fix: move _initialized flag to end of GLContext.__init__

Prevents '_vao' attribute error when init fails partway through
and subsequent calls skip initialization due to early _initialized flag.

.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"

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:
 

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(abs(sat), 0.0, 1.0), clamp(l, 0.0, 1.0));
+    return hsl2rgb(hsl);
+}
+
+//=============================================================================
+// Main
+//=============================================================================
+
+void main() {
+    vec4 original = texture(u_image0, v_texCoord);
+
+    float hueShift   = u_float0 / 360.0;   // -180..180 -> -0.5..0.5
+    float satAmount  = u_float1 / 100.0;   // -100..100 -> -1..1
+    float lightAmount= u_float2 / 100.0;   // -100..100 -> -1..1
+    float overlap    = u_float3 / 100.0;   // 0..100 -> 0..1
+
+    vec3 result;
+
+    if (u_int0 == MODE_COLORIZE) {
+        result = colorize(original.rgb, hueShift, satAmount, lightAmount);
+        fragColor = vec4(result, original.a);
+        return;
+    }
+
+    vec3 hsx = (u_int1 == COLORSPACE_HSL)
+        ? rgb2hsl(original.rgb)
+        : rgb2hsb(original.rgb);
+
+    float weight = getModeWeight(hsx.x, u_int0, overlap);
+
+    if (u_int0 != MODE_MASTER && hsx.y < EPSILON) {
+        weight = 0.0;
+    }
+
+    if (weight > EPSILON) {
+        float h = fract(hsx.x + hueShift * weight);
+        float s = clamp(adjustSaturation(hsx.y, satAmount * weight), 0.0, 1.0);
+        float v = (u_int1 == COLORSPACE_HSL)
+            ? clamp(adjustLightness(hsx.z, lightAmount * weight), 0.0, 1.0)
+            : clamp(adjustBrightness(hsx.z, lightAmount * weight), 0.0, 1.0);
+
+        vec3 adjusted = vec3(h, s, v);
+        result = (u_int1 == COLORSPACE_HSL)
+            ? hsl2rgb(adjusted)
+            : hsb2rgb(adjusted);
+    } else {
+        result = original.rgb;
+    }
+
+    fragColor = vec4(result, original.a);
+}

blueprints/.glsl/Image_Blur_1.frag

@@ -0,0 +1,111 @@
+#version 300 es
+#pragma passes 2
+precision highp float;
+
+// Blur type constants
+const int BLUR_GAUSSIAN = 0;
+const int BLUR_BOX = 1;
+const int BLUR_RADIAL = 2;
+
+// Radial blur config
+const int RADIAL_SAMPLES = 12;
+const float RADIAL_STRENGTH = 0.0003;
+
+uniform sampler2D u_image0;
+uniform vec2 u_resolution;
+uniform int u_int0;      // Blur type (BLUR_GAUSSIAN, BLUR_BOX, BLUR_RADIAL)
+uniform float u_float0;  // Blur radius/amount
+uniform int u_pass;      // Pass index (0 = horizontal, 1 = vertical)
+
+in vec2 v_texCoord;
+layout(location = 0) out vec4 fragColor0;
+
+float gaussian(float x, float sigma) {
+    return exp(-(x * x) / (2.0 * sigma * sigma));
+}
+
+void main() {
+    vec2 texelSize = 1.0 / u_resolution;
+    float radius = max(u_float0, 0.0);
+
+    // Radial (angular) blur - single pass, doesn't use separable
+    if (u_int0 == BLUR_RADIAL) {
+        // Only execute on first pass
+        if (u_pass > 0) {
+            fragColor0 = texture(u_image0, v_texCoord);
+            return;
+        }
+
+        vec2 center = vec2(0.5);
+        vec2 dir = v_texCoord - center;
+        float dist = length(dir);
+
+        if (dist < 1e-4) {
+            fragColor0 = texture(u_image0, v_texCoord);
+            return;
+        }
+
+        vec4 sum = vec4(0.0);
+        float totalWeight = 0.0;
+        float angleStep = radius * RADIAL_STRENGTH;
+
+        dir /= dist;
+
+        float cosStep = cos(angleStep);
+        float sinStep = sin(angleStep);
+
+        float negAngle = -float(RADIAL_SAMPLES) * angleStep;
+        vec2 rotDir = vec2(
+            dir.x * cos(negAngle) - dir.y * sin(negAngle),
+            dir.x * sin(negAngle) + dir.y * cos(negAngle)
+        );
+
+        for (int i = -RADIAL_SAMPLES; i <= RADIAL_SAMPLES; i++) {
+            vec2 uv = center + rotDir * dist;
+            float w = 1.0 - abs(float(i)) / float(RADIAL_SAMPLES);
+            sum += texture(u_image0, uv) * w;
+            totalWeight += w;
+
+            rotDir = vec2(
+                rotDir.x * cosStep - rotDir.y * sinStep,
+                rotDir.x * sinStep + rotDir.y * cosStep
+            );
+        }
+
+        fragColor0 = sum / max(totalWeight, 0.001);
+        return;
+    }
+
+    // Separable Gaussian / Box blur
+    int samples = int(ceil(radius));
+
+    if (samples == 0) {
+        fragColor0 = texture(u_image0, v_texCoord);
+        return;
+    }
+
+    // Direction: pass 0 = horizontal, pass 1 = vertical
+    vec2 dir = (u_pass == 0) ? vec2(1.0, 0.0) : vec2(0.0, 1.0);
+
+    vec4 color = vec4(0.0);
+    float totalWeight = 0.0;
+    float sigma = radius / 2.0;
+
+    for (int i = -samples; i <= samples; i++) {
+        vec2 offset = dir * float(i) * texelSize;
+        vec4 sample_color = texture(u_image0, v_texCoord + offset);
+
+        float weight;
+        if (u_int0 == BLUR_GAUSSIAN) {
+            weight = gaussian(float(i), sigma);
+        } else {
+            // BLUR_BOX
+            weight = 1.0;
+        }
+
+        color += sample_color * weight;
+        totalWeight += weight;
+    }
+
+    fragColor0 = color / totalWeight;
+}

