Merge branch 'master' into add-coderabbit-config

Amp-Thread-ID: https://ampcode.com/threads/T-019c7915-2abf-743c-9c74-b93d87d63055

.coderabbit.yaml

@@ -0,0 +1,114 @@
+# yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json
+language: "en-US"
+early_access: false
+
+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: "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: true
+
+  finishing_touches:
+    docstrings:
+      enabled: false
+    unit_tests:
+      enabled: false
+
+  tools:
+    ruff:
+      enabled: true
+    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/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

@@ -227,7 +227,7 @@ Put your VAE in: models/vae
 
 AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version:
 
-```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/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:
 

app/node_replace_manager.py

@@ -0,0 +1,105 @@
+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():
+            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())

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/Brightness and Contrast.json

@@ -0,0 +1 @@
+{"revision":0,"last_node_id":140,"last_link_id":0,"nodes":[{"id":140,"type":"916dff42-6166-4d45-b028-04eaf69fbb35","pos":[500,1440],"size":[250,178],"flags":{},"order":2,"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":[]}],"properties":{"proxyWidgets":[["4","value"],["5","value"]]},"widgets_values":[],"title":"Brightness and Contrast"}],"links":[],"version":0.4,"definitions":{"subgraphs":[{"id":"916dff42-6166-4d45-b028-04eaf69fbb35","version":1,"state":{"lastGroupId":0,"lastNodeId":143,"lastLinkId":118,"lastRerouteId":0},"revision":0,"config":{},"name":"Brightness and Contrast","inputNode":{"id":-10,"bounding":[360,-176,120,60]},"outputNode":{"id":-20,"bounding":[1410,-176,120,60]},"inputs":[{"id":"a5aae7ea-b511-4045-b5da-94101e269cd7","name":"images.image0","type":"IMAGE","linkIds":[117],"localized_name":"images.image0","label":"image","pos":[460,-156]}],"outputs":[{"id":"30b72604-69b3-4944-b253-a9099bbd73a9","name":"IMAGE0","type":"IMAGE","linkIds":[118],"localized_name":"IMAGE0","label":"IMAGE","pos":[1430,-156]}],"widgets":[],"nodes":[{"id":4,"type":"PrimitiveFloat","pos":[540,-280],"size":[270,58],"flags":{},"order":0,"mode":0,"inputs":[{"label":"brightness","localized_name":"value","name":"value","type":"FLOAT","widget":{"name":"value"},"link":null}],"outputs":[{"localized_name":"FLOAT","name":"FLOAT","type":"FLOAT","links":[115]}],"properties":{"Node name for S&R":"PrimitiveFloat","min":0,"max":100,"precision":1,"step":1},"widgets_values":[50]},{"id":5,"type":"PrimitiveFloat","pos":[540,-170],"size":[270,58],"flags":{},"order":1,"mode":0,"inputs":[{"label":"contrast","localized_name":"value","name":"value","type":"FLOAT","widget":{"name":"value"},"link":null}],"outputs":[{"localized_name":"FLOAT","name":"FLOAT","type":"FLOAT","links":[116]}],"properties":{"Node name for S&R":"PrimitiveFloat","min":0,"max":100,"precision":1,"step":1},"widgets_values":[0]},{"id":143,"type":"GLSLShader","pos":[840,-280],"size":[400,212],"flags":{},"order":2,"mode":0,"inputs":[{"label":"image0","localized_name":"images.image0","name":"images.image0","type":"IMAGE","link":117},{"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":115},{"label":"u_float1","localized_name":"floats.u_float1","name":"floats.u_float1","shape":7,"type":"FLOAT","link":116},{"label":"u_float2","localized_name":"floats.u_float2","name":"floats.u_float2","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":[118]},{"localized_name":"IMAGE1","name":"IMAGE1","type":"IMAGE","links":null},{"localized_name":"IMAGE2","name":"IMAGE2","type":"IMAGE","links":null},{"localized_name":"IMAGE3","name":"IMAGE3","type":"IMAGE","links":null}],"properties":{"Node name for S&R":"GLSLShader"},"widgets_values":["#version 300 es\nprecision highp float;\n\nuniform sampler2D u_image0;\nuniform float u_float0; // Brightness slider -100..100\nuniform float u_float1; // Contrast slider -100..100\n\nin vec2 v_texCoord;\nout vec4 fragColor;\n\nconst float MID_GRAY = 0.18;  // 18% reflectance\n\n// sRGB gamma 2.2 approximation\nvec3 srgbToLinear(vec3 c) {\n    return pow(max(c, 0.0), vec3(2.2));\n}\n\nvec3 linearToSrgb(vec3 c) {\n    return pow(max(c, 0.0), vec3(1.0/2.2));\n}\n\nfloat mapBrightness(float b) {\n    return clamp(b / 100.0, -1.0, 1.0);\n}\n\nfloat mapContrast(float c) {\n    return clamp(c / 100.0 + 1.0, 0.0, 2.0);\n}\n\nvoid main() {\n    vec4 orig = texture(u_image0, v_texCoord);\n\n    float brightness = mapBrightness(u_float0);\n    float contrast   = mapContrast(u_float1);\n\n    vec3 lin = srgbToLinear(orig.rgb);\n\n    lin = (lin - MID_GRAY) * contrast + brightness + MID_GRAY;\n\n    // Convert back to sRGB\n    vec3 result = linearToSrgb(clamp(lin, 0.0, 1.0));\n\n    fragColor = vec4(result, orig.a);\n}\n","from_input"]}],"groups":[],"links":[{"id":115,"origin_id":4,"origin_slot":0,"target_id":143,"target_slot":2,"type":"FLOAT"},{"id":116,"origin_id":5,"origin_slot":0,"target_id":143,"target_slot":3,"type":"FLOAT"},{"id":117,"origin_id":-10,"origin_slot":0,"target_id":143,"target_slot":0,"type":"IMAGE"},{"id":118,"origin_id":143,"origin_slot":0,"target_id":-20,"target_slot":0,"type":"IMAGE"}],"extra":{"workflowRendererVersion":"LG"}}]},"extra":{}}

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"}}]}}

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"}}]}}

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/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/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)

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/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/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/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/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/lora.py

@@ -374,6 +374,31 @@ 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/model_base.py

@@ -178,10 +178,7 @@ 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
@@ -218,6 +215,13 @@ 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
 
@@ -372,9 +376,7 @@ 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)
@@ -1165,7 +1167,7 @@ class Anima(BaseModel):
                 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()), t5xxl_ids.to(device=device), t5xxl_weights=t5xxl_weights.to(device=device, dtype=self.get_dtype()))
+                    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)

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]
 

comfy/model_management.py

@@ -1213,8 +1213,12 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str
 
         signature = comfy_aimdo.model_vbar.vbar_fault(weight._v)
         if signature is not None:
-            v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, weight._v_tensor)[0]
-            if not comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
+            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)

comfy/model_patcher.py

@@ -406,13 +406,16 @@ 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)
@@ -679,18 +682,19 @@ 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"):
@@ -1495,7 +1499,7 @@ class ModelPatcherDynamic(ModelPatcher):
             #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:
@@ -1513,8 +1517,10 @@ class ModelPatcherDynamic(ModelPatcher):
 
                     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:
@@ -1525,10 +1531,16 @@ 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
@@ -1536,14 +1548,19 @@ 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)
-                        m._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(m._v, device_to)
-                    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:
@@ -1552,16 +1569,17 @@ class ModelPatcherDynamic(ModelPatcher):
                         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._v_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device_to)
                             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.")
 
             self.model.device = device_to
@@ -1581,7 +1599,7 @@ 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)
@@ -1602,6 +1620,13 @@ class ModelPatcherDynamic(ModelPatcher):
         if unpatch_weights:
             self.partially_unload_ram(1e32)
             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/ops.py

@@ -21,7 +21,6 @@ import logging
 import comfy.model_management
 from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
 import comfy.float
-import comfy.rmsnorm
 import json
 import comfy.memory_management
 import comfy.pinned_memory
@@ -80,17 +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 = s._v_tensor
     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 ]
@@ -140,9 +143,13 @@ def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compu
                     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)
@@ -163,10 +170,10 @@ 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
             elif update_weight:
@@ -182,13 +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.
@@ -206,7 +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)):
@@ -456,7 +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
@@ -468,8 +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
 
@@ -845,8 +850,8 @@ 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
@@ -876,8 +881,7 @@ 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/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/sd.py

@@ -423,6 +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
@@ -1182,6 +1193,7 @@ class TEModel(Enum):
     JINA_CLIP_2 = 19
     QWEN3_8B = 20
     QWEN3_06B = 21
+    GEMMA_3_4B_VISION = 22
 
 
 def detect_te_model(sd):
@@ -1210,7 +1222,10 @@ 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']
@@ -1270,6 +1285,8 @@ 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()
@@ -1335,6 +1352,14 @@ 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)

comfy/sd1_clip.py

@@ -171,8 +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
 
@@ -186,15 +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
@@ -301,6 +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 = ""
@@ -557,6 +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)]
@@ -656,6 +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:
@@ -679,6 +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)
@@ -715,3 +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

@@ -710,6 +710,15 @@ 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."]
 
@@ -898,11 +907,13 @@ 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
 
@@ -1264,6 +1275,15 @@ 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)
@@ -1341,6 +1361,14 @@ 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)

comfy/text_encoders/ace15.py

@@ -3,7 +3,6 @@ import comfy.text_encoders.llama
 from comfy import sd1_clip
 import torch
 import math
-from tqdm.auto import trange
 import yaml
 import comfy.utils
 
@@ -11,12 +10,12 @@ import comfy.utils
 def sample_manual_loop_no_classes(
     model,
     ids=None,
-    paddings=[],
     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,
@@ -36,9 +35,6 @@ def sample_manual_loop_no_classes(
 
     embeds, attention_mask, num_tokens, embeds_info = model.process_tokens(ids, device)
     embeds_batch = embeds.shape[0]
-    for i, t in enumerate(paddings):
-        attention_mask[i, :t] = 0
-        attention_mask[i, t:] = 1
 
     output_audio_codes = []
     past_key_values = []
@@ -52,7 +48,7 @@ def sample_manual_loop_no_classes(
 
     progress_bar = comfy.utils.ProgressBar(max_new_tokens)
 
-    for step in trange(max_new_tokens, desc="LM sampling"):
+    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]
@@ -81,6 +77,12 @@ def sample_manual_loop_no_classes(
             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)
@@ -111,7 +113,7 @@ def sample_manual_loop_no_classes(
     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):
+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]
 
@@ -129,13 +131,11 @@ def generate_audio_codes(model, positive, negative, min_tokens=1, max_tokens=102
             pos_pad = (len(negative) - len(positive))
             positive = [model.special_tokens["pad"]] * pos_pad + positive
 
-        paddings = [pos_pad, neg_pad]
         ids = [positive, negative]
     else:
-        paddings = []
         ids = [positive]
 
-    return sample_manual_loop_no_classes(model, ids, paddings, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k, seed=seed, min_tokens=min_tokens, max_new_tokens=max_tokens)
+    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):
@@ -193,6 +193,7 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer):
         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
@@ -240,6 +241,7 @@ class ACE15Tokenizer(sd1_clip.SD1Tokenizer):
                               "temperature": temperature,
                               "top_p": top_p,
                               "top_k": top_k,
+                              "min_p": min_p,
                               }
         return out
 
@@ -300,7 +302,7 @@ class ACE15TEModel(torch.nn.Module):
 
         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["max_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"])
+            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

comfy/text_encoders/anima.py

@@ -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/llama.py

@@ -3,6 +3,8 @@ import torch.nn as nn
 from dataclasses import dataclass
 from typing import Optional, Any, Tuple
 import math
+from tqdm import tqdm
+import comfy.utils
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
@@ -103,6 +105,7 @@ class Qwen3_06BConfig:
     rope_scale = None
     final_norm: bool = True
     lm_head: bool = False
+    stop_tokens = [151643, 151645]
 
 @dataclass
 class Qwen3_06B_ACE15_Config:
@@ -126,6 +129,7 @@ class Qwen3_06B_ACE15_Config:
     rope_scale = None
     final_norm: bool = True
     lm_head: bool = False
+    stop_tokens = [151643, 151645]
 
 @dataclass
 class Qwen3_2B_ACE15_lm_Config:
@@ -149,6 +153,7 @@ class Qwen3_2B_ACE15_lm_Config:
     rope_scale = None
     final_norm: bool = True
     lm_head: bool = False
+    stop_tokens = [151643, 151645]
 
 @dataclass
 class Qwen3_4B_ACE15_lm_Config:
@@ -172,6 +177,7 @@ class Qwen3_4B_ACE15_lm_Config:
     rope_scale = None
     final_norm: bool = True
     lm_head: bool = False
+    stop_tokens = [151643, 151645]
 
 @dataclass
 class Qwen3_4BConfig:
@@ -195,6 +201,7 @@ class Qwen3_4BConfig:
     rope_scale = None
     final_norm: bool = True
     lm_head: bool = False
+    stop_tokens = [151643, 151645]
 
 @dataclass
 class Qwen3_8BConfig:
@@ -218,6 +225,7 @@ class Qwen3_8BConfig:
     rope_scale = None
     final_norm: bool = True
     lm_head: bool = False
+    stop_tokens = [151643, 151645]
 
 @dataclass
 class Ovis25_2BConfig:
@@ -288,6 +296,7 @@ class Gemma2_2B_Config:
     rope_scale = None
     final_norm: bool = True
     lm_head: bool = False
+    stop_tokens = [1]
 
 @dataclass
 class Gemma3_4B_Config:
@@ -312,6 +321,14 @@ class Gemma3_4B_Config:
     rope_scale = [8.0, 1.0]
     final_norm: bool = True
     lm_head: bool = False
+    stop_tokens = [1, 106]
+
+GEMMA3_VISION_CONFIG = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14}
+
+@dataclass
+class Gemma3_4B_Vision_Config(Gemma3_4B_Config):
+    vision_config = GEMMA3_VISION_CONFIG
+    mm_tokens_per_image = 256
 
 @dataclass
 class Gemma3_12B_Config:
@@ -336,8 +353,9 @@ class Gemma3_12B_Config:
     rope_scale = [8.0, 1.0]
     final_norm: bool = True
     lm_head: bool = False
-    vision_config = {"num_channels": 3, "hidden_act": "gelu_pytorch_tanh", "hidden_size": 1152, "image_size": 896, "intermediate_size": 4304, "model_type": "siglip_vision_model", "num_attention_heads": 16, "num_hidden_layers": 27, "patch_size": 14}
+    vision_config = GEMMA3_VISION_CONFIG
     mm_tokens_per_image = 256
+    stop_tokens = [1, 106]
 
 class RMSNorm(nn.Module):
     def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
@@ -355,13 +373,6 @@ class RMSNorm(nn.Module):
 
 
 
-def rotate_half(x):
-    """Rotates half the hidden dims of the input."""
-    x1 = x[..., : x.shape[-1] // 2]
-    x2 = x[..., x.shape[-1] // 2 :]
-    return torch.cat((-x2, x1), dim=-1)
-
-
 def precompute_freqs_cis(head_dim, position_ids, theta, rope_scale=None, rope_dims=None, device=None):
     if not isinstance(theta, list):
         theta = [theta]
@@ -390,20 +401,30 @@ def precompute_freqs_cis(head_dim, position_ids, theta, rope_scale=None, rope_di
         else:
             cos = cos.unsqueeze(1)
             sin = sin.unsqueeze(1)
-        out.append((cos, sin))
+        sin_split = sin.shape[-1] // 2
+        out.append((cos, sin[..., : sin_split], -sin[..., sin_split :]))
 
     if len(out) == 1:
         return out[0]
 
     return out
 
-
 def apply_rope(xq, xk, freqs_cis):
     org_dtype = xq.dtype
     cos = freqs_cis[0]
     sin = freqs_cis[1]
-    q_embed = (xq * cos) + (rotate_half(xq) * sin)
-    k_embed = (xk * cos) + (rotate_half(xk) * sin)
+    nsin = freqs_cis[2]
+
+    q_embed = (xq * cos)
+    q_split = q_embed.shape[-1] // 2
+    q_embed[..., : q_split].addcmul_(xq[..., q_split :], nsin)
+    q_embed[..., q_split :].addcmul_(xq[..., : q_split], sin)
+
+    k_embed = (xk * cos)
+    k_split = k_embed.shape[-1] // 2
+    k_embed[..., : k_split].addcmul_(xk[..., k_split :], nsin)
+    k_embed[..., k_split :].addcmul_(xk[..., : k_split], sin)
+
     return q_embed.to(org_dtype), k_embed.to(org_dtype)
 
 
@@ -438,8 +459,10 @@ class Attention(nn.Module):
         freqs_cis: Optional[torch.Tensor] = None,
         optimized_attention=None,
         past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        sliding_window: Optional[int] = None,
     ):
         batch_size, seq_length, _ = hidden_states.shape
+
         xq = self.q_proj(hidden_states)
         xk = self.k_proj(hidden_states)
         xv = self.v_proj(hidden_states)
@@ -474,6 +497,11 @@ class Attention(nn.Module):
             else:
                 present_key_value = (xk, xv, index + num_tokens)
 
+            if sliding_window is not None and xk.shape[2] > sliding_window:
+                xk = xk[:, :, -sliding_window:]
+                xv = xv[:, :, -sliding_window:]
+                attention_mask = attention_mask[..., -sliding_window:] if attention_mask is not None else None
+
         xk = xk.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
         xv = xv.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
 
@@ -556,10 +584,12 @@ class TransformerBlockGemma2(nn.Module):
         optimized_attention=None,
         past_key_value: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
     ):
+        sliding_window = None
         if self.transformer_type == 'gemma3':
             if self.sliding_attention:
+                sliding_window = self.sliding_attention
                 if x.shape[1] > self.sliding_attention:
-                    sliding_mask = torch.full((x.shape[1], x.shape[1]), float("-inf"), device=x.device, dtype=x.dtype)
+                    sliding_mask = torch.full((x.shape[1], x.shape[1]), torch.finfo(x.dtype).min, device=x.device, dtype=x.dtype)
                     sliding_mask.tril_(diagonal=-self.sliding_attention)
                     if attention_mask is not None:
                         attention_mask = attention_mask + sliding_mask
@@ -578,6 +608,7 @@ class TransformerBlockGemma2(nn.Module):
             freqs_cis=freqs_cis,
             optimized_attention=optimized_attention,
             past_key_value=past_key_value,
+            sliding_window=sliding_window,
         )
 
         x = self.post_attention_layernorm(x)
@@ -762,6 +793,107 @@ class BaseLlama:
     def forward(self, input_ids, *args, **kwargs):
         return self.model(input_ids, *args, **kwargs)
 
+class BaseGenerate:
+    def logits(self, x):
+        input = x[:, -1:]
+        if hasattr(self.model, "lm_head"):
+            module = self.model.lm_head
+        else:
+            module = self.model.embed_tokens
+
+        offload_stream = None
+        if module.comfy_cast_weights:
+            weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True)
+        else:
+            weight = self.model.embed_tokens.weight.to(x)
+
+        x = torch.nn.functional.linear(input, weight, None)
+
+        comfy.ops.uncast_bias_weight(module, weight, None, offload_stream)
+        return x
+
+    def generate(self, embeds=None, do_sample=True, max_length=256, temperature=1.0, top_k=50, top_p=0.9, min_p=0.0, repetition_penalty=1.0, seed=42, stop_tokens=None, initial_tokens=[], execution_dtype=None, min_tokens=0):
+        device = embeds.device
+        model_config = self.model.config
+
+        if stop_tokens is None:
+            stop_tokens = self.model.config.stop_tokens
+
+        if execution_dtype is None:
+            if comfy.model_management.should_use_bf16(device):
+                execution_dtype = torch.bfloat16
+            else:
+                execution_dtype = torch.float32
+        embeds = embeds.to(execution_dtype)
+
+        if embeds.ndim == 2:
+            embeds = embeds.unsqueeze(0)
+
+        past_key_values = [] #kv_cache init
+        max_cache_len = embeds.shape[1] + max_length
+        for x in range(model_config.num_hidden_layers):
+            past_key_values.append((torch.empty([embeds.shape[0], model_config.num_key_value_heads, max_cache_len, model_config.head_dim], device=device, dtype=execution_dtype),
+                                    torch.empty([embeds.shape[0], model_config.num_key_value_heads, max_cache_len, model_config.head_dim], device=device, dtype=execution_dtype), 0))
+
+        generator = torch.Generator(device=device).manual_seed(seed) if do_sample else None
+
+        generated_token_ids = []
+        pbar = comfy.utils.ProgressBar(max_length)
+
+        # Generation loop
+        for step in tqdm(range(max_length), desc="Generating tokens"):
+            x, _, past_key_values = self.model.forward(None, embeds=embeds, attention_mask=None, past_key_values=past_key_values)
+            logits = self.logits(x)[:, -1]
+            next_token = self.sample_token(logits, temperature, top_k, top_p, min_p, repetition_penalty, initial_tokens + generated_token_ids, generator, do_sample=do_sample)
+            token_id = next_token[0].item()
+            generated_token_ids.append(token_id)
+
+            embeds = self.model.embed_tokens(next_token).to(execution_dtype)
+            pbar.update(1)
+
+            if token_id in stop_tokens:
+                break
+
+        return generated_token_ids
+
+    def sample_token(self, logits, temperature, top_k, top_p, min_p, repetition_penalty, token_history, generator, do_sample=True):
+
+        if not do_sample or temperature == 0.0:
+            return torch.argmax(logits, dim=-1, keepdim=True)
+
+        # Sampling mode
+        if repetition_penalty != 1.0:
+            for i in range(logits.shape[0]):
+                for token_id in set(token_history):
+                    logits[i, token_id] *= repetition_penalty if logits[i, token_id] < 0 else 1/repetition_penalty
+
+        if temperature != 1.0:
+            logits = logits / temperature
+
+        if top_k > 0:
+            indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
+            logits[indices_to_remove] = torch.finfo(logits.dtype).min
+
+        if min_p > 0.0:
+            probs_before_filter = torch.nn.functional.softmax(logits, dim=-1)
+            top_probs, _ = probs_before_filter.max(dim=-1, keepdim=True)
+            min_threshold = min_p * top_probs
+            indices_to_remove = probs_before_filter < min_threshold
+            logits[indices_to_remove] = torch.finfo(logits.dtype).min
+
+        if top_p < 1.0:
+            sorted_logits, sorted_indices = torch.sort(logits, descending=True)
+            cumulative_probs = torch.cumsum(torch.nn.functional.softmax(sorted_logits, dim=-1), dim=-1)
+            sorted_indices_to_remove = cumulative_probs > top_p
+            sorted_indices_to_remove[..., 0] = False
+            indices_to_remove = torch.zeros_like(logits, dtype=torch.bool)
+            indices_to_remove.scatter_(1, sorted_indices, sorted_indices_to_remove)
+            logits[indices_to_remove] = torch.finfo(logits.dtype).min
+
+        probs = torch.nn.functional.softmax(logits, dim=-1)
+
+        return torch.multinomial(probs, num_samples=1, generator=generator)
+
 class BaseQwen3:
     def logits(self, x):
         input = x[:, -1:]
@@ -805,7 +937,7 @@ class Qwen25_3B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_06B(BaseLlama, BaseQwen3, torch.nn.Module):
+class Qwen3_06B(BaseLlama, BaseQwen3, BaseGenerate, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_06BConfig(**config_dict)
@@ -832,7 +964,7 @@ class Qwen3_2B_ACE15_lm(BaseLlama, BaseQwen3, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_4B(BaseLlama, BaseQwen3, torch.nn.Module):
+class Qwen3_4B(BaseLlama, BaseQwen3, BaseGenerate, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_4BConfig(**config_dict)
@@ -850,7 +982,7 @@ class Qwen3_4B_ACE15_lm(BaseLlama, BaseQwen3, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_8B(BaseLlama, BaseQwen3, torch.nn.Module):
+class Qwen3_8B(BaseLlama, BaseQwen3, BaseGenerate, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_8BConfig(**config_dict)
@@ -868,7 +1000,7 @@ class Ovis25_2B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
+class Qwen25_7BVLI(BaseLlama, BaseGenerate, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen25_7BVLI_Config(**config_dict)
@@ -878,6 +1010,9 @@ class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
         self.visual = qwen_vl.Qwen2VLVisionTransformer(hidden_size=1280, output_hidden_size=config.hidden_size, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
+        # todo: should this be tied or not?
+        #self.lm_head = operations.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)
+
     def preprocess_embed(self, embed, device):
         if embed["type"] == "image":
             image, grid = qwen_vl.process_qwen2vl_images(embed["data"])
@@ -911,7 +1046,7 @@ class Qwen25_7BVLI(BaseLlama, torch.nn.Module):
 
         return super().forward(x, attention_mask=attention_mask, embeds=embeds, num_tokens=num_tokens, intermediate_output=intermediate_output, final_layer_norm_intermediate=final_layer_norm_intermediate, dtype=dtype, position_ids=position_ids)
 
