_views
 
-aimdo_allocator = None
+aimdo_enabled = False

comfy/model_base.py

@@ -50,6 +50,7 @@ import comfy.ldm.omnigen.omnigen2
 import comfy.ldm.qwen_image.model
 import comfy.ldm.kandinsky5.model
 import comfy.ldm.anima.model
+import comfy.ldm.ace.ace_step15
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -75,6 +76,7 @@ class ModelType(Enum):
     FLUX = 8
     IMG_TO_IMG = 9
     FLOW_COSMOS = 10
+    IMG_TO_IMG_FLOW = 11
 
 
 def model_sampling(model_config, model_type):
@@ -107,6 +109,8 @@ def model_sampling(model_config, model_type):
     elif model_type == ModelType.FLOW_COSMOS:
         c = comfy.model_sampling.COSMOS_RFLOW
         s = comfy.model_sampling.ModelSamplingCosmosRFlow
+    elif model_type == ModelType.IMG_TO_IMG_FLOW:
+        c = comfy.model_sampling.IMG_TO_IMG_FLOW
 
     class ModelSampling(s, c):
         pass
@@ -146,11 +150,11 @@ class BaseModel(torch.nn.Module):
                 self.diffusion_model.to(memory_format=torch.channels_last)
                 logging.debug("using channels last mode for diffusion model")
             logging.info("model weight dtype {}, manual cast: {}".format(self.get_dtype(), self.manual_cast_dtype))
+            comfy.model_management.archive_model_dtypes(self.diffusion_model)
+
         self.model_type = model_type
         self.model_sampling = model_sampling(model_config, model_type)
 
-        comfy.model_management.archive_model_dtypes(self.diffusion_model)
-
         self.adm_channels = unet_config.get("adm_in_channels", None)
         if self.adm_channels is None:
             self.adm_channels = 0
@@ -177,10 +181,7 @@ class BaseModel(torch.nn.Module):
             xc = torch.cat([xc] + [comfy.model_management.cast_to_device(c_concat, xc.device, xc.dtype)], dim=1)
 
         context = c_crossattn
-        dtype = self.get_dtype()
-
-        if self.manual_cast_dtype is not None:
-            dtype = self.manual_cast_dtype
+        dtype = self.get_dtype_inference()
 
         xc = xc.to(dtype)
         device = xc.device
@@ -217,6 +218,13 @@ class BaseModel(torch.nn.Module):
     def get_dtype(self):
         return self.diffusion_model.dtype
 
+    def get_dtype_inference(self):
+        dtype = self.get_dtype()
+
+        if self.manual_cast_dtype is not None:
+            dtype = self.manual_cast_dtype
+        return dtype
+
     def encode_adm(self, **kwargs):
         return None
 
@@ -371,9 +379,7 @@ class BaseModel(torch.nn.Module):
                     input_shapes += shape
 
         if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
-            dtype = self.get_dtype()
-            if self.manual_cast_dtype is not None:
-                dtype = self.manual_cast_dtype
+            dtype = self.get_dtype_inference()
             #TODO: this needs to be tweaked
             area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes))
             return (area * comfy.model_management.dtype_size(dtype) * 0.01 * self.memory_usage_factor) * (1024 * 1024)
@@ -968,6 +974,10 @@ class LTXV(BaseModel):
         if keyframe_idxs is not None:
             out['keyframe_idxs'] = comfy.conds.CONDRegular(keyframe_idxs)
 
+        guide_attention_entries = kwargs.get("guide_attention_entries", None)
+        if guide_attention_entries is not None:
+            out['guide_attention_entries'] = comfy.conds.CONDConstant(guide_attention_entries)
+
         return out
 
     def process_timestep(self, timestep, x, denoise_mask=None, **kwargs):
@@ -985,10 +995,14 @@ class LTXAV(BaseModel):
     def extra_conds(self, **kwargs):
         out = super().extra_conds(**kwargs)
         attention_mask = kwargs.get("attention_mask", None)
+        device = kwargs["device"]
+
         if attention_mask is not None:
             out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
         cross_attn = kwargs.get("cross_attn", None)
         if cross_attn is not None:
+            if hasattr(self.diffusion_model, "preprocess_text_embeds"):
+                cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype_inference()))
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
 
         out['frame_rate'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", 25))
@@ -1016,6 +1030,10 @@ class LTXAV(BaseModel):
         if latent_shapes is not None:
             out['latent_shapes'] = comfy.conds.CONDConstant(latent_shapes)
 
+        guide_attention_entries = kwargs.get("guide_attention_entries", None)
+        if guide_attention_entries is not None:
+            out['guide_attention_entries'] = comfy.conds.CONDConstant(guide_attention_entries)
+
         return out
 
     def process_timestep(self, timestep, x, denoise_mask=None, audio_denoise_mask=None, **kwargs):
@@ -1159,12 +1177,16 @@ class Anima(BaseModel):
         device = kwargs["device"]
         if cross_attn is not None:
             if t5xxl_ids is not None:
-                cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.unsqueeze(0).to(device=device))
                 if t5xxl_weights is not None:
-                    cross_attn *= t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn)
+                    t5xxl_weights = t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn)
+                t5xxl_ids = t5xxl_ids.unsqueeze(0)
+
+                if torch.is_inference_mode_enabled():  # if not we are training
+                    cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype_inference()), t5xxl_ids.to(device=device), t5xxl_weights=t5xxl_weights.to(device=device, dtype=self.get_dtype_inference()))
+                else:
+                    out['t5xxl_ids'] = comfy.conds.CONDRegular(t5xxl_ids)
+                    out['t5xxl_weights'] = comfy.conds.CONDRegular(t5xxl_weights)
 
-                if cross_attn.shape[1] < 512:
-                    cross_attn = torch.nn.functional.pad(cross_attn, (0, 0, 0, 512 - cross_attn.shape[1]))
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
 
@@ -1455,6 +1477,12 @@ class WAN22(WAN21):
     def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
         return latent_image
 
+class WAN21_FlowRVS(WAN21):
+    def __init__(self, model_config, model_type=ModelType.IMG_TO_IMG_FLOW, image_to_video=False, device=None):
+        model_config.unet_config["model_type"] = "t2v"
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
+        self.image_to_video = image_to_video
+
 class Hunyuan3Dv2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan3d.model.Hunyuan3Dv2)
@@ -1540,6 +1568,49 @@ class ACEStep(BaseModel):
         out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0))
         return out
 
+class ACEStep15(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ace.ace_step15.AceStepConditionGenerationModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        device = kwargs["device"]
+        noise = kwargs["noise"]
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            if torch.count_nonzero(cross_attn) == 0:
+                out['replace_with_null_embeds'] = comfy.conds.CONDConstant(True)
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
+        if cross_attn is not None:
+            out['lyric_embed'] = comfy.conds.CONDRegular(conditioning_lyrics)
+
+        refer_audio = kwargs.get("reference_audio_timbre_latents", None)
+        if refer_audio is None or len(refer_audio) == 0:
+            refer_audio = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device)
+            pass_audio_codes = True
+        else:
+            refer_audio = refer_audio[-1][:, :, :noise.shape[2]]
+            out['is_covers'] = comfy.conds.CONDConstant(True)
+            pass_audio_codes = False
+
+        if pass_audio_codes:
+            audio_codes = kwargs.get("audio_codes", None)
+            if audio_codes is not None:
+                out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device))
+                refer_audio = refer_audio[:, :, :750]
+            else:
+                out['is_covers'] = comfy.conds.CONDConstant(False)
+
+        if refer_audio.shape[2] < noise.shape[2]:
+            pad = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device)
+            refer_audio = torch.cat([refer_audio.to(pad), pad[:, :, refer_audio.shape[2]:]], dim=2)
+
+        out['refer_audio'] = comfy.conds.CONDRegular(refer_audio)
+        return out
+
 class Omnigen2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel)

comfy/model_detection.py

@@ -19,6 +19,12 @@ def count_blocks(state_dict_keys, prefix_string):
         count += 1
     return count
 
+def any_suffix_in(keys, prefix, main, suffix_list=[]):
+    for x in suffix_list:
+        if "{}{}{}".format(prefix, main, x) in keys:
+            return True
+    return False
+
 def calculate_transformer_depth(prefix, state_dict_keys, state_dict):
     context_dim = None
     use_linear_in_transformer = False
@@ -186,7 +192,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["meanflow_sum"] = False
         return dit_config
 
-    if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight)
+    if any_suffix_in(state_dict_keys, key_prefix, 'double_blocks.0.img_attn.norm.key_norm.', ["weight", "scale"]) and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or any_suffix_in(state_dict_keys, key_prefix, 'distilled_guidance_layer.norms.0.', ["weight", "scale"])): #Flux, Chroma or Chroma Radiance (has no img_in.weight)
         dit_config = {}
         if '{}double_stream_modulation_img.lin.weight'.format(key_prefix) in state_dict_keys:
             dit_config["image_model"] = "flux2"
@@ -241,7 +247,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
         dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
         dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
-        if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma
+
+        if any_suffix_in(state_dict_keys, key_prefix, 'distilled_guidance_layer.0.norms.0.', ["weight", "scale"]) or any_suffix_in(state_dict_keys, key_prefix, 'distilled_guidance_layer.norms.0.', ["weight", "scale"]): #Chroma
             dit_config["image_model"] = "chroma"
             dit_config["in_channels"] = 64
             dit_config["out_channels"] = 64
@@ -249,7 +256,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["out_dim"] = 3072
             dit_config["hidden_dim"] = 5120
             dit_config["n_layers"] = 5
-            if f"{key_prefix}nerf_blocks.0.norm.scale" in state_dict_keys: #Chroma Radiance
+
+            if any_suffix_in(state_dict_keys, key_prefix, 'nerf_blocks.0.norm.', ["weight", "scale"]): #Chroma Radiance
                 dit_config["image_model"] = "chroma_radiance"
                 dit_config["in_channels"] = 3
                 dit_config["out_channels"] = 3
@@ -259,7 +267,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["nerf_depth"] = 4
                 dit_config["nerf_max_freqs"] = 8
                 dit_config["nerf_tile_size"] = 512
-                dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear"
+                dit_config["nerf_final_head_type"] = "conv" if any_suffix_in(state_dict_keys, key_prefix, 'nerf_final_layer_conv.norm.', ["weight", "scale"]) else "linear"
                 dit_config["nerf_embedder_dtype"] = torch.float32
                 if "{}__x0__".format(key_prefix) in state_dict_keys: # x0 pred
                     dit_config["use_x0"] = True
@@ -268,7 +276,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         else:
             dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
             dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys
-            dit_config["txt_norm"] = "{}txt_norm.scale".format(key_prefix) in state_dict_keys
+            dit_config["txt_norm"] = any_suffix_in(state_dict_keys, key_prefix, 'txt_norm.', ["weight", "scale"])
             if dit_config["yak_mlp"] and dit_config["txt_norm"]:  # Ovis model
                 dit_config["txt_ids_dims"] = [1, 2]
 