blueprints/.glsl/Image_Channels_23.frag

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

blueprints/.glsl/Image_Levels_1.frag

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

blueprints/.glsl/README.md

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

blueprints/.glsl/Sharpen_23.frag

@@ -0,0 +1,28 @@
+#version 300 es
+precision highp float;
+
+uniform sampler2D u_image0;
+uniform vec2 u_resolution;
+uniform float u_float0;  // strength [0.0 – 2.0] typical: 0.3–1.0
+
+in vec2 v_texCoord;
+layout(location = 0) out vec4 fragColor0;
+
+void main() {
+    vec2 texel = 1.0 / u_resolution;
+    
+    // Sample center and neighbors
+    vec4 center = texture(u_image0, v_texCoord);
+    vec4 top    = texture(u_image0, v_texCoord + vec2( 0.0, -texel.y));
+    vec4 bottom = texture(u_image0, v_texCoord + vec2( 0.0,  texel.y));
+    vec4 left   = texture(u_image0, v_texCoord + vec2(-texel.x,  0.0));
+    vec4 right  = texture(u_image0, v_texCoord + vec2( texel.x,  0.0));
+    
+    // Edge enhancement (Laplacian)
+    vec4 edges = center * 4.0 - top - bottom - left - right;
+    
+    // Add edges back scaled by strength
+    vec4 sharpened = center + edges * u_float0;
+    
+    fragColor0 = vec4(clamp(sharpened.rgb, 0.0, 1.0), center.a);
+}

blueprints/.glsl/Unsharp_Mask_26.frag

@@ -0,0 +1,61 @@
+#version 300 es
+precision highp float;
+
+uniform sampler2D u_image0;
+uniform vec2 u_resolution;
+uniform float u_float0;  // amount    [0.0 - 3.0]  typical: 0.5-1.5
+uniform float u_float1;  // radius    [0.5 - 10.0] blur radius in pixels
+uniform float u_float2;  // threshold [0.0 - 0.1]  min difference to sharpen
+
+in vec2 v_texCoord;
+layout(location = 0) out vec4 fragColor0;
+
+float gaussian(float x, float sigma) {
+    return exp(-(x * x) / (2.0 * sigma * sigma));
+}
+
+float getLuminance(vec3 color) {
+    return dot(color, vec3(0.2126, 0.7152, 0.0722));
+}
+
+void main() {
+    vec2 texel = 1.0 / u_resolution;
+    float radius = max(u_float1, 0.5);
+    float amount = u_float0;
+    float threshold = u_float2;
+
+    vec4 original = texture(u_image0, v_texCoord);
+
+    // Gaussian blur for the "unsharp" mask
+    int samples = int(ceil(radius));
+    float sigma = radius / 2.0;
+
+    vec4 blurred = vec4(0.0);
+    float totalWeight = 0.0;
+
+    for (int x = -samples; x <= samples; x++) {
+        for (int y = -samples; y <= samples; y++) {
+            vec2 offset = vec2(float(x), float(y)) * texel;
+            vec4 sample_color = texture(u_image0, v_texCoord + offset);
+
+            float dist = length(vec2(float(x), float(y)));
+            float weight = gaussian(dist, sigma);
+            blurred += sample_color * weight;
+            totalWeight += weight;
+        }
+    }
+    blurred /= totalWeight;
+
+    // Unsharp mask = original - blurred
+    vec3 mask = original.rgb - blurred.rgb;
+
+    // Luminance-based threshold with smooth falloff
+    float lumaDelta = abs(getLuminance(original.rgb) - getLuminance(blurred.rgb));
+    float thresholdScale = smoothstep(0.0, threshold, lumaDelta);
+    mask *= thresholdScale;
+
+    // Sharpen: original + mask * amount
+    vec3 sharpened = original.rgb + mask * amount;
+
+    fragColor0 = vec4(clamp(sharpened, 0.0, 1.0), original.a);
+}

blueprints/.glsl/update_blueprints.py

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

blueprints/Image Channels.json

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

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/audio_encoders/audio_encoders.py