-class Gemma2_2B(BaseLlama, torch.nn.Module):
+class Gemma2_2B(BaseLlama, BaseGenerate, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Gemma2_2B_Config(**config_dict)
@@ -920,7 +1055,7 @@ class Gemma2_2B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Gemma3_4B(BaseLlama, torch.nn.Module):
+class Gemma3_4B(BaseLlama, BaseGenerate, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Gemma3_4B_Config(**config_dict)
@@ -929,7 +1064,25 @@ class Gemma3_4B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Gemma3_12B(BaseLlama, torch.nn.Module):
+class Gemma3_4B_Vision(BaseLlama, BaseGenerate, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Gemma3_4B_Vision_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+        self.multi_modal_projector = Gemma3MultiModalProjector(config, dtype, device, operations)
+        self.vision_model = comfy.clip_model.CLIPVision(config.vision_config, dtype, device, operations)
+        self.image_size = config.vision_config["image_size"]
+
+    def preprocess_embed(self, embed, device):
+        if embed["type"] == "image":
+            image = comfy.clip_model.clip_preprocess(embed["data"], size=self.image_size, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], crop=True)
+            return self.multi_modal_projector(self.vision_model(image.to(device, dtype=torch.float32))[0]), None
+        return None, None
+
+class Gemma3_12B(BaseLlama, BaseGenerate, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Gemma3_12B_Config(**config_dict)

comfy/text_encoders/lt.py

@@ -6,6 +6,7 @@ import comfy.text_encoders.genmo
 from comfy.ldm.lightricks.embeddings_connector import Embeddings1DConnector
 import torch
 import comfy.utils
+import math
 
 class T5XXLTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -22,40 +23,79 @@ def ltxv_te(*args, **kwargs):
     return comfy.text_encoders.genmo.mochi_te(*args, **kwargs)
 
 
-class Gemma3_12BTokenizer(sd1_clip.SDTokenizer):
-    def __init__(self, embedding_directory=None, tokenizer_data={}):
-        tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, disable_weights=True, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
-
+class Gemma3_Tokenizer():
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}
 
+    def tokenize_with_weights(self, text, return_word_ids=False, image=None, llama_template=None, skip_template=True, **kwargs):
+        self.llama_template = "<start_of_turn>system\nYou are a helpful assistant.<end_of_turn>\n<start_of_turn>user\n{}<end_of_turn>\n<start_of_turn>model\n"
+        self.llama_template_images = "<start_of_turn>system\nYou are a helpful assistant.<end_of_turn>\n<start_of_turn>user\n\n<image_soft_token>{}<end_of_turn>\n\n<start_of_turn>model\n"
+
+        if image is None:
+            images = []
+        else:
+            samples = image.movedim(-1, 1)
+            total = int(896 * 896)
+
+            scale_by = math.sqrt(total / (samples.shape[3] * samples.shape[2]))
+            width = round(samples.shape[3] * scale_by)
+            height = round(samples.shape[2] * scale_by)
+
+            s = comfy.utils.common_upscale(samples, width, height, "area", "disabled").movedim(1, -1)
+            images = [s[:, :, :, :3]]
+
+        if text.startswith('<start_of_turn>'):
+            skip_template = True
+
+        if skip_template:
+            llama_text = text
+        else:
+            if llama_template is None:
+                if len(images) > 0:
+                    llama_text = self.llama_template_images.format(text)
+                else:
+                    llama_text = self.llama_template.format(text)
+            else:
+                llama_text = llama_template.format(text)
+
+        text_tokens = super().tokenize_with_weights(llama_text, return_word_ids)
+
+        if len(images) > 0:
+            embed_count = 0
+            for r in text_tokens:
+                for i, token in enumerate(r):
+                    if token[0] == 262144 and embed_count < len(images):
+                        r[i] = ({"type": "image", "data": images[embed_count]},) + token[1:]
+                        embed_count += 1
+        return text_tokens
+
+class Gemma3_12BTokenizer(Gemma3_Tokenizer, sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = tokenizer_data.get("spiece_model", None)
+        special_tokens = {"<image_soft_token>": 262144, "<end_of_turn>": 106}
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=3840, embedding_key='gemma3_12b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_left=True, disable_weights=True, tokenizer_args={"add_bos": True, "add_eos": False, "special_tokens": special_tokens}, tokenizer_data=tokenizer_data)
+
+
 class LTXAVGemmaTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, name="gemma3_12b", tokenizer=Gemma3_12BTokenizer)
 
+
 class Gemma3_12BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="all", 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
-
+        self.dtypes = set()
+        self.dtypes.add(dtype)
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_12B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
-    def tokenize_with_weights(self, text, return_word_ids=False, llama_template="{}", image_embeds=None, **kwargs):
-        text = llama_template.format(text)
-        text_tokens = super().tokenize_with_weights(text, return_word_ids)
-        embed_count = 0
-        for k in text_tokens:
-            tt = text_tokens[k]
-            for r in tt:
-                for i in range(len(r)):
-                    if r[i][0] == 262144:
-                        if image_embeds is not None and embed_count < image_embeds.shape[0]:
-                            r[i] = ({"type": "embedding", "data": image_embeds[embed_count], "original_type": "image"},) + r[i][1:]
-                            embed_count += 1
-        return text_tokens
+    def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed):
+        tokens_only = [[t[0] for t in b] for b in tokens]
+        embeds, _, _, embeds_info = self.process_tokens(tokens_only, self.execution_device)
+        comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5)
+        return self.transformer.generate(embeds, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed, stop_tokens=[106])  # 106 is <end_of_turn>
 
 class LTXAVTEModel(torch.nn.Module):
     def __init__(self, dtype_llama=None, device="cpu", dtype=None, model_options={}):
@@ -97,6 +137,7 @@ class LTXAVTEModel(torch.nn.Module):
         token_weight_pairs = token_weight_pairs["gemma3_12b"]
 
         out, pooled, extra = self.gemma3_12b.encode_token_weights(token_weight_pairs)
+        out = out[:, :, -torch.sum(extra["attention_mask"]).item():]
         out_device = out.device
         if comfy.model_management.should_use_bf16(self.execution_device):
             out = out.to(device=self.execution_device, dtype=torch.bfloat16)
@@ -111,6 +152,9 @@ class LTXAVTEModel(torch.nn.Module):
 
         return out.to(out_device), pooled
 
+    def generate(self, tokens, do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed):
+        return self.gemma3_12b.generate(tokens["gemma3_12b"], do_sample, max_length, temperature, top_k, top_p, min_p, repetition_penalty, seed)
+
     def load_sd(self, sd):
         if "model.layers.47.self_attn.q_norm.weight" in sd:
             return self.gemma3_12b.load_sd(sd)
@@ -138,6 +182,7 @@ class LTXAVTEModel(torch.nn.Module):
 
         token_weight_pairs = token_weight_pairs.get("gemma3_12b", [])
         num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
+        num_tokens = max(num_tokens, 64)
         return num_tokens * constant * 1024 * 1024
 
 def ltxav_te(dtype_llama=None, llama_quantization_metadata=None):
@@ -150,3 +195,14 @@ def ltxav_te(dtype_llama=None, llama_quantization_metadata=None):
                 dtype = dtype_llama
             super().__init__(dtype_llama=dtype_llama, device=device, dtype=dtype, model_options=model_options)
     return LTXAVTEModel_
+
+def gemma3_te(dtype_llama=None, llama_quantization_metadata=None):
+    class Gemma3_12BModel_(Gemma3_12BModel):
+        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
+            if dtype_llama is not None:
+                dtype = dtype_llama
+            super().__init__(device=device, dtype=dtype, model_options=model_options)
+    return Gemma3_12BModel_

comfy/text_encoders/lumina2.py

@@ -1,23 +1,23 @@
 from comfy import sd1_clip
 from .spiece_tokenizer import SPieceTokenizer
 import comfy.text_encoders.llama
-
+from comfy.text_encoders.lt import Gemma3_Tokenizer
+import comfy.utils
 
 class Gemma2BTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=2304, embedding_key='gemma2_2b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
+        special_tokens = {"<end_of_turn>": 107}
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=2304, embedding_key='gemma2_2b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False, "special_tokens": special_tokens}, tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}
 
-class Gemma3_4BTokenizer(sd1_clip.SDTokenizer):
+class Gemma3_4BTokenizer(Gemma3_Tokenizer, sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, disable_weights=True, tokenizer_data=tokenizer_data)
-
-    def state_dict(self):
-        return {"spiece_model": self.tokenizer.serialize_model()}
+        special_tokens = {"<image_soft_token>": 262144, "<end_of_turn>": 106}
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False, "special_tokens": special_tokens}, disable_weights=True, tokenizer_data=tokenizer_data)
 
 class LuminaTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
@@ -40,6 +40,20 @@ class Gemma3_4BModel(sd1_clip.SDClipModel):
 
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
+class Gemma3_4B_Vision_Model(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="hidden", layer_idx=-2, 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={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B_Vision, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
+    def process_tokens(self, tokens, device):
+        embeds, _, _, embeds_info = super().process_tokens(tokens, device)
+        comfy.utils.normalize_image_embeddings(embeds, embeds_info, self.transformer.model.config.hidden_size ** 0.5)
+        return embeds
+
 class LuminaModel(sd1_clip.SD1ClipModel):
     def __init__(self, device="cpu", dtype=None, model_options={}, name="gemma2_2b", clip_model=Gemma2_2BModel):
         super().__init__(device=device, dtype=dtype, name=name, clip_model=clip_model, model_options=model_options)
@@ -50,6 +64,8 @@ def te(dtype_llama=None, llama_quantization_metadata=None, model_type="gemma2_2b
         model = Gemma2_2BModel
     elif model_type == "gemma3_4b":
         model = Gemma3_4BModel
+    elif model_type == "gemma3_4b_vision":
+        model = Gemma3_4B_Vision_Model
 
     class LuminaTEModel_(LuminaModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):

comfy/text_encoders/spiece_tokenizer.py

@@ -6,9 +6,10 @@ class SPieceTokenizer:
     def from_pretrained(path, **kwargs):
         return SPieceTokenizer(path, **kwargs)
 
-    def __init__(self, tokenizer_path, add_bos=False, add_eos=True):
+    def __init__(self, tokenizer_path, add_bos=False, add_eos=True, special_tokens=None):
         self.add_bos = add_bos
         self.add_eos = add_eos
+        self.special_tokens = special_tokens
         import sentencepiece
         if torch.is_tensor(tokenizer_path):
             tokenizer_path = tokenizer_path.numpy().tobytes()
@@ -27,8 +28,32 @@ class SPieceTokenizer:
         return out
 
     def __call__(self, string):
+        if self.special_tokens is not None:
+            import re
+            special_tokens_pattern = '|'.join(re.escape(token) for token in self.special_tokens.keys())
+            if special_tokens_pattern and re.search(special_tokens_pattern, string):
+                parts = re.split(f'({special_tokens_pattern})', string)
+                result = []
+                for part in parts:
+                    if not part:
+                        continue
+                    if part in self.special_tokens:
+                        result.append(self.special_tokens[part])
+                    else:
+                        encoded = self.tokenizer.encode(part, add_bos=False, add_eos=False)
+                        result.extend(encoded)
+                return {"input_ids": result}
+
         out = self.tokenizer.encode(string)
         return {"input_ids": out}
 
+    def decode(self, token_ids, skip_special_tokens=False):
+
+        if skip_special_tokens and self.special_tokens:
+            special_token_ids = set(self.special_tokens.values())
+            token_ids = [tid for tid in token_ids if tid not in special_token_ids]
+
+        return self.tokenizer.decode(token_ids)
+
     def serialize_model(self):
         return torch.ByteTensor(list(self.tokenizer.serialized_model_proto()))

comfy/utils.py

@@ -20,13 +20,14 @@
 import torch
 import math
 import struct
-import comfy.checkpoint_pickle
+import comfy.memory_management
 import safetensors.torch
 import numpy as np
 from PIL import Image
 import logging
 import itertools
 from torch.nn.functional import interpolate
+from tqdm.auto import trange
 from einops import rearrange
 from comfy.cli_args import args, enables_dynamic_vram
 import json
@@ -37,26 +38,26 @@ import warnings
 MMAP_TORCH_FILES = args.mmap_torch_files
 DISABLE_MMAP = args.disable_mmap
 
-ALWAYS_SAFE_LOAD = False
-if hasattr(torch.serialization, "add_safe_globals"):  # TODO: this was added in pytorch 2.4, the unsafe path should be removed once earlier versions are deprecated
+
+if True:  # ckpt/pt file whitelist for safe loading of old sd files
     class ModelCheckpoint:
         pass
     ModelCheckpoint.__module__ = "pytorch_lightning.callbacks.model_checkpoint"
 
     def scalar(*args, **kwargs):
-        from numpy.core.multiarray import scalar as sc
-        return sc(*args, **kwargs)
+        return None
     scalar.__module__ = "numpy.core.multiarray"
 
     from numpy import dtype
     from numpy.dtypes import Float64DType
-    from _codecs import encode
+
+    def encode(*args, **kwargs):  # no longer necessary on newer torch
+        return None
+    encode.__module__ = "_codecs"
 
     torch.serialization.add_safe_globals([ModelCheckpoint, scalar, dtype, Float64DType, encode])
-    ALWAYS_SAFE_LOAD = True
     logging.info("Checkpoint files will always be loaded safely.")
-else:
-    logging.warning("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended as older versions of pytorch are no longer supported.")
+
 
 # Current as of safetensors 0.7.0
 _TYPES = {
@@ -139,11 +140,8 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
         if MMAP_TORCH_FILES:
             torch_args["mmap"] = True
 
-        if safe_load or ALWAYS_SAFE_LOAD:
-            pl_sd = torch.load(ckpt, map_location=device, weights_only=True, **torch_args)
-        else:
-            logging.warning("WARNING: loading {} unsafely, upgrade your pytorch to 2.4 or newer to load this file safely.".format(ckpt))
-            pl_sd = torch.load(ckpt, map_location=device, pickle_module=comfy.checkpoint_pickle)
+        pl_sd = torch.load(ckpt, map_location=device, weights_only=True, **torch_args)
+
         if "state_dict" in pl_sd:
             sd = pl_sd["state_dict"]
         else:
@@ -674,10 +672,10 @@ def flux_to_diffusers(mmdit_config, output_prefix=""):
                         "ff_context.linear_in.bias": "txt_mlp.0.bias",
                         "ff_context.linear_out.weight": "txt_mlp.2.weight",
                         "ff_context.linear_out.bias": "txt_mlp.2.bias",
-                        "attn.norm_q.weight": "img_attn.norm.query_norm.scale",
-                        "attn.norm_k.weight": "img_attn.norm.key_norm.scale",
-                        "attn.norm_added_q.weight": "txt_attn.norm.query_norm.scale",
-                        "attn.norm_added_k.weight": "txt_attn.norm.key_norm.scale",
+                        "attn.norm_q.weight": "img_attn.norm.query_norm.weight",
+                        "attn.norm_k.weight": "img_attn.norm.key_norm.weight",
+                        "attn.norm_added_q.weight": "txt_attn.norm.query_norm.weight",
+                        "attn.norm_added_k.weight": "txt_attn.norm.key_norm.weight",
                     }
 
         for k in block_map:
@@ -700,8 +698,8 @@ def flux_to_diffusers(mmdit_config, output_prefix=""):
                         "norm.linear.bias": "modulation.lin.bias",
                         "proj_out.weight": "linear2.weight",
                         "proj_out.bias": "linear2.bias",
-                        "attn.norm_q.weight": "norm.query_norm.scale",
-                        "attn.norm_k.weight": "norm.key_norm.scale",
+                        "attn.norm_q.weight": "norm.query_norm.weight",
+                        "attn.norm_k.weight": "norm.key_norm.weight",
                         "attn.to_qkv_mlp_proj.weight": "linear1.weight", # Flux 2
                         "attn.to_out.weight": "linear2.weight", # Flux 2
                     }
@@ -1155,6 +1153,32 @@ def tiled_scale_multidim(samples, function, tile=(64, 64), overlap=8, upscale_am
 def tiled_scale(samples, function, tile_x=64, tile_y=64, overlap = 8, upscale_amount = 4, out_channels = 3, output_device="cpu", pbar = None):
     return tiled_scale_multidim(samples, function, (tile_y, tile_x), overlap=overlap, upscale_amount=upscale_amount, out_channels=out_channels, output_device=output_device, pbar=pbar)
 
+def model_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
+
 PROGRESS_BAR_ENABLED = True
 def set_progress_bar_enabled(enabled):
     global PROGRESS_BAR_ENABLED
@@ -1394,3 +1418,11 @@ def deepcopy_list_dict(obj, memo=None):
 
     memo[obj_id] = res
     return res
+
+def normalize_image_embeddings(embeds, embeds_info, scale_factor):
+    """Normalize image embeddings to match text embedding scale"""
+    for info in embeds_info:
+        if info.get("type") == "image":
+            start_idx = info["index"]
+            end_idx = start_idx + info["size"]
+            embeds[:, start_idx:end_idx, :] /= scale_factor

comfy/weight_adapter/base.py

@@ -49,6 +49,12 @@ class WeightAdapterBase:
         """
         raise NotImplementedError
 
+    def calculate_shape(
+        self,
+        key
+    ):
+        return None
+
     def calculate_weight(
         self,
         weight,

comfy/weight_adapter/lora.py

@@ -214,6 +214,13 @@ class LoRAAdapter(WeightAdapterBase):
         else:
             return None
 
+    def calculate_shape(
+        self,
+        key
+    ):
+        reshape = self.weights[5]
+        return tuple(reshape) if reshape is not None else None
+
     def calculate_weight(
         self,
         weight,

comfy_api/feature_flags.py

@@ -14,6 +14,7 @@ SERVER_FEATURE_FLAGS: dict[str, Any] = {
     "supports_preview_metadata": True,
     "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes
     "extension": {"manager": {"supports_v4": True}},
+    "node_replacements": True,
 }
 
 

comfy_api/latest/__init__.py

@@ -21,6 +21,17 @@ class ComfyAPI_latest(ComfyAPIBase):
     VERSION = "latest"
     STABLE = False
 
+    def __init__(self):
+        super().__init__()
+        self.node_replacement = self.NodeReplacement()
+        self.execution = self.Execution()
+
+    class NodeReplacement(ProxiedSingleton):
+        async def register(self, node_replace: io.NodeReplace) -> None:
+            """Register a node replacement mapping."""
+            from server import PromptServer
+            PromptServer.instance.node_replace_manager.register(node_replace)
+
     class Execution(ProxiedSingleton):
         async def set_progress(
             self,
@@ -73,8 +84,6 @@ class ComfyAPI_latest(ComfyAPIBase):
                 image=to_display,
             )
 
-    execution: Execution
-
 class ComfyExtension(ABC):
     async def on_load(self) -> None:
         """

comfy_api/latest/_io.py

@@ -30,46 +30,6 @@ from comfy_execution.graph_utils import ExecutionBlocker
 from ._util import MESH, VOXEL, SVG as _SVG, File3D
 
 
-class EmptyInputSentinel:
-    """
-    Sentinel class indicating an empty/missing input.
-
-    Use the class itself (not an instance) as the sentinel.
-    Compare using 'is' or 'is not' only.
-    """
-
-    def __new__(cls):
-        raise TypeError("EmptyInputSentinel cannot be instantiated, use the class itself")
-
-    def __init_subclass__(cls, **kwargs):
-        raise TypeError("EmptyInputSentinel cannot be subclassed")
-
-    @classmethod
-    def __class_getitem__(cls, item):
-        raise TypeError("EmptyInputSentinel cannot be subscripted")
-
-    def __repr__(self):
-        return "<EmptyInput>"
-
-    def __bool__(self):
-        raise TypeError("EmptyInputSentinel cannot be used in boolean context")
-
-    def __eq__(self, other):
-        raise TypeError("EmptyInputSentinel cannot be compared with ==, use 'is' instead")
-
-    def __ne__(self, other):
-        raise TypeError("EmptyInputSentinel cannot be compared with !=, use 'is not' instead")
-
-    def __hash__(self):
-        raise TypeError("EmptyInputSentinel cannot be hashed")
-
-    def __iter__(self):
-        raise TypeError("EmptyInputSentinel cannot be iterated")
-
-    def __len__(self):
-        raise TypeError("EmptyInputSentinel has no length")
-
-
 class FolderType(str, Enum):
     input = "input"
     output = "output"
@@ -115,6 +75,12 @@ class NumberDisplay(str, Enum):
     slider = "slider"
 
 
+class ControlAfterGenerate(str, Enum):
+    fixed = "fixed"
+    increment = "increment"
+    decrement = "decrement"
+    randomize = "randomize"
+
 class _ComfyType(ABC):
     Type = Any
     io_type: str = None
@@ -303,7 +269,7 @@ class Int(ComfyTypeIO):
     class Input(WidgetInput):
         '''Integer input.'''
         def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
-                    default: int=None, min: int=None, max: int=None, step: int=None, control_after_generate: bool=None,
+                    default: int=None, min: int=None, max: int=None, step: int=None, control_after_generate: bool | ControlAfterGenerate=None,
                     display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
             super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced)
             self.min = min
@@ -385,7 +351,7 @@ class Combo(ComfyTypeIO):
             tooltip: str=None,
             lazy: bool=None,
             default: str | int | Enum = None,
-            control_after_generate: bool=None,
+            control_after_generate: bool | ControlAfterGenerate=None,
             upload: UploadType=None,
             image_folder: FolderType=None,
             remote: RemoteOptions=None,
@@ -429,7 +395,7 @@ class MultiCombo(ComfyTypeI):
     Type = list[str]
     class Input(Combo.Input):
         def __init__(self, id: str, options: list[str], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
-                    default: list[str]=None, placeholder: str=None, chip: bool=None, control_after_generate: bool=None,
+                    default: list[str]=None, placeholder: str=None, chip: bool=None, control_after_generate: bool | ControlAfterGenerate=None,
                     socketless: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
             super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless, extra_dict=extra_dict, raw_link=raw_link, advanced=advanced)
             self.multiselect = True
@@ -1243,6 +1209,30 @@ class Color(ComfyTypeIO):
       def as_dict(self):
           return super().as_dict()
 
+@comfytype(io_type="BOUNDING_BOX")
+class BoundingBox(ComfyTypeIO):
+    class BoundingBoxDict(TypedDict):
+        x: int
+        y: int
+        width: int
+        height: int
+    Type = BoundingBoxDict
+
+    class Input(WidgetInput):
+        def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None,
+                     socketless: bool=True, default: dict=None, component: str=None):
+            super().__init__(id, display_name, optional, tooltip, None, default, socketless)
+            self.component = component
+            if default is None:
+                self.default = {"x": 0, "y": 0, "width": 512, "height": 512}
+
+        def as_dict(self):
+            d = super().as_dict()
+            if self.component:
+                d["component"] = self.component
+            return d
+
+
 DYNAMIC_INPUT_LOOKUP: dict[str, Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]] = {}
 def register_dynamic_input_func(io_type: str, func: Callable[[dict[str, Any], dict[str, Any], tuple[str, dict[str, Any]], str, list[str] | None], None]):
     DYNAMIC_INPUT_LOOKUP[io_type] = func
@@ -1349,6 +1339,7 @@ class NodeInfoV1:
     api_node: bool=None
     price_badge: dict | None = None
     search_aliases: list[str]=None
+    essentials_category: str=None
 
 
 @dataclass
@@ -1470,6 +1461,8 @@ class Schema:
     """Flags a node as expandable, allowing NodeOutput to include 'expand' property."""
     accept_all_inputs: bool=False
     """When True, all inputs from the prompt will be passed to the node as kwargs, even if not defined in the schema."""
+    essentials_category: str | None = None
+    """Optional category for the Essentials tab. Path-based like category field (e.g., 'Basic', 'Image Tools/Editing')."""
 
     def validate(self):
         '''Validate the schema:
@@ -1576,6 +1569,7 @@ class Schema:
             python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes"),
             price_badge=self.price_badge.as_dict(self.inputs) if self.price_badge is not None else None,
             search_aliases=self.search_aliases if self.search_aliases else None,
+            essentials_category=self.essentials_category,
         )
         return info
 
@@ -2070,11 +2064,74 @@ class _UIOutput(ABC):
         ...
 