@@ -501,6 +509,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         if ref_conv_weight is not None:
             dit_config["in_dim_ref_conv"] = ref_conv_weight.shape[1]
 
+        if metadata is not None and "config" in metadata:
+            dit_config.update(json.loads(metadata["config"]).get("transformer", {}))
+
         return dit_config
 
     if '{}latent_in.weight'.format(key_prefix) in state_dict_keys:  # Hunyuan 3D
@@ -655,6 +666,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.')
         return dit_config
 
+    if '{}encoder.lyric_encoder.layers.0.input_layernorm.weight'.format(key_prefix) in state_dict_keys:
+        dit_config = {}
+        dit_config["audio_model"] = "ace1.5"
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 
@@ -779,6 +795,10 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         unet_config["use_temporal_resblock"] = False
         unet_config["use_temporal_attention"] = False
 
+    heatmap_key = '{}heatmap_head.conv_layers.0.weight'.format(key_prefix)
+    if heatmap_key in state_dict_keys:
+        unet_config["heatmap_head"] = True
+
     return unet_config
 
 def model_config_from_unet_config(unet_config, state_dict=None):
@@ -999,7 +1019,7 @@ def unet_config_from_diffusers_unet(state_dict, dtype=None):
 
     LotusD = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': 4,
             'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
-            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024, 'num_heads': 8,
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024, 'num_head_channels': 64,
             'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
             'use_temporal_attention': False, 'use_temporal_resblock': False}
 

comfy/model_management.py

@@ -20,6 +20,7 @@ import psutil
 import logging
 from enum import Enum
 from comfy.cli_args import args, PerformanceFeature
+import threading
 import torch
 import sys
 import platform
@@ -54,6 +55,11 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
+
+# Training Related State
+in_training = False
+
+
 def get_supported_float8_types():
     float8_types = []
     try:
@@ -344,7 +350,7 @@ AMD_ENABLE_MIOPEN_ENV = 'COMFYUI_ENABLE_MIOPEN'
 
 try:
     if is_amd():
-        arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
+        arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName.split(':')[0]
         if not (any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH)):
             if os.getenv(AMD_ENABLE_MIOPEN_ENV) != '1':
                 torch.backends.cudnn.enabled = False  # Seems to improve things a lot on AMD
@@ -372,7 +378,7 @@ try:
         if args.use_split_cross_attention == False and args.use_quad_cross_attention == False:
             if aotriton_supported(arch):  # AMD efficient attention implementation depends on aotriton.
                 if torch_version_numeric >= (2, 7):  # works on 2.6 but doesn't actually seem to improve much
-                    if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
+                    if any((a in arch) for a in ["gfx90a", "gfx942", "gfx950", "gfx1100", "gfx1101", "gfx1151"]):  # TODO: more arches, TODO: gfx950
                         ENABLE_PYTORCH_ATTENTION = True
                 if rocm_version >= (7, 0):
                    if any((a in arch) for a in ["gfx1200", "gfx1201"]):
@@ -830,7 +836,7 @@ def unet_inital_load_device(parameters, dtype):
 
     mem_dev = get_free_memory(torch_dev)
     mem_cpu = get_free_memory(cpu_dev)
-    if mem_dev > mem_cpu and model_size < mem_dev and comfy.memory_management.aimdo_allocator is None:
+    if mem_dev > mem_cpu and model_size < mem_dev and comfy.memory_management.aimdo_enabled:
         return torch_dev
     else:
         return cpu_dev
@@ -1112,11 +1118,10 @@ def get_cast_buffer(offload_stream, device, size, ref):
             return None
         if cast_buffer is not None and cast_buffer.numel() > 50 * (1024 ** 2):
             #I want my wrongly sized 50MB+ of VRAM back from the caching allocator right now
-            torch.cuda.synchronize()
+            synchronize()
             del STREAM_CAST_BUFFERS[offload_stream]
             del cast_buffer
-            #FIXME: This doesn't work in Aimdo because mempool cant clear cache
-            torch.cuda.empty_cache()
+            soft_empty_cache()
         with wf_context:
             cast_buffer = torch.empty((size), dtype=torch.int8, device=device)
             STREAM_CAST_BUFFERS[offload_stream] = cast_buffer
@@ -1132,9 +1137,7 @@ def reset_cast_buffers():
     for offload_stream in STREAM_CAST_BUFFERS:
         offload_stream.synchronize()
     STREAM_CAST_BUFFERS.clear()
-    if comfy.memory_management.aimdo_allocator is None:
-        #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist
-        torch.cuda.empty_cache()
+    soft_empty_cache()
 
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
@@ -1202,27 +1205,35 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str
         assert r is None
         assert stream is None
 
-        r = torch.empty_like(weight, dtype=weight._model_dtype, device=device)
+        cast_geometry = comfy.memory_management.tensors_to_geometries([ weight ])
+
+        if dtype is None:
+            dtype = weight._model_dtype
 
         signature = comfy_aimdo.model_vbar.vbar_fault(weight._v)
         if signature is not None:
-            raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device)
-            v_tensor = comfy.memory_management.interpret_gathered_like([r], raw_tensor)[0]
+            if comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
+                v_tensor = weight._v_tensor
+            else:
+                raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device)
+                v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, raw_tensor)[0]
+                weight._v_tensor = v_tensor
+                weight._v_signature = signature
+                #Send it over
+                v_tensor.copy_(weight, non_blocking=non_blocking)
+            return v_tensor.to(dtype=dtype)
 
-        if comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
-            #always take a deep copy even if _v is good, as we have no reasonable point to unpin
-            #a non comfy weight
-            r.copy_(v_tensor)
-            comfy_aimdo.model_vbar.vbar_unpin(weight._v)
-            return r
+        r = torch.empty_like(weight, dtype=dtype, device=device)
 
+        if weight.dtype != r.dtype and weight.dtype != weight._model_dtype:
+            #Offloaded casting could skip this, however it would make the quantizations
+            #inconsistent between loaded and offloaded weights. So force the double casting
+            #that would happen in regular flow to make offload deterministic.
+            cast_buffer = torch.empty_like(weight, dtype=weight._model_dtype, device=device)
+            cast_buffer.copy_(weight, non_blocking=non_blocking)
+            weight = cast_buffer
         r.copy_(weight, non_blocking=non_blocking)
 
-        if signature is not None:
-            weight._v_signature = signature
-            v_tensor.copy_(r)
-            comfy_aimdo.model_vbar.vbar_unpin(weight._v)
-
         return r
 
     if device is None or weight.device == device:
@@ -1275,7 +1286,7 @@ def discard_cuda_async_error():
         a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
         b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
         _ = a + b
-        torch.cuda.synchronize()
+        synchronize()
     except torch.AcceleratorError:
         #Dump it! We already know about it from the synchronous return
         pass
@@ -1679,6 +1690,12 @@ def lora_compute_dtype(device):
     LORA_COMPUTE_DTYPES[device] = dtype
     return dtype
 
+def synchronize():
+    if is_intel_xpu():
+        torch.xpu.synchronize()
+    elif torch.cuda.is_available():
+        torch.cuda.synchronize()
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:
@@ -1690,11 +1707,9 @@ def soft_empty_cache(force=False):
     elif is_mlu():
         torch.mlu.empty_cache()
     elif torch.cuda.is_available():
-        if comfy.memory_management.aimdo_allocator is None:
-            #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist
-            torch.cuda.synchronize()
-            torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
+        torch.cuda.synchronize()
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
 
 def unload_all_models():
     free_memory(1e30, get_torch_device())
@@ -1704,9 +1719,6 @@ def debug_memory_summary():
         return torch.cuda.memory.memory_summary()
     return ""
 
-#TODO: might be cleaner to put this somewhere else
-import threading
-
 class InterruptProcessingException(Exception):
     pass
 

comfy/model_patcher.py

@@ -19,7 +19,6 @@
 from __future__ import annotations
 
 import collections
-import copy
 import inspect
 import logging
 import math
@@ -111,6 +110,10 @@ def move_weight_functions(m, device):
                 memory += f.move_to(device=device)
     return memory
 
+def string_to_seed(data):
+    logging.warning("WARNING: string_to_seed has moved from comfy.model_patcher to comfy.utils")
+    return comfy.utils.string_to_seed(data)
+
 class LowVramPatch:
     def __init__(self, key, patches, convert_func=None, set_func=None):
         self.key = key
@@ -157,6 +160,11 @@ def get_key_weight(model, key):
 
     return weight, set_func, convert_func
 
+def key_param_name_to_key(key, param):
+    if len(key) == 0:
+        return param
+    return "{}.{}".format(key, param)
+
 class AutoPatcherEjector:
     def __init__(self, model: 'ModelPatcher', skip_and_inject_on_exit_only=False):
         self.model = model
@@ -263,6 +271,7 @@ class ModelPatcher:
         self.is_clip = False
         self.hook_mode = comfy.hooks.EnumHookMode.MaxSpeed
 
+        self.cached_patcher_init: tuple[Callable, tuple] | None = None
         if not hasattr(self.model, 'model_loaded_weight_memory'):
             self.model.model_loaded_weight_memory = 0
 
@@ -299,8 +308,15 @@ class ModelPatcher:
     def get_free_memory(self, device):
         return comfy.model_management.get_free_memory(device)
 