@@ -25,11 +25,11 @@ class AudioEncoderModel():
         elif model_type == "whisper3":
             self.model = WhisperLargeV3(**model_config)
         self.model.eval()
-        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
         self.model_sample_rate = 16000
 
     def load_sd(self, sd):
-        return self.model.load_state_dict(sd, strict=False)
+        return self.model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
 
     def get_sd(self):
         return self.model.state_dict()

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/cli_args.py

@@ -159,6 +159,7 @@ class PerformanceFeature(enum.Enum):
     Fp8MatrixMultiplication = "fp8_matrix_mult"
     CublasOps = "cublas_ops"
     AutoTune = "autotune"
+    DynamicVRAM = "dynamic_vram"
 
 parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
 
@@ -257,3 +258,6 @@ elif args.fast == []:
 # '--fast' is provided with a list of performance features, use that list
 else:
     args.fast = set(args.fast)
+
+def enables_dynamic_vram():
+    return PerformanceFeature.DynamicVRAM in args.fast and not args.highvram and not args.gpu_only

comfy/clip_vision.py

@@ -47,10 +47,10 @@ class ClipVisionModel():
         self.model = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast)
         self.model.eval()
 
-        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
 
     def load_sd(self, sd):
-        return self.model.load_state_dict(sd, strict=False)
+        return self.model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
 
     def get_sd(self):
         return self.model.state_dict()

comfy/controlnet.py

@@ -203,7 +203,7 @@ class ControlNet(ControlBase):
         self.control_model = control_model
         self.load_device = load_device
         if control_model is not None:
-            self.control_model_wrapped = comfy.model_patcher.ModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
+            self.control_model_wrapped = comfy.model_patcher.CoreModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
 
         self.compression_ratio = compression_ratio
         self.global_average_pooling = global_average_pooling
@@ -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

@@ -5,7 +5,7 @@ from scipy import integrate
 import torch
 from torch import nn
 import torchsde
-from tqdm.auto import trange, tqdm
+from tqdm.auto import tqdm
 
 from . import utils
 from . import deis
@@ -13,6 +13,9 @@ from . import sa_solver
 import comfy.model_patcher
 import comfy.model_sampling
 
+import comfy.memory_management
+from comfy.utils import model_trange as trange
+
 def append_zero(x):
     return torch.cat([x, x.new_zeros([1])])
 

comfy/latent_formats.py

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

comfy/ldm/anima/model.py

@@ -195,8 +195,20 @@ class Anima(MiniTrainDIT):
         super().__init__(*args, **kwargs)
         self.llm_adapter = LLMAdapter(device=kwargs.get("device"), dtype=kwargs.get("dtype"), operations=kwargs.get("operations"))
 
-    def preprocess_text_embeds(self, text_embeds, text_ids):
+    def preprocess_text_embeds(self, text_embeds, text_ids, t5xxl_weights=None):
         if text_ids is not None:
-            return self.llm_adapter(text_embeds, text_ids)
+            out = self.llm_adapter(text_embeds, text_ids)
+            if t5xxl_weights is not None:
+                out = out * t5xxl_weights
+
+            if out.shape[1] < 512:
+                out = torch.nn.functional.pad(out, (0, 0, 0, 512 - out.shape[1]))
+            return out
         else:
             return text_embeds
+
+    def forward(self, x, timesteps, context, **kwargs):
+        t5xxl_ids = kwargs.pop("t5xxl_ids", None)
+        if t5xxl_ids is not None:
+            context = self.preprocess_text_embeds(context, t5xxl_ids, t5xxl_weights=kwargs.pop("t5xxl_weights", None))
+        return super().forward(x, timesteps, context, **kwargs)

comfy/ldm/chroma/layers.py

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

comfy/ldm/chroma_radiance/layers.py

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

comfy/ldm/cosmos/predict2.py

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

comfy/ldm/flux/layers.py

@@ -5,9 +5,9 @@ import torch
 from torch import Tensor, nn
 
 from .math import attention, rope
-import comfy.ops
-import comfy.ldm.common_dit
 
+# Fix import for some custom nodes, TODO: delete eventually.
+RMSNorm = None
 
 class EmbedND(nn.Module):
     def __init__(self, dim: int, theta: int, axes_dim: list):
@@ -87,20 +87,12 @@ def build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=False, yak_mlp=False, dt
             operations.Linear(mlp_hidden_dim, hidden_size, bias=True, dtype=dtype, device=device),
         )
 