 
+class InputMapOldId(TypedDict):
+    """Map an old node input to a new node input by ID."""
+    new_id: str
+    old_id: str
+
+class InputMapSetValue(TypedDict):
+    """Set a specific value for a new node input."""
+    new_id: str
+    set_value: Any
+
+InputMap = InputMapOldId | InputMapSetValue
+"""
+Input mapping for node replacement. Type is inferred by dictionary keys:
+- {"new_id": str, "old_id": str} - maps old input to new input
+- {"new_id": str, "set_value": Any} - sets a specific value for new input
+"""
+
+class OutputMap(TypedDict):
+    """Map outputs of node replacement via indexes."""
+    new_idx: int
+    old_idx: int
+
+class NodeReplace:
+    """
+    Defines a possible node replacement, mapping inputs and outputs of the old node to the new node.
+
+    Also supports assigning specific values to the input widgets of the new node.
+
+    Args:
+        new_node_id: The class name of the new replacement node.
+        old_node_id: The class name of the deprecated node.
+        old_widget_ids: Ordered list of input IDs for widgets that may not have an input slot
+            connected. The workflow JSON stores widget values by their relative position index,
+            not by ID. This list maps those positional indexes to input IDs, enabling the
+            replacement system to correctly identify widget values during node migration.
+        input_mapping: List of input mappings from old node to new node.
+        output_mapping: List of output mappings from old node to new node.
+    """
+    def __init__(self,
+        new_node_id: str,
+        old_node_id: str,
+        old_widget_ids: list[str] | None=None,
+        input_mapping: list[InputMap] | None=None,
+        output_mapping: list[OutputMap] | None=None,
+    ):
+        self.new_node_id = new_node_id
+        self.old_node_id = old_node_id
+        self.old_widget_ids = old_widget_ids
+        self.input_mapping = input_mapping
+        self.output_mapping = output_mapping
+
+    def as_dict(self):
+        """Create serializable representation of the node replacement."""
+        return {
+            "new_node_id": self.new_node_id,
+            "old_node_id": self.old_node_id,
+            "old_widget_ids": self.old_widget_ids,
+            "input_mapping": list(self.input_mapping) if self.input_mapping else None,
+            "output_mapping": list(self.output_mapping) if self.output_mapping else None,
+        }
+
+
 __all__ = [
     "FolderType",
     "UploadType",
     "RemoteOptions",
     "NumberDisplay",
+    "ControlAfterGenerate",
 
     "comfytype",
     "Custom",
@@ -2150,7 +2207,6 @@ __all__ = [
     "DynamicCombo",
     "Autogrow",
     # Other classes
-    "EmptyInputSentinel",
     "HiddenHolder",
     "Hidden",
     "NodeInfoV1",
@@ -2162,4 +2218,6 @@ __all__ = [
     "ImageCompare",
     "PriceBadgeDepends",
     "PriceBadge",
+    "BoundingBox",
+    "NodeReplace",
 ]

comfy_api_nodes/apis/bria.py

@@ -45,17 +45,55 @@ class BriaEditImageRequest(BaseModel):
     )
 
 
+class BriaRemoveBackgroundRequest(BaseModel):
+    image: str = Field(...)
+    sync: bool = Field(False)
+    visual_input_content_moderation: bool = Field(
+        False, description="If true, returns 422 on input image moderation failure."
+    )
+    visual_output_content_moderation: bool = Field(
+        False, description="If true, returns 422 on visual output moderation failure."
+    )
+    seed: int = Field(...)
+
+
 class BriaStatusResponse(BaseModel):
     request_id: str = Field(...)
     status_url: str = Field(...)
     warning: str | None = Field(None)
 
 
-class BriaResult(BaseModel):
+class BriaRemoveBackgroundResult(BaseModel):
+    image_url: str = Field(...)
+
+
+class BriaRemoveBackgroundResponse(BaseModel):
+    status: str = Field(...)
+    result: BriaRemoveBackgroundResult | None = Field(None)
+
+
+class BriaImageEditResult(BaseModel):
     structured_prompt: str = Field(...)
     image_url: str = Field(...)
 
 
-class BriaResponse(BaseModel):
+class BriaImageEditResponse(BaseModel):
     status: str = Field(...)
-    result: BriaResult | None = Field(None)
+    result: BriaImageEditResult | None = Field(None)
+
+
+class BriaRemoveVideoBackgroundRequest(BaseModel):
+    video: str = Field(...)
+    background_color: str = Field(default="transparent", description="Background color for the output video.")
+    output_container_and_codec: str = Field(...)
+    preserve_audio: bool = Field(True)
+    seed: int = Field(...)
+
+
+class BriaRemoveVideoBackgroundResult(BaseModel):
+    video_url: str = Field(...)
+
+
+class BriaRemoveVideoBackgroundResponse(BaseModel):
+    status: str = Field(...)
+    result: BriaRemoveVideoBackgroundResult | None = Field(None)

comfy_api_nodes/apis/elevenlabs.py

@@ -0,0 +1,88 @@
+from pydantic import BaseModel, Field
+
+
+class SpeechToTextRequest(BaseModel):
+    model_id: str = Field(...)
+    cloud_storage_url: str = Field(...)
+    language_code: str | None = Field(None, description="ISO-639-1 or ISO-639-3 language code")
+    tag_audio_events: bool | None = Field(None, description="Annotate sounds like (laughter) in transcript")
+    num_speakers: int | None = Field(None, description="Max speakers predicted")
+    timestamps_granularity: str = Field(default="word", description="Timing precision: none, word, or character")
+    diarize: bool | None = Field(None, description="Annotate which speaker is talking")
+    diarization_threshold: float | None = Field(None, description="Speaker separation sensitivity")
+    temperature: float | None = Field(None, description="Randomness control")
+    seed: int = Field(..., description="Seed for deterministic sampling")
+
+
+class SpeechToTextWord(BaseModel):
+    text: str = Field(..., description="The word text")
+    type: str = Field(default="word", description="Type of text element (word, spacing, etc.)")
+    start: float | None = Field(None, description="Start time in seconds (when timestamps enabled)")
+    end: float | None = Field(None, description="End time in seconds (when timestamps enabled)")
+    speaker_id: str | None = Field(None, description="Speaker identifier when diarization is enabled")
+    logprob: float | None = Field(None, description="Log probability of the word")
+
+
+class SpeechToTextResponse(BaseModel):
+    language_code: str = Field(..., description="Detected or specified language code")
+    language_probability: float | None = Field(None, description="Confidence of language detection")
+    text: str = Field(..., description="Full transcript text")
+    words: list[SpeechToTextWord] | None = Field(None, description="Word-level timing information")
+
+
+class TextToSpeechVoiceSettings(BaseModel):
+    stability: float | None = Field(None, description="Voice stability")
+    similarity_boost: float | None = Field(None, description="Similarity boost")
+    style: float | None = Field(None, description="Style exaggeration")
+    use_speaker_boost: bool | None = Field(None, description="Boost similarity to original speaker")
+    speed: float | None = Field(None, description="Speech speed")
+
+
+class TextToSpeechRequest(BaseModel):
+    text: str = Field(..., description="Text to convert to speech")
+    model_id: str = Field(..., description="Model ID for TTS")
+    language_code: str | None = Field(None, description="ISO-639-1 or ISO-639-3 language code")
+    voice_settings: TextToSpeechVoiceSettings | None = Field(None, description="Voice settings")
+    seed: int = Field(..., description="Seed for deterministic sampling")
+    apply_text_normalization: str | None = Field(None, description="Text normalization mode: auto, on, off")
+
+
+class TextToSoundEffectsRequest(BaseModel):
+    text: str = Field(..., description="Text prompt to convert into a sound effect")
+    duration_seconds: float = Field(..., description="Duration of generated sound in seconds")
+    prompt_influence: float = Field(..., description="How closely generation follows the prompt")
+    loop: bool | None = Field(None, description="Whether to create a smoothly looping sound effect")
+
+
+class AddVoiceRequest(BaseModel):
+    name: str = Field(..., description="Name that identifies the voice")
+    remove_background_noise: bool = Field(..., description="Remove background noise from voice samples")
+
+
+class AddVoiceResponse(BaseModel):
+    voice_id: str = Field(..., description="The newly created voice's unique identifier")
+
+
+class SpeechToSpeechRequest(BaseModel):
+    model_id: str = Field(..., description="Model ID for speech-to-speech")
+    voice_settings: str = Field(..., description="JSON string of voice settings")
+    seed: int = Field(..., description="Seed for deterministic sampling")
+    remove_background_noise: bool = Field(..., description="Remove background noise from input audio")
+
+
+class DialogueInput(BaseModel):
+    text: str = Field(..., description="Text content to convert to speech")
+    voice_id: str = Field(..., description="Voice identifier for this dialogue segment")
+
+
+class DialogueSettings(BaseModel):
+    stability: float | None = Field(None, description="Voice stability (0-1)")
+
+
+class TextToDialogueRequest(BaseModel):
+    inputs: list[DialogueInput] = Field(..., description="List of dialogue segments")
+    model_id: str = Field(..., description="Model ID for dialogue generation")
+    language_code: str | None = Field(None, description="ISO-639-1 language code")
+    settings: DialogueSettings | None = Field(None, description="Voice settings")
+    seed: int | None = Field(None, description="Seed for deterministic sampling")
+    apply_text_normalization: str | None = Field(None, description="Text normalization mode: auto, on, off")

comfy_api_nodes/apis/gemini.py

@@ -116,9 +116,15 @@ class GeminiGenerationConfig(BaseModel):
     topP: float | None = Field(None, ge=0.0, le=1.0)
 
 
+class GeminiImageOutputOptions(BaseModel):
+    mimeType: str = Field("image/png")
+    compressionQuality: int | None = Field(None)
+
+
 class GeminiImageConfig(BaseModel):
     aspectRatio: str | None = Field(None)
     imageSize: str | None = Field(None)
+    imageOutputOptions: GeminiImageOutputOptions = Field(default_factory=GeminiImageOutputOptions)
 
 
 class GeminiImageGenerationConfig(GeminiGenerationConfig):

comfy_api_nodes/apis/hunyuan3d.py

@@ -64,3 +64,23 @@ class To3DProTaskResultResponse(BaseModel):
 
 class To3DProTaskQueryRequest(BaseModel):
     JobId: str = Field(...)
+
+
+class To3DUVFileInput(BaseModel):
+    Type: str = Field(..., description="File type: GLB, OBJ, or FBX")
+    Url: str = Field(...)
+
+
+class To3DUVTaskRequest(BaseModel):
+    File: To3DUVFileInput = Field(...)
+
+
+class TextureEditImageInfo(BaseModel):
+    Url: str = Field(...)
+
+
+class TextureEditTaskRequest(BaseModel):
+    File3D: To3DUVFileInput = Field(...)
+    Image: TextureEditImageInfo | None = Field(None)
+    Prompt: str | None = Field(None)
+    EnablePBR: bool | None = Field(None)

comfy_api_nodes/apis/recraft.py

@@ -198,11 +198,6 @@ dict_recraft_substyles_v3 = {
 }
 
 
-class RecraftModel(str, Enum):
-    recraftv3 = 'recraftv3'
-    recraftv2 = 'recraftv2'
-
-
 class RecraftImageSize(str, Enum):
     res_1024x1024 = '1024x1024'
     res_1365x1024 = '1365x1024'
@@ -221,6 +216,41 @@ class RecraftImageSize(str, Enum):
     res_1707x1024 = '1707x1024'
 
 
+RECRAFT_V4_SIZES = [
+    "1024x1024",
+    "1536x768",
+    "768x1536",
+    "1280x832",
+    "832x1280",
+    "1216x896",
+    "896x1216",
+    "1152x896",
+    "896x1152",
+    "832x1344",
+    "1280x896",
+    "896x1280",
+    "1344x768",
+    "768x1344",
+]
+
+RECRAFT_V4_PRO_SIZES = [
+    "2048x2048",
+    "3072x1536",
+    "1536x3072",
+    "2560x1664",
+    "1664x2560",
+    "2432x1792",
+    "1792x2432",
+    "2304x1792",
+    "1792x2304",
+    "1664x2688",
+    "1434x1024",
+    "1024x1434",
+    "2560x1792",
+    "1792x2560",
+]
+
+
 class RecraftColorObject(BaseModel):
     rgb: list[int] = Field(..., description='An array of 3 integer values in range of 0...255 defining RGB Color Model')
 
@@ -234,17 +264,16 @@ class RecraftControlsObject(BaseModel):
 
 class RecraftImageGenerationRequest(BaseModel):
     prompt: str = Field(..., description='The text prompt describing the image to generate')
-    size: RecraftImageSize | None = Field(None, description='The size of the generated image (e.g., "1024x1024")')
+    size: str | None = Field(None, description='The size of the generated image (e.g., "1024x1024")')
     n: int = Field(..., description='The number of images to generate')
     negative_prompt: str | None = Field(None, description='A text description of undesired elements on an image')
-    model: RecraftModel | None = Field(RecraftModel.recraftv3, description='The model to use for generation (e.g., "recraftv3")')
+    model: str = Field(...)
     style: str | None = Field(None, description='The style to apply to the generated image (e.g., "digital_illustration")')
     substyle: str | None = Field(None, description='The substyle to apply to the generated image, depending on the style input')
     controls: RecraftControlsObject | None = Field(None, description='A set of custom parameters to tweak generation process')
     style_id: str | None = Field(None, description='Use a previously uploaded style as a reference; UUID')
     strength: float | None = Field(None, description='Defines the difference with the original image, should lie in [0, 1], where 0 means almost identical, and 1 means miserable similarity')
     random_seed: int | None = Field(None, description="Seed for video generation")
-    # text_layout
 
 
 class RecraftReturnedObject(BaseModel):

comfy_api_nodes/nodes_bfl.py

@@ -57,6 +57,7 @@ class FluxProUltraImageNode(IO.ComfyNode):
                     tooltip="Whether to perform upsampling on the prompt. "
                     "If active, automatically modifies the prompt for more creative generation, "
                     "but results are nondeterministic (same seed will not produce exactly the same result).",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",
@@ -200,6 +201,7 @@ class FluxKontextProImageNode(IO.ComfyNode):
                     "prompt_upsampling",
                     default=False,
                     tooltip="Whether to perform upsampling on the prompt. If active, automatically modifies the prompt for more creative generation, but results are nondeterministic (same seed will not produce exactly the same result).",
+                    advanced=True,
                 ),
                 IO.Image.Input(
                     "input_image",
@@ -296,6 +298,7 @@ class FluxProExpandNode(IO.ComfyNode):
                     tooltip="Whether to perform upsampling on the prompt. "
                     "If active, automatically modifies the prompt for more creative generation, "
                     "but results are nondeterministic (same seed will not produce exactly the same result).",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "top",
@@ -433,6 +436,7 @@ class FluxProFillNode(IO.ComfyNode):
                     tooltip="Whether to perform upsampling on the prompt. "
                     "If active, automatically modifies the prompt for more creative generation, "
                     "but results are nondeterministic (same seed will not produce exactly the same result).",
+                    advanced=True,
                 ),
                 IO.Float.Input(
                     "guidance",
@@ -577,6 +581,7 @@ class Flux2ProImageNode(IO.ComfyNode):
                     default=True,
                     tooltip="Whether to perform upsampling on the prompt. "
                     "If active, automatically modifies the prompt for more creative generation.",
+                    advanced=True,
                 ),
                 IO.Image.Input("images", optional=True, tooltip="Up to 9 images to be used as references."),
             ],

comfy_api_nodes/nodes_bria.py

@@ -3,7 +3,11 @@ from typing_extensions import override
 from comfy_api.latest import IO, ComfyExtension, Input
 from comfy_api_nodes.apis.bria import (
     BriaEditImageRequest,
-    BriaResponse,
+    BriaRemoveBackgroundRequest,
+    BriaRemoveBackgroundResponse,
+    BriaRemoveVideoBackgroundRequest,
+    BriaRemoveVideoBackgroundResponse,
+    BriaImageEditResponse,
     BriaStatusResponse,
     InputModerationSettings,
 )
@@ -11,10 +15,12 @@ from comfy_api_nodes.util import (
     ApiEndpoint,
     convert_mask_to_image,
     download_url_to_image_tensor,
-    get_number_of_images,
+    download_url_to_video_output,
     poll_op,
     sync_op,
-    upload_images_to_comfyapi,
+    upload_image_to_comfyapi,
+    upload_video_to_comfyapi,
+    validate_video_duration,
 )
 
 
@@ -73,21 +79,15 @@ class BriaImageEditNode(IO.ComfyNode):
                 IO.DynamicCombo.Input(
                     "moderation",
                     options=[
+                        IO.DynamicCombo.Option("false", []),
                         IO.DynamicCombo.Option(
                             "true",
                             [
-                                IO.Boolean.Input(
-                                    "prompt_content_moderation", default=False
-                                ),
-                                IO.Boolean.Input(
-                                    "visual_input_moderation", default=False
-                                ),
-                                IO.Boolean.Input(
-                                    "visual_output_moderation", default=True
-                                ),
+                                IO.Boolean.Input("prompt_content_moderation", default=False),
+                                IO.Boolean.Input("visual_input_moderation", default=False),
+                                IO.Boolean.Input("visual_output_moderation", default=True),
                             ],
                         ),
-                        IO.DynamicCombo.Option("false", []),
                     ],
                     tooltip="Moderation settings",
                 ),
@@ -127,50 +127,26 @@ class BriaImageEditNode(IO.ComfyNode):
         mask: Input.Image | None = None,
     ) -> IO.NodeOutput:
         if not prompt and not structured_prompt:
-            raise ValueError(
-                "One of prompt or structured_prompt is required to be non-empty."
-            )
-        if get_number_of_images(image) != 1:
-            raise ValueError("Exactly one input image is required.")
+            raise ValueError("One of prompt or structured_prompt is required to be non-empty.")
         mask_url = None
         if mask is not None:
-            mask_url = (
-                await upload_images_to_comfyapi(
-                    cls,
-                    convert_mask_to_image(mask),
-                    max_images=1,
-                    mime_type="image/png",
-                    wait_label="Uploading mask",
-                )
-            )[0]
+            mask_url = await upload_image_to_comfyapi(cls, convert_mask_to_image(mask), wait_label="Uploading mask")
         response = await sync_op(
             cls,
             ApiEndpoint(path="proxy/bria/v2/image/edit", method="POST"),
             data=BriaEditImageRequest(
                 instruction=prompt if prompt else None,
                 structured_instruction=structured_prompt if structured_prompt else None,
-                images=await upload_images_to_comfyapi(
-                    cls,
-                    image,
-                    max_images=1,
-                    mime_type="image/png",
-                    wait_label="Uploading image",
-                ),
+                images=[await upload_image_to_comfyapi(cls, image, wait_label="Uploading image")],
                 mask=mask_url,
                 negative_prompt=negative_prompt if negative_prompt else None,
                 guidance_scale=guidance_scale,
                 seed=seed,
                 model_version=model,
                 steps_num=steps,
-                prompt_content_moderation=moderation.get(
-                    "prompt_content_moderation", False
-                ),
-                visual_input_content_moderation=moderation.get(
-                    "visual_input_moderation", False
-                ),
-                visual_output_content_moderation=moderation.get(
-                    "visual_output_moderation", False
-                ),
+                prompt_content_moderation=moderation.get("prompt_content_moderation", False),
+                visual_input_content_moderation=moderation.get("visual_input_moderation", False),
+                visual_output_content_moderation=moderation.get("visual_output_moderation", False),
             ),
             response_model=BriaStatusResponse,
         )
@@ -178,7 +154,7 @@ class BriaImageEditNode(IO.ComfyNode):
             cls,
             ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"),
             status_extractor=lambda r: r.status,
-            response_model=BriaResponse,
+            response_model=BriaImageEditResponse,
         )
         return IO.NodeOutput(
             await download_url_to_image_tensor(response.result.image_url),
@@ -186,11 +162,167 @@ class BriaImageEditNode(IO.ComfyNode):
         )
 
 
+class BriaRemoveImageBackground(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="BriaRemoveImageBackground",
+            display_name="Bria Remove Image Background",
+            category="api node/image/Bria",
+            description="Remove the background from an image using Bria RMBG 2.0.",
+            inputs=[
+                IO.Image.Input("image"),
+                IO.DynamicCombo.Input(
+                    "moderation",
+                    options=[
+                        IO.DynamicCombo.Option("false", []),
+                        IO.DynamicCombo.Option(
+                            "true",
+                            [
+                                IO.Boolean.Input("visual_input_moderation", default=False),
+                                IO.Boolean.Input("visual_output_moderation", default=True),
+                            ],
+                        ),
+                    ],
+                    tooltip="Moderation settings",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+            ],
+            outputs=[IO.Image.Output()],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.018}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        image: Input.Image,
+        moderation: dict,
+        seed: int,
+    ) -> IO.NodeOutput:
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/bria/v2/image/edit/remove_background", method="POST"),
+            data=BriaRemoveBackgroundRequest(
+                image=await upload_image_to_comfyapi(cls, image, wait_label="Uploading image"),
+                sync=False,
+                visual_input_content_moderation=moderation.get("visual_input_moderation", False),
+                visual_output_content_moderation=moderation.get("visual_output_moderation", False),
+                seed=seed,
+            ),
+            response_model=BriaStatusResponse,
+        )
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"),
+            status_extractor=lambda r: r.status,
+            response_model=BriaRemoveBackgroundResponse,
+        )
+        return IO.NodeOutput(await download_url_to_image_tensor(response.result.image_url))
+
+
+class BriaRemoveVideoBackground(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="BriaRemoveVideoBackground",
+            display_name="Bria Remove Video Background",
+            category="api node/video/Bria",
+            description="Remove the background from a video using Bria. ",
+            inputs=[
+                IO.Video.Input("video"),
+                IO.Combo.Input(
+                    "background_color",
+                    options=[
+                        "Black",
+                        "White",
+                        "Gray",
+                        "Red",
+                        "Green",
+                        "Blue",
+                        "Yellow",
+                        "Cyan",
+                        "Magenta",
+                        "Orange",
+                    ],
+                    tooltip="Background color for the output video.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+            ],
+            outputs=[IO.Video.Output()],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.14,"format":{"suffix":"/second"}}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        video: Input.Video,
+        background_color: str,
+        seed: int,
+    ) -> IO.NodeOutput:
+        validate_video_duration(video, max_duration=60.0)
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/bria/v2/video/edit/remove_background", method="POST"),
+            data=BriaRemoveVideoBackgroundRequest(
+                video=await upload_video_to_comfyapi(cls, video),
+                background_color=background_color,
+                output_container_and_codec="mp4_h264",
+                seed=seed,
+            ),
+            response_model=BriaStatusResponse,
+        )
+        response = await poll_op(
+            cls,
+            ApiEndpoint(path=f"/proxy/bria/v2/status/{response.request_id}"),
+            status_extractor=lambda r: r.status,
+            response_model=BriaRemoveVideoBackgroundResponse,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(response.result.video_url))
+
+
 class BriaExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
         return [
             BriaImageEditNode,
+            BriaRemoveImageBackground,
+            BriaRemoveVideoBackground,
         ]
 