-    def clone(self):
-        n = self.__class__(self.model, self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update)
+    def clone(self, disable_dynamic=False):
+        class_ = self.__class__
+        model = self.model
+        if self.is_dynamic() and disable_dynamic:
+            class_ = ModelPatcher
+            temp_model_patcher = self.cached_patcher_init[0](*self.cached_patcher_init[1], disable_dynamic=True)
+            model = temp_model_patcher.model
+
+        n = class_(model, self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update)
         n.patches = {}
         for k in self.patches:
             n.patches[k] = self.patches[k][:]
@@ -308,7 +324,7 @@ class ModelPatcher:
 
         n.object_patches = self.object_patches.copy()
         n.weight_wrapper_patches = self.weight_wrapper_patches.copy()
-        n.model_options = copy.deepcopy(self.model_options)
+        n.model_options = comfy.utils.deepcopy_list_dict(self.model_options)
         n.backup = self.backup
         n.object_patches_backup = self.object_patches_backup
         n.parent = self
@@ -354,6 +370,8 @@ class ModelPatcher:
         n.is_clip = self.is_clip
         n.hook_mode = self.hook_mode
 
+        n.cached_patcher_init = self.cached_patcher_init
+
         for callback in self.get_all_callbacks(CallbacksMP.ON_CLONE):
             callback(self, n)
         return n
@@ -398,13 +416,16 @@ class ModelPatcher:
     def memory_required(self, input_shape):
         return self.model.memory_required(input_shape=input_shape)
 
+    def disable_model_cfg1_optimization(self):
+        self.model_options["disable_cfg1_optimization"] = True
+
     def set_model_sampler_cfg_function(self, sampler_cfg_function, disable_cfg1_optimization=False):
         if len(inspect.signature(sampler_cfg_function).parameters) == 3:
             self.model_options["sampler_cfg_function"] = lambda args: sampler_cfg_function(args["cond"], args["uncond"], args["cond_scale"]) #Old way
         else:
             self.model_options["sampler_cfg_function"] = sampler_cfg_function
         if disable_cfg1_optimization:
-            self.model_options["disable_cfg1_optimization"] = True
+            self.disable_model_cfg1_optimization()
 
     def set_model_sampler_post_cfg_function(self, post_cfg_function, disable_cfg1_optimization=False):
         self.model_options = set_model_options_post_cfg_function(self.model_options, post_cfg_function, disable_cfg1_optimization)
@@ -671,18 +692,19 @@ class ModelPatcher:
         for key in list(self.pinned):
             self.unpin_weight(key)
 
-    def _load_list(self, prio_comfy_cast_weights=False):
+    def _load_list(self, prio_comfy_cast_weights=False, default_device=None):
         loading = []
         for n, m in self.model.named_modules():
-            params = []
-            skip = False
-            for name, param in m.named_parameters(recurse=False):
-                params.append(name)
+            default = False
+            params = { name: param for name, param in m.named_parameters(recurse=False) }
             for name, param in m.named_parameters(recurse=True):
                 if name not in params:
-                    skip = True # skip random weights in non leaf modules
+                    default = True # default random weights in non leaf modules
                     break
-            if not skip and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
+            if default and default_device is not None:
+                for param in params.values():
+                    param.data = param.data.to(device=default_device)
+            if not default and (hasattr(m, "comfy_cast_weights") or len(params) > 0):
                 module_mem = comfy.model_management.module_size(m)
                 module_offload_mem = module_mem
                 if hasattr(m, "comfy_cast_weights"):
@@ -791,7 +813,7 @@ class ModelPatcher:
                         continue
 
                 for param in params:
-                    key = "{}.{}".format(n, param)
+                    key = key_param_name_to_key(n, param)
                     self.unpin_weight(key)
                     self.patch_weight_to_device(key, device_to=device_to)
                 if comfy.model_management.is_device_cuda(device_to):
@@ -807,7 +829,7 @@ class ModelPatcher:
                 n = x[1]
                 params = x[3]
                 for param in params:
-                    self.pin_weight_to_device("{}.{}".format(n, param))
+                    self.pin_weight_to_device(key_param_name_to_key(n, param))
 
             usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
@@ -913,7 +935,7 @@ class ModelPatcher:
                 if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
                     move_weight = True
                     for param in params:
-                        key = "{}.{}".format(n, param)
+                        key = key_param_name_to_key(n, param)
                         bk = self.backup.get(key, None)
                         if bk is not None:
                             if not lowvram_possible:
@@ -964,7 +986,7 @@ class ModelPatcher:
                         logging.debug("freed {}".format(n))
 
                         for param in params:
-                            self.pin_weight_to_device("{}.{}".format(n, param))
+                            self.pin_weight_to_device(key_param_name_to_key(n, param))
 
 
             self.model.model_lowvram = True
@@ -1391,7 +1413,7 @@ class ModelPatcher:
                 continue
             key = "diffusion_model." + k
             unet_state_dict[k] = LazyCastingParam(self, key, comfy.utils.get_attr(self.model, key))
-        return self.model.state_dict_for_saving(unet_state_dict)
+        return self.model.state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
 
     def __del__(self):
         self.unpin_all_weights()
@@ -1483,9 +1505,11 @@ class ModelPatcherDynamic(ModelPatcher):
             if vbar is not None:
                 vbar.prioritize()
 
-            #We have way more tools for acceleration on comfy weight offloading, so always
+            #We force reserve VRAM for the non comfy-weight so we dont have to deal
+            #with pin and unpin syncrhonization which can be expensive for small weights
+            #with a high layer rate (e.g. autoregressive LLMs).
             #prioritize the non-comfy weights (note the order reverse).
-            loading = self._load_list(prio_comfy_cast_weights=True)
+            loading = self._load_list(prio_comfy_cast_weights=True, default_device=device_to)
             loading.sort(reverse=True)
 
             for x in loading:
@@ -1497,14 +1521,16 @@ class ModelPatcherDynamic(ModelPatcher):
 
                 def setup_param(self, m, n, param_key):
                     nonlocal num_patches
-                    key = "{}.{}".format(n, param_key)
+                    key = key_param_name_to_key(n, param_key)
 
                     weight_function = []
 
                     weight, _, _ = get_key_weight(self.model, key)
                     if weight is None:
-                        return 0
+                        return (False, 0)
                     if key in self.patches:
+                        if comfy.lora.calculate_shape(self.patches[key], weight, key) != weight.shape:
+                            return (True, 0)
                         setattr(m, param_key + "_lowvram_function", LowVramPatch(key, self.patches))
                         num_patches += 1
                     else:
@@ -1515,10 +1541,16 @@ class ModelPatcherDynamic(ModelPatcher):
                     setattr(m, param_key + "_function", weight_function)
                     geometry = weight
                     if not isinstance(weight, QuantizedTensor):
-                        model_dtype = getattr(m, param_key + "_comfy_model_dtype", weight.dtype)
+                        model_dtype = getattr(m, param_key + "_comfy_model_dtype", None) or weight.dtype
                         weight._model_dtype = model_dtype
                         geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype)
-                    return comfy.memory_management.vram_aligned_size(geometry)
+                    return (False, comfy.memory_management.vram_aligned_size(geometry))
+
+                def force_load_param(self, param_key, device_to):
+                    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)
 
                 if hasattr(m, "comfy_cast_weights"):
                     m.comfy_cast_weights = True
@@ -1526,28 +1558,37 @@ class ModelPatcherDynamic(ModelPatcher):
                     m.seed_key = n
                     set_dirty(m, dirty)
 
-                    v_weight_size = 0
-                    v_weight_size += setup_param(self, m, n, "weight")
-                    v_weight_size += setup_param(self, m, n, "bias")
+                    force_load, v_weight_size = setup_param(self, m, n, "weight")
+                    force_load_bias, v_weight_bias = setup_param(self, m, n, "bias")
+                    force_load = force_load or force_load_bias
+                    v_weight_size += v_weight_bias
 
-                    if vbar is not None and not hasattr(m, "_v"):
-                        m._v = vbar.alloc(v_weight_size)
-                    allocated_size += v_weight_size
+                    if force_load:
+                        logging.info(f"Module {n} has resizing Lora - force loading")
+                        force_load_param(self, "weight", device_to)
+                        force_load_param(self, "bias", device_to)
+                    else:
+                        if vbar is not None and not hasattr(m, "_v"):
+                            m._v = vbar.alloc(v_weight_size)
+                        allocated_size += v_weight_size
 
                 else:
                     for param in params:
-                        key = "{}.{}".format(n, param)
+                        key = key_param_name_to_key(n, param)
                         weight, _, _ = get_key_weight(self.model, key)
                         weight.seed_key = key
                         set_dirty(weight, dirty)
                         geometry = weight
-                        model_dtype = getattr(m, param + "_comfy_model_dtype", weight.dtype)
+                        model_dtype = getattr(m, param + "_comfy_model_dtype", None) or weight.dtype
                         geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype)
                         weight_size = geometry.numel() * geometry.element_size()
                         if vbar is not None and not hasattr(weight, "_v"):
                             weight._v = vbar.alloc(weight_size)
                             weight._model_dtype = model_dtype
                         allocated_size += weight_size
+                    vbar.set_watermark_limit(allocated_size)
+
+                move_weight_functions(m, device_to)
 
             logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.")
 