-class RMSNorm(torch.nn.Module):
-    def __init__(self, dim: int, dtype=None, device=None, operations=None):
-        super().__init__()
-        self.scale = nn.Parameter(torch.empty((dim), dtype=dtype, device=device))
-
-    def forward(self, x: Tensor):
-        return comfy.ldm.common_dit.rms_norm(x, self.scale, 1e-6)
-
 
 class QKNorm(torch.nn.Module):
     def __init__(self, dim: int, dtype=None, device=None, operations=None):
         super().__init__()
-        self.query_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
-        self.key_norm = RMSNorm(dim, dtype=dtype, device=device, operations=operations)
+        self.query_norm = operations.RMSNorm(dim, dtype=dtype, device=device)
+        self.key_norm = operations.RMSNorm(dim, dtype=dtype, device=device)
 
     def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple:
         q = self.query_norm(q)
@@ -169,7 +161,7 @@ class SiLUActivation(nn.Module):
 
 
 class DoubleStreamBlock(nn.Module):
-    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, flipped_img_txt=False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False, modulation=True, mlp_silu_act=False, proj_bias=True, yak_mlp=False, dtype=None, device=None, operations=None):
         super().__init__()
 
         mlp_hidden_dim = int(hidden_size * mlp_ratio)
@@ -197,8 +189,6 @@ class DoubleStreamBlock(nn.Module):
 
         self.txt_mlp = build_mlp(hidden_size, mlp_hidden_dim, mlp_silu_act=mlp_silu_act, yak_mlp=yak_mlp, dtype=dtype, device=device, operations=operations)
 
-        self.flipped_img_txt = flipped_img_txt
-
     def forward(self, img: Tensor, txt: Tensor, vec: Tensor, pe: Tensor, attn_mask=None, modulation_dims_img=None, modulation_dims_txt=None, transformer_options={}):
         if self.modulation:
             img_mod1, img_mod2 = self.img_mod(vec)
@@ -224,32 +214,17 @@ 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
-
-            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)

comfy/ldm/flux/math.py

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

comfy/ldm/flux/model.py

@@ -16,7 +16,6 @@ from .layers import (
     SingleStreamBlock,
     timestep_embedding,
     Modulation,
-    RMSNorm
 )
 
 @dataclass
@@ -81,7 +80,7 @@ class Flux(nn.Module):
         self.txt_in = operations.Linear(params.context_in_dim, self.hidden_size, bias=params.ops_bias, dtype=dtype, device=device)
 
         if params.txt_norm:
-            self.txt_norm = RMSNorm(params.context_in_dim, dtype=dtype, device=device, operations=operations)
+            self.txt_norm = operations.RMSNorm(params.context_in_dim, dtype=dtype, device=device)
         else:
             self.txt_norm = None
 

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)
@@ -378,14 +377,14 @@ 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
 
@@ -413,7 +412,7 @@ 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"
@@ -435,9 +434,9 @@ 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/hunyuan_video/upsampler.py

@@ -109,10 +109,10 @@ class HunyuanVideo15SRModel():
         self.model_class = UPSAMPLERS.get(model_type)
         self.model = self.model_class(**config).eval()
 
-        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
 
     def load_sd(self, sd):
-        return self.model.load_state_dict(sd, strict=True)
+        return self.model.load_state_dict(sd, strict=True, assign=self.patcher.is_dynamic())
 
     def get_sd(self):
         return self.model.state_dict()

comfy/ldm/modules/attention.py

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

comfy/ldm/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

@@ -332,6 +332,12 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["{}".format(key_lora)] = k
                 key_map["transformer.{}".format(key_lora)] = k
 
+    if isinstance(model, comfy.model_base.ACEStep15):
+        for k in sdk:
+            if k.startswith("diffusion_model.decoder.") and k.endswith(".weight"):
+                key_lora = k[len("diffusion_model.decoder."):-len(".weight")]
+                key_map["base_model.model.{}".format(key_lora)] = k  # Official base model loras
+
     return key_map
 
 