 

comfy_api_nodes/nodes_bytedance.py

@@ -114,6 +114,7 @@ class ByteDanceImageNode(IO.ComfyNode):
                     default=False,
                     tooltip='Whether to add an "AI generated" watermark to the image',
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -259,12 +260,14 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                     default=False,
                     tooltip='Whether to add an "AI generated" watermark to the image.',
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "fail_on_partial",
                     default=True,
                     tooltip="If enabled, abort execution if any requested images are missing or return an error.",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -432,18 +435,21 @@ class ByteDanceTextToVideoNode(IO.ComfyNode):
                     tooltip="Specifies whether to fix the camera. The platform appends an instruction "
                     "to fix the camera to your prompt, but does not guarantee the actual effect.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "watermark",
                     default=False,
                     tooltip='Whether to add an "AI generated" watermark to the video.',
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "generate_audio",
                     default=False,
                     tooltip="This parameter is ignored for any model except seedance-1-5-pro.",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -561,18 +567,21 @@ class ByteDanceImageToVideoNode(IO.ComfyNode):
                     tooltip="Specifies whether to fix the camera. The platform appends an instruction "
                     "to fix the camera to your prompt, but does not guarantee the actual effect.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "watermark",
                     default=False,
                     tooltip='Whether to add an "AI generated" watermark to the video.',
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "generate_audio",
                     default=False,
                     tooltip="This parameter is ignored for any model except seedance-1-5-pro.",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -694,18 +703,21 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode):
                     tooltip="Specifies whether to fix the camera. The platform appends an instruction "
                     "to fix the camera to your prompt, but does not guarantee the actual effect.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "watermark",
                     default=False,
                     tooltip='Whether to add an "AI generated" watermark to the video.',
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "generate_audio",
                     default=False,
                     tooltip="This parameter is ignored for any model except seedance-1-5-pro.",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -834,6 +846,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
                     default=False,
                     tooltip='Whether to add an "AI generated" watermark to the video.',
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[

comfy_api_nodes/nodes_elevenlabs.py

@@ -0,0 +1,924 @@
+import json
+import uuid
+
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api_nodes.apis.elevenlabs import (
+    AddVoiceRequest,
+    AddVoiceResponse,
+    DialogueInput,
+    DialogueSettings,
+    SpeechToSpeechRequest,
+    SpeechToTextRequest,
+    SpeechToTextResponse,
+    TextToDialogueRequest,
+    TextToSoundEffectsRequest,
+    TextToSpeechRequest,
+    TextToSpeechVoiceSettings,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    audio_bytes_to_audio_input,
+    audio_ndarray_to_bytesio,
+    audio_tensor_to_contiguous_ndarray,
+    sync_op,
+    sync_op_raw,
+    upload_audio_to_comfyapi,
+    validate_string,
+)
+
+ELEVENLABS_MUSIC_SECTIONS = "ELEVENLABS_MUSIC_SECTIONS"  # Custom type for music sections
+ELEVENLABS_COMPOSITION_PLAN = "ELEVENLABS_COMPOSITION_PLAN"  # Custom type for composition plan
+ELEVENLABS_VOICE = "ELEVENLABS_VOICE"  # Custom type for voice selection
+
+# Predefined ElevenLabs voices: (voice_id, display_name, gender, accent)
+ELEVENLABS_VOICES = [
+    ("CwhRBWXzGAHq8TQ4Fs17", "Roger", "male", "american"),
+    ("EXAVITQu4vr4xnSDxMaL", "Sarah", "female", "american"),
+    ("FGY2WhTYpPnrIDTdsKH5", "Laura", "female", "american"),
+    ("IKne3meq5aSn9XLyUdCD", "Charlie", "male", "australian"),
+    ("JBFqnCBsd6RMkjVDRZzb", "George", "male", "british"),
+    ("N2lVS1w4EtoT3dr4eOWO", "Callum", "male", "american"),
+    ("SAz9YHcvj6GT2YYXdXww", "River", "neutral", "american"),
+    ("SOYHLrjzK2X1ezoPC6cr", "Harry", "male", "american"),
+    ("TX3LPaxmHKxFdv7VOQHJ", "Liam", "male", "american"),
+    ("Xb7hH8MSUJpSbSDYk0k2", "Alice", "female", "british"),
+    ("XrExE9yKIg1WjnnlVkGX", "Matilda", "female", "american"),
+    ("bIHbv24MWmeRgasZH58o", "Will", "male", "american"),
+    ("cgSgspJ2msm6clMCkdW9", "Jessica", "female", "american"),
+    ("cjVigY5qzO86Huf0OWal", "Eric", "male", "american"),
+    ("hpp4J3VqNfWAUOO0d1Us", "Bella", "female", "american"),
+    ("iP95p4xoKVk53GoZ742B", "Chris", "male", "american"),
+    ("nPczCjzI2devNBz1zQrb", "Brian", "male", "american"),
+    ("onwK4e9ZLuTAKqWW03F9", "Daniel", "male", "british"),
+    ("pFZP5JQG7iQjIQuC4Bku", "Lily", "female", "british"),
+    ("pNInz6obpgDQGcFmaJgB", "Adam", "male", "american"),
+    ("pqHfZKP75CvOlQylNhV4", "Bill", "male", "american"),
+]
+
+ELEVENLABS_VOICE_OPTIONS = [f"{name} ({gender}, {accent})" for _, name, gender, accent in ELEVENLABS_VOICES]
+ELEVENLABS_VOICE_MAP = {
+    f"{name} ({gender}, {accent})": voice_id for voice_id, name, gender, accent in ELEVENLABS_VOICES
+}
+
+
+class ElevenLabsSpeechToText(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ElevenLabsSpeechToText",
+            display_name="ElevenLabs Speech to Text",
+            category="api node/audio/ElevenLabs",
+            description="Transcribe audio to text. "
+            "Supports automatic language detection, speaker diarization, and audio event tagging.",
+            inputs=[
+                IO.Audio.Input(
+                    "audio",
+                    tooltip="Audio to transcribe.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "scribe_v2",
+                            [
+                                IO.Boolean.Input(
+                                    "tag_audio_events",
+                                    default=False,
+                                    tooltip="Annotate sounds like (laughter), (music), etc. in transcript.",
+                                ),
+                                IO.Boolean.Input(
+                                    "diarize",
+                                    default=False,
+                                    tooltip="Annotate which speaker is talking.",
+                                ),
+                                IO.Float.Input(
+                                    "diarization_threshold",
+                                    default=0.22,
+                                    min=0.1,
+                                    max=0.4,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Speaker separation sensitivity. "
+                                    "Lower values are more sensitive to speaker changes.",
+                                ),
+                                IO.Float.Input(
+                                    "temperature",
+                                    default=0.0,
+                                    min=0.0,
+                                    max=2.0,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Randomness control. "
+                                    "0.0 uses model default. Higher values increase randomness.",
+                                ),
+                                IO.Combo.Input(
+                                    "timestamps_granularity",
+                                    options=["word", "character", "none"],
+                                    default="word",
+                                    tooltip="Timing precision for transcript words.",
+                                ),
+                            ],
+                        ),
+                    ],
+                    tooltip="Model to use for transcription.",
+                ),
+                IO.String.Input(
+                    "language_code",
+                    default="",
+                    tooltip="ISO-639-1 or ISO-639-3 language code (e.g., 'en', 'es', 'fra'). "
+                    "Leave empty for automatic detection.",
+                ),
+                IO.Int.Input(
+                    "num_speakers",
+                    default=0,
+                    min=0,
+                    max=32,
+                    display_mode=IO.NumberDisplay.slider,
+                    tooltip="Maximum number of speakers to predict. Set to 0 for automatic detection.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=1,
+                    min=0,
+                    max=2147483647,
+                    tooltip="Seed for reproducibility (determinism not guaranteed).",
+                ),
+            ],
+            outputs=[
+                IO.String.Output(display_name="text"),
+                IO.String.Output(display_name="language_code"),
+                IO.String.Output(display_name="words_json"),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.0073,"format":{"approximate":true,"suffix":"/minute"}}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        audio: Input.Audio,
+        model: dict,
+        language_code: str,
+        num_speakers: int,
+        seed: int,
+    ) -> IO.NodeOutput:
+        if model["diarize"] and num_speakers:
+            raise ValueError(
+                "Number of speakers cannot be specified when diarization is enabled. "
+                "Either disable diarization or set num_speakers to 0."
+            )
+        request = SpeechToTextRequest(
+            model_id=model["model"],
+            cloud_storage_url=await upload_audio_to_comfyapi(
+                cls, audio, container_format="mp4", codec_name="aac", mime_type="audio/mp4"
+            ),
+            language_code=language_code if language_code.strip() else None,
+            tag_audio_events=model["tag_audio_events"],
+            num_speakers=num_speakers if num_speakers > 0 else None,
+            timestamps_granularity=model["timestamps_granularity"],
+            diarize=model["diarize"],
+            diarization_threshold=model["diarization_threshold"] if model["diarize"] else None,
+            seed=seed,
+            temperature=model["temperature"],
+        )
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/elevenlabs/v1/speech-to-text", method="POST"),
+            response_model=SpeechToTextResponse,
+            data=request,
+            content_type="multipart/form-data",
+        )
+        words_json = json.dumps(
+            [w.model_dump(exclude_none=True) for w in response.words] if response.words else [],
+            indent=2,
+        )
+        return IO.NodeOutput(response.text, response.language_code, words_json)
+
+
+class ElevenLabsVoiceSelector(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ElevenLabsVoiceSelector",
+            display_name="ElevenLabs Voice Selector",
+            category="api node/audio/ElevenLabs",
+            description="Select a predefined ElevenLabs voice for text-to-speech generation.",
+            inputs=[
+                IO.Combo.Input(
+                    "voice",
+                    options=ELEVENLABS_VOICE_OPTIONS,
+                    tooltip="Choose a voice from the predefined ElevenLabs voices.",
+                ),
+            ],
+            outputs=[
+                IO.Custom(ELEVENLABS_VOICE).Output(display_name="voice"),
+            ],
+            is_api_node=False,
+        )
+
+    @classmethod
+    def execute(cls, voice: str) -> IO.NodeOutput:
+        voice_id = ELEVENLABS_VOICE_MAP.get(voice)
+        if not voice_id:
+            raise ValueError(f"Unknown voice: {voice}")
+        return IO.NodeOutput(voice_id)
+
+
+class ElevenLabsTextToSpeech(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ElevenLabsTextToSpeech",
+            display_name="ElevenLabs Text to Speech",
+            category="api node/audio/ElevenLabs",
+            description="Convert text to speech.",
+            inputs=[
+                IO.Custom(ELEVENLABS_VOICE).Input(
+                    "voice",
+                    tooltip="Voice to use for speech synthesis. Connect from Voice Selector or Instant Voice Clone.",
+                ),
+                IO.String.Input(
+                    "text",
+                    multiline=True,
+                    default="",
+                    tooltip="The text to convert to speech.",
+                ),
+                IO.Float.Input(
+                    "stability",
+                    default=0.5,
+                    min=0.0,
+                    max=1.0,
+                    step=0.01,
+                    display_mode=IO.NumberDisplay.slider,
+                    tooltip="Voice stability. Lower values give broader emotional range, "
+                    "higher values produce more consistent but potentially monotonous speech.",
+                ),
+                IO.Combo.Input(
+                    "apply_text_normalization",
+                    options=["auto", "on", "off"],
+                    tooltip="Text normalization mode. 'auto' lets the system decide, "
+                    "'on' always applies normalization, 'off' skips it.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "eleven_multilingual_v2",
+                            [
+                                IO.Float.Input(
+                                    "speed",
+                                    default=1.0,
+                                    min=0.7,
+                                    max=1.3,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Speech speed. 1.0 is normal, <1.0 slower, >1.0 faster.",
+                                ),
+                                IO.Float.Input(
+                                    "similarity_boost",
+                                    default=0.75,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Similarity boost. Higher values make the voice more similar to the original.",
+                                ),
+                                IO.Boolean.Input(
+                                    "use_speaker_boost",
+                                    default=False,
+                                    tooltip="Boost similarity to the original speaker voice.",
+                                ),
+                                IO.Float.Input(
+                                    "style",
+                                    default=0.0,
+                                    min=0.0,
+                                    max=0.2,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Style exaggeration. Higher values increase stylistic expression "
+                                    "but may reduce stability.",
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "eleven_v3",
+                            [
+                                IO.Float.Input(
+                                    "speed",
+                                    default=1.0,
+                                    min=0.7,
+                                    max=1.3,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Speech speed. 1.0 is normal, <1.0 slower, >1.0 faster.",
+                                ),
+                                IO.Float.Input(
+                                    "similarity_boost",
+                                    default=0.75,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Similarity boost. Higher values make the voice more similar to the original.",
+                                ),
+                            ],
+                        ),
+                    ],
+                    tooltip="Model to use for text-to-speech.",
+                ),
+                IO.String.Input(
+                    "language_code",
+                    default="",
+                    tooltip="ISO-639-1 or ISO-639-3 language code (e.g., 'en', 'es', 'fra'). "
+                    "Leave empty for automatic detection.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=1,
+                    min=0,
+                    max=2147483647,
+                    tooltip="Seed for reproducibility (determinism not guaranteed).",
+                ),
+                IO.Combo.Input(
+                    "output_format",
+                    options=["mp3_44100_192", "opus_48000_192"],
+                    tooltip="Audio output format.",
+                ),
+            ],
+            outputs=[
+                IO.Audio.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.24,"format":{"approximate":true,"suffix":"/1K chars"}}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        voice: str,
+        text: str,
+        stability: float,
+        apply_text_normalization: str,
+        model: dict,
+        language_code: str,
+        seed: int,
+        output_format: str,
+    ) -> IO.NodeOutput:
+        validate_string(text, min_length=1)
+        request = TextToSpeechRequest(
+            text=text,
+            model_id=model["model"],
+            language_code=language_code if language_code.strip() else None,
+            voice_settings=TextToSpeechVoiceSettings(
+                stability=stability,
+                similarity_boost=model["similarity_boost"],
+                speed=model["speed"],
+                use_speaker_boost=model.get("use_speaker_boost", None),
+                style=model.get("style", None),
+            ),
+            seed=seed,
+            apply_text_normalization=apply_text_normalization,
+        )
+        response = await sync_op_raw(
+            cls,
+            ApiEndpoint(
+                path=f"/proxy/elevenlabs/v1/text-to-speech/{voice}",
+                method="POST",
+                query_params={"output_format": output_format},
+            ),
+            data=request,
+            as_binary=True,
+        )
+        return IO.NodeOutput(audio_bytes_to_audio_input(response))
+
+
+class ElevenLabsAudioIsolation(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ElevenLabsAudioIsolation",
+            display_name="ElevenLabs Voice Isolation",
+            category="api node/audio/ElevenLabs",
+            description="Remove background noise from audio, isolating vocals or speech.",
+            inputs=[
+                IO.Audio.Input(
+                    "audio",
+                    tooltip="Audio to process for background noise removal.",
+                ),
+            ],
+            outputs=[
+                IO.Audio.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.24,"format":{"approximate":true,"suffix":"/minute"}}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        audio: Input.Audio,
+    ) -> IO.NodeOutput:
+        audio_data_np = audio_tensor_to_contiguous_ndarray(audio["waveform"])
+        audio_bytes_io = audio_ndarray_to_bytesio(audio_data_np, audio["sample_rate"], "mp4", "aac")
+        response = await sync_op_raw(
+            cls,
+            ApiEndpoint(path="/proxy/elevenlabs/v1/audio-isolation", method="POST"),
+            files={"audio": ("audio.mp4", audio_bytes_io, "audio/mp4")},
+            content_type="multipart/form-data",
+            as_binary=True,
+        )
+        return IO.NodeOutput(audio_bytes_to_audio_input(response))
+
+
+class ElevenLabsTextToSoundEffects(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ElevenLabsTextToSoundEffects",
+            display_name="ElevenLabs Text to Sound Effects",
+            category="api node/audio/ElevenLabs",
+            description="Generate sound effects from text descriptions.",
+            inputs=[
+                IO.String.Input(
+                    "text",
+                    multiline=True,
+                    default="",
+                    tooltip="Text description of the sound effect to generate.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "eleven_sfx_v2",
+                            [
+                                IO.Float.Input(
+                                    "duration",
+                                    default=5.0,
+                                    min=0.5,
+                                    max=30.0,
+                                    step=0.1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Duration of generated sound in seconds.",
+                                ),
+                                IO.Boolean.Input(
+                                    "loop",
+                                    default=False,
+                                    tooltip="Create a smoothly looping sound effect.",
+                                ),
+                                IO.Float.Input(
+                                    "prompt_influence",
+                                    default=0.3,
+                                    min=0.0,
+                                    max=1.0,
+                                    step=0.01,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="How closely generation follows the prompt. "
+                                    "Higher values make the sound follow the text more closely.",
+                                ),
+                            ],
+                        ),
+                    ],
+                    tooltip="Model to use for sound effect generation.",
+                ),
+                IO.Combo.Input(
+                    "output_format",
+                    options=["mp3_44100_192", "opus_48000_192"],
+                    tooltip="Audio output format.",
+                ),
+            ],
+            outputs=[
+                IO.Audio.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.14,"format":{"approximate":true,"suffix":"/minute"}}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        text: str,
+        model: dict,
+        output_format: str,
+    ) -> IO.NodeOutput:
+        validate_string(text, min_length=1)
+        response = await sync_op_raw(
+            cls,
+            ApiEndpoint(
+                path="/proxy/elevenlabs/v1/sound-generation",
+                method="POST",
+                query_params={"output_format": output_format},
+            ),
+            data=TextToSoundEffectsRequest(
+                text=text,
+                duration_seconds=model["duration"],
+                prompt_influence=model["prompt_influence"],
+                loop=model.get("loop", None),
+            ),
+            as_binary=True,
+        )
+        return IO.NodeOutput(audio_bytes_to_audio_input(response))
+
+
+class ElevenLabsInstantVoiceClone(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ElevenLabsInstantVoiceClone",
+            display_name="ElevenLabs Instant Voice Clone",
+            category="api node/audio/ElevenLabs",
+            description="Create a cloned voice from audio samples. "
+            "Provide 1-8 audio recordings of the voice to clone.",
+            inputs=[
+                IO.Autogrow.Input(
+                    "files",
+                    template=IO.Autogrow.TemplatePrefix(
+                        IO.Audio.Input("audio"),
+                        prefix="audio",
+                        min=1,
+                        max=8,
+                    ),
+                    tooltip="Audio recordings for voice cloning.",
+                ),
+                IO.Boolean.Input(
+                    "remove_background_noise",
+                    default=False,
+                    tooltip="Remove background noise from voice samples using audio isolation.",
+                ),
+            ],
+            outputs=[
+                IO.Custom(ELEVENLABS_VOICE).Output(display_name="voice"),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(expr="""{"type":"usd","usd":0.15}"""),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        files: IO.Autogrow.Type,
+        remove_background_noise: bool,
+    ) -> IO.NodeOutput:
+        file_tuples: list[tuple[str, tuple[str, bytes, str]]] = []
+        for key in files:
+            audio = files[key]
+            sample_rate: int = audio["sample_rate"]
+            waveform = audio["waveform"]
+            audio_data_np = audio_tensor_to_contiguous_ndarray(waveform)
+            audio_bytes_io = audio_ndarray_to_bytesio(audio_data_np, sample_rate, "mp4", "aac")
+            file_tuples.append(("files", (f"{key}.mp4", audio_bytes_io.getvalue(), "audio/mp4")))
+
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/elevenlabs/v1/voices/add", method="POST"),
+            response_model=AddVoiceResponse,
+            data=AddVoiceRequest(
+                name=str(uuid.uuid4()),
+                remove_background_noise=remove_background_noise,
+            ),
+            files=file_tuples,
+            content_type="multipart/form-data",
+        )
+        return IO.NodeOutput(response.voice_id)
+
+
+ELEVENLABS_STS_VOICE_SETTINGS = [
+    IO.Float.Input(
+        "speed",
+        default=1.0,
+        min=0.7,
+        max=1.3,
+        step=0.01,
+        display_mode=IO.NumberDisplay.slider,
+        tooltip="Speech speed. 1.0 is normal, <1.0 slower, >1.0 faster.",
+    ),
+    IO.Float.Input(
+        "similarity_boost",
+        default=0.75,
+        min=0.0,
+        max=1.0,
+        step=0.01,
+        display_mode=IO.NumberDisplay.slider,
+        tooltip="Similarity boost. Higher values make the voice more similar to the original.",
+    ),
+    IO.Boolean.Input(
+        "use_speaker_boost",
+        default=False,
+        tooltip="Boost similarity to the original speaker voice.",
+    ),
+    IO.Float.Input(
+        "style",
+        default=0.0,
+        min=0.0,
+        max=0.2,
+        step=0.01,
+        display_mode=IO.NumberDisplay.slider,
+        tooltip="Style exaggeration. Higher values increase stylistic expression but may reduce stability.",
+    ),
+]
+
+
+class ElevenLabsSpeechToSpeech(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ElevenLabsSpeechToSpeech",
+            display_name="ElevenLabs Speech to Speech",
+            category="api node/audio/ElevenLabs",
+            description="Transform speech from one voice to another while preserving the original content and emotion.",
+            inputs=[
+                IO.Custom(ELEVENLABS_VOICE).Input(
+                    "voice",
+                    tooltip="Target voice for the transformation. "
+                    "Connect from Voice Selector or Instant Voice Clone.",
+                ),
+                IO.Audio.Input(
+                    "audio",
+                    tooltip="Source audio to transform.",
+                ),
+                IO.Float.Input(
+                    "stability",
+                    default=0.5,
+                    min=0.0,
+                    max=1.0,
+                    step=0.01,
+                    display_mode=IO.NumberDisplay.slider,
+                    tooltip="Voice stability. Lower values give broader emotional range, "
+                    "higher values produce more consistent but potentially monotonous speech.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "eleven_multilingual_sts_v2",
+                            ELEVENLABS_STS_VOICE_SETTINGS,
+                        ),
+                        IO.DynamicCombo.Option(
+                            "eleven_english_sts_v2",
+                            ELEVENLABS_STS_VOICE_SETTINGS,
+                        ),
+                    ],
+                    tooltip="Model to use for speech-to-speech transformation.",
+                ),
+                IO.Combo.Input(
+                    "output_format",
+                    options=["mp3_44100_192", "opus_48000_192"],
+                    tooltip="Audio output format.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=4294967295,
+                    tooltip="Seed for reproducibility.",
+                ),
+                IO.Boolean.Input(
+                    "remove_background_noise",
+                    default=False,
+                    tooltip="Remove background noise from input audio using audio isolation.",
+                ),
+            ],
+            outputs=[
+                IO.Audio.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.24,"format":{"approximate":true,"suffix":"/minute"}}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        voice: str,
+        audio: Input.Audio,
+        stability: float,
+        model: dict,
+        output_format: str,
+        seed: int,
+        remove_background_noise: bool,
+    ) -> IO.NodeOutput:
+        audio_data_np = audio_tensor_to_contiguous_ndarray(audio["waveform"])
+        audio_bytes_io = audio_ndarray_to_bytesio(audio_data_np, audio["sample_rate"], "mp4", "aac")
+        voice_settings = TextToSpeechVoiceSettings(
+            stability=stability,
+            similarity_boost=model["similarity_boost"],
+            style=model["style"],
+            use_speaker_boost=model["use_speaker_boost"],
+            speed=model["speed"],
+        )
+        response = await sync_op_raw(
+            cls,
+            ApiEndpoint(
+                path=f"/proxy/elevenlabs/v1/speech-to-speech/{voice}",
+                method="POST",
+                query_params={"output_format": output_format},
+            ),
+            data=SpeechToSpeechRequest(
+                model_id=model["model"],
+                voice_settings=voice_settings.model_dump_json(exclude_none=True),
+                seed=seed,
+                remove_background_noise=remove_background_noise,
+            ),
+            files={"audio": ("audio.mp4", audio_bytes_io.getvalue(), "audio/mp4")},
+            content_type="multipart/form-data",
+            as_binary=True,
+        )
+        return IO.NodeOutput(audio_bytes_to_audio_input(response))
+
+
+def _generate_dialogue_inputs(count: int) -> list:
+    """Generate input widgets for a given number of dialogue entries."""
+    inputs = []
+    for i in range(1, count + 1):
+        inputs.extend(
+            [
+                IO.String.Input(
+                    f"text{i}",
+                    multiline=True,
+                    default="",
+                    tooltip=f"Text content for dialogue entry {i}.",
+                ),
+                IO.Custom(ELEVENLABS_VOICE).Input(
+                    f"voice{i}",
+                    tooltip=f"Voice for dialogue entry {i}. Connect from Voice Selector or Instant Voice Clone.",
+                ),
+            ]
+        )
+    return inputs
+
+
+class ElevenLabsTextToDialogue(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls) -> IO.Schema:
+        return IO.Schema(
+            node_id="ElevenLabsTextToDialogue",
+            display_name="ElevenLabs Text to Dialogue",
+            category="api node/audio/ElevenLabs",
+            description="Generate multi-speaker dialogue from text. Each dialogue entry has its own text and voice.",
+            inputs=[
+                IO.Float.Input(
+                    "stability",
+                    default=0.5,
+                    min=0.0,
+                    max=1.0,
+                    step=0.5,
+                    display_mode=IO.NumberDisplay.slider,
+                    tooltip="Voice stability. Lower values give broader emotional range, "
+                    "higher values produce more consistent but potentially monotonous speech.",
+                ),
+                IO.Combo.Input(
+                    "apply_text_normalization",
+                    options=["auto", "on", "off"],
+                    tooltip="Text normalization mode. 'auto' lets the system decide, "
+                    "'on' always applies normalization, 'off' skips it.",
+                ),
+                IO.Combo.Input(
+                    "model",
+                    options=["eleven_v3"],
+                    tooltip="Model to use for dialogue generation.",
+                ),
+                IO.DynamicCombo.Input(
+                    "inputs",
+                    options=[
+                        IO.DynamicCombo.Option("1", _generate_dialogue_inputs(1)),
+                        IO.DynamicCombo.Option("2", _generate_dialogue_inputs(2)),
+                        IO.DynamicCombo.Option("3", _generate_dialogue_inputs(3)),
+                        IO.DynamicCombo.Option("4", _generate_dialogue_inputs(4)),
+                        IO.DynamicCombo.Option("5", _generate_dialogue_inputs(5)),
+                        IO.DynamicCombo.Option("6", _generate_dialogue_inputs(6)),
+                        IO.DynamicCombo.Option("7", _generate_dialogue_inputs(7)),
+                        IO.DynamicCombo.Option("8", _generate_dialogue_inputs(8)),
+                        IO.DynamicCombo.Option("9", _generate_dialogue_inputs(9)),
+                        IO.DynamicCombo.Option("10", _generate_dialogue_inputs(10)),
+                    ],
+                    tooltip="Number of dialogue entries.",
+                ),
+                IO.String.Input(
+                    "language_code",
+                    default="",
+                    tooltip="ISO-639-1 or ISO-639-3 language code (e.g., 'en', 'es', 'fra'). "
+                    "Leave empty for automatic detection.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=1,
+                    min=0,
+                    max=4294967295,
+                    tooltip="Seed for reproducibility.",
+                ),
+                IO.Combo.Input(
+                    "output_format",
+                    options=["mp3_44100_192", "opus_48000_192"],
+                    tooltip="Audio output format.",
+                ),
+            ],
+            outputs=[
+                IO.Audio.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.24,"format":{"approximate":true,"suffix":"/1K chars"}}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        stability: float,
+        apply_text_normalization: str,
+        model: str,
+        inputs: dict,
+        language_code: str,
+        seed: int,
+        output_format: str,
+    ) -> IO.NodeOutput:
+        num_entries = int(inputs["inputs"])
+        dialogue_inputs: list[DialogueInput] = []
+        for i in range(1, num_entries + 1):
+            text = inputs[f"text{i}"]
+            voice_id = inputs[f"voice{i}"]
+            validate_string(text, min_length=1)
+            dialogue_inputs.append(DialogueInput(text=text, voice_id=voice_id))
+        request = TextToDialogueRequest(
+            inputs=dialogue_inputs,
+            model_id=model,
+            language_code=language_code if language_code.strip() else None,
+            settings=DialogueSettings(stability=stability),
+            seed=seed,
+            apply_text_normalization=apply_text_normalization,
+        )
+        response = await sync_op_raw(
+            cls,
+            ApiEndpoint(
+                path="/proxy/elevenlabs/v1/text-to-dialogue",
+                method="POST",
+                query_params={"output_format": output_format},
+            ),
+            data=request,
+            as_binary=True,
+        )
+        return IO.NodeOutput(audio_bytes_to_audio_input(response))
+
+
+class ElevenLabsExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            ElevenLabsSpeechToText,
+            ElevenLabsVoiceSelector,
+            ElevenLabsTextToSpeech,
+            ElevenLabsAudioIsolation,
+            ElevenLabsTextToSoundEffects,
+            ElevenLabsInstantVoiceClone,
+            ElevenLabsSpeechToSpeech,
+            ElevenLabsTextToDialogue,
+        ]
+
+
+async def comfy_entrypoint() -> ElevenLabsExtension:
+    return ElevenLabsExtension()

comfy_api_nodes/nodes_gemini.py

@@ -6,6 +6,7 @@ See: https://cloud.google.com/vertex-ai/generative-ai/docs/model-reference/infer
 import base64
 import os
 from enum import Enum
+from fnmatch import fnmatch
 from io import BytesIO
 from typing import Literal
 