@@ -1568,7 +1609,7 @@ class ModelPatcherDynamic(ModelPatcher):
         return 0 if vbar is None else vbar.free_memory(memory_to_free)
 
     def partially_unload_ram(self, ram_to_unload):
-        loading = self._load_list(prio_comfy_cast_weights=True)
+        loading = self._load_list(prio_comfy_cast_weights=True, default_device=self.offload_device)
         for x in loading:
             _, _, _, _, m, _ = x
             ram_to_unload -= comfy.pinned_memory.unpin_memory(m)
@@ -1588,7 +1629,14 @@ class ModelPatcherDynamic(ModelPatcher):
 
         if unpatch_weights:
             self.partially_unload_ram(1e32)
-            self.partially_unload(None)
+            self.partially_unload(None, 1e32)
+            for m in self.model.modules():
+                move_weight_functions(m, device_to)
+
+            keys = list(self.backup.keys())
+            for k in keys:
+                bk = self.backup[k]
+                comfy.utils.set_attr_param(self.model, k, bk.weight)
 
     def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
         assert not force_patch_weights #See above

comfy/model_sampling.py

@@ -83,6 +83,16 @@ class IMG_TO_IMG(X0):
     def calculate_input(self, sigma, noise):
         return noise
 
+class IMG_TO_IMG_FLOW(CONST):
+    def calculate_denoised(self, sigma, model_output, model_input):
+        return model_output
+
+    def noise_scaling(self, sigma, noise, latent_image, max_denoise=False):
+        return latent_image
+
+    def inverse_noise_scaling(self, sigma, latent):
+        return 1.0 - latent
+
 class COSMOS_RFLOW:
     def calculate_input(self, sigma, noise):
         sigma = (sigma / (sigma + 1))

comfy/ops.py

@@ -19,9 +19,8 @@
 import torch
 import logging
 import comfy.model_management
-from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
+from comfy.cli_args import args, PerformanceFeature
 import comfy.float
-import comfy.rmsnorm
 import json
 import comfy.memory_management
 import comfy.pinned_memory
@@ -54,6 +53,8 @@ try:
             SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
 
             def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+                if q.nelement() < 1024 * 128:  # arbitrary number, for small inputs cudnn attention seems slower
+                    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
                 with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True):
                     return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
         else:
@@ -78,17 +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):
+def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant):
     offload_stream = None
     xfer_dest = None
-    cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
 
     signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
-    if signature is not None:
-        xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
     resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
+    if signature is not None:
+        if resident:
+            weight = s._v_weight
+            bias = s._v_bias
+        else:
+            xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
 
     if not resident:
+        cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
         cast_dest = None
 
         xfer_source = [ s.weight, s.bias ]
@@ -96,16 +101,16 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
         pin = comfy.pinned_memory.get_pin(s)
         if pin is not None:
             xfer_source = [ pin ]
-        else:
-            for data, geometry in zip([ s.weight, s.bias ], cast_geometry):
-                if data is None:
-                    continue
-                if data.dtype != geometry.dtype:
-                    cast_dest = xfer_dest
-                    if cast_dest is None:
-                        cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device)
-                    xfer_dest = None
-                    break
+
+        for data, geometry in zip([ s.weight, s.bias ], cast_geometry):
+            if data is None:
+                continue
+            if data.dtype != geometry.dtype:
+                cast_dest = xfer_dest
+                if cast_dest is None:
+                    cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device)
+                xfer_dest = None
+                break
 
         dest_size = comfy.memory_management.vram_aligned_size(xfer_source)
         offload_stream = comfy.model_management.get_offload_stream(device)
@@ -132,15 +137,19 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
         comfy.model_management.sync_stream(device, offload_stream)
 
         if cast_dest is not None:
-            for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like(xfer_source, xfer_dest),
+            for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like([s.weight, s.bias ], xfer_dest),
                                            comfy.memory_management.interpret_gathered_like(cast_geometry, cast_dest)):
                 if post_cast is not None:
                     post_cast.copy_(pre_cast)
             xfer_dest = cast_dest
 
-    params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest)
-    weight = params[0]
-    bias = params[1]
+        params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest)
+        weight = params[0]
+        bias = params[1]
+        if signature is not None:
+            s._v_weight = weight
+            s._v_bias = bias
+        s._v_signature=signature
 
     def post_cast(s, param_key, x, dtype, resident, update_weight):
         lowvram_fn = getattr(s, param_key + "_lowvram_function", None)
@@ -161,14 +170,14 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
             #FIXME: this is not accurate, we need to be sensitive to the compute dtype
             x = lowvram_fn(x)
             if (isinstance(orig, QuantizedTensor) and
-                (orig.dtype == dtype and len(fns) == 0 or update_weight)):
+                (want_requant and len(fns) == 0 or update_weight)):
                 seed = comfy.utils.string_to_seed(s.seed_key)
                 y = QuantizedTensor.from_float(x, s.layout_type, scale="recalculate", stochastic_rounding=seed)
-                if orig.dtype == dtype and len(fns) == 0:
+                if want_requant and len(fns) == 0:
                     #The layer actually wants our freshly saved QT
                     x = y
-            else:
-                y = x
+            elif update_weight:
+                y = comfy.float.stochastic_rounding(x, orig.dtype, seed = comfy.utils.string_to_seed(s.seed_key))
             if update_weight:
                 orig.copy_(y)
         for f in fns:
@@ -180,13 +189,12 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
     weight = post_cast(s, "weight", weight, dtype, resident, update_weight)
     if s.bias is not None:
         bias = post_cast(s, "bias", bias, bias_dtype, resident, update_weight)
-    s._v_signature=signature
 
     #FIXME: weird offload return protocol
     return weight, bias, (offload_stream, device if signature is not None else None, None)
 
 
-def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False, compute_dtype=None):
+def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False, compute_dtype=None, want_requant=False):
     # NOTE: offloadable=False is a a legacy and if you are a custom node author reading this please pass
     # offloadable=True and call uncast_bias_weight() after your last usage of the weight/bias. This
     # will add async-offload support to your cast and improve performance.
@@ -204,7 +212,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
     non_blocking = comfy.model_management.device_supports_non_blocking(device)
 
     if hasattr(s, "_v"):
-        return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype)
+        return cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype, want_requant)
 
     if offloadable and (device != s.weight.device or
                         (s.bias is not None and device != s.bias.device)):
@@ -288,7 +296,7 @@ class disable_weight_init:
     class Linear(torch.nn.Linear, CastWeightBiasOp):
 
         def __init__(self, in_features, out_features, bias=True, device=None, dtype=None):
-            if not comfy.model_management.WINDOWS or not enables_dynamic_vram():
+            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
                 super().__init__(in_features, out_features, bias, device, dtype)
                 return
 
@@ -309,7 +317,7 @@ class disable_weight_init:
         def _load_from_state_dict(self, state_dict, prefix, local_metadata,
                                 strict, missing_keys, unexpected_keys, error_msgs):
 
-            if not comfy.model_management.WINDOWS or not enables_dynamic_vram():
+            if not comfy.model_management.WINDOWS or not comfy.memory_management.aimdo_enabled:
                 return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
                                                      missing_keys, unexpected_keys, error_msgs)
             assign_to_params_buffers = local_metadata.get("assign_to_params_buffers", False)
@@ -454,7 +462,7 @@ class disable_weight_init:
             else:
                 return super().forward(*args, **kwargs)
 
-    class RMSNorm(comfy.rmsnorm.RMSNorm, CastWeightBiasOp):
+    class RMSNorm(torch.nn.RMSNorm, CastWeightBiasOp):
         def reset_parameters(self):
             self.bias = None
             return None
@@ -466,8 +474,7 @@ class disable_weight_init:
                 weight = None
                 bias = None
                 offload_stream = None
-            x = comfy.rmsnorm.rms_norm(input, weight, self.eps)  # TODO: switch to commented out line when old torch is deprecated
-            # x = torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
+            x = torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
             uncast_bias_weight(self, weight, bias, offload_stream)
             return x
 
@@ -820,6 +827,10 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                 else:
                     sd = {}
 
+                if not hasattr(self, 'weight'):
+                    logging.warning("Warning: state dict on uninitialized op {}".format(prefix))
+                    return sd
+
                 if self.bias is not None:
                     sd["{}bias".format(prefix)] = self.bias
 
@@ -843,8 +854,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
             def _forward(self, input, weight, bias):
                 return torch.nn.functional.linear(input, weight, bias)
 
-            def forward_comfy_cast_weights(self, input, compute_dtype=None):
-                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True, compute_dtype=compute_dtype)
+            def forward_comfy_cast_weights(self, input, compute_dtype=None, want_requant=False):
+                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True, compute_dtype=compute_dtype, want_requant=want_requant)
                 x = self._forward(input, weight, bias)
                 uncast_bias_weight(self, weight, bias, offload_stream)
                 return x
@@ -874,8 +885,7 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                             scale = comfy.model_management.cast_to_device(scale, input.device, None)
                         input = QuantizedTensor.from_float(input_reshaped, self.layout_type, scale=scale)
 
-
-                output = self.forward_comfy_cast_weights(input, compute_dtype)
+                output = self.forward_comfy_cast_weights(input, compute_dtype, want_requant=isinstance(input, QuantizedTensor))
 
                 # Reshape output back to 3D if input was 3D
                 if reshaped_3d:

comfy/pinned_memory.py

@@ -11,8 +11,7 @@ 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
-    params = comfy.memory_management.tensors_to_geometries([ module.weight, module.bias ])
-    size = comfy.memory_management.vram_aligned_size(params)
+    size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ])
     pin = torch.empty((size,), dtype=torch.uint8)
     if comfy.model_management.pin_memory(pin):
         module._pin = pin

comfy/rmsnorm.py

@@ -1,57 +1,10 @@
 import torch
 import comfy.model_management
-import numbers
-import logging
-
-RMSNorm = None
-
-try:
-    rms_norm_torch = torch.nn.functional.rms_norm
-    RMSNorm = torch.nn.RMSNorm
-except:
-    rms_norm_torch = None
-    logging.warning("Please update pytorch to use native RMSNorm")
 
+RMSNorm = torch.nn.RMSNorm
 
 def rms_norm(x, weight=None, eps=1e-6):
-    if rms_norm_torch is not None and not (torch.jit.is_tracing() or torch.jit.is_scripting()):
-        if weight is None:
-            return rms_norm_torch(x, (x.shape[-1],), eps=eps)
-        else:
-            return rms_norm_torch(x, weight.shape, weight=comfy.model_management.cast_to(weight, dtype=x.dtype, device=x.device), eps=eps)
+    if weight is None:
+        return torch.nn.functional.rms_norm(x, (x.shape[-1],), eps=eps)
     else:
-        r = x * torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + eps)
-        if weight is None:
-            return r
-        else:
-            return r * comfy.model_management.cast_to(weight, dtype=x.dtype, device=x.device)
-
-
-if RMSNorm is None:
-    class RMSNorm(torch.nn.Module):
-        def __init__(
-            self,
-            normalized_shape,
-            eps=1e-6,
-            elementwise_affine=True,
-            device=None,
-            dtype=None,
-        ):
-            factory_kwargs = {"device": device, "dtype": dtype}
-            super().__init__()
-            if isinstance(normalized_shape, numbers.Integral):
-                # mypy error: incompatible types in assignment
-                normalized_shape = (normalized_shape,)  # type: ignore[assignment]
-            self.normalized_shape = tuple(normalized_shape)  # type: ignore[arg-type]
-            self.eps = eps
-            self.elementwise_affine = elementwise_affine
-            if self.elementwise_affine:
-                self.weight = torch.nn.Parameter(
-                    torch.empty(self.normalized_shape, **factory_kwargs)
-                )
-            else:
-                self.register_parameter("weight", None)
-            self.bias = None
-
-        def forward(self, x):
-            return rms_norm(x, self.weight, self.eps)
+        return torch.nn.functional.rms_norm(x, weight.shape, weight=comfy.model_management.cast_to(weight, dtype=x.dtype, device=x.device), eps=eps)

comfy/sampler_helpers.py

@@ -122,20 +122,26 @@ def estimate_memory(model, noise_shape, conds):
     minimum_memory_required = model.model.memory_required([noise_shape[0]] + list(noise_shape[1:]), cond_shapes=cond_shapes_min)
     return memory_required, minimum_memory_required
 