comfy/lora_convert.py

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

comfy/memory_management.py

@@ -0,0 +1,81 @@
+import math
+import torch
+from typing import NamedTuple
+
+from comfy.quant_ops import QuantizedTensor
+
+class TensorGeometry(NamedTuple):
+    shape: any
+    dtype: torch.dtype
+
+    def element_size(self):
+        info = torch.finfo(self.dtype) if self.dtype.is_floating_point else torch.iinfo(self.dtype)
+        return info.bits // 8
+
+    def numel(self):
+        return math.prod(self.shape)
+
+def tensors_to_geometries(tensors, dtype=None):
+    geometries = []
+    for t in tensors:
+        if t is None or isinstance(t, QuantizedTensor):
+            geometries.append(t)
+            continue
+        tdtype = t.dtype
+        if hasattr(t, "_model_dtype"):
+            tdtype = t._model_dtype
+        if dtype is not None:
+            tdtype = dtype
+        geometries.append(TensorGeometry(shape=t.shape, dtype=tdtype))
+    return geometries
+
+def vram_aligned_size(tensor):
+    if isinstance(tensor, list):
+        return sum([vram_aligned_size(t) for t in tensor])
+
+    if isinstance(tensor, QuantizedTensor):
+        inner_tensors, _ = tensor.__tensor_flatten__()
+        return vram_aligned_size([ getattr(tensor, attr) for attr in inner_tensors ])
+
+    if tensor is None:
+        return 0
+
+    size = tensor.numel() * tensor.element_size()
+    aligment_req = 1024
+    return (size + aligment_req - 1) // aligment_req * aligment_req
+
+def interpret_gathered_like(tensors, gathered):
+    offset = 0
+    dest_views = []
+
+    if gathered.dim() != 1 or gathered.element_size() != 1:
+        raise ValueError(f"Buffer must be 1D and single-byte (got {gathered.dim()}D {gathered.dtype})")
+
+    for tensor in tensors:
+
+        if tensor is None:
+            dest_views.append(None)
+            continue
+
+        if isinstance(tensor, QuantizedTensor):
+            inner_tensors, qt_ctx = tensor.__tensor_flatten__()
+            templates = { attr: getattr(tensor, attr) for attr in inner_tensors }
+        else:
+            templates = { "data": tensor }
+
+        actuals = {}
+        for attr, template in templates.items():
+            size = template.numel() * template.element_size()
+            if offset + size > gathered.numel():
+                raise ValueError(f"Buffer too small: needs {offset + size} bytes, but only has {gathered.numel()}. ")
+            actuals[attr] = gathered[offset:offset+size].view(dtype=template.dtype).view(template.shape)
+            offset += vram_aligned_size(template)
+
+        if isinstance(tensor, QuantizedTensor):
+            dest_views.append(QuantizedTensor.__tensor_unflatten__(actuals, qt_ctx, 0, 0))
+        else:
+            dest_views.append(actuals["data"])
+
+    return dest_views
+
+aimdo_allocator = None

comfy/model_base.py

@@ -50,6 +50,7 @@ import comfy.ldm.omnigen.omnigen2
 import comfy.ldm.qwen_image.model
 import comfy.ldm.kandinsky5.model
 import comfy.ldm.anima.model
+import comfy.ldm.ace.ace_step15
 
 import comfy.model_management
 import comfy.patcher_extension
@@ -146,6 +147,8 @@ class BaseModel(torch.nn.Module):
                 self.diffusion_model.to(memory_format=torch.channels_last)
                 logging.debug("using channels last mode for diffusion model")
             logging.info("model weight dtype {}, manual cast: {}".format(self.get_dtype(), self.manual_cast_dtype))
+            comfy.model_management.archive_model_dtypes(self.diffusion_model)
+
         self.model_type = model_type
         self.model_sampling = model_sampling(model_config, model_type)
 
@@ -299,7 +302,7 @@ class BaseModel(torch.nn.Module):
 
         return out
 
-    def load_model_weights(self, sd, unet_prefix=""):
+    def load_model_weights(self, sd, unet_prefix="", assign=False):
         to_load = {}
         keys = list(sd.keys())
         for k in keys:
@@ -307,7 +310,7 @@ class BaseModel(torch.nn.Module):
                 to_load[k[len(unet_prefix):]] = sd.pop(k)
 
         to_load = self.model_config.process_unet_state_dict(to_load)
-        m, u = self.diffusion_model.load_state_dict(to_load, strict=False)
+        m, u = self.diffusion_model.load_state_dict(to_load, strict=False, assign=assign)
         if len(m) > 0:
             logging.warning("unet missing: {}".format(m))
 
@@ -322,7 +325,7 @@ class BaseModel(torch.nn.Module):
     def process_latent_out(self, latent):
         return self.latent_format.process_out(latent)
 
-    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+    def state_dict_for_saving(self, unet_state_dict, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
         extra_sds = []
         if clip_state_dict is not None:
             extra_sds.append(self.model_config.process_clip_state_dict_for_saving(clip_state_dict))
@@ -330,10 +333,7 @@ class BaseModel(torch.nn.Module):
             extra_sds.append(self.model_config.process_vae_state_dict_for_saving(vae_state_dict))
         if clip_vision_state_dict is not None:
             extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict))
-
-        unet_state_dict = self.diffusion_model.state_dict()
         unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict)
-
         if self.model_type == ModelType.V_PREDICTION:
             unet_state_dict["v_pred"] = torch.tensor([])
 
@@ -776,8 +776,8 @@ class StableAudio1(BaseModel):
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
 