@@ -119,6 +120,13 @@ async def create_image_parts(
     return image_parts
 
 
+def _mime_matches(mime: GeminiMimeType | None, pattern: str) -> bool:
+    """Check if a MIME type matches a pattern. Supports fnmatch globs (e.g. 'image/*')."""
+    if mime is None:
+        return False
+    return fnmatch(mime.value, pattern)
+
+
 def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Literal["text"] | str) -> list[GeminiPart]:
     """
     Filter response parts by their type.
@@ -151,9 +159,9 @@ def get_parts_by_type(response: GeminiGenerateContentResponse, part_type: Litera
         for part in candidate.content.parts:
             if part_type == "text" and part.text:
                 parts.append(part)
-            elif part.inlineData and part.inlineData.mimeType == part_type:
+            elif part.inlineData and _mime_matches(part.inlineData.mimeType, part_type):
                 parts.append(part)
-            elif part.fileData and part.fileData.mimeType == part_type:
+            elif part.fileData and _mime_matches(part.fileData.mimeType, part_type):
                 parts.append(part)
 
     if not parts and blocked_reasons:
@@ -178,7 +186,7 @@ def get_text_from_response(response: GeminiGenerateContentResponse) -> str:
 
 async def get_image_from_response(response: GeminiGenerateContentResponse) -> Input.Image:
     image_tensors: list[Input.Image] = []
-    parts = get_parts_by_type(response, "image/png")
+    parts = get_parts_by_type(response, "image/*")
     for part in parts:
         if part.inlineData:
             image_data = base64.b64decode(part.inlineData.data)
@@ -308,6 +316,7 @@ class GeminiNode(IO.ComfyNode):
                     default="",
                     optional=True,
                     tooltip="Foundational instructions that dictate an AI's behavior.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -585,6 +594,7 @@ class GeminiImage(IO.ComfyNode):
                     tooltip="Choose 'IMAGE' for image-only output, or "
                     "'IMAGE+TEXT' to return both the generated image and a text response.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.String.Input(
                     "system_prompt",
@@ -592,6 +602,7 @@ class GeminiImage(IO.ComfyNode):
                     default=GEMINI_IMAGE_SYS_PROMPT,
                     optional=True,
                     tooltip="Foundational instructions that dictate an AI's behavior.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -626,7 +637,7 @@ class GeminiImage(IO.ComfyNode):
 
         if not aspect_ratio:
             aspect_ratio = "auto"  # for backward compatability with old workflows; to-do remove this in December
-        image_config = GeminiImageConfig(aspectRatio=aspect_ratio)
+        image_config = GeminiImageConfig() if aspect_ratio == "auto" else GeminiImageConfig(aspectRatio=aspect_ratio)
 
         if images is not None:
             parts.extend(await create_image_parts(cls, images))
@@ -646,7 +657,7 @@ class GeminiImage(IO.ComfyNode):
                 ],
                 generationConfig=GeminiImageGenerationConfig(
                     responseModalities=(["IMAGE"] if response_modalities == "IMAGE" else ["TEXT", "IMAGE"]),
-                    imageConfig=None if aspect_ratio == "auto" else image_config,
+                    imageConfig=image_config,
                 ),
                 systemInstruction=gemini_system_prompt,
             ),
@@ -706,6 +717,7 @@ class GeminiImage2(IO.ComfyNode):
                     options=["IMAGE+TEXT", "IMAGE"],
                     tooltip="Choose 'IMAGE' for image-only output, or "
                     "'IMAGE+TEXT' to return both the generated image and a text response.",
+                    advanced=True,
                 ),
                 IO.Image.Input(
                     "images",
@@ -725,6 +737,7 @@ class GeminiImage2(IO.ComfyNode):
                     default=GEMINI_IMAGE_SYS_PROMPT,
                     optional=True,
                     tooltip="Foundational instructions that dictate an AI's behavior.",
+                    advanced=True,
                 ),
             ],
             outputs=[

comfy_api_nodes/nodes_hunyuan3d.py

@@ -1,31 +1,48 @@
 from typing_extensions import override
 
-from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api.latest import IO, ComfyExtension, Input, Types
 from comfy_api_nodes.apis.hunyuan3d import (
     Hunyuan3DViewImage,
     InputGenerateType,
     ResultFile3D,
+    TextureEditTaskRequest,
     To3DProTaskCreateResponse,
     To3DProTaskQueryRequest,
     To3DProTaskRequest,
     To3DProTaskResultResponse,
+    To3DUVFileInput,
+    To3DUVTaskRequest,
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
     download_url_to_file_3d,
+    download_url_to_image_tensor,
     downscale_image_tensor_by_max_side,
     poll_op,
     sync_op,
+    upload_3d_model_to_comfyapi,
     upload_image_to_comfyapi,
     validate_image_dimensions,
     validate_string,
 )
 
 
-def get_file_from_response(response_objs: list[ResultFile3D], file_type: str) -> ResultFile3D | None:
+def _is_tencent_rate_limited(status: int, body: object) -> bool:
+    return (
+        status == 400
+        and isinstance(body, dict)
+        and "RequestLimitExceeded" in str(body.get("Response", {}).get("Error", {}).get("Code", ""))
+    )
+
+
+def get_file_from_response(
+    response_objs: list[ResultFile3D], file_type: str, raise_if_not_found: bool = True
+) -> ResultFile3D | None:
     for i in response_objs:
         if i.Type.lower() == file_type.lower():
             return i
+    if raise_if_not_found:
+        raise ValueError(f"'{file_type}' file type is not found in the response.")
     return None
 
 
@@ -35,8 +52,9 @@ class TencentTextToModelNode(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="TencentTextToModelNode",
-            display_name="Hunyuan3D: Text to Model (Pro)",
+            display_name="Hunyuan3D: Text to Model",
             category="api node/3d/Tencent",
+            essentials_category="3D",
             inputs=[
                 IO.Combo.Input(
                     "model",
@@ -120,6 +138,7 @@ class TencentTextToModelNode(IO.ComfyNode):
                 EnablePBR=generate_type.get("pbr", None),
                 PolygonType=generate_type.get("polygon_type", None),
             ),
+            is_rate_limited=_is_tencent_rate_limited,
         )
         if response.Error:
             raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
@@ -131,11 +150,14 @@ class TencentTextToModelNode(IO.ComfyNode):
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
-        glb_result = get_file_from_response(result.ResultFile3Ds, "glb")
-        obj_result = get_file_from_response(result.ResultFile3Ds, "obj")
-        file_glb = await download_url_to_file_3d(glb_result.Url, "glb", task_id=task_id) if glb_result else None
         return IO.NodeOutput(
-            file_glb, file_glb, await download_url_to_file_3d(obj_result.Url, "obj", task_id=task_id) if obj_result else None
+            f"{task_id}.glb",
+            await download_url_to_file_3d(
+                get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb", task_id=task_id
+            ),
+            await download_url_to_file_3d(
+                get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj", task_id=task_id
+            ),
         )
 
 
@@ -145,8 +167,9 @@ class TencentImageToModelNode(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="TencentImageToModelNode",
-            display_name="Hunyuan3D: Image(s) to Model (Pro)",
+            display_name="Hunyuan3D: Image(s) to Model",
             category="api node/3d/Tencent",
+            essentials_category="3D",
             inputs=[
                 IO.Combo.Input(
                     "model",
@@ -268,6 +291,7 @@ class TencentImageToModelNode(IO.ComfyNode):
                 EnablePBR=generate_type.get("pbr", None),
                 PolygonType=generate_type.get("polygon_type", None),
             ),
+            is_rate_limited=_is_tencent_rate_limited,
         )
         if response.Error:
             raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
@@ -279,11 +303,257 @@ class TencentImageToModelNode(IO.ComfyNode):
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
-        glb_result = get_file_from_response(result.ResultFile3Ds, "glb")
-        obj_result = get_file_from_response(result.ResultFile3Ds, "obj")
-        file_glb = await download_url_to_file_3d(glb_result.Url, "glb", task_id=task_id) if glb_result else None
         return IO.NodeOutput(
-            file_glb, file_glb, await download_url_to_file_3d(obj_result.Url, "obj", task_id=task_id) if obj_result else None
+            f"{task_id}.glb",
+            await download_url_to_file_3d(
+                get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb", task_id=task_id
+            ),
+            await download_url_to_file_3d(
+                get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj", task_id=task_id
+            ),
+        )
+
+
+class TencentModelTo3DUVNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="TencentModelTo3DUVNode",
+            display_name="Hunyuan3D: Model to UV",
+            category="api node/3d/Tencent",
+            description="Perform UV unfolding on a 3D model to generate UV texture. "
+            "Input model must have less than 30000 faces.",
+            inputs=[
+                IO.MultiType.Input(
+                    "model_3d",
+                    types=[IO.File3DGLB, IO.File3DOBJ, IO.File3DFBX, IO.File3DAny],
+                    tooltip="Input 3D model (GLB, OBJ, or FBX)",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=1,
+                    min=0,
+                    max=2147483647,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+            ],
+            outputs=[
+                IO.File3DOBJ.Output(display_name="OBJ"),
+                IO.File3DFBX.Output(display_name="FBX"),
+                IO.Image.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(expr='{"type":"usd","usd":0.2}'),
+        )
+
+    SUPPORTED_FORMATS = {"glb", "obj", "fbx"}
+
+    @classmethod
+    async def execute(
+        cls,
+        model_3d: Types.File3D,
+        seed: int,
+    ) -> IO.NodeOutput:
+        _ = seed
+        file_format = model_3d.format.lower()
+        if file_format not in cls.SUPPORTED_FORMATS:
+            raise ValueError(
+                f"Unsupported file format: '{file_format}'. "
+                f"Supported formats: {', '.join(sorted(cls.SUPPORTED_FORMATS))}."
+            )
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-uv", method="POST"),
+            response_model=To3DProTaskCreateResponse,
+            data=To3DUVTaskRequest(
+                File=To3DUVFileInput(
+                    Type=file_format.upper(),
+                    Url=await upload_3d_model_to_comfyapi(cls, model_3d, file_format),
+                )
+            ),
+            is_rate_limited=_is_tencent_rate_limited,
+        )
+        if response.Error:
+            raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
+        result = await poll_op(
+            cls,
+            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-uv/query", method="POST"),
+            data=To3DProTaskQueryRequest(JobId=response.JobId),
+            response_model=To3DProTaskResultResponse,
+            status_extractor=lambda r: r.Status,
+        )
+        return IO.NodeOutput(
+            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj"),
+            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"),
+            await download_url_to_image_tensor(get_file_from_response(result.ResultFile3Ds, "image").Url),
+        )
+
+
+class Tencent3DTextureEditNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="Tencent3DTextureEditNode",
+            display_name="Hunyuan3D: 3D Texture Edit",
+            category="api node/3d/Tencent",
+            description="After inputting the 3D model, perform 3D model texture redrawing.",
+            inputs=[
+                IO.MultiType.Input(
+                    "model_3d",
+                    types=[IO.File3DFBX, IO.File3DAny],
+                    tooltip="3D model in FBX format. Model should have less than 100000 faces.",
+                ),
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    default="",
+                    tooltip="Describes texture editing. Supports up to 1024 UTF-8 characters.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+            ],
+            outputs=[
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd": 0.6}""",
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model_3d: Types.File3D,
+        prompt: str,
+        seed: int,
+    ) -> IO.NodeOutput:
+        _ = seed
+        file_format = model_3d.format.lower()
+        if file_format != "fbx":
+            raise ValueError(f"Unsupported file format: '{file_format}'. Only FBX format is supported.")
+        validate_string(prompt, field_name="prompt", min_length=1, max_length=1024)
+        model_url = await upload_3d_model_to_comfyapi(cls, model_3d, file_format)
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-texture-edit", method="POST"),
+            response_model=To3DProTaskCreateResponse,
+            data=TextureEditTaskRequest(
+                File3D=To3DUVFileInput(Type=file_format.upper(), Url=model_url),
+                Prompt=prompt,
+                EnablePBR=True,
+            ),
+            is_rate_limited=_is_tencent_rate_limited,
+        )
+        if response.Error:
+            raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
+
+        result = await poll_op(
+            cls,
+            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-texture-edit/query", method="POST"),
+            data=To3DProTaskQueryRequest(JobId=response.JobId),
+            response_model=To3DProTaskResultResponse,
+            status_extractor=lambda r: r.Status,
+        )
+        return IO.NodeOutput(
+            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb"),
+            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"),
+        )
+
+
+class Tencent3DPartNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="Tencent3DPartNode",
+            display_name="Hunyuan3D: 3D Part",
+            category="api node/3d/Tencent",
+            description="Automatically perform component identification and generation based on the model structure.",
+            inputs=[
+                IO.MultiType.Input(
+                    "model_3d",
+                    types=[IO.File3DFBX, IO.File3DAny],
+                    tooltip="3D model in FBX format. Model should have less than 30000 faces.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=2147483647,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                    tooltip="Seed controls whether the node should re-run; "
+                    "results are non-deterministic regardless of seed.",
+                ),
+            ],
+            outputs=[
+                IO.File3DFBX.Output(display_name="FBX"),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(expr='{"type":"usd","usd":0.6}'),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model_3d: Types.File3D,
+        seed: int,
+    ) -> IO.NodeOutput:
+        _ = seed
+        file_format = model_3d.format.lower()
+        if file_format != "fbx":
+            raise ValueError(f"Unsupported file format: '{file_format}'. Only FBX format is supported.")
+        model_url = await upload_3d_model_to_comfyapi(cls, model_3d, file_format)
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-part", method="POST"),
+            response_model=To3DProTaskCreateResponse,
+            data=To3DUVTaskRequest(
+                File=To3DUVFileInput(Type=file_format.upper(), Url=model_url),
+            ),
+            is_rate_limited=_is_tencent_rate_limited,
+        )
+        if response.Error:
+            raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
+        result = await poll_op(
+            cls,
+            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-part/query", method="POST"),
+            data=To3DProTaskQueryRequest(JobId=response.JobId),
+            response_model=To3DProTaskResultResponse,
+            status_extractor=lambda r: r.Status,
+        )
+        return IO.NodeOutput(
+            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"),
         )
 
 
@@ -293,6 +563,9 @@ class TencentHunyuan3DExtension(ComfyExtension):
         return [
             TencentTextToModelNode,
             TencentImageToModelNode,
+            # TencentModelTo3DUVNode,
+            # Tencent3DTextureEditNode,
+            Tencent3DPartNode,
         ]
 
 

comfy_api_nodes/nodes_ideogram.py

@@ -261,6 +261,7 @@ class IdeogramV1(IO.ComfyNode):
                     default="AUTO",
                     tooltip="Determine if MagicPrompt should be used in generation",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",
@@ -394,6 +395,7 @@ class IdeogramV2(IO.ComfyNode):
                     default="AUTO",
                     tooltip="Determine if MagicPrompt should be used in generation",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",
@@ -411,6 +413,7 @@ class IdeogramV2(IO.ComfyNode):
                     default="NONE",
                     tooltip="Style type for generation (V2 only)",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.String.Input(
                     "negative_prompt",
@@ -564,6 +567,7 @@ class IdeogramV3(IO.ComfyNode):
                     default="AUTO",
                     tooltip="Determine if MagicPrompt should be used in generation",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",
@@ -590,6 +594,7 @@ class IdeogramV3(IO.ComfyNode):
                     default="DEFAULT",
                     tooltip="Controls the trade-off between generation speed and quality",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Image.Input(
                     "character_image",

comfy_api_nodes/nodes_kling.py

@@ -2262,6 +2262,7 @@ class KlingLipSyncAudioToVideoNode(IO.ComfyNode):
             node_id="KlingLipSyncAudioToVideoNode",
             display_name="Kling Lip Sync Video with Audio",
             category="api node/video/Kling",
+            essentials_category="Video Generation",
             description="Kling Lip Sync Audio to Video Node. Syncs mouth movements in a video file to the audio content of an audio file. When using, ensure that the audio contains clearly distinguishable vocals and that the video contains a distinct face. The audio file should not be larger than 5MB. The video file should not be larger than 100MB, should have height/width between 720px and 1920px, and should be between 2s and 10s in length.",
             inputs=[
                 IO.Video.Input("video"),
@@ -2333,6 +2334,7 @@ class KlingLipSyncTextToVideoNode(IO.ComfyNode):
                     max=2.0,
                     display_mode=IO.NumberDisplay.slider,
                     tooltip="Speech Rate. Valid range: 0.8~2.0, accurate to one decimal place.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -2454,6 +2456,7 @@ class KlingImageGenerationNode(IO.ComfyNode):
                 IO.Combo.Input(
                     "image_type",
                     options=[i.value for i in KlingImageGenImageReferenceType],
+                    advanced=True,
                 ),
                 IO.Float.Input(
                     "image_fidelity",
@@ -2463,6 +2466,7 @@ class KlingImageGenerationNode(IO.ComfyNode):
                     step=0.01,
                     display_mode=IO.NumberDisplay.slider,
                     tooltip="Reference intensity for user-uploaded images",
+                    advanced=True,
                 ),
                 IO.Float.Input(
                     "human_fidelity",
@@ -2472,6 +2476,7 @@ class KlingImageGenerationNode(IO.ComfyNode):
                     step=0.01,
                     display_mode=IO.NumberDisplay.slider,
                     tooltip="Subject reference similarity",
+                    advanced=True,
                 ),
                 IO.Combo.Input("model_name", options=["kling-v3", "kling-v2", "kling-v1-5"]),
                 IO.Combo.Input(
@@ -2587,7 +2592,7 @@ class TextToVideoWithAudio(IO.ComfyNode):
                 IO.Combo.Input("mode", options=["pro"]),
                 IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "1:1"]),
                 IO.Combo.Input("duration", options=[5, 10]),
-                IO.Boolean.Input("generate_audio", default=True),
+                IO.Boolean.Input("generate_audio", default=True, advanced=True),
             ],
             outputs=[
                 IO.Video.Output(),
@@ -2655,7 +2660,7 @@ class ImageToVideoWithAudio(IO.ComfyNode):
                 IO.String.Input("prompt", multiline=True, tooltip="Positive text prompt."),
                 IO.Combo.Input("mode", options=["pro"]),
                 IO.Combo.Input("duration", options=[5, 10]),
-                IO.Boolean.Input("generate_audio", default=True),
+                IO.Boolean.Input("generate_audio", default=True, advanced=True),
             ],
             outputs=[
                 IO.Video.Output(),

comfy_api_nodes/nodes_ltxv.py

@@ -74,6 +74,7 @@ class TextToVideoNode(IO.ComfyNode):
                     default=False,
                     optional=True,
                     tooltip="When true, the generated video will include AI-generated audio matching the scene.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -151,6 +152,7 @@ class ImageToVideoNode(IO.ComfyNode):
                     default=False,
                     optional=True,
                     tooltip="When true, the generated video will include AI-generated audio matching the scene.",
+                    advanced=True,
                 ),
             ],
             outputs=[

comfy_api_nodes/nodes_magnific.py

@@ -30,6 +30,30 @@ from comfy_api_nodes.util import (
     validate_image_dimensions,
 )
 
+_EUR_TO_USD = 1.19
+
+
+def _tier_price_eur(megapixels: float) -> float:
+    """Price in EUR for a single Magnific upscaling step based on input megapixels."""
+    if megapixels <= 1.3:
+        return 0.143
+    if megapixels <= 3.0:
+        return 0.286
+    if megapixels <= 6.4:
+        return 0.429
+    return 1.716
+
+
+def _calculate_magnific_upscale_price_usd(width: int, height: int, scale: int) -> float:
+    """Calculate total Magnific upscale price in USD for given input dimensions and scale factor."""
+    num_steps = int(math.log2(scale))
+    total_eur = 0.0
+    pixels = width * height
+    for _ in range(num_steps):
+        total_eur += _tier_price_eur(pixels / 1_000_000)
+        pixels *= 4
+    return round(total_eur * _EUR_TO_USD, 2)
+
 
 class MagnificImageUpscalerCreativeNode(IO.ComfyNode):
     @classmethod
@@ -86,11 +110,13 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode):
                 IO.Combo.Input(
                     "engine",
                     options=["automatic", "magnific_illusio", "magnific_sharpy", "magnific_sparkle"],
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "auto_downscale",
                     default=False,
                     tooltip="Automatically downscale input image if output would exceed maximum pixel limit.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -103,11 +129,20 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["scale_factor"]),
+                depends_on=IO.PriceBadgeDepends(widgets=["scale_factor", "auto_downscale"]),
                 expr="""
                 (
-                  $max := widgets.scale_factor = "2x" ? 1.326 : 1.657;
-                  {"type": "range_usd", "min_usd": 0.11, "max_usd": $max}
+                  $ad := widgets.auto_downscale;
+                  $mins := $ad
+                    ? {"2x": 0.172, "4x": 0.343, "8x": 0.515, "16x": 0.515}
+                    : {"2x": 0.172, "4x": 0.343, "8x": 0.515, "16x": 0.844};
+                  $maxs := {"2x": 0.515, "4x": 0.844, "8x": 1.015, "16x": 1.187};
+                  {
+                    "type": "range_usd",
+                    "min_usd": $lookup($mins, widgets.scale_factor),
+                    "max_usd": $lookup($maxs, widgets.scale_factor),
+                    "format": { "approximate": true }
+                  }
                 )
                 """,
             ),
@@ -168,6 +203,10 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode):
                     f"Use a smaller input image or lower scale factor."
                 )
 
+        final_height, final_width = get_image_dimensions(image)
+        actual_scale = int(scale_factor.rstrip("x"))
+        price_usd = _calculate_magnific_upscale_price_usd(final_width, final_height, actual_scale)
+
         initial_res = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/freepik/v1/ai/image-upscaler", method="POST"),
@@ -189,6 +228,7 @@ class MagnificImageUpscalerCreativeNode(IO.ComfyNode):
             ApiEndpoint(path=f"/proxy/freepik/v1/ai/image-upscaler/{initial_res.task_id}"),
             response_model=TaskResponse,
             status_extractor=lambda x: x.status,
+            price_extractor=lambda _: price_usd,
             poll_interval=10.0,
             max_poll_attempts=480,
         )
@@ -242,6 +282,7 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode):
                     "auto_downscale",
                     default=False,
                     tooltip="Automatically downscale input image if output would exceed maximum resolution.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -257,8 +298,14 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode):
                 depends_on=IO.PriceBadgeDepends(widgets=["scale_factor"]),
                 expr="""
                 (
-                  $max := widgets.scale_factor = "2x" ? 1.326 : 1.657;
-                  {"type": "range_usd", "min_usd": 0.11, "max_usd": $max}
+                  $mins := {"2x": 0.172, "4x": 0.343, "8x": 0.515, "16x": 0.844};
+                  $maxs := {"2x": 2.045, "4x": 2.545, "8x": 2.889, "16x": 3.06};
+                  {
+                    "type": "range_usd",
+                    "min_usd": $lookup($mins, widgets.scale_factor),
+                    "max_usd": $lookup($maxs, widgets.scale_factor),
+                    "format": { "approximate": true }
+                  }
                 )
                 """,
             ),
@@ -321,6 +368,9 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode):
                     f"Use a smaller input image or lower scale factor."
                 )
 