-def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False):
+def prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False):
     executor = comfy.patcher_extension.WrapperExecutor.new_executor(
         _prepare_sampling,
         comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.PREPARE_SAMPLING, model_options, is_model_options=True)
     )
-    return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load)
+    return executor.execute(model, noise_shape, conds, model_options=model_options, force_full_load=force_full_load, force_offload=force_offload)
 
-def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False):
+def _prepare_sampling(model: ModelPatcher, noise_shape, conds, model_options=None, force_full_load=False, force_offload=False):
     real_model: BaseModel = None
     models, inference_memory = get_additional_models(conds, model.model_dtype())
     models += get_additional_models_from_model_options(model_options)
     models += model.get_nested_additional_models()  # TODO: does this require inference_memory update?
-    memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
-    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required + inference_memory, minimum_memory_required=minimum_memory_required + inference_memory, force_full_load=force_full_load)
+    if force_offload: # In training + offload enabled, we want to force prepare sampling to trigger partial load
+        memory_required = 1e20
+        minimum_memory_required = None
+    else:
+        memory_required, minimum_memory_required = estimate_memory(model, noise_shape, conds)
+        memory_required += inference_memory
+        minimum_memory_required += inference_memory
+    comfy.model_management.load_models_gpu([model] + models, memory_required=memory_required, minimum_memory_required=minimum_memory_required, force_full_load=force_full_load)
     real_model = model.model
 
     return real_model, conds, models

comfy/sd.py

@@ -59,6 +59,7 @@ import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.jina_clip_2
 import comfy.text_encoders.newbie
 import comfy.text_encoders.anima
+import comfy.text_encoders.ace15
 
 import comfy.model_patcher
 import comfy.lora
@@ -422,6 +423,17 @@ class CLIP:
     def get_key_patches(self):
         return self.patcher.get_key_patches()
 
+    def generate(self, tokens, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.95, min_p=0.0, repetition_penalty=1.0, seed=None):
+        self.cond_stage_model.reset_clip_options()
+
+        self.load_model()
+        self.cond_stage_model.set_clip_options({"layer": None})
+        self.cond_stage_model.set_clip_options({"execution_device": self.patcher.load_device})
+        return self.cond_stage_model.generate(tokens, do_sample=do_sample, max_length=max_length, temperature=temperature, top_k=top_k, top_p=top_p, min_p=min_p, repetition_penalty=repetition_penalty, seed=seed)
+
+    def decode(self, token_ids, skip_special_tokens=True):
+        return self.tokenizer.decode(token_ids, skip_special_tokens=skip_special_tokens)
+
 class VAE:
     def __init__(self, sd=None, device=None, config=None, dtype=None, metadata=None):
         if 'decoder.up_blocks.0.resnets.0.norm1.weight' in sd.keys(): #diffusers format
@@ -452,6 +464,8 @@ class VAE:
         self.extra_1d_channel = None
         self.crop_input = True
 
+        self.audio_sample_rate = 44100
+
         if config is None:
             if "decoder.mid.block_1.mix_factor" in sd:
                 encoder_config = {'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}
@@ -549,14 +563,27 @@ class VAE:
                                                                     encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig},
                                                                     decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig})
             elif "decoder.layers.1.layers.0.beta" in sd:
-                self.first_stage_model = AudioOobleckVAE()
+                config = {}
+                param_key = None
+                self.upscale_ratio = 2048
+                self.downscale_ratio = 2048
+                if "decoder.layers.2.layers.1.weight_v" in sd:
+                    param_key = "decoder.layers.2.layers.1.weight_v"
+                if "decoder.layers.2.layers.1.parametrizations.weight.original1" in sd:
+                    param_key = "decoder.layers.2.layers.1.parametrizations.weight.original1"
+                if param_key is not None:
+                    if sd[param_key].shape[-1] == 12:
+                        config["strides"] = [2, 4, 4, 6, 10]
+                        self.audio_sample_rate = 48000
+                        self.upscale_ratio = 1920
+                        self.downscale_ratio = 1920
+
+                self.first_stage_model = AudioOobleckVAE(**config)
                 self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype)
                 self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * 2048) * model_management.dtype_size(dtype)
                 self.latent_channels = 64
                 self.output_channels = 2
                 self.pad_channel_value = "replicate"
-                self.upscale_ratio = 2048
-                self.downscale_ratio =  2048
                 self.latent_dim = 1
                 self.process_output = lambda audio: audio
                 self.process_input = lambda audio: audio
@@ -667,8 +694,9 @@ class VAE:
                     self.latent_dim = 3
                     self.latent_channels = 16
                     self.output_channels = sd["encoder.conv1.weight"].shape[1]
+                    self.conv_out_channels = sd["decoder.head.2.weight"].shape[0]
                     self.pad_channel_value = 1.0
-                    ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "image_channels": self.output_channels, "dropout": 0.0}
+                    ddconfig = {"dim": dim, "z_dim": self.latent_channels, "dim_mult": [1, 2, 4, 4], "num_res_blocks": 2, "attn_scales": [], "temperal_downsample": [False, True, True], "image_channels": self.output_channels, "conv_out_channels": self.conv_out_channels, "dropout": 0.0}
                     self.first_stage_model = comfy.ldm.wan.vae.WanVAE(**ddconfig)
                     self.working_dtypes = [torch.bfloat16, torch.float16, torch.float32]
                     self.memory_used_encode = lambda shape, dtype: (1500 if shape[2]<=4 else 6000) * shape[3] * shape[4] * model_management.dtype_size(dtype)
@@ -777,8 +805,6 @@ class VAE:
             self.first_stage_model = AutoencoderKL(**(config['params']))
         self.first_stage_model = self.first_stage_model.eval()
 
-        model_management.archive_model_dtypes(self.first_stage_model)
-
         if device is None:
             device = model_management.vae_device()
         self.device = device
@@ -787,6 +813,7 @@ class VAE:
             dtype = model_management.vae_dtype(self.device, self.working_dtypes)
         self.vae_dtype = dtype
         self.first_stage_model.to(self.vae_dtype)
+        model_management.archive_model_dtypes(self.first_stage_model)
         self.output_device = model_management.intermediate_device()
 
         mp = comfy.model_patcher.CoreModelPatcher
@@ -856,7 +883,7 @@ class VAE:
             / 3.0)
         return output
 
-    def decode_tiled_1d(self, samples, tile_x=128, overlap=32):
+    def decode_tiled_1d(self, samples, tile_x=256, overlap=32):
         if samples.ndim == 3:
             decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
         else:
@@ -960,7 +987,7 @@ class VAE:
         if overlap is not None:
             args["overlap"] = overlap
 
-        if dims == 1:
+        if dims == 1 or self.extra_1d_channel is not None:
             args.pop("tile_y")
             output = self.decode_tiled_1d(samples, **args)
         elif dims == 2:
@@ -1167,6 +1194,7 @@ class TEModel(Enum):
     JINA_CLIP_2 = 19
     QWEN3_8B = 20
     QWEN3_06B = 21
+    GEMMA_3_4B_VISION = 22
 
 
 def detect_te_model(sd):
@@ -1195,7 +1223,10 @@ def detect_te_model(sd):
         if 'model.layers.47.self_attn.q_norm.weight' in sd:
             return TEModel.GEMMA_3_12B
         if 'model.layers.0.self_attn.q_norm.weight' in sd:
-            return TEModel.GEMMA_3_4B
+            if 'vision_model.embeddings.patch_embedding.weight' in sd:
+                return TEModel.GEMMA_3_4B_VISION
+            else:
+                return TEModel.GEMMA_3_4B
         return TEModel.GEMMA_2_2B
     if 'model.layers.0.self_attn.k_proj.bias' in sd:
         weight = sd['model.layers.0.self_attn.k_proj.bias']
@@ -1255,6 +1286,8 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         else:
             if "text_projection" in clip_data[i]:
                 clip_data[i]["text_projection.weight"] = clip_data[i]["text_projection"].transpose(0, 1) #old models saved with the CLIPSave node
+        if "lm_head.weight" in clip_data[i]:
+            clip_data[i]["model.lm_head.weight"] = clip_data[i].pop("lm_head.weight") # prefix missing in some models
 
     tokenizer_data = {}
     clip_target = EmptyClass()
@@ -1320,6 +1353,14 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data), model_type="gemma3_4b")
             clip_target.tokenizer = comfy.text_encoders.lumina2.NTokenizer
             tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
+        elif te_model == TEModel.GEMMA_3_4B_VISION:
+            clip_target.clip = comfy.text_encoders.lumina2.te(**llama_detect(clip_data), model_type="gemma3_4b_vision")
+            clip_target.tokenizer = comfy.text_encoders.lumina2.NTokenizer
+            tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
+        elif te_model == TEModel.GEMMA_3_12B:
+            clip_target.clip = comfy.text_encoders.lt.gemma3_te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.lt.Gemma3_12BTokenizer
+            tokenizer_data["spiece_model"] = clip_data[0].get("spiece_model", None)
         elif te_model == TEModel.LLAMA3_8:
             clip_target.clip = comfy.text_encoders.hidream.hidream_clip(**llama_detect(clip_data),
                                                                         clip_l=False, clip_g=False, t5=False, llama=True, dtype_t5=None)
@@ -1427,6 +1468,14 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
                 clip_data_jina = clip_data[0]
             tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None)
             tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None)
+        elif clip_type == CLIPType.ACE:
+            te_models = [detect_te_model(clip_data[0]), detect_te_model(clip_data[1])]
+            if TEModel.QWEN3_4B in te_models:
+                model_type = "qwen3_4b"
+            else:
+                model_type = "qwen3_2b"
+            clip_target.clip = comfy.text_encoders.ace15.te(lm_model=model_type, **llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.ace15.ACE15Tokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer
@@ -1482,14 +1531,24 @@ def load_checkpoint(config_path=None, ckpt_path=None, output_vae=True, output_cl
 
     return (model, clip, vae)
 