-    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
-        sd = super().state_dict_for_saving(clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
+    def state_dict_for_saving(self, unet_state_dict, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+        sd = super().state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
         d = {"conditioner.conditioners.seconds_start.": self.seconds_start_embedder.state_dict(), "conditioner.conditioners.seconds_total.": self.seconds_total_embedder.state_dict()}
         for k in d:
             s = d[k]
@@ -1160,12 +1160,16 @@ class Anima(BaseModel):
         device = kwargs["device"]
         if cross_attn is not None:
             if t5xxl_ids is not None:
-                cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.unsqueeze(0).to(device=device))
                 if t5xxl_weights is not None:
-                    cross_attn *= t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn)
+                    t5xxl_weights = t5xxl_weights.unsqueeze(0).unsqueeze(-1).to(cross_attn)
+                t5xxl_ids = t5xxl_ids.unsqueeze(0)
+
+                if torch.is_inference_mode_enabled():  # if not we are training
+                    cross_attn = self.diffusion_model.preprocess_text_embeds(cross_attn.to(device=device, dtype=self.get_dtype()), t5xxl_ids.to(device=device), t5xxl_weights=t5xxl_weights.to(device=device, dtype=self.get_dtype()))
+                else:
+                    out['t5xxl_ids'] = comfy.conds.CONDRegular(t5xxl_ids)
+                    out['t5xxl_weights'] = comfy.conds.CONDRegular(t5xxl_weights)
 
-                if cross_attn.shape[1] < 512:
-                    cross_attn = torch.nn.functional.pad(cross_attn, (0, 0, 0, 512 - cross_attn.shape[1]))
             out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
         return out
 
@@ -1541,6 +1545,49 @@ class ACEStep(BaseModel):
         out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0))
         return out
 
+class ACEStep15(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ace.ace_step15.AceStepConditionGenerationModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        device = kwargs["device"]
+        noise = kwargs["noise"]
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            if torch.count_nonzero(cross_attn) == 0:
+                out['replace_with_null_embeds'] = comfy.conds.CONDConstant(True)
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
+        if cross_attn is not None:
+            out['lyric_embed'] = comfy.conds.CONDRegular(conditioning_lyrics)
+
+        refer_audio = kwargs.get("reference_audio_timbre_latents", None)
+        if refer_audio is None or len(refer_audio) == 0:
+            refer_audio = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device)
+            pass_audio_codes = True
+        else:
+            refer_audio = refer_audio[-1][:, :, :noise.shape[2]]
+            out['is_covers'] = comfy.conds.CONDConstant(True)
+            pass_audio_codes = False
+
+        if pass_audio_codes:
+            audio_codes = kwargs.get("audio_codes", None)
+            if audio_codes is not None:
+                out['audio_codes'] = comfy.conds.CONDRegular(torch.tensor(audio_codes, device=device))
+                refer_audio = refer_audio[:, :, :750]
+            else:
+                out['is_covers'] = comfy.conds.CONDConstant(False)
+
+        if refer_audio.shape[2] < noise.shape[2]:
+            pad = comfy.ldm.ace.ace_step15.get_silence_latent(noise.shape[2], device)
+            refer_audio = torch.cat([refer_audio.to(pad), pad[:, :, refer_audio.shape[2]:]], dim=2)
+
+        out['refer_audio'] = comfy.conds.CONDRegular(refer_audio)
+        return out
+
 class Omnigen2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel)

comfy/model_detection.py

@@ -19,6 +19,12 @@ def count_blocks(state_dict_keys, prefix_string):
         count += 1
     return count
 
+def any_suffix_in(keys, prefix, main, suffix_list=[]):
+    for x in suffix_list:
+        if "{}{}{}".format(prefix, main, x) in keys:
+            return True
+    return False
+
 def calculate_transformer_depth(prefix, state_dict_keys, state_dict):
     context_dim = None
     use_linear_in_transformer = False
@@ -186,7 +192,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["meanflow_sum"] = False
         return dit_config
 
-    if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or f"{key_prefix}distilled_guidance_layer.norms.0.scale" in state_dict_keys): #Flux, Chroma or Chroma Radiance (has no img_in.weight)
+    if any_suffix_in(state_dict_keys, key_prefix, 'double_blocks.0.img_attn.norm.key_norm.', ["weight", "scale"]) and ('{}img_in.weight'.format(key_prefix) in state_dict_keys or any_suffix_in(state_dict_keys, key_prefix, 'distilled_guidance_layer.norms.0.', ["weight", "scale"])): #Flux, Chroma or Chroma Radiance (has no img_in.weight)
         dit_config = {}
         if '{}double_stream_modulation_img.lin.weight'.format(key_prefix) in state_dict_keys:
             dit_config["image_model"] = "flux2"
@@ -241,7 +247,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
 
         dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
         dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
-        if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma
+
+        if any_suffix_in(state_dict_keys, key_prefix, 'distilled_guidance_layer.0.norms.0.', ["weight", "scale"]) or any_suffix_in(state_dict_keys, key_prefix, 'distilled_guidance_layer.norms.0.', ["weight", "scale"]): #Chroma
             dit_config["image_model"] = "chroma"
             dit_config["in_channels"] = 64
             dit_config["out_channels"] = 64
@@ -249,7 +256,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["out_dim"] = 3072
             dit_config["hidden_dim"] = 5120
             dit_config["n_layers"] = 5
-            if f"{key_prefix}nerf_blocks.0.norm.scale" in state_dict_keys: #Chroma Radiance
+
+            if any_suffix_in(state_dict_keys, key_prefix, 'nerf_blocks.0.norm.', ["weight", "scale"]): #Chroma Radiance
                 dit_config["image_model"] = "chroma_radiance"
                 dit_config["in_channels"] = 3
                 dit_config["out_channels"] = 3
@@ -259,7 +267,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
                 dit_config["nerf_depth"] = 4
                 dit_config["nerf_max_freqs"] = 8
                 dit_config["nerf_tile_size"] = 512
-                dit_config["nerf_final_head_type"] = "conv" if f"{key_prefix}nerf_final_layer_conv.norm.scale" in state_dict_keys else "linear"
+                dit_config["nerf_final_head_type"] = "conv" if any_suffix_in(state_dict_keys, key_prefix, 'nerf_final_layer_conv.norm.', ["weight", "scale"]) else "linear"
                 dit_config["nerf_embedder_dtype"] = torch.float32
                 if "{}__x0__".format(key_prefix) in state_dict_keys: # x0 pred
                     dit_config["use_x0"] = True
@@ -268,7 +276,7 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         else:
             dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
             dit_config["yak_mlp"] = '{}double_blocks.0.img_mlp.gate_proj.weight'.format(key_prefix) in state_dict_keys
-            dit_config["txt_norm"] = "{}txt_norm.scale".format(key_prefix) in state_dict_keys
+            dit_config["txt_norm"] = any_suffix_in(state_dict_keys, key_prefix, 'txt_norm.', ["weight", "scale"])
             if dit_config["yak_mlp"] and dit_config["txt_norm"]:  # Ovis model
                 dit_config["txt_ids_dims"] = [1, 2]
 
@@ -655,6 +663,11 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["num_visual_blocks"] = count_blocks(state_dict_keys, '{}visual_transformer_blocks.'.format(key_prefix) + '{}.')
         return dit_config
 
+    if '{}encoder.lyric_encoder.layers.0.input_layernorm.weight'.format(key_prefix) in state_dict_keys:
+        dit_config = {}
+        dit_config["audio_model"] = "ace1.5"
+        return dit_config
+
     if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
         return None
 

comfy/model_management.py

@@ -20,12 +20,20 @@ import psutil
 import logging
 from enum import Enum
 from comfy.cli_args import args, PerformanceFeature
+import threading
 import torch
 import sys
 import platform
 import weakref
 import gc
 import os
+from contextlib import nullcontext
+import comfy.memory_management
+import comfy.utils
+import comfy.quant_ops
+
+import comfy_aimdo.torch
+import comfy_aimdo.model_vbar
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -47,6 +55,11 @@ cpu_state = CPUState.GPU
 
 total_vram = 0
 
+
+# Training Related State
+in_training = False
+
+
 def get_supported_float8_types():
     float8_types = []
     try:
@@ -578,9 +591,15 @@ WINDOWS = any(platform.win32_ver())
 
 EXTRA_RESERVED_VRAM = 400 * 1024 * 1024
 if WINDOWS:
+    import comfy.windows
     EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue
     if total_vram > (15 * 1024):  # more extra reserved vram on 16GB+ cards
         EXTRA_RESERVED_VRAM += 100 * 1024 * 1024
+    def get_free_ram():
+        return comfy.windows.get_free_ram()
+else:
+    def get_free_ram():
+        return psutil.virtual_memory().available
 
 if args.reserve_vram is not None:
     EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
@@ -592,7 +611,7 @@ def extra_reserved_memory():
 def minimum_inference_memory():
     return (1024 * 1024 * 1024) * 0.8 + extra_reserved_memory()
 