+        final_height, final_width = get_image_dimensions(image)
+        price_usd = _calculate_magnific_upscale_price_usd(final_width, final_height, requested_scale)
+
         initial_res = await sync_op(
             cls,
             ApiEndpoint(path="/proxy/freepik/v1/ai/image-upscaler-precision-v2", method="POST"),
@@ -339,6 +389,7 @@ class MagnificImageUpscalerPreciseV2Node(IO.ComfyNode):
             ApiEndpoint(path=f"/proxy/freepik/v1/ai/image-upscaler-precision-v2/{initial_res.task_id}"),
             response_model=TaskResponse,
             status_extractor=lambda x: x.status,
+            price_extractor=lambda _: price_usd,
             poll_interval=10.0,
             max_poll_attempts=480,
         )
@@ -392,6 +443,7 @@ class MagnificImageStyleTransferNode(IO.ComfyNode):
                         "softy",
                     ],
                     tooltip="Processing engine selection.",
+                    advanced=True,
                 ),
                 IO.DynamicCombo.Input(
                     "portrait_mode",
@@ -420,6 +472,7 @@ class MagnificImageStyleTransferNode(IO.ComfyNode):
                     default=True,
                     tooltip="When disabled, expect each generation to introduce a degree of randomness, "
                     "leading to more diverse outcomes.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -534,16 +587,19 @@ class MagnificImageRelightNode(IO.ComfyNode):
                     "interpolate_from_original",
                     default=False,
                     tooltip="Restricts generation freedom to match original more closely.",
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "change_background",
                     default=True,
                     tooltip="Modifies background based on prompt/reference.",
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "preserve_details",
                     default=True,
                     tooltip="Maintains texture and fine details from original.",
+                    advanced=True,
                 ),
                 IO.DynamicCombo.Input(
                     "advanced_settings",
@@ -877,8 +933,8 @@ class MagnificExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
         return [
-            # MagnificImageUpscalerCreativeNode,
-            # MagnificImageUpscalerPreciseV2Node,
+            MagnificImageUpscalerCreativeNode,
+            MagnificImageUpscalerPreciseV2Node,
             MagnificImageStyleTransferNode,
             MagnificImageRelightNode,
             MagnificImageSkinEnhancerNode,

comfy_api_nodes/nodes_meshy.py

@@ -58,11 +58,12 @@ class MeshyTextToModelNode(IO.ComfyNode):
                     ],
                     tooltip="When set to false, returns an unprocessed triangular mesh.",
                 ),
-                IO.Combo.Input("symmetry_mode", options=["auto", "on", "off"]),
+                IO.Combo.Input("symmetry_mode", options=["auto", "on", "off"], advanced=True),
                 IO.Combo.Input(
                     "pose_mode",
                     options=["", "A-pose", "T-pose"],
                     tooltip="Specify the pose mode for the generated model.",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",
@@ -155,6 +156,7 @@ class MeshyRefineNode(IO.ComfyNode):
                     tooltip="Generate PBR Maps (metallic, roughness, normal) in addition to the base color. "
                     "Note: this should be set to false when using Sculpture style, "
                     "as Sculpture style generates its own set of PBR maps.",
+                    advanced=True,
                 ),
                 IO.String.Input(
                     "texture_prompt",
@@ -299,6 +301,7 @@ class MeshyImageToModelNode(IO.ComfyNode):
                     "pose_mode",
                     options=["", "A-pose", "T-pose"],
                     tooltip="Specify the pose mode for the generated model.",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",
@@ -429,7 +432,7 @@ class MeshyMultiImageToModelNode(IO.ComfyNode):
                     ],
                     tooltip="When set to false, returns an unprocessed triangular mesh.",
                 ),
-                IO.Combo.Input("symmetry_mode", options=["auto", "on", "off"]),
+                IO.Combo.Input("symmetry_mode", options=["auto", "on", "off"], advanced=True),
                 IO.DynamicCombo.Input(
                     "should_texture",
                     options=[
@@ -466,6 +469,7 @@ class MeshyMultiImageToModelNode(IO.ComfyNode):
                     "pose_mode",
                     options=["", "A-pose", "T-pose"],
                     tooltip="Specify the pose mode for the generated model.",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",
@@ -728,8 +732,9 @@ class MeshyTextureNode(IO.ComfyNode):
                     tooltip="Use the original UV of the model instead of generating new UVs. "
                     "When enabled, Meshy preserves existing textures from the uploaded model. "
                     "If the model has no original UV, the quality of the output might not be as good.",
+                    advanced=True,
                 ),
-                IO.Boolean.Input("pbr", default=False),
+                IO.Boolean.Input("pbr", default=False, advanced=True),
                 IO.String.Input(
                     "text_style_prompt",
                     default="",

comfy_api_nodes/nodes_openai.py

@@ -43,7 +43,6 @@ class SupportedOpenAIModel(str, Enum):
     o1 = "o1"
     o3 = "o3"
     o1_pro = "o1-pro"
-    gpt_4o = "gpt-4o"
     gpt_4_1 = "gpt-4.1"
     gpt_4_1_mini = "gpt-4.1-mini"
     gpt_4_1_nano = "gpt-4.1-nano"
@@ -576,6 +575,7 @@ class OpenAIChatNode(IO.ComfyNode):
             node_id="OpenAIChatNode",
             display_name="OpenAI ChatGPT",
             category="api node/text/OpenAI",
+            essentials_category="Text Generation",
             description="Generate text responses from an OpenAI model.",
             inputs=[
                 IO.String.Input(
@@ -588,6 +588,7 @@ class OpenAIChatNode(IO.ComfyNode):
                     "persist_context",
                     default=False,
                     tooltip="This parameter is deprecated and has no effect.",
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "model",
@@ -649,11 +650,6 @@ class OpenAIChatNode(IO.ComfyNode):
                     "usd": [0.01, 0.04],
                     "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" }
                   }
-                  : $contains($m, "gpt-4o") ? {
-                    "type": "list_usd",
-                    "usd": [0.0025, 0.01],
-                    "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" }
-                  }
                   : $contains($m, "gpt-4.1-nano") ? {
                     "type": "list_usd",
                     "usd": [0.0001, 0.0004],
@@ -862,6 +858,7 @@ class OpenAIChatConfig(IO.ComfyNode):
                     options=["auto", "disabled"],
                     default="auto",
                     tooltip="The truncation strategy to use for the model response. auto: If the context of this response and previous ones exceeds the model's context window size, the model will truncate the response to fit the context window by dropping input items in the middle of the conversation.disabled: If a model response will exceed the context window size for a model, the request will fail with a 400 error",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "max_output_tokens",
@@ -870,6 +867,7 @@ class OpenAIChatConfig(IO.ComfyNode):
                     max=16384,
                     tooltip="An upper bound for the number of tokens that can be generated for a response, including visible output tokens",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.String.Input(
                     "instructions",

comfy_api_nodes/nodes_recraft.py

@@ -1,5 +1,4 @@
 from io import BytesIO
-from typing import Optional, Union
 
 import aiohttp
 import torch
@@ -9,6 +8,8 @@ from typing_extensions import override
 from comfy.utils import ProgressBar
 from comfy_api.latest import IO, ComfyExtension
 from comfy_api_nodes.apis.recraft import (
+    RECRAFT_V4_PRO_SIZES,
+    RECRAFT_V4_SIZES,
     RecraftColor,
     RecraftColorChain,
     RecraftControls,
@@ -18,7 +19,6 @@ from comfy_api_nodes.apis.recraft import (
     RecraftImageGenerationResponse,
     RecraftImageSize,
     RecraftIO,
-    RecraftModel,
     RecraftStyle,
     RecraftStyleV3,
     get_v3_substyles,
@@ -39,7 +39,7 @@ async def handle_recraft_file_request(
     cls: type[IO.ComfyNode],
     image: torch.Tensor,
     path: str,
-    mask: Optional[torch.Tensor] = None,
+    mask: torch.Tensor | None = None,
     total_pixels: int = 4096 * 4096,
     timeout: int = 1024,
     request=None,
@@ -73,11 +73,11 @@ async def handle_recraft_file_request(
 def recraft_multipart_parser(
     data,
     parent_key=None,
-    formatter: Optional[type[callable]] = None,
-    converted_to_check: Optional[list[list]] = None,
+    formatter: type[callable] | None = None,
+    converted_to_check: list[list] | None = None,
     is_list: bool = False,
     return_mode: str = "formdata",  # "dict" | "formdata"
-) -> Union[dict, aiohttp.FormData]:
+) -> dict | aiohttp.FormData:
     """
     Formats data such that multipart/form-data will work with aiohttp library when both files and data are present.
 
@@ -309,7 +309,7 @@ class RecraftStyleInfiniteStyleLibrary(IO.ComfyNode):
             node_id="RecraftStyleV3InfiniteStyleLibrary",
             display_name="Recraft Style - Infinite Style Library",
             category="api node/image/Recraft",
-            description="Select style based on preexisting UUID from Recraft's Infinite Style Library.",
+            description="Choose style based on preexisting UUID from Recraft's Infinite Style Library.",
             inputs=[
                 IO.String.Input("style_id", default="", tooltip="UUID of style from Infinite Style Library."),
             ],
@@ -485,7 +485,7 @@ class RecraftTextToImageNode(IO.ComfyNode):
             data=RecraftImageGenerationRequest(
                 prompt=prompt,
                 negative_prompt=negative_prompt,
-                model=RecraftModel.recraftv3,
+                model="recraftv3",
                 size=size,
                 n=n,
                 style=recraft_style.style,
@@ -598,7 +598,7 @@ class RecraftImageToImageNode(IO.ComfyNode):
         request = RecraftImageGenerationRequest(
             prompt=prompt,
             negative_prompt=negative_prompt,
-            model=RecraftModel.recraftv3,
+            model="recraftv3",
             n=n,
             strength=round(strength, 2),
             style=recraft_style.style,
@@ -698,7 +698,7 @@ class RecraftImageInpaintingNode(IO.ComfyNode):
         request = RecraftImageGenerationRequest(
             prompt=prompt,
             negative_prompt=negative_prompt,
-            model=RecraftModel.recraftv3,
+            model="recraftv3",
             n=n,
             style=recraft_style.style,
             substyle=recraft_style.substyle,
@@ -810,7 +810,7 @@ class RecraftTextToVectorNode(IO.ComfyNode):
             data=RecraftImageGenerationRequest(
                 prompt=prompt,
                 negative_prompt=negative_prompt,
-                model=RecraftModel.recraftv3,
+                model="recraftv3",
                 size=size,
                 n=n,
                 style=recraft_style.style,
@@ -933,7 +933,7 @@ class RecraftReplaceBackgroundNode(IO.ComfyNode):
         request = RecraftImageGenerationRequest(
             prompt=prompt,
             negative_prompt=negative_prompt,
-            model=RecraftModel.recraftv3,
+            model="recraftv3",
             n=n,
             style=recraft_style.style,
             substyle=recraft_style.substyle,
@@ -963,6 +963,7 @@ class RecraftRemoveBackgroundNode(IO.ComfyNode):
             node_id="RecraftRemoveBackgroundNode",
             display_name="Recraft Remove Background",
             category="api node/image/Recraft",
+            essentials_category="Image Tools",
             description="Remove background from image, and return processed image and mask.",
             inputs=[
                 IO.Image.Input("image"),
@@ -1078,6 +1079,252 @@ class RecraftCreativeUpscaleNode(RecraftCrispUpscaleNode):
         )
 
 
+class RecraftV4TextToImageNode(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="RecraftV4TextToImageNode",
+            display_name="Recraft V4 Text to Image",
+            category="api node/image/Recraft",
+            description="Generates images using Recraft V4 or V4 Pro models.",
+            inputs=[
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="Prompt for the image generation. Maximum 10,000 characters.",
+                ),
+                IO.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    tooltip="An optional text description of undesired elements on an image.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "recraftv4",
+                            [
+                                IO.Combo.Input(
+                                    "size",
+                                    options=RECRAFT_V4_SIZES,
+                                    default="1024x1024",
+                                    tooltip="The size of the generated image.",
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "recraftv4_pro",
+                            [
+                                IO.Combo.Input(
+                                    "size",
+                                    options=RECRAFT_V4_PRO_SIZES,
+                                    default="2048x2048",
+                                    tooltip="The size of the generated image.",
+                                ),
+                            ],
+                        ),
+                    ],
+                    tooltip="The model to use for generation.",
+                ),
+                IO.Int.Input(
+                    "n",
+                    default=1,
+                    min=1,
+                    max=6,
+                    tooltip="The number of images to generate.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=0xFFFFFFFFFFFFFFFF,
+                    control_after_generate=True,
+                    tooltip="Seed to determine if node should re-run; "
+                    "actual results are nondeterministic regardless of seed.",
+                ),
+                IO.Custom(RecraftIO.CONTROLS).Input(
+                    "recraft_controls",
+                    tooltip="Optional additional controls over the generation via the Recraft Controls node.",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                IO.Image.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "n"]),
+                expr="""
+                (
+                    $prices := {"recraftv4": 0.04, "recraftv4_pro": 0.25};
+                    {"type":"usd","usd": $lookup($prices, widgets.model) * widgets.n}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        prompt: str,
+        negative_prompt: str,
+        model: dict,
+        n: int,
+        seed: int,
+        recraft_controls: RecraftControls | None = None,
+    ) -> IO.NodeOutput:
+        validate_string(prompt, strip_whitespace=False, min_length=1, max_length=10000)
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/recraft/image_generation", method="POST"),
+            response_model=RecraftImageGenerationResponse,
+            data=RecraftImageGenerationRequest(
+                prompt=prompt,
+                negative_prompt=negative_prompt if negative_prompt else None,
+                model=model["model"],
+                size=model["size"],
+                n=n,
+                controls=recraft_controls.create_api_model() if recraft_controls else None,
+            ),
+            max_retries=1,
+        )
+        images = []
+        for data in response.data:
+            with handle_recraft_image_output():
+                image = bytesio_to_image_tensor(await download_url_as_bytesio(data.url, timeout=1024))
+            if len(image.shape) < 4:
+                image = image.unsqueeze(0)
+            images.append(image)
+        return IO.NodeOutput(torch.cat(images, dim=0))
+
+
+class RecraftV4TextToVectorNode(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="RecraftV4TextToVectorNode",
+            display_name="Recraft V4 Text to Vector",
+            category="api node/image/Recraft",
+            description="Generates SVG using Recraft V4 or V4 Pro models.",
+            inputs=[
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="Prompt for the image generation. Maximum 10,000 characters.",
+                ),
+                IO.String.Input(
+                    "negative_prompt",
+                    multiline=True,
+                    tooltip="An optional text description of undesired elements on an image.",
+                ),
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "recraftv4",
+                            [
+                                IO.Combo.Input(
+                                    "size",
+                                    options=RECRAFT_V4_SIZES,
+                                    default="1024x1024",
+                                    tooltip="The size of the generated image.",
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "recraftv4_pro",
+                            [
+                                IO.Combo.Input(
+                                    "size",
+                                    options=RECRAFT_V4_PRO_SIZES,
+                                    default="2048x2048",
+                                    tooltip="The size of the generated image.",
+                                ),
+                            ],
+                        ),
+                    ],
+                    tooltip="The model to use for generation.",
+                ),
+                IO.Int.Input(
+                    "n",
+                    default=1,
+                    min=1,
+                    max=6,
+                    tooltip="The number of images to generate.",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=0,
+                    min=0,
+                    max=0xFFFFFFFFFFFFFFFF,
+                    control_after_generate=True,
+                    tooltip="Seed to determine if node should re-run; "
+                    "actual results are nondeterministic regardless of seed.",
+                ),
+                IO.Custom(RecraftIO.CONTROLS).Input(
+                    "recraft_controls",
+                    tooltip="Optional additional controls over the generation via the Recraft Controls node.",
+                    optional=True,
+                ),
+            ],
+            outputs=[
+                IO.SVG.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "n"]),
+                expr="""
+                (
+                    $prices := {"recraftv4": 0.08, "recraftv4_pro": 0.30};
+                    {"type":"usd","usd": $lookup($prices, widgets.model) * widgets.n}
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        prompt: str,
+        negative_prompt: str,
+        model: dict,
+        n: int,
+        seed: int,
+        recraft_controls: RecraftControls | None = None,
+    ) -> IO.NodeOutput:
+        validate_string(prompt, strip_whitespace=False, min_length=1, max_length=10000)
+        response = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/recraft/image_generation", method="POST"),
+            response_model=RecraftImageGenerationResponse,
+            data=RecraftImageGenerationRequest(
+                prompt=prompt,
+                negative_prompt=negative_prompt if negative_prompt else None,
+                model=model["model"],
+                size=model["size"],
+                n=n,
+                style="vector_illustration",
+                substyle=None,
+                controls=recraft_controls.create_api_model() if recraft_controls else None,
+            ),
+            max_retries=1,
+        )
+        svg_data = []
+        for data in response.data:
+            svg_data.append(await download_url_as_bytesio(data.url, timeout=1024))
+        return IO.NodeOutput(SVG(svg_data))
+
+
 class RecraftExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -1098,6 +1345,8 @@ class RecraftExtension(ComfyExtension):
             RecraftCreateStyleNode,
             RecraftColorRGBNode,
             RecraftControlsNode,
+            RecraftV4TextToImageNode,
+            RecraftV4TextToVectorNode,
         ]
 
 

comfy_api_nodes/nodes_rodin.py

@@ -493,7 +493,7 @@ class Rodin3D_Gen2(IO.ComfyNode):
                     default="500K-Triangle",
                     optional=True,
                 ),
-                IO.Boolean.Input("TAPose", default=False),
+                IO.Boolean.Input("TAPose", default=False, advanced=True),
             ],
             outputs=[
                 IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
@@ -505,6 +505,9 @@ class Rodin3D_Gen2(IO.ComfyNode):
                 IO.Hidden.unique_id,
             ],
             is_api_node=True,
+            price_badge=IO.PriceBadge(
+                expr="""{"type":"usd","usd":0.4}""",
+            ),
         )
 
     @classmethod

comfy_api_nodes/nodes_stability.py

@@ -86,6 +86,7 @@ class StabilityStableImageUltraNode(IO.ComfyNode):
                     "style_preset",
                     options=get_stability_style_presets(),
                     tooltip="Optional desired style of generated image.",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",
@@ -107,6 +108,7 @@ class StabilityStableImageUltraNode(IO.ComfyNode):
                     tooltip="A blurb of text describing what you do not wish to see in the output image. This is an advanced feature.",
                     force_input=True,
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Float.Input(
                     "image_denoise",
@@ -218,6 +220,7 @@ class StabilityStableImageSD_3_5Node(IO.ComfyNode):
                     "style_preset",
                     options=get_stability_style_presets(),
                     tooltip="Optional desired style of generated image.",
+                    advanced=True,
                 ),
                 IO.Float.Input(
                     "cfg_scale",
@@ -247,6 +250,7 @@ class StabilityStableImageSD_3_5Node(IO.ComfyNode):
                     tooltip="Keywords of what you do not wish to see in the output image. This is an advanced feature.",
                     force_input=True,
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Float.Input(
                     "image_denoise",
@@ -384,6 +388,7 @@ class StabilityUpscaleConservativeNode(IO.ComfyNode):
                     tooltip="Keywords of what you do not wish to see in the output image. This is an advanced feature.",
                     force_input=True,
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -474,6 +479,7 @@ class StabilityUpscaleCreativeNode(IO.ComfyNode):
                     "style_preset",
                     options=get_stability_style_presets(),
                     tooltip="Optional desired style of generated image.",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",
@@ -491,6 +497,7 @@ class StabilityUpscaleCreativeNode(IO.ComfyNode):
                     tooltip="Keywords of what you do not wish to see in the output image. This is an advanced feature.",
                     force_input=True,
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -624,6 +631,7 @@ class StabilityTextToAudio(IO.ComfyNode):
             node_id="StabilityTextToAudio",
             display_name="Stability AI Text To Audio",
             category="api node/audio/Stability AI",
+            essentials_category="Audio",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Combo.Input(
@@ -659,6 +667,7 @@ class StabilityTextToAudio(IO.ComfyNode):
                     step=1,
                     tooltip="Controls the number of sampling steps.",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -736,6 +745,7 @@ class StabilityAudioToAudio(IO.ComfyNode):
                     step=1,
                     tooltip="Controls the number of sampling steps.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Float.Input(
                     "strength",
@@ -829,6 +839,7 @@ class StabilityAudioInpaint(IO.ComfyNode):
                     step=1,
                     tooltip="Controls the number of sampling steps.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "mask_start",
@@ -837,6 +848,7 @@ class StabilityAudioInpaint(IO.ComfyNode):
                     max=190,
                     step=1,
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "mask_end",
@@ -845,6 +857,7 @@ class StabilityAudioInpaint(IO.ComfyNode):
                     max=190,
                     step=1,
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[

comfy_api_nodes/nodes_topaz.py

@@ -63,12 +63,14 @@ class TopazImageEnhance(IO.ComfyNode):
                     "subject_detection",
                     options=["All", "Foreground", "Background"],
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "face_enhancement",
                     default=True,
                     optional=True,
                     tooltip="Enhance faces (if present) during processing.",
+                    advanced=True,
                 ),
                 IO.Float.Input(
                     "face_enhancement_creativity",
@@ -79,6 +81,7 @@ class TopazImageEnhance(IO.ComfyNode):
                     display_mode=IO.NumberDisplay.number,
                     optional=True,
                     tooltip="Set the creativity level for face enhancement.",
+                    advanced=True,
                 ),
                 IO.Float.Input(
                     "face_enhancement_strength",
@@ -89,6 +92,7 @@ class TopazImageEnhance(IO.ComfyNode):
                     display_mode=IO.NumberDisplay.number,
                     optional=True,
                     tooltip="Controls how sharp enhanced faces are relative to the background.",
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "crop_to_fill",
@@ -96,6 +100,7 @@ class TopazImageEnhance(IO.ComfyNode):
                     optional=True,
                     tooltip="By default, the image is letterboxed when the output aspect ratio differs. "
                     "Enable to crop the image to fill the output dimensions.",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "output_width",
@@ -106,6 +111,7 @@ class TopazImageEnhance(IO.ComfyNode):
                     display_mode=IO.NumberDisplay.number,
                     optional=True,
                     tooltip="Zero value means to calculate automatically (usually it will be original size or output_height if specified).",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "output_height",
@@ -116,6 +122,7 @@ class TopazImageEnhance(IO.ComfyNode):
                     display_mode=IO.NumberDisplay.number,
                     optional=True,
                     tooltip="Zero value means to output in the same height as original or output width.",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "creativity",
@@ -131,12 +138,14 @@ class TopazImageEnhance(IO.ComfyNode):
                     default=True,
                     optional=True,
                     tooltip="Preserve subjects' facial identity.",
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "color_preservation",
                     default=True,
                     optional=True,
                     tooltip="Preserve the original colors.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -234,9 +243,10 @@ class TopazVideoEnhance(IO.ComfyNode):
                     default="low",
                     tooltip="Creativity level (applies only to Starlight (Astra) Creative).",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input("interpolation_enabled", default=False, optional=True),
-                IO.Combo.Input("interpolation_model", options=["apo-8"], default="apo-8", optional=True),
+                IO.Combo.Input("interpolation_model", options=["apo-8"], default="apo-8", optional=True, advanced=True),
                 IO.Int.Input(
                     "interpolation_slowmo",
                     default=1,
@@ -246,6 +256,7 @@ class TopazVideoEnhance(IO.ComfyNode):
                     tooltip="Slow-motion factor applied to the input video. "
                     "For example, 2 makes the output twice as slow and doubles the duration.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "interpolation_frame_rate",
@@ -261,6 +272,7 @@ class TopazVideoEnhance(IO.ComfyNode):
                     default=False,
                     tooltip="Analyze the input for duplicate frames and remove them.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Float.Input(
                     "interpolation_duplicate_threshold",
@@ -271,6 +283,7 @@ class TopazVideoEnhance(IO.ComfyNode):
                     display_mode=IO.NumberDisplay.number,
                     tooltip="Detection sensitivity for duplicate frames.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "dynamic_compression_level",
@@ -278,6 +291,7 @@ class TopazVideoEnhance(IO.ComfyNode):
                     default="Low",
                     tooltip="CQP level.",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[