-def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}):
+def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}, disable_dynamic=False):
     sd, metadata = comfy.utils.load_torch_file(ckpt_path, return_metadata=True)
-    out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options, metadata=metadata)
+    out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options, metadata=metadata, disable_dynamic=disable_dynamic)
     if out is None:
         raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(ckpt_path, model_detection_error_hint(ckpt_path, sd)))
+    if output_model:
+        out[0].cached_patcher_init = (load_checkpoint_guess_config_model_only, (ckpt_path, embedding_directory, model_options, te_model_options))
     return out
 
-def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}, metadata=None):
+def load_checkpoint_guess_config_model_only(ckpt_path, embedding_directory=None, model_options={}, te_model_options={}, disable_dynamic=False):
+    model, *_ = load_checkpoint_guess_config(ckpt_path, False, False, False,
+            embedding_directory=embedding_directory,
+            model_options=model_options,
+            te_model_options=te_model_options,
+            disable_dynamic=disable_dynamic)
+    return model
+
+def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}, metadata=None, disable_dynamic=False):
     clip = None
     clipvision = None
     vae = None
@@ -1538,7 +1597,8 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
     if output_model:
         inital_load_device = model_management.unet_inital_load_device(parameters, unet_dtype)
         model = model_config.get_model(sd, diffusion_model_prefix, device=inital_load_device)
-        model_patcher = comfy.model_patcher.CoreModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device())
+        ModelPatcher = comfy.model_patcher.ModelPatcher if disable_dynamic else comfy.model_patcher.CoreModelPatcher
+        model_patcher = ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device())
         model.load_model_weights(sd, diffusion_model_prefix, assign=model_patcher.is_dynamic())
 
     if output_vae:
@@ -1589,7 +1649,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
     return (model_patcher, clip, vae, clipvision)
 
 
-def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
+def load_diffusion_model_state_dict(sd, model_options={}, metadata=None, disable_dynamic=False):
     """
     Loads a UNet diffusion model from a state dictionary, supporting both diffusers and regular formats.
 
@@ -1673,7 +1733,8 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
         model_config.optimizations["fp8"] = True
 
     model = model_config.get_model(new_sd, "")
-    model_patcher = comfy.model_patcher.CoreModelPatcher(model, load_device=load_device, offload_device=offload_device)
+    ModelPatcher = comfy.model_patcher.ModelPatcher if disable_dynamic else comfy.model_patcher.CoreModelPatcher
+    model_patcher = ModelPatcher(model, load_device=load_device, offload_device=offload_device)
     if not model_management.is_device_cpu(offload_device):
         model.to(offload_device)
     model.load_model_weights(new_sd, "", assign=model_patcher.is_dynamic())
@@ -1682,12 +1743,13 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
         logging.info("left over keys in diffusion model: {}".format(left_over))
     return model_patcher
 
-def load_diffusion_model(unet_path, model_options={}):
+def load_diffusion_model(unet_path, model_options={}, disable_dynamic=False):
     sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True)
-    model = load_diffusion_model_state_dict(sd, model_options=model_options, metadata=metadata)
+    model = load_diffusion_model_state_dict(sd, model_options=model_options, metadata=metadata, disable_dynamic=disable_dynamic)
     if model is None:
         logging.error("ERROR UNSUPPORTED DIFFUSION MODEL {}".format(unet_path))
         raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(unet_path, model_detection_error_hint(unet_path, sd)))
+    model.cached_patcher_init = (load_diffusion_model, (unet_path, model_options))
     return model
 
 def load_unet(unet_path, dtype=None):

comfy/sd1_clip.py

@@ -155,6 +155,8 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         self.execution_device = options.get("execution_device", self.execution_device)
         if isinstance(self.layer, list) or self.layer == "all":
             pass
+        elif isinstance(layer_idx, list):
+            self.layer = layer_idx
         elif layer_idx is None or abs(layer_idx) > self.num_layers:
             self.layer = "last"
         else:
@@ -169,8 +171,9 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
 
     def process_tokens(self, tokens, device):
         end_token = self.special_tokens.get("end", None)
+        pad_token = self.special_tokens.get("pad", -1)
         if end_token is None:
-            cmp_token = self.special_tokens.get("pad", -1)
+            cmp_token = pad_token
         else:
             cmp_token = end_token
 
@@ -184,15 +187,21 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
             other_embeds = []
             eos = False
             index = 0
+            left_pad = False
             for y in x:
                 if isinstance(y, numbers.Integral):
-                    if eos:
+                    token = int(y)
+                    if index == 0 and token == pad_token:
+                        left_pad = True
+
+                    if eos or (left_pad and token == pad_token):
                         attention_mask.append(0)
                     else:
                         attention_mask.append(1)
-                    token = int(y)
+                        left_pad = False
+
                     tokens_temp += [token]
-                    if not eos and token == cmp_token:
+                    if not eos and token == cmp_token and not left_pad:
                         if end_token is None:
                             attention_mask[-1] = 0
                         eos = True
@@ -299,6 +308,15 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
     def load_sd(self, sd):
         return self.transformer.load_state_dict(sd, strict=False, assign=getattr(self, "can_assign_sd", False))
 
+    def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed):
+        if isinstance(tokens, dict):
+            tokens_only = next(iter(tokens.values())) # todo: get this better?
+        else:
+            tokens_only = tokens
+        tokens_only = [[t[0] for t in b] for b in tokens_only]
+        embeds = self.process_tokens(tokens_only, device=self.execution_device)[0]
+        return self.transformer.generate(embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed)
+
 def parse_parentheses(string):
     result = []
     current_item = ""
@@ -555,6 +573,8 @@ class SDTokenizer:
         min_length = tokenizer_options.get("{}_min_length".format(self.embedding_key), self.min_length)
         min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding)
 
+        min_length = kwargs.get("min_length", min_length)
+
         text = escape_important(text)
         if kwargs.get("disable_weights", self.disable_weights):
             parsed_weights = [(text, 1.0)]
@@ -654,6 +674,9 @@ class SDTokenizer:
     def state_dict(self):
         return {}
 
+    def decode(self, token_ids, skip_special_tokens=True):
+        return self.tokenizer.decode(token_ids, skip_special_tokens=skip_special_tokens)
+
 class SD1Tokenizer:
     def __init__(self, embedding_directory=None, tokenizer_data={}, clip_name="l", tokenizer=SDTokenizer, name=None):
         if name is not None:
@@ -677,6 +700,9 @@ class SD1Tokenizer:
     def state_dict(self):
         return getattr(self, self.clip).state_dict()
 
+    def decode(self, token_ids, skip_special_tokens=True):
+        return getattr(self, self.clip).decode(token_ids, skip_special_tokens=skip_special_tokens)
+
 class SD1CheckpointClipModel(SDClipModel):
     def __init__(self, device="cpu", dtype=None, model_options={}):
         super().__init__(device=device, return_projected_pooled=False, dtype=dtype, model_options=model_options)
@@ -713,3 +739,6 @@ class SD1ClipModel(torch.nn.Module):
 
     def load_sd(self, sd):
         return getattr(self, self.clip).load_sd(sd)
+
+    def generate(self, tokens, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.95, min_p=0.0, repetition_penalty=1.0, seed=None):
+        return getattr(self, self.clip).generate(tokens, do_sample=do_sample, max_length=max_length, temperature=temperature, top_k=top_k, top_p=top_p, min_p=min_p, repetition_penalty=repetition_penalty, seed=seed)

comfy/supported_models.py

@@ -24,6 +24,7 @@ import comfy.text_encoders.hunyuan_image
 import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.z_image
 import comfy.text_encoders.anima
+import comfy.text_encoders.ace15
 
 from . import supported_models_base
 from . import latent_formats
@@ -524,7 +525,8 @@ class LotusD(SD20):
     }
 
     unet_extra_config = {
-        "num_classes": 'sequential'
+        "num_classes": 'sequential',
+        "num_head_channels": 64,
     }
 
     def get_model(self, state_dict, prefix="", device=None):
@@ -709,6 +711,15 @@ class Flux(supported_models_base.BASE):
 
     supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
+    def process_unet_state_dict(self, state_dict):
+        out_sd = {}
+        for k in list(state_dict.keys()):
+            key_out = k
+            if key_out.endswith("_norm.scale"):
+                key_out = "{}.weight".format(key_out[:-len(".scale")])
+            out_sd[key_out] = state_dict[k]
+        return out_sd
+
     vae_key_prefix = ["vae."]
     text_encoder_key_prefix = ["text_encoders."]
 
@@ -897,11 +908,13 @@ class HunyuanVideo(supported_models_base.BASE):
             key_out = key_out.replace("txt_in.c_embedder.linear_1.", "txt_in.c_embedder.in_layer.").replace("txt_in.c_embedder.linear_2.", "txt_in.c_embedder.out_layer.")
             key_out = key_out.replace("_mod.linear.", "_mod.lin.").replace("_attn_qkv.", "_attn.qkv.")
             key_out = key_out.replace("mlp.fc1.", "mlp.0.").replace("mlp.fc2.", "mlp.2.")
-            key_out = key_out.replace("_attn_q_norm.weight", "_attn.norm.query_norm.scale").replace("_attn_k_norm.weight", "_attn.norm.key_norm.scale")
-            key_out = key_out.replace(".q_norm.weight", ".norm.query_norm.scale").replace(".k_norm.weight", ".norm.key_norm.scale")
+            key_out = key_out.replace("_attn_q_norm.weight", "_attn.norm.query_norm.weight").replace("_attn_k_norm.weight", "_attn.norm.key_norm.weight")
+            key_out = key_out.replace(".q_norm.weight", ".norm.query_norm.weight").replace(".k_norm.weight", ".norm.key_norm.weight")
             key_out = key_out.replace("_attn_proj.", "_attn.proj.")
             key_out = key_out.replace(".modulation.linear.", ".modulation.lin.")
             key_out = key_out.replace("_in.mlp.2.", "_in.out_layer.").replace("_in.mlp.0.", "_in.in_layer.")
+            if key_out.endswith(".scale"):
+                key_out = "{}.weight".format(key_out[:-len(".scale")])
             out_sd[key_out] = state_dict[k]
         return out_sd
 
@@ -992,7 +1005,7 @@ class CosmosT2IPredict2(supported_models_base.BASE):
 
     memory_usage_factor = 1.0
 
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
@@ -1022,11 +1035,7 @@ class Anima(supported_models_base.BASE):
 
     memory_usage_factor = 1.0
 
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
-
-    def __init__(self, unet_config):
-        super().__init__(unet_config)
-        self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.95
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.Anima(self, device=device)
@@ -1037,6 +1046,12 @@ class Anima(supported_models_base.BASE):
         detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_06b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.anima.AnimaTokenizer, comfy.text_encoders.anima.te(**detect))
 
+    def set_inference_dtype(self, dtype, manual_cast_dtype, **kwargs):
+        self.memory_usage_factor = (self.unet_config.get("model_channels", 2048) / 2048) * 0.95
+        if dtype is torch.float16:
+            self.memory_usage_factor *= 1.4
+        return super().set_inference_dtype(dtype, manual_cast_dtype, **kwargs)
+
 class CosmosI2VPredict2(CosmosT2IPredict2):
     unet_config = {
         "image_model": "cosmos_predict2",
@@ -1242,6 +1257,16 @@ class WAN22_T2V(WAN21_T2V):
         out = model_base.WAN22(self, image_to_video=True, device=device)
         return out
 
+class WAN21_FlowRVS(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "flow_rvs",
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21_FlowRVS(self, image_to_video=True, device=device)
+        return out
+
 class Hunyuan3Dv2(supported_models_base.BASE):
     unet_config = {
         "image_model": "hunyuan3d2",
@@ -1261,6 +1286,15 @@ class Hunyuan3Dv2(supported_models_base.BASE):
 
     latent_format = latent_formats.Hunyuan3Dv2
 
+    def process_unet_state_dict(self, state_dict):
+        out_sd = {}
+        for k in list(state_dict.keys()):
+            key_out = k
+            if key_out.endswith(".scale"):
+                key_out = "{}.weight".format(key_out[:-len(".scale")])
+            out_sd[key_out] = state_dict[k]
+        return out_sd
+
     def process_unet_state_dict_for_saving(self, state_dict):
         replace_prefix = {"": "model."}
         return utils.state_dict_prefix_replace(state_dict, replace_prefix)
@@ -1338,6 +1372,14 @@ class Chroma(supported_models_base.BASE):
 
     supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
+    def process_unet_state_dict(self, state_dict):
+        out_sd = {}
+        for k in list(state_dict.keys()):
+            key_out = k
+            if key_out.endswith(".scale"):
+                key_out = "{}.weight".format(key_out[:-len(".scale")])
+            out_sd[key_out] = state_dict[k]
+        return out_sd
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.Chroma(self, device=device)
@@ -1596,6 +1638,46 @@ class Kandinsky5Image(Kandinsky5):
         return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
+class ACEStep15(supported_models_base.BASE):
+    unet_config = {
+        "audio_model": "ace1.5",
+    }
+
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 3.0,
+    }
+
+    latent_format = comfy.latent_formats.ACEAudio15
+
+    memory_usage_factor = 4.7
+
+    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.ACEStep15(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        detect_2b = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref))
+        detect_4b = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref))
+        if "dtype_llama" in detect_2b:
+            detect = detect_2b
+            detect["lm_model"] = "qwen3_2b"
+        elif "dtype_llama" in detect_4b:
+            detect = detect_4b
+            detect["lm_model"] = "qwen3_4b"
+
+        return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**detect))
+
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, WAN21_FlowRVS, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
 
 models += [SVD_img2vid]