comfy_api_nodes/nodes_tripo.py

@@ -89,13 +89,13 @@ class TripoTextToModelNode(IO.ComfyNode):
                 IO.Combo.Input("style", options=TripoStyle, default="None", optional=True),
                 IO.Boolean.Input("texture", default=True, optional=True),
                 IO.Boolean.Input("pbr", default=True, optional=True),
-                IO.Int.Input("image_seed", default=42, optional=True),
-                IO.Int.Input("model_seed", default=42, optional=True),
-                IO.Int.Input("texture_seed", default=42, optional=True),
-                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True),
-                IO.Int.Input("face_limit", default=-1, min=-1, max=2000000, optional=True),
-                IO.Boolean.Input("quad", default=False, optional=True),
-                IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True),
+                IO.Int.Input("image_seed", default=42, optional=True, advanced=True),
+                IO.Int.Input("model_seed", default=42, optional=True, advanced=True),
+                IO.Int.Input("texture_seed", default=42, optional=True, advanced=True),
+                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
+                IO.Int.Input("face_limit", default=-1, min=-1, max=2000000, optional=True, advanced=True),
+                IO.Boolean.Input("quad", default=False, optional=True, advanced=True),
+                IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
             ],
             outputs=[
                 IO.String.Output(display_name="model_file"),  # for backward compatibility only
@@ -210,18 +210,18 @@ class TripoImageToModelNode(IO.ComfyNode):
                 IO.Combo.Input("style", options=TripoStyle, default="None", optional=True),
                 IO.Boolean.Input("texture", default=True, optional=True),
                 IO.Boolean.Input("pbr", default=True, optional=True),
-                IO.Int.Input("model_seed", default=42, optional=True),
+                IO.Int.Input("model_seed", default=42, optional=True, advanced=True),
                 IO.Combo.Input(
-                    "orientation", options=TripoOrientation, default=TripoOrientation.DEFAULT, optional=True
+                    "orientation", options=TripoOrientation, default=TripoOrientation.DEFAULT, optional=True, advanced=True
                 ),
-                IO.Int.Input("texture_seed", default=42, optional=True),
-                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True),
+                IO.Int.Input("texture_seed", default=42, optional=True, advanced=True),
+                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
                 IO.Combo.Input(
-                    "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True
+                    "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True, advanced=True
                 ),
-                IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True),
-                IO.Boolean.Input("quad", default=False, optional=True),
-                IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True),
+                IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True, advanced=True),
+                IO.Boolean.Input("quad", default=False, optional=True, advanced=True),
+                IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
             ],
             outputs=[
                 IO.String.Output(display_name="model_file"),  # for backward compatibility only
@@ -347,18 +347,19 @@ class TripoMultiviewToModelNode(IO.ComfyNode):
                     options=TripoOrientation,
                     default=TripoOrientation.DEFAULT,
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input("texture", default=True, optional=True),
                 IO.Boolean.Input("pbr", default=True, optional=True),
-                IO.Int.Input("model_seed", default=42, optional=True),
-                IO.Int.Input("texture_seed", default=42, optional=True),
-                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True),
+                IO.Int.Input("model_seed", default=42, optional=True, advanced=True),
+                IO.Int.Input("texture_seed", default=42, optional=True, advanced=True),
+                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
                 IO.Combo.Input(
-                    "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True
+                    "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True, advanced=True
                 ),
-                IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True),
-                IO.Boolean.Input("quad", default=False, optional=True),
-                IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True),
+                IO.Int.Input("face_limit", default=-1, min=-1, max=500000, optional=True, advanced=True),
+                IO.Boolean.Input("quad", default=False, optional=True, advanced=True),
+                IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
             ],
             outputs=[
                 IO.String.Output(display_name="model_file"),  # for backward compatibility only
@@ -474,10 +475,10 @@ class TripoTextureNode(IO.ComfyNode):
                 IO.Custom("MODEL_TASK_ID").Input("model_task_id"),
                 IO.Boolean.Input("texture", default=True, optional=True),
                 IO.Boolean.Input("pbr", default=True, optional=True),
-                IO.Int.Input("texture_seed", default=42, optional=True),
-                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True),
+                IO.Int.Input("texture_seed", default=42, optional=True, advanced=True),
+                IO.Combo.Input("texture_quality", default="standard", options=["standard", "detailed"], optional=True, advanced=True),
                 IO.Combo.Input(
-                    "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True
+                    "texture_alignment", default="original_image", options=["original_image", "geometry"], optional=True, advanced=True
                 ),
             ],
             outputs=[
@@ -682,13 +683,14 @@ class TripoConversionNode(IO.ComfyNode):
             inputs=[
                 IO.Custom("MODEL_TASK_ID,RIG_TASK_ID,RETARGET_TASK_ID").Input("original_model_task_id"),
                 IO.Combo.Input("format", options=["GLTF", "USDZ", "FBX", "OBJ", "STL", "3MF"]),
-                IO.Boolean.Input("quad", default=False, optional=True),
+                IO.Boolean.Input("quad", default=False, optional=True, advanced=True),
                 IO.Int.Input(
                     "face_limit",
                     default=-1,
                     min=-1,
                     max=2000000,
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "texture_size",
@@ -696,47 +698,53 @@ class TripoConversionNode(IO.ComfyNode):
                     min=128,
                     max=4096,
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "texture_format",
                     options=["BMP", "DPX", "HDR", "JPEG", "OPEN_EXR", "PNG", "TARGA", "TIFF", "WEBP"],
                     default="JPEG",
                     optional=True,
+                    advanced=True,
                 ),
-                IO.Boolean.Input("force_symmetry", default=False, optional=True),
-                IO.Boolean.Input("flatten_bottom", default=False, optional=True),
+                IO.Boolean.Input("force_symmetry", default=False, optional=True, advanced=True),
+                IO.Boolean.Input("flatten_bottom", default=False, optional=True, advanced=True),
                 IO.Float.Input(
                     "flatten_bottom_threshold",
                     default=0.0,
                     min=0.0,
                     max=1.0,
                     optional=True,
+                    advanced=True,
                 ),
-                IO.Boolean.Input("pivot_to_center_bottom", default=False, optional=True),
+                IO.Boolean.Input("pivot_to_center_bottom", default=False, optional=True, advanced=True),
                 IO.Float.Input(
                     "scale_factor",
                     default=1.0,
                     min=0.0,
                     optional=True,
+                    advanced=True,
                 ),
-                IO.Boolean.Input("with_animation", default=False, optional=True),
-                IO.Boolean.Input("pack_uv", default=False, optional=True),
-                IO.Boolean.Input("bake", default=False, optional=True),
-                IO.String.Input("part_names", default="", optional=True),  # comma-separated list
+                IO.Boolean.Input("with_animation", default=False, optional=True, advanced=True),
+                IO.Boolean.Input("pack_uv", default=False, optional=True, advanced=True),
+                IO.Boolean.Input("bake", default=False, optional=True, advanced=True),
+                IO.String.Input("part_names", default="", optional=True, advanced=True),  # comma-separated list
                 IO.Combo.Input(
                     "fbx_preset",
                     options=["blender", "mixamo", "3dsmax"],
                     default="blender",
                     optional=True,
+                    advanced=True,
                 ),
-                IO.Boolean.Input("export_vertex_colors", default=False, optional=True),
+                IO.Boolean.Input("export_vertex_colors", default=False, optional=True, advanced=True),
                 IO.Combo.Input(
                     "export_orientation",
                     options=["align_image", "default"],
                     default="default",
                     optional=True,
+                    advanced=True,
                 ),
-                IO.Boolean.Input("animate_in_place", default=False, optional=True),
+                IO.Boolean.Input("animate_in_place", default=False, optional=True, advanced=True),
             ],
             outputs=[],
             hidden=[

comfy_api_nodes/nodes_veo2.py

@@ -81,6 +81,7 @@ class VeoVideoGenerationNode(IO.ComfyNode):
                     default=True,
                     tooltip="Whether to enhance the prompt with AI assistance",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "person_generation",
@@ -88,6 +89,7 @@ class VeoVideoGenerationNode(IO.ComfyNode):
                     default="ALLOW",
                     tooltip="Whether to allow generating people in the video",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",
@@ -299,6 +301,7 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode):
                     default=True,
                     tooltip="This parameter is deprecated and ignored.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "person_generation",
@@ -306,6 +309,7 @@ class Veo3VideoGenerationNode(VeoVideoGenerationNode):
                     default="ALLOW",
                     tooltip="Whether to allow generating people in the video",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "seed",

comfy_api_nodes/nodes_vidu.py

@@ -54,6 +54,7 @@ async def execute_task(
         response_model=TaskStatusResponse,
         status_extractor=lambda r: r.state,
         progress_extractor=lambda r: r.progress,
+        price_extractor=lambda r: r.credits * 0.005 if r.credits is not None else None,
         max_poll_attempts=max_poll_attempts,
     )
     if not response.creations:
@@ -111,12 +112,14 @@ class ViduTextToVideoNode(IO.ComfyNode):
                     options=["1080p"],
                     tooltip="Supported values may vary by model & duration",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "movement_amplitude",
                     options=["auto", "small", "medium", "large"],
                     tooltip="The movement amplitude of objects in the frame",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -207,12 +210,14 @@ class ViduImageToVideoNode(IO.ComfyNode):
                     options=["1080p"],
                     tooltip="Supported values may vary by model & duration",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "movement_amplitude",
                     options=["auto", "small", "medium", "large"],
                     tooltip="The movement amplitude of objects in the frame",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -313,12 +318,14 @@ class ViduReferenceVideoNode(IO.ComfyNode):
                     options=["1080p"],
                     tooltip="Supported values may vary by model & duration",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "movement_amplitude",
                     options=["auto", "small", "medium", "large"],
                     tooltip="The movement amplitude of objects in the frame",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -425,12 +432,14 @@ class ViduStartEndToVideoNode(IO.ComfyNode):
                     options=["1080p"],
                     tooltip="Supported values may vary by model & duration",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "movement_amplitude",
                     options=["auto", "small", "medium", "large"],
                     tooltip="The movement amplitude of objects in the frame",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -510,11 +519,12 @@ class Vidu2TextToVideoNode(IO.ComfyNode):
                     control_after_generate=True,
                 ),
                 IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "3:4", "4:3", "1:1"]),
-                IO.Combo.Input("resolution", options=["720p", "1080p"]),
+                IO.Combo.Input("resolution", options=["720p", "1080p"], advanced=True),
                 IO.Boolean.Input(
                     "background_music",
                     default=False,
                     tooltip="Whether to add background music to the generated video.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -608,11 +618,13 @@ class Vidu2ImageToVideoNode(IO.ComfyNode):
                 IO.Combo.Input(
                     "resolution",
                     options=["720p", "1080p"],
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "movement_amplitude",
                     options=["auto", "small", "medium", "large"],
                     tooltip="The movement amplitude of objects in the frame.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -726,6 +738,7 @@ class Vidu2ReferenceVideoNode(IO.ComfyNode):
                     "audio",
                     default=False,
                     tooltip="When enabled video will contain generated speech and background music based on the prompt.",
+                    advanced=True,
                 ),
                 IO.Int.Input(
                     "duration",
@@ -745,11 +758,12 @@ class Vidu2ReferenceVideoNode(IO.ComfyNode):
                     control_after_generate=True,
                 ),
                 IO.Combo.Input("aspect_ratio", options=["16:9", "9:16", "4:3", "3:4", "1:1"]),
-                IO.Combo.Input("resolution", options=["720p", "1080p"]),
+                IO.Combo.Input("resolution", options=["720p", "1080p"], advanced=True),
                 IO.Combo.Input(
                     "movement_amplitude",
                     options=["auto", "small", "medium", "large"],
                     tooltip="The movement amplitude of objects in the frame.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -863,11 +877,12 @@ class Vidu2StartEndToVideoNode(IO.ComfyNode):
                     display_mode=IO.NumberDisplay.number,
                     control_after_generate=True,
                 ),
-                IO.Combo.Input("resolution", options=["720p", "1080p"]),
+                IO.Combo.Input("resolution", options=["720p", "1080p"], advanced=True),
                 IO.Combo.Input(
                     "movement_amplitude",
                     options=["auto", "small", "medium", "large"],
                     tooltip="The movement amplitude of objects in the frame.",
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -1306,6 +1321,36 @@ class Vidu3TextToVideoNode(IO.ComfyNode):
                                 ),
                             ],
                         ),
+                        IO.DynamicCombo.Option(
+                            "viduq3-turbo",
+                            [
+                                IO.Combo.Input(
+                                    "aspect_ratio",
+                                    options=["16:9", "9:16", "3:4", "4:3", "1:1"],
+                                    tooltip="The aspect ratio of the output video.",
+                                ),
+                                IO.Combo.Input(
+                                    "resolution",
+                                    options=["720p", "1080p"],
+                                    tooltip="Resolution of the output video.",
+                                ),
+                                IO.Int.Input(
+                                    "duration",
+                                    default=5,
+                                    min=1,
+                                    max=16,
+                                    step=1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Duration of the output video in seconds.",
+                                ),
+                                IO.Boolean.Input(
+                                    "audio",
+                                    default=False,
+                                    tooltip="When enabled, outputs video with sound "
+                                    "(including dialogue and sound effects).",
+                                ),
+                            ],
+                        ),
                     ],
                     tooltip="Model to use for video generation.",
                 ),
@@ -1334,13 +1379,20 @@ class Vidu3TextToVideoNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["model.duration", "model.resolution"]),
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.duration", "model.resolution"]),
                 expr="""
                 (
                   $res := $lookup(widgets, "model.resolution");
-                  $base := $lookup({"720p": 0.075, "1080p": 0.1}, $res);
-                  $perSec := $lookup({"720p": 0.025, "1080p": 0.05}, $res);
-                  {"type":"usd","usd": $base + $perSec * ($lookup(widgets, "model.duration") - 1)}
+                  $d := $lookup(widgets, "model.duration");
+                  $contains(widgets.model, "turbo")
+                    ? (
+                        $rate := $lookup({"720p": 0.06, "1080p": 0.08}, $res);
+                        {"type":"usd","usd": $rate * $d}
+                      )
+                    : (
+                        $rate := $lookup({"720p": 0.15, "1080p": 0.16}, $res);
+                        {"type":"usd","usd": $rate * $d}
+                      )
                 )
                 """,
             ),
@@ -1409,6 +1461,31 @@ class Vidu3ImageToVideoNode(IO.ComfyNode):
                                 ),
                             ],
                         ),
+                        IO.DynamicCombo.Option(
+                            "viduq3-turbo",
+                            [
+                                IO.Combo.Input(
+                                    "resolution",
+                                    options=["720p", "1080p"],
+                                    tooltip="Resolution of the output video.",
+                                ),
+                                IO.Int.Input(
+                                    "duration",
+                                    default=5,
+                                    min=1,
+                                    max=16,
+                                    step=1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Duration of the output video in seconds.",
+                                ),
+                                IO.Boolean.Input(
+                                    "audio",
+                                    default=False,
+                                    tooltip="When enabled, outputs video with sound "
+                                    "(including dialogue and sound effects).",
+                                ),
+                            ],
+                        ),
                     ],
                     tooltip="Model to use for video generation.",
                 ),
@@ -1442,13 +1519,20 @@ class Vidu3ImageToVideoNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["model.duration", "model.resolution"]),
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.duration", "model.resolution"]),
                 expr="""
                 (
                   $res := $lookup(widgets, "model.resolution");
-                  $base := $lookup({"720p": 0.075, "1080p": 0.275, "2k": 0.35}, $res);
-                  $perSec := $lookup({"720p": 0.05, "1080p": 0.075, "2k": 0.075}, $res);
-                  {"type":"usd","usd": $base + $perSec * ($lookup(widgets, "model.duration") - 1)}
+                  $d := $lookup(widgets, "model.duration");
+                  $contains(widgets.model, "turbo")
+                    ? (
+                        $rate := $lookup({"720p": 0.06, "1080p": 0.08}, $res);
+                        {"type":"usd","usd": $rate * $d}
+                      )
+                    : (
+                        $rate := $lookup({"720p": 0.15, "1080p": 0.16, "2k": 0.2}, $res);
+                        {"type":"usd","usd": $rate * $d}
+                      )
                 )
                 """,
             ),
@@ -1481,6 +1565,145 @@ class Vidu3ImageToVideoNode(IO.ComfyNode):
         return IO.NodeOutput(await download_url_to_video_output(results[0].url))
 
 
+class Vidu3StartEndToVideoNode(IO.ComfyNode):
+
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="Vidu3StartEndToVideoNode",
+            display_name="Vidu Q3 Start/End Frame-to-Video Generation",
+            category="api node/video/Vidu",
+            description="Generate a video from a start frame, an end frame, and a prompt.",
+            inputs=[
+                IO.DynamicCombo.Input(
+                    "model",
+                    options=[
+                        IO.DynamicCombo.Option(
+                            "viduq3-pro",
+                            [
+                                IO.Combo.Input(
+                                    "resolution",
+                                    options=["720p", "1080p"],
+                                    tooltip="Resolution of the output video.",
+                                ),
+                                IO.Int.Input(
+                                    "duration",
+                                    default=5,
+                                    min=1,
+                                    max=16,
+                                    step=1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Duration of the output video in seconds.",
+                                ),
+                                IO.Boolean.Input(
+                                    "audio",
+                                    default=False,
+                                    tooltip="When enabled, outputs video with sound "
+                                    "(including dialogue and sound effects).",
+                                ),
+                            ],
+                        ),
+                        IO.DynamicCombo.Option(
+                            "viduq3-turbo",
+                            [
+                                IO.Combo.Input(
+                                    "resolution",
+                                    options=["720p", "1080p"],
+                                    tooltip="Resolution of the output video.",
+                                ),
+                                IO.Int.Input(
+                                    "duration",
+                                    default=5,
+                                    min=1,
+                                    max=16,
+                                    step=1,
+                                    display_mode=IO.NumberDisplay.slider,
+                                    tooltip="Duration of the output video in seconds.",
+                                ),
+                                IO.Boolean.Input(
+                                    "audio",
+                                    default=False,
+                                    tooltip="When enabled, outputs video with sound "
+                                    "(including dialogue and sound effects).",
+                                ),
+                            ],
+                        ),
+                    ],
+                    tooltip="Model to use for video generation.",
+                ),
+                IO.Image.Input("first_frame"),
+                IO.Image.Input("end_frame"),
+                IO.String.Input(
+                    "prompt",
+                    multiline=True,
+                    tooltip="Prompt description (max 2000 characters).",
+                ),
+                IO.Int.Input(
+                    "seed",
+                    default=1,
+                    min=0,
+                    max=2147483647,
+                    step=1,
+                    display_mode=IO.NumberDisplay.number,
+                    control_after_generate=True,
+                ),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.duration", "model.resolution"]),
+                expr="""
+                (
+                  $res := $lookup(widgets, "model.resolution");
+                  $d := $lookup(widgets, "model.duration");
+                  $contains(widgets.model, "turbo")
+                    ? (
+                        $rate := $lookup({"720p": 0.06, "1080p": 0.08}, $res);
+                        {"type":"usd","usd": $rate * $d}
+                      )
+                    : (
+                        $rate := $lookup({"720p": 0.15, "1080p": 0.16}, $res);
+                        {"type":"usd","usd": $rate * $d}
+                      )
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: dict,
+        first_frame: Input.Image,
+        end_frame: Input.Image,
+        prompt: str,
+        seed: int,
+    ) -> IO.NodeOutput:
+        validate_string(prompt, max_length=2000)
+        validate_images_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False)
+        payload = TaskCreationRequest(
+            model=model["model"],
+            prompt=prompt,
+            duration=model["duration"],
+            seed=seed,
+            resolution=model["resolution"],
+            audio=model["audio"],
+            images=[
+                (await upload_images_to_comfyapi(cls, frame, max_images=1, mime_type="image/png"))[0]
+                for frame in (first_frame, end_frame)
+            ],
+        )
+        results = await execute_task(cls, VIDU_START_END_VIDEO, payload)
+        return IO.NodeOutput(await download_url_to_video_output(results[0].url))
+
+
 class ViduExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -1497,6 +1720,7 @@ class ViduExtension(ComfyExtension):
             ViduMultiFrameVideoNode,
             Vidu3TextToVideoNode,
             Vidu3ImageToVideoNode,
+            Vidu3StartEndToVideoNode,
         ]
 
 

comfy_api_nodes/nodes_wan.py

@@ -227,12 +227,14 @@ class WanTextToImageApi(IO.ComfyNode):
                     default=True,
                     tooltip="Whether to enhance the prompt with AI assistance.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "watermark",
                     default=False,
                     tooltip="Whether to add an AI-generated watermark to the result.",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -355,6 +357,7 @@ class WanImageToImageApi(IO.ComfyNode):
                     default=False,
                     tooltip="Whether to add an AI-generated watermark to the result.",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -495,18 +498,21 @@ class WanTextToVideoApi(IO.ComfyNode):
                     default=False,
                     optional=True,
                     tooltip="If no audio input is provided, generate audio automatically.",
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "prompt_extend",
                     default=True,
                     tooltip="Whether to enhance the prompt with AI assistance.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "watermark",
                     default=False,
                     tooltip="Whether to add an AI-generated watermark to the result.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "shot_type",
@@ -515,6 +521,7 @@ class WanTextToVideoApi(IO.ComfyNode):
                     "single continuous shot or multiple shots with cuts. "
                     "This parameter takes effect only when prompt_extend is True.",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -667,18 +674,21 @@ class WanImageToVideoApi(IO.ComfyNode):
                     default=False,
                     optional=True,
                     tooltip="If no audio input is provided, generate audio automatically.",
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "prompt_extend",
                     default=True,
                     tooltip="Whether to enhance the prompt with AI assistance.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "watermark",
                     default=False,
                     tooltip="Whether to add an AI-generated watermark to the result.",
                     optional=True,
+                    advanced=True,
                 ),
                 IO.Combo.Input(
                     "shot_type",
@@ -687,6 +697,7 @@ class WanImageToVideoApi(IO.ComfyNode):
                     "single continuous shot or multiple shots with cuts. "
                     "This parameter takes effect only when prompt_extend is True.",
                     optional=True,
+                    advanced=True,
                 ),
             ],
             outputs=[
@@ -839,11 +850,13 @@ class WanReferenceVideoApi(IO.ComfyNode):
                     options=["single", "multi"],
                     tooltip="Specifies the shot type for the generated video, that is, whether the video is a "
                     "single continuous shot or multiple shots with cuts.",
+                    advanced=True,
                 ),
                 IO.Boolean.Input(
                     "watermark",
                     default=False,
                     tooltip="Whether to add an AI-generated watermark to the result.",
+                    advanced=True,
                 ),
             ],
             outputs=[

comfy_api_nodes/util/__init__.py

@@ -9,6 +9,8 @@ from .client import (
 from .conversions import (
     audio_bytes_to_audio_input,
     audio_input_to_mp3,
+    audio_ndarray_to_bytesio,
+    audio_tensor_to_contiguous_ndarray,
     audio_to_base64_string,
     bytesio_to_image_tensor,
     convert_mask_to_image,
@@ -33,6 +35,7 @@ from .download_helpers import (
     download_url_to_video_output,
 )
 from .upload_helpers import (
+    upload_3d_model_to_comfyapi,
     upload_audio_to_comfyapi,
     upload_file_to_comfyapi,
     upload_image_to_comfyapi,
@@ -62,6 +65,7 @@ __all__ = [
     "sync_op",
     "sync_op_raw",
     # Upload helpers
+    "upload_3d_model_to_comfyapi",
     "upload_audio_to_comfyapi",
     "upload_file_to_comfyapi",
     "upload_image_to_comfyapi",
@@ -76,6 +80,8 @@ __all__ = [
     # Conversions
     "audio_bytes_to_audio_input",
     "audio_input_to_mp3",
+    "audio_ndarray_to_bytesio",
+    "audio_tensor_to_contiguous_ndarray",
     "audio_to_base64_string",
     "bytesio_to_image_tensor",
     "convert_mask_to_image",

comfy_api_nodes/util/client.py

@@ -57,6 +57,7 @@ class _RequestConfig:
     files: dict[str, Any] | list[tuple[str, Any]] | None
     multipart_parser: Callable | None
     max_retries: int
+    max_retries_on_rate_limit: int
     retry_delay: float
     retry_backoff: float
     wait_label: str = "Waiting"
@@ -65,6 +66,7 @@ class _RequestConfig:
     final_label_on_success: str | None = "Completed"
     progress_origin_ts: float | None = None
     price_extractor: Callable[[dict[str, Any]], float | None] | None = None
+    is_rate_limited: Callable[[int, Any], bool] | None = None
 
 
 @dataclass
@@ -78,7 +80,7 @@ class _PollUIState:
     active_since: float | None = None  # start time of current active interval (None if queued)
 