comfy/text_encoders/ace15.py

@@ -0,0 +1,348 @@
+from .anima import Qwen3Tokenizer
+import comfy.text_encoders.llama
+from comfy import sd1_clip
+import torch
+import math
+import yaml
+import comfy.utils
+
+
+def sample_manual_loop_no_classes(
+    model,
+    ids=None,
+    execution_dtype=None,
+    cfg_scale: float = 2.0,
+    temperature: float = 0.85,
+    top_p: float = 0.9,
+    top_k: int = None,
+    min_p: float = 0.000,
+    seed: int = 1,
+    min_tokens: int = 1,
+    max_new_tokens: int = 2048,
+    audio_start_id: int = 151669,  # The cutoff ID for audio codes
+    audio_end_id: int = 215669,
+    eos_token_id: int = 151645,
+):
+    if ids is None:
+        return []
+    device = model.execution_device
+
+    if execution_dtype is None:
+        if comfy.model_management.should_use_bf16(device):
+            execution_dtype = torch.bfloat16
+        else:
+            execution_dtype = torch.float32
+
+    embeds, attention_mask, num_tokens, embeds_info = model.process_tokens(ids, device)
+    embeds_batch = embeds.shape[0]
+
+    output_audio_codes = []
+    past_key_values = []
+    generator = torch.Generator(device=device)
+    generator.manual_seed(seed)
+    model_config = model.transformer.model.config
+    past_kv_shape = [embeds_batch, model_config.num_key_value_heads, embeds.shape[1] + min_tokens, model_config.head_dim]
+
+    for x in range(model_config.num_hidden_layers):
+        past_key_values.append((torch.empty(past_kv_shape, device=device, dtype=execution_dtype), torch.empty(past_kv_shape, device=device, dtype=execution_dtype), 0))
+
+    progress_bar = comfy.utils.ProgressBar(max_new_tokens)
+
+    for step in comfy.utils.model_trange(max_new_tokens, desc="LM sampling"):
+        outputs = model.transformer(None, attention_mask, embeds=embeds.to(execution_dtype), num_tokens=num_tokens, intermediate_output=None, dtype=execution_dtype, embeds_info=embeds_info, past_key_values=past_key_values)
+        next_token_logits = model.transformer.logits(outputs[0])[:, -1]
+        past_key_values = outputs[2]
+
+        if cfg_scale != 1.0:
+            cond_logits = next_token_logits[0:1]
+            uncond_logits = next_token_logits[1:2]
+            cfg_logits = uncond_logits + cfg_scale * (cond_logits - uncond_logits)
+        else:
+            cfg_logits = next_token_logits[0:1]
+
+        use_eos_score = eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step
+        if use_eos_score:
+            eos_score = cfg_logits[:, eos_token_id].clone()
+
+        remove_logit_value = torch.finfo(cfg_logits.dtype).min
+        # Only generate audio tokens
+        cfg_logits[:, :audio_start_id] = remove_logit_value
+        cfg_logits[:, audio_end_id:] = remove_logit_value
+
+        if use_eos_score:
+            cfg_logits[:, eos_token_id] = eos_score
+
+        if top_k is not None and top_k > 0:
+            top_k_vals, _ = torch.topk(cfg_logits, top_k)
+            min_val = top_k_vals[..., -1, None]
+            cfg_logits[cfg_logits < min_val] = remove_logit_value
+
+        if min_p is not None and min_p > 0:
+            probs = torch.softmax(cfg_logits, dim=-1)
+            p_max = probs.max(dim=-1, keepdim=True).values
+            indices_to_remove = probs < (min_p * p_max)
+            cfg_logits[indices_to_remove] = remove_logit_value
+
+        if top_p is not None and top_p < 1.0:
+            sorted_logits, sorted_indices = torch.sort(cfg_logits, descending=True)
+            cumulative_probs = torch.cumsum(torch.softmax(sorted_logits, dim=-1), dim=-1)
+            sorted_indices_to_remove = cumulative_probs > top_p
+            sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
+            sorted_indices_to_remove[..., 0] = 0
+            indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
+            cfg_logits[indices_to_remove] = remove_logit_value
+
+        if temperature > 0:
+            cfg_logits = cfg_logits / temperature
+            next_token = torch.multinomial(torch.softmax(cfg_logits, dim=-1), num_samples=1, generator=generator).squeeze(1)
+        else:
+            next_token = torch.argmax(cfg_logits, dim=-1)
+
+        token = next_token.item()
+
+        if token == eos_token_id:
+            break
+
+        embed, _, _, _ = model.process_tokens([[token]], device)
+        embeds = embed.repeat(embeds_batch, 1, 1)
+        attention_mask = torch.cat([attention_mask, torch.ones((embeds_batch, 1), device=device, dtype=attention_mask.dtype)], dim=1)
+
+        output_audio_codes.append(token - audio_start_id)
+        progress_bar.update_absolute(step)
+
+    return output_audio_codes
+
+
+def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=1024, seed=0, cfg_scale=2.0, temperature=0.85, top_p=0.9, top_k=0, min_p=0.000):
+    positive = [[token for token, _ in inner_list] for inner_list in positive]
+    positive = positive[0]
+
+    if cfg_scale != 1.0:
+        negative = [[token for token, _ in inner_list] for inner_list in negative]
+        negative = negative[0]
+
+        neg_pad = 0
+        if len(negative) < len(positive):
+            neg_pad = (len(positive) - len(negative))
+            negative = [model.special_tokens["pad"]] * neg_pad + negative
+
+        pos_pad = 0
+        if len(negative) > len(positive):
+            pos_pad = (len(negative) - len(positive))
+            positive = [model.special_tokens["pad"]] * pos_pad + positive
+
+        ids = [positive, negative]
+    else:
+        ids = [positive]
+
+    return sample_manual_loop_no_classes(model, ids, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, min_p=min_p, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens)
+
+
+class ACE15Tokenizer(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="qwen3_06b", tokenizer=Qwen3Tokenizer)
+
+    def _metas_to_cot(self, *, return_yaml: bool = False, **kwargs) -> str:
+        user_metas = {
+            k: kwargs.pop(k)
+            for k in ("bpm", "duration", "keyscale", "timesignature")
+            if k in kwargs
+        }
+        timesignature = user_metas.get("timesignature")
+        if isinstance(timesignature, str) and timesignature.endswith("/4"):
+            user_metas["timesignature"] = timesignature[:-2]
+        user_metas = {
+            k: v if not isinstance(v, str) or not v.isdigit() else int(v)
+            for k, v in user_metas.items()
+            if v not in {"unspecified", None}
+        }
+        if len(user_metas):
+            meta_yaml = yaml.dump(user_metas, allow_unicode=True, sort_keys=True).strip()
+        else:
+            meta_yaml = ""
+        return f"<think>\n{meta_yaml}\n</think>" if not return_yaml else meta_yaml
+
+    def _metas_to_cap(self, **kwargs) -> str:
+        use_keys = ("bpm", "timesignature", "keyscale", "duration")
+        user_metas = { k: kwargs.pop(k, "N/A") for k in use_keys }
+        timesignature = user_metas.get("timesignature")
+        if isinstance(timesignature, str) and timesignature.endswith("/4"):
+            user_metas["timesignature"] = timesignature[:-2]
+        duration = user_metas["duration"]
+        if duration == "N/A":
+            user_metas["duration"] = "30 seconds"
+        elif isinstance(duration, (str, int, float)):
+            user_metas["duration"] = f"{math.ceil(float(duration))} seconds"
+        else:
+            raise TypeError("Unexpected type for duration key, must be str, int or float")
+        return "\n".join(f"- {k}: {user_metas[k]}" for k in use_keys)
+
+    def tokenize_with_weights(self, text, return_word_ids=False, **kwargs):
+        text = text.strip()
+        text_negative = kwargs.get("caption_negative", text).strip()
+        lyrics = kwargs.get("lyrics", "")
+        lyrics_negative = kwargs.get("lyrics_negative", lyrics)
+        duration = kwargs.get("duration", 120)
+        if isinstance(duration, str):
+            duration = float(duration.split(None, 1)[0])
+        language = kwargs.get("language")
+        seed = kwargs.get("seed", 0)
+
+        generate_audio_codes = kwargs.get("generate_audio_codes", True)
+        cfg_scale = kwargs.get("cfg_scale", 2.0)
+        temperature = kwargs.get("temperature", 0.85)
+        top_p = kwargs.get("top_p", 0.9)
+        top_k = kwargs.get("top_k", 0.0)
+        min_p = kwargs.get("min_p", 0.000)
+
+        duration = math.ceil(duration)
+        kwargs["duration"] = duration
+        tokens_duration = duration * 5
+        min_tokens = int(kwargs.get("min_tokens", tokens_duration))
+        max_tokens = int(kwargs.get("max_tokens", tokens_duration))
+
+        metas_negative = {
+            k.rsplit("_", 1)[0]: kwargs.pop(k)
+            for k in ("bpm_negative", "duration_negative", "keyscale_negative", "timesignature_negative", "language_negative", "caption_negative")
+            if k in kwargs
+        }
+        if not kwargs.get("use_negative_caption"):
+            _ = metas_negative.pop("caption", None)
+
+        cot_text = self._metas_to_cot(caption=text, **kwargs)
+        cot_text_negative = "<think>\n\n</think>" if not metas_negative else self._metas_to_cot(**metas_negative)
+        meta_cap = self._metas_to_cap(**kwargs)
+
+        lm_template = "<|im_start|>system\n# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n<|im_end|>\n<|im_start|>user\n# Caption\n{}\n\n# Lyric\n{}\n<|im_end|>\n<|im_start|>assistant\n{}\n\n<|im_end|>\n"
+        lyrics_template = "# Languages\n{}\n\n# Lyric\n{}<|endoftext|><|endoftext|>"
+        qwen3_06b_template = "# Instruction\nGenerate audio semantic tokens based on the given conditions:\n\n# Caption\n{}\n\n# Metas\n{}\n<|endoftext|>\n<|endoftext|>"
+
+        llm_prompts = {
+            "lm_prompt": lm_template.format(text, lyrics.strip(), cot_text),
+            "lm_prompt_negative": lm_template.format(text_negative, lyrics_negative.strip(), cot_text_negative),
+            "lyrics": lyrics_template.format(language if language is not None else "", lyrics),
+            "qwen3_06b": qwen3_06b_template.format(text, meta_cap),
+        }
+
+        out = {
+            prompt_key: self.qwen3_06b.tokenize_with_weights(
+                prompt,
+                prompt_key == "qwen3_06b" and return_word_ids,
+                disable_weights = True,
+                **kwargs,
+            )
+            for prompt_key, prompt in llm_prompts.items()
+        }
+        out["lm_metadata"] = {"min_tokens": min_tokens,
+                              "max_tokens": max_tokens,
+                              "seed": seed,
+                              "generate_audio_codes": generate_audio_codes,
+                              "cfg_scale": cfg_scale,
+                              "temperature": temperature,
+                              "top_p": top_p,
+                              "top_k": top_k,
+                              "min_p": min_p,
+                              }
+        return out
+
+
+class Qwen3_06BModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_06B_ACE15, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class Qwen3_2B_ACE15(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_2B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class Qwen3_4B_ACE15(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_4B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+class ACE15TEModel(torch.nn.Module):
+    def __init__(self, device="cpu", dtype=None, dtype_llama=None, lm_model=None, model_options={}):
+        super().__init__()
+        if dtype_llama is None:
+            dtype_llama = dtype
+
+        model = None
+        self.constant = 0.4375
+        if lm_model == "qwen3_4b":
+            model = Qwen3_4B_ACE15
+            self.constant = 0.5625
+        elif lm_model == "qwen3_2b":
+            model = Qwen3_2B_ACE15
+
+        self.lm_model = lm_model
+        self.qwen3_06b = Qwen3_06BModel(device=device, dtype=dtype, model_options=model_options)
+        if model is not None:
+            setattr(self, self.lm_model, model(device=device, dtype=dtype_llama, model_options=model_options))
+
+        self.dtypes = set([dtype, dtype_llama])
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_base = token_weight_pairs["qwen3_06b"]
+        token_weight_pairs_lyrics = token_weight_pairs["lyrics"]
+
+        self.qwen3_06b.set_clip_options({"layer": None})
+        base_out, _, extra = self.qwen3_06b.encode_token_weights(token_weight_pairs_base)
+        self.qwen3_06b.set_clip_options({"layer": [0]})
+        lyrics_embeds, _, extra_l = self.qwen3_06b.encode_token_weights(token_weight_pairs_lyrics)
+
+        out = {"conditioning_lyrics": lyrics_embeds[:, 0]}
+
+        lm_metadata = token_weight_pairs["lm_metadata"]
+        if lm_metadata["generate_audio_codes"]:
+            audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"], cfg_scale=lm_metadata["cfg_scale"], temperature=lm_metadata["temperature"], top_p=lm_metadata["top_p"], top_k=lm_metadata["top_k"], min_p=lm_metadata["min_p"])
+            out["audio_codes"] = [audio_codes]
+
+        return base_out, None, out
+
+    def set_clip_options(self, options):
+        self.qwen3_06b.set_clip_options(options)
+        lm_model = getattr(self, self.lm_model, None)
+        if lm_model is not None:
+            lm_model.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.qwen3_06b.reset_clip_options()
+        lm_model = getattr(self, self.lm_model, None)
+        if lm_model is not None:
+            lm_model.reset_clip_options()
+
+    def load_sd(self, sd):
+        if "model.layers.0.post_attention_layernorm.weight" in sd:
+            shape = sd["model.layers.0.post_attention_layernorm.weight"].shape
+            if shape[0] == 1024:
+                return self.qwen3_06b.load_sd(sd)
+            else:
+                return getattr(self, self.lm_model).load_sd(sd)
+
+    def memory_estimation_function(self, token_weight_pairs, device=None):
+        lm_metadata = token_weight_pairs["lm_metadata"]
+        constant = self.constant
+        if comfy.model_management.should_use_bf16(device):
+            constant *= 0.5
+
+        token_weight_pairs = token_weight_pairs.get("lm_prompt", [])
+        num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
+        num_tokens += lm_metadata['min_tokens']
+        return num_tokens * constant * 1024 * 1024
+
+def te(dtype_llama=None, llama_quantization_metadata=None, lm_model="qwen3_2b"):
+    class ACE15TEModel_(ACE15TEModel):
+        def __init__(self, device="cpu", dtype=None, model_options={}):
+            if llama_quantization_metadata is not None:
+                model_options = model_options.copy()
+                model_options["llama_quantization_metadata"] = llama_quantization_metadata
+            super().__init__(device=device, dtype_llama=dtype_llama, lm_model=lm_model, dtype=dtype, model_options=model_options)
+    return ACE15TEModel_

comfy/text_encoders/anima.py

@@ -8,7 +8,7 @@ import torch
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=1024, embedding_key='qwen3_06b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -23,7 +23,7 @@ class AnimaTokenizer:
     def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
         out = {}
         qwen_ids = self.qwen3_06b.tokenize_with_weights(text, return_word_ids, **kwargs)
-        out["qwen3_06b"] = [[(token, 1.0) for token, _ in inner_list] for inner_list in qwen_ids]  # Set weights to 1.0
+        out["qwen3_06b"] = [[(k[0], 1.0, k[2]) if return_word_ids else (k[0], 1.0) for k in inner_list] for inner_list in qwen_ids]  # Set weights to 1.0
         out["t5xxl"] = self.t5xxl.tokenize_with_weights(text, return_word_ids, **kwargs)
         return out
 
@@ -33,6 +33,8 @@ class AnimaTokenizer:
     def state_dict(self):
         return {}
 
+    def decode(self, token_ids, **kwargs):
+        return self.qwen3_06b.decode(token_ids, **kwargs)
 
 class Qwen3_06BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):

comfy/text_encoders/flux.py

@@ -118,7 +118,7 @@ class MistralTokenizerClass:
 class Mistral3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         self.tekken_data = tokenizer_data.get("tekken_model", None)
-        super().__init__("", pad_with_end=False, 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=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)
 
     def state_dict(self):
         return {"tekken_model": self.tekken_data}
@@ -176,12 +176,12 @@ def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False):
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class Qwen3Tokenizer8B(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class KleinTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}, name="qwen3_4b"):