 
-_RETRY_STATUS = {408, 429, 500, 502, 503, 504}
+_RETRY_STATUS = {408, 500, 502, 503, 504}  # status 429 is handled separately
 COMPLETED_STATUSES = ["succeeded", "succeed", "success", "completed", "finished", "done", "complete"]
 FAILED_STATUSES = ["cancelled", "canceled", "canceling", "fail", "failed", "error"]
 QUEUED_STATUSES = ["created", "queued", "queueing", "submitted", "initializing"]
@@ -103,6 +105,8 @@ async def sync_op(
     final_label_on_success: str | None = "Completed",
     progress_origin_ts: float | None = None,
     monitor_progress: bool = True,
+    max_retries_on_rate_limit: int = 16,
+    is_rate_limited: Callable[[int, Any], bool] | None = None,
 ) -> M:
     raw = await sync_op_raw(
         cls,
@@ -122,6 +126,8 @@ async def sync_op(
         final_label_on_success=final_label_on_success,
         progress_origin_ts=progress_origin_ts,
         monitor_progress=monitor_progress,
+        max_retries_on_rate_limit=max_retries_on_rate_limit,
+        is_rate_limited=is_rate_limited,
     )
     if not isinstance(raw, dict):
         raise Exception("Expected JSON response to validate into a Pydantic model, got non-JSON (binary or text).")
@@ -143,9 +149,9 @@ async def poll_op(
     poll_interval: float = 5.0,
     max_poll_attempts: int = 160,
     timeout_per_poll: float = 120.0,
-    max_retries_per_poll: int = 3,
+    max_retries_per_poll: int = 10,
     retry_delay_per_poll: float = 1.0,
-    retry_backoff_per_poll: float = 2.0,
+    retry_backoff_per_poll: float = 1.4,
     estimated_duration: int | None = None,
     cancel_endpoint: ApiEndpoint | None = None,
     cancel_timeout: float = 10.0,
@@ -194,6 +200,8 @@ async def sync_op_raw(
     final_label_on_success: str | None = "Completed",
     progress_origin_ts: float | None = None,
     monitor_progress: bool = True,
+    max_retries_on_rate_limit: int = 16,
+    is_rate_limited: Callable[[int, Any], bool] | None = None,
 ) -> dict[str, Any] | bytes:
     """
     Make a single network request.
@@ -222,6 +230,8 @@ async def sync_op_raw(
         final_label_on_success=final_label_on_success,
         progress_origin_ts=progress_origin_ts,
         price_extractor=price_extractor,
+        max_retries_on_rate_limit=max_retries_on_rate_limit,
+        is_rate_limited=is_rate_limited,
     )
     return await _request_base(cfg, expect_binary=as_binary)
 
@@ -240,9 +250,9 @@ async def poll_op_raw(
     poll_interval: float = 5.0,
     max_poll_attempts: int = 160,
     timeout_per_poll: float = 120.0,
-    max_retries_per_poll: int = 3,
+    max_retries_per_poll: int = 10,
     retry_delay_per_poll: float = 1.0,
-    retry_backoff_per_poll: float = 2.0,
+    retry_backoff_per_poll: float = 1.4,
     estimated_duration: int | None = None,
     cancel_endpoint: ApiEndpoint | None = None,
     cancel_timeout: float = 10.0,
@@ -506,7 +516,7 @@ def _friendly_http_message(status: int, body: Any) -> str:
     if status == 409:
         return "There is a problem with your account. Please contact support@comfy.org."
     if status == 429:
-        return "Rate Limit Exceeded: Please try again later."
+        return "Rate Limit Exceeded: The server returned 429 after all retry attempts. Please wait and try again."
     try:
         if isinstance(body, dict):
             err = body.get("error")
@@ -586,6 +596,8 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool):
     start_time = cfg.progress_origin_ts if cfg.progress_origin_ts is not None else time.monotonic()
     attempt = 0
     delay = cfg.retry_delay
+    rate_limit_attempts = 0
+    rate_limit_delay = cfg.retry_delay
     operation_succeeded: bool = False
     final_elapsed_seconds: int | None = None
     extracted_price: float | None = None
@@ -653,17 +665,14 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool):
                 payload_headers["Content-Type"] = "application/json"
                 payload_kw["json"] = cfg.data or {}
 
-            try:
-                request_logger.log_request_response(
-                    operation_id=operation_id,
-                    request_method=method,
-                    request_url=url,
-                    request_headers=dict(payload_headers) if payload_headers else None,
-                    request_params=dict(params) if params else None,
-                    request_data=request_body_log,
-                )
-            except Exception as _log_e:
-                logging.debug("[DEBUG] request logging failed: %s", _log_e)
+            request_logger.log_request_response(
+                operation_id=operation_id,
+                request_method=method,
+                request_url=url,
+                request_headers=dict(payload_headers) if payload_headers else None,
+                request_params=dict(params) if params else None,
+                request_data=request_body_log,
+            )
 
             req_coro = sess.request(method, url, params=params, **payload_kw)
             req_task = asyncio.create_task(req_coro)
@@ -688,41 +697,33 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool):
                         body = await resp.json()
                     except (ContentTypeError, json.JSONDecodeError):
                         body = await resp.text()
-                    if resp.status in _RETRY_STATUS and attempt <= cfg.max_retries:
+                    should_retry = False
+                    wait_time = 0.0
+                    retry_label = ""
+                    is_rl = resp.status == 429 or (
+                        cfg.is_rate_limited is not None and cfg.is_rate_limited(resp.status, body)
+                    )
+                    if is_rl and rate_limit_attempts < cfg.max_retries_on_rate_limit:
+                        rate_limit_attempts += 1
+                        wait_time = min(rate_limit_delay, 30.0)
+                        rate_limit_delay *= cfg.retry_backoff
+                        retry_label = f"rate-limit retry {rate_limit_attempts} of {cfg.max_retries_on_rate_limit}"
+                        should_retry = True
+                    elif resp.status in _RETRY_STATUS and (attempt - rate_limit_attempts) <= cfg.max_retries:
+                        wait_time = delay
+                        delay *= cfg.retry_backoff
+                        retry_label = f"retry {attempt - rate_limit_attempts} of {cfg.max_retries}"
+                        should_retry = True
+
+                    if should_retry:
                         logging.warning(
-                            "HTTP %s %s -> %s. Retrying in %.2fs (retry %d of %d).",
+                            "HTTP %s %s -> %s. Waiting %.2fs (%s).",
                             method,
                             url,
                             resp.status,
-                            delay,
-                            attempt,
-                            cfg.max_retries,
+                            wait_time,
+                            retry_label,
                         )
-                        try:
-                            request_logger.log_request_response(
-                                operation_id=operation_id,
-                                request_method=method,
-                                request_url=url,
-                                response_status_code=resp.status,
-                                response_headers=dict(resp.headers),
-                                response_content=body,
-                                error_message=_friendly_http_message(resp.status, body),
-                            )
-                        except Exception as _log_e:
-                            logging.debug("[DEBUG] response logging failed: %s", _log_e)
-
-                        await sleep_with_interrupt(
-                            delay,
-                            cfg.node_cls,
-                            cfg.wait_label if cfg.monitor_progress else None,
-                            start_time if cfg.monitor_progress else None,
-                            cfg.estimated_total,
-                            display_callback=_display_time_progress if cfg.monitor_progress else None,
-                        )
-                        delay *= cfg.retry_backoff
-                        continue
-                    msg = _friendly_http_message(resp.status, body)
-                    try:
                         request_logger.log_request_response(
                             operation_id=operation_id,
                             request_method=method,
@@ -730,10 +731,27 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool):
                             response_status_code=resp.status,
                             response_headers=dict(resp.headers),
                             response_content=body,
-                            error_message=msg,
+                            error_message=f"HTTP {resp.status} ({retry_label}, will retry in {wait_time:.1f}s)",
                         )
-                    except Exception as _log_e:
-                        logging.debug("[DEBUG] response logging failed: %s", _log_e)
+                        await sleep_with_interrupt(
+                            wait_time,
+                            cfg.node_cls,
+                            cfg.wait_label if cfg.monitor_progress else None,
+                            start_time if cfg.monitor_progress else None,
+                            cfg.estimated_total,
+                            display_callback=_display_time_progress if cfg.monitor_progress else None,
+                        )
+                        continue
+                    msg = _friendly_http_message(resp.status, body)
+                    request_logger.log_request_response(
+                        operation_id=operation_id,
+                        request_method=method,
+                        request_url=url,
+                        response_status_code=resp.status,
+                        response_headers=dict(resp.headers),
+                        response_content=body,
+                        error_message=msg,
+                    )
                     raise Exception(msg)
 
                 if expect_binary:
@@ -753,17 +771,14 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool):
                     bytes_payload = bytes(buff)
                     operation_succeeded = True
                     final_elapsed_seconds = int(time.monotonic() - start_time)
-                    try:
-                        request_logger.log_request_response(
-                            operation_id=operation_id,
-                            request_method=method,
-                            request_url=url,
-                            response_status_code=resp.status,
-                            response_headers=dict(resp.headers),
-                            response_content=bytes_payload,
-                        )
-                    except Exception as _log_e:
-                        logging.debug("[DEBUG] response logging failed: %s", _log_e)
+                    request_logger.log_request_response(
+                        operation_id=operation_id,
+                        request_method=method,
+                        request_url=url,
+                        response_status_code=resp.status,
+                        response_headers=dict(resp.headers),
+                        response_content=bytes_payload,
+                    )
                     return bytes_payload
                 else:
                     try:
@@ -780,45 +795,39 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool):
                         extracted_price = cfg.price_extractor(payload) if cfg.price_extractor else None
                     operation_succeeded = True
                     final_elapsed_seconds = int(time.monotonic() - start_time)
-                    try:
-                        request_logger.log_request_response(
-                            operation_id=operation_id,
-                            request_method=method,
-                            request_url=url,
-                            response_status_code=resp.status,
-                            response_headers=dict(resp.headers),
-                            response_content=response_content_to_log,
-                        )
-                    except Exception as _log_e:
-                        logging.debug("[DEBUG] response logging failed: %s", _log_e)
+                    request_logger.log_request_response(
+                        operation_id=operation_id,
+                        request_method=method,
+                        request_url=url,
+                        response_status_code=resp.status,
+                        response_headers=dict(resp.headers),
+                        response_content=response_content_to_log,
+                    )
                     return payload
 
         except ProcessingInterrupted:
             logging.debug("Polling was interrupted by user")
             raise
         except (ClientError, OSError) as e:
-            if attempt <= cfg.max_retries:
+            if (attempt - rate_limit_attempts) <= cfg.max_retries:
                 logging.warning(
                     "Connection error calling %s %s. Retrying in %.2fs (%d/%d): %s",
                     method,
                     url,
                     delay,
-                    attempt,
+                    attempt - rate_limit_attempts,
                     cfg.max_retries,
                     str(e),
                 )
-                try:
-                    request_logger.log_request_response(
-                        operation_id=operation_id,
-                        request_method=method,
-                        request_url=url,
-                        request_headers=dict(payload_headers) if payload_headers else None,
-                        request_params=dict(params) if params else None,
-                        request_data=request_body_log,
-                        error_message=f"{type(e).__name__}: {str(e)} (will retry)",
-                    )
-                except Exception as _log_e:
-                    logging.debug("[DEBUG] request error logging failed: %s", _log_e)
+                request_logger.log_request_response(
+                    operation_id=operation_id,
+                    request_method=method,
+                    request_url=url,
+                    request_headers=dict(payload_headers) if payload_headers else None,
+                    request_params=dict(params) if params else None,
+                    request_data=request_body_log,
+                    error_message=f"{type(e).__name__}: {str(e)} (will retry)",
+                )
                 await sleep_with_interrupt(
                     delay,
                     cfg.node_cls,
@@ -831,23 +840,6 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool):
                 continue
             diag = await _diagnose_connectivity()
             if not diag["internet_accessible"]:
-                try:
-                    request_logger.log_request_response(
-                        operation_id=operation_id,
-                        request_method=method,
-                        request_url=url,
-                        request_headers=dict(payload_headers) if payload_headers else None,
-                        request_params=dict(params) if params else None,
-                        request_data=request_body_log,
-                        error_message=f"LocalNetworkError: {str(e)}",
-                    )
-                except Exception as _log_e:
-                    logging.debug("[DEBUG] final error logging failed: %s", _log_e)
-                raise LocalNetworkError(
-                    "Unable to connect to the API server due to local network issues. "
-                    "Please check your internet connection and try again."
-                ) from e
-            try:
                 request_logger.log_request_response(
                     operation_id=operation_id,
                     request_method=method,
@@ -855,10 +847,21 @@ async def _request_base(cfg: _RequestConfig, expect_binary: bool):
                     request_headers=dict(payload_headers) if payload_headers else None,
                     request_params=dict(params) if params else None,
                     request_data=request_body_log,
-                    error_message=f"ApiServerError: {str(e)}",
+                    error_message=f"LocalNetworkError: {str(e)}",
                 )
-            except Exception as _log_e:
-                logging.debug("[DEBUG] final error logging failed: %s", _log_e)
+                raise LocalNetworkError(
+                    "Unable to connect to the API server due to local network issues. "
+                    "Please check your internet connection and try again."
+                ) from e
+            request_logger.log_request_response(
+                operation_id=operation_id,
+                request_method=method,
+                request_url=url,
+                request_headers=dict(payload_headers) if payload_headers else None,
+                request_params=dict(params) if params else None,
+                request_data=request_body_log,
+                error_message=f"ApiServerError: {str(e)}",
+            )
             raise ApiServerError(
                 f"The API server at {default_base_url()} is currently unreachable. "
                 f"The service may be experiencing issues."

comfy_api_nodes/util/conversions.py

@@ -57,7 +57,7 @@ def tensor_to_bytesio(
     image: torch.Tensor,
     *,
     total_pixels: int | None = 2048 * 2048,
-    mime_type: str = "image/png",
+    mime_type: str | None = "image/png",
 ) -> BytesIO:
     """Converts a torch.Tensor image to a named BytesIO object.
 

comfy_api_nodes/util/download_helpers.py

@@ -167,27 +167,25 @@ async def download_url_to_bytesio(
                     with contextlib.suppress(Exception):
                         dest.seek(0)
 
-                with contextlib.suppress(Exception):
-                    request_logger.log_request_response(
-                        operation_id=op_id,
-                        request_method="GET",
-                        request_url=url,
-                        response_status_code=resp.status,
-                        response_headers=dict(resp.headers),
-                        response_content=f"[streamed {written} bytes to dest]",
-                    )
+                request_logger.log_request_response(
+                    operation_id=op_id,
+                    request_method="GET",
+                    request_url=url,
+                    response_status_code=resp.status,
+                    response_headers=dict(resp.headers),
+                    response_content=f"[streamed {written} bytes to dest]",
+                )
                 return
         except asyncio.CancelledError:
             raise ProcessingInterrupted("Task cancelled") from None
         except (ClientError, OSError) as e:
             if attempt <= max_retries:
-                with contextlib.suppress(Exception):
-                    request_logger.log_request_response(
-                        operation_id=op_id,
-                        request_method="GET",
-                        request_url=url,
-                        error_message=f"{type(e).__name__}: {str(e)} (will retry)",
-                    )
+                request_logger.log_request_response(
+                    operation_id=op_id,
+                    request_method="GET",
+                    request_url=url,
+                    error_message=f"{type(e).__name__}: {str(e)} (will retry)",
+                )
                 await sleep_with_interrupt(delay, cls, None, None, None)
                 delay *= retry_backoff
                 continue

comfy_api_nodes/util/request_logger.py

@@ -8,7 +8,6 @@ from typing import Any
 
 import folder_paths
 
-# Get the logger instance
 logger = logging.getLogger(__name__)
 
 
@@ -91,38 +90,41 @@ def log_request_response(
     Filenames are sanitized and length-limited for cross-platform safety.
     If we still fail to write, we fall back to appending into api.log.
     """
-    log_dir = get_log_directory()
-    filepath = _build_log_filepath(log_dir, operation_id, request_url)
-
-    log_content: list[str] = []
-    log_content.append(f"Timestamp: {datetime.datetime.now().isoformat()}")
-    log_content.append(f"Operation ID: {operation_id}")
-    log_content.append("-" * 30 + " REQUEST " + "-" * 30)
-    log_content.append(f"Method: {request_method}")
-    log_content.append(f"URL: {request_url}")
-    if request_headers:
-        log_content.append(f"Headers:\n{_format_data_for_logging(request_headers)}")
-    if request_params:
-        log_content.append(f"Params:\n{_format_data_for_logging(request_params)}")
-    if request_data is not None:
-        log_content.append(f"Data/Body:\n{_format_data_for_logging(request_data)}")
-
-    log_content.append("\n" + "-" * 30 + " RESPONSE " + "-" * 30)
-    if response_status_code is not None:
-        log_content.append(f"Status Code: {response_status_code}")
-    if response_headers:
-        log_content.append(f"Headers:\n{_format_data_for_logging(response_headers)}")
-    if response_content is not None:
-        log_content.append(f"Content:\n{_format_data_for_logging(response_content)}")
-    if error_message:
-        log_content.append(f"Error:\n{error_message}")
-
     try:
-        with open(filepath, "w", encoding="utf-8") as f:
-            f.write("\n".join(log_content))
-        logger.debug("API log saved to: %s", filepath)
-    except Exception as e:
-        logger.error("Error writing API log to %s: %s", filepath, str(e))
+        log_dir = get_log_directory()
+        filepath = _build_log_filepath(log_dir, operation_id, request_url)
+
+        log_content: list[str] = []
+        log_content.append(f"Timestamp: {datetime.datetime.now().isoformat()}")
+        log_content.append(f"Operation ID: {operation_id}")
+        log_content.append("-" * 30 + " REQUEST " + "-" * 30)
+        log_content.append(f"Method: {request_method}")
+        log_content.append(f"URL: {request_url}")
+        if request_headers:
+            log_content.append(f"Headers:\n{_format_data_for_logging(request_headers)}")
+        if request_params:
+            log_content.append(f"Params:\n{_format_data_for_logging(request_params)}")
+        if request_data is not None:
+            log_content.append(f"Data/Body:\n{_format_data_for_logging(request_data)}")
+
+        log_content.append("\n" + "-" * 30 + " RESPONSE " + "-" * 30)
+        if response_status_code is not None:
+            log_content.append(f"Status Code: {response_status_code}")
+        if response_headers:
+            log_content.append(f"Headers:\n{_format_data_for_logging(response_headers)}")
+        if response_content is not None:
+            log_content.append(f"Content:\n{_format_data_for_logging(response_content)}")
+        if error_message:
+            log_content.append(f"Error:\n{error_message}")
+
+        try:
+            with open(filepath, "w", encoding="utf-8") as f:
+                f.write("\n".join(log_content))
+            logger.debug("API log saved to: %s", filepath)
+        except Exception as e:
+            logger.error("Error writing API log to %s: %s", filepath, str(e))
+    except Exception as _log_e:
+        logging.debug("[DEBUG] log_request_response failed: %s", _log_e)
 
 
 if __name__ == '__main__':

comfy_api_nodes/util/upload_helpers.py

@@ -164,6 +164,27 @@ async def upload_video_to_comfyapi(
     return await upload_file_to_comfyapi(cls, video_bytes_io, filename, upload_mime_type, wait_label)
 
 
+_3D_MIME_TYPES = {
+    "glb": "model/gltf-binary",
+    "obj": "model/obj",
+    "fbx": "application/octet-stream",
+}
+
+
+async def upload_3d_model_to_comfyapi(
+    cls: type[IO.ComfyNode],
+    model_3d: Types.File3D,
+    file_format: str,
+) -> str:
+    """Uploads a 3D model file to ComfyUI API and returns its download URL."""
+    return await upload_file_to_comfyapi(
+        cls,
+        model_3d.get_data(),
+        f"{uuid.uuid4()}.{file_format}",
+        _3D_MIME_TYPES.get(file_format, "application/octet-stream"),
+    )
+
+
 async def upload_file_to_comfyapi(
     cls: type[IO.ComfyNode],
     file_bytes_io: BytesIO,
@@ -255,17 +276,14 @@ async def upload_file(
         monitor_task = asyncio.create_task(_monitor())
         sess: aiohttp.ClientSession | None = None
         try:
-            try:
-                request_logger.log_request_response(
-                    operation_id=operation_id,
-                    request_method="PUT",
-                    request_url=upload_url,
-                    request_headers=headers or None,
-                    request_params=None,
-                    request_data=f"[File data {len(data)} bytes]",
-                )
-            except Exception as e:
-                logging.debug("[DEBUG] upload request logging failed: %s", e)
+            request_logger.log_request_response(
+                operation_id=operation_id,
+                request_method="PUT",
+                request_url=upload_url,
+                request_headers=headers or None,
+                request_params=None,
+                request_data=f"[File data {len(data)} bytes]",
+            )
 
             sess = aiohttp.ClientSession(timeout=timeout)
             req = sess.put(upload_url, data=data, headers=headers, skip_auto_headers=skip_auto_headers)
@@ -311,31 +329,27 @@ async def upload_file(
                         delay *= retry_backoff
                         continue
                     raise Exception(f"Failed to upload (HTTP {resp.status}).")
-                try:
-                    request_logger.log_request_response(
-                        operation_id=operation_id,
-                        request_method="PUT",
-                        request_url=upload_url,
-                        response_status_code=resp.status,
-                        response_headers=dict(resp.headers),
-                        response_content="File uploaded successfully.",
-                    )
-                except Exception as e:
-                    logging.debug("[DEBUG] upload response logging failed: %s", e)
+                request_logger.log_request_response(
+                    operation_id=operation_id,
+                    request_method="PUT",
+                    request_url=upload_url,
+                    response_status_code=resp.status,
+                    response_headers=dict(resp.headers),
+                    response_content="File uploaded successfully.",
+                )
                 return
         except asyncio.CancelledError:
             raise ProcessingInterrupted("Task cancelled") from None
         except (aiohttp.ClientError, OSError) as e:
             if attempt <= max_retries:
-                with contextlib.suppress(Exception):
-                    request_logger.log_request_response(
-                        operation_id=operation_id,
-                        request_method="PUT",
-                        request_url=upload_url,
-                        request_headers=headers or None,
-                        request_data=f"[File data {len(data)} bytes]",
-                        error_message=f"{type(e).__name__}: {str(e)} (will retry)",
-                    )
+                request_logger.log_request_response(
+                    operation_id=operation_id,
+                    request_method="PUT",
+                    request_url=upload_url,
+                    request_headers=headers or None,
+                    request_data=f"[File data {len(data)} bytes]",
+                    error_message=f"{type(e).__name__}: {str(e)} (will retry)",
+                )
                 await sleep_with_interrupt(
                     delay,
                     cls,

comfy_extras/nodes_ace.py

@@ -49,13 +49,14 @@ class TextEncodeAceStepAudio15(io.ComfyNode):
                 io.Float.Input("temperature", default=0.85, min=0.0, max=2.0, step=0.01, advanced=True),
                 io.Float.Input("top_p", default=0.9, min=0.0, max=2000.0, step=0.01, advanced=True),
                 io.Int.Input("top_k", default=0, min=0, max=100, advanced=True),
+                io.Float.Input("min_p", default=0.000, min=0.0, max=1.0, step=0.001, advanced=True),
             ],
             outputs=[io.Conditioning.Output()],
         )
 
     @classmethod
-    def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale, generate_audio_codes, cfg_scale, temperature, top_p, top_k) -> io.NodeOutput:
-        tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, seed=seed, generate_audio_codes=generate_audio_codes, cfg_scale=cfg_scale, temperature=temperature, top_p=top_p, top_k=top_k)
+    def execute(cls, clip, tags, lyrics, seed, bpm, duration, timesignature, language, keyscale, generate_audio_codes, cfg_scale, temperature, top_p, top_k, min_p) -> io.NodeOutput:
+        tokens = clip.tokenize(tags, lyrics=lyrics, bpm=bpm, duration=duration, timesignature=int(timesignature), language=language, keyscale=keyscale, 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)
         conditioning = clip.encode_from_tokens_scheduled(tokens)
         return io.NodeOutput(conditioning)
 

comfy_extras/nodes_advanced_samplers.py

@@ -47,8 +47,8 @@ class SamplerLCMUpscale(io.ComfyNode):
             node_id="SamplerLCMUpscale",
             category="sampling/custom_sampling/samplers",
             inputs=[
-                io.Float.Input("scale_ratio", default=1.0, min=0.1, max=20.0, step=0.01),
-                io.Int.Input("scale_steps", default=-1, min=-1, max=1000, step=1),
+                io.Float.Input("scale_ratio", default=1.0, min=0.1, max=20.0, step=0.01, advanced=True),
+                io.Int.Input("scale_steps", default=-1, min=-1, max=1000, step=1, advanced=True),
                 io.Combo.Input("upscale_method", options=cls.UPSCALE_METHODS),
             ],
             outputs=[io.Sampler.Output()],
@@ -94,7 +94,7 @@ class SamplerEulerCFGpp(io.ComfyNode):
             display_name="SamplerEulerCFG++",
             category="_for_testing",  # "sampling/custom_sampling/samplers"
             inputs=[
-                io.Combo.Input("version", options=["regular", "alternative"]),
+                io.Combo.Input("version", options=["regular", "alternative"], advanced=True),
             ],
             outputs=[io.Sampler.Output()],
             is_experimental=True,

comfy_extras/nodes_apg.py

@@ -26,6 +26,7 @@ class APG(io.ComfyNode):
                     max=10.0,
                     step=0.01,
                     tooltip="Controls the scale of the parallel guidance vector. Default CFG behavior at a setting of 1.",
+                    advanced=True,
                 ),
                 io.Float.Input(
                     "norm_threshold",
@@ -34,6 +35,7 @@ class APG(io.ComfyNode):
                     max=50.0,
                     step=0.1,
                     tooltip="Normalize guidance vector to this value, normalization disable at a setting of 0.",
+                    advanced=True,
                 ),
                 io.Float.Input(
                     "momentum",
@@ -42,6 +44,7 @@ class APG(io.ComfyNode):
                     max=1.0,
                     step=0.01,
                     tooltip="Controls a running average of guidance during diffusion, disabled at a setting of 0.",
+                    advanced=True,
                 ),
             ],
             outputs=[io.Model.Output()],