-def free_memory(memory_required, device, keep_loaded=[]):
+def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_required=0):
     cleanup_models_gc()
     unloaded_model = []
     can_unload = []
@@ -607,15 +626,23 @@ def free_memory(memory_required, device, keep_loaded=[]):
 
     for x in sorted(can_unload):
         i = x[-1]
-        memory_to_free = None
+        memory_to_free = 1e32
+        ram_to_free = 1e32
         if not DISABLE_SMART_MEMORY:
-            free_mem = get_free_memory(device)
-            if free_mem > memory_required:
-                break
-            memory_to_free = memory_required - free_mem
-        logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
-        if current_loaded_models[i].model_unload(memory_to_free):
+            memory_to_free = memory_required - get_free_memory(device)
+            ram_to_free = ram_required - get_free_ram()
+
+        if current_loaded_models[i].model.is_dynamic() and for_dynamic:
+            #don't actually unload dynamic models for the sake of other dynamic models
+            #as that works on-demand.
+            memory_required -= current_loaded_models[i].model.loaded_size()
+            memory_to_free = 0
+        if memory_to_free > 0 and current_loaded_models[i].model_unload(memory_to_free):
+            logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
             unloaded_model.append(i)
+        if ram_to_free > 0:
+            logging.debug(f"RAM Unloading {current_loaded_models[i].model.model.__class__.__name__}")
+            current_loaded_models[i].model.partially_unload_ram(ram_to_free)
 
     for i in sorted(unloaded_model, reverse=True):
         unloaded_models.append(current_loaded_models.pop(i))
@@ -650,7 +677,10 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
 
     models_to_load = []
 
+    free_for_dynamic=True
     for x in models:
+        if not x.is_dynamic():
+            free_for_dynamic = False
         loaded_model = LoadedModel(x)
         try:
             loaded_model_index = current_loaded_models.index(loaded_model)
@@ -676,19 +706,25 @@ def load_models_gpu(models, memory_required=0, force_patch_weights=False, minimu
             model_to_unload.model.detach(unpatch_all=False)
             model_to_unload.model_finalizer.detach()
 
+
     total_memory_required = {}
+    total_ram_required = {}
     for loaded_model in models_to_load:
         total_memory_required[loaded_model.device] = total_memory_required.get(loaded_model.device, 0) + loaded_model.model_memory_required(loaded_model.device)
+        #x2, one to make sure the OS can fit the model for loading in disk cache, and for us to do any pinning we
+        #want to do.
+        #FIXME: This should subtract off the to_load current pin consumption.
+        total_ram_required[loaded_model.device] = total_ram_required.get(loaded_model.device, 0) + loaded_model.model_memory() * 2
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
-            free_memory(total_memory_required[device] * 1.1 + extra_mem, device)
+            free_memory(total_memory_required[device] * 1.1 + extra_mem, device, for_dynamic=free_for_dynamic, ram_required=total_ram_required[device])
 
     for device in total_memory_required:
         if device != torch.device("cpu"):
             free_mem = get_free_memory(device)
             if free_mem < minimum_memory_required:
-                models_l = free_memory(minimum_memory_required, device)
+                models_l = free_memory(minimum_memory_required, device, for_dynamic=free_for_dynamic)
                 logging.info("{} models unloaded.".format(len(models_l)))
 
     for loaded_model in models_to_load:
@@ -732,6 +768,9 @@ def loaded_models(only_currently_used=False):
 
 def cleanup_models_gc():
     do_gc = False
+
+    reset_cast_buffers()
+
     for i in range(len(current_loaded_models)):
         cur = current_loaded_models[i]
         if cur.is_dead():
@@ -749,6 +788,11 @@ def cleanup_models_gc():
                 logging.warning("WARNING, memory leak with model {}. Please make sure it is not being referenced from somewhere.".format(cur.real_model().__class__.__name__))
 
 
+def archive_model_dtypes(model):
+    for name, module in model.named_modules():
+        for param_name, param in module.named_parameters(recurse=False):
+            setattr(module, f"{param_name}_comfy_model_dtype", param.dtype)
+
 
 def cleanup_models():
     to_delete = []
@@ -792,7 +836,7 @@ def unet_inital_load_device(parameters, dtype):
 
     mem_dev = get_free_memory(torch_dev)
     mem_cpu = get_free_memory(cpu_dev)
-    if mem_dev > mem_cpu and model_size < mem_dev:
+    if mem_dev > mem_cpu and model_size < mem_dev and comfy.memory_management.aimdo_allocator is None:
         return torch_dev
     else:
         return cpu_dev
@@ -1051,6 +1095,51 @@ def current_stream(device):
         return None
 
 stream_counters = {}
+
+STREAM_CAST_BUFFERS = {}
+LARGEST_CASTED_WEIGHT = (None, 0)
+
+def get_cast_buffer(offload_stream, device, size, ref):
+    global LARGEST_CASTED_WEIGHT
+
+    if offload_stream is not None:
+        wf_context = offload_stream
+        if hasattr(wf_context, "as_context"):
+            wf_context = wf_context.as_context(offload_stream)
+    else:
+        wf_context = nullcontext()
+
+    cast_buffer = STREAM_CAST_BUFFERS.get(offload_stream, None)
+    if cast_buffer is None or cast_buffer.numel() < size:
+        if ref is LARGEST_CASTED_WEIGHT[0]:
+            #If there is one giant weight we do not want both streams to
+            #allocate a buffer for it. It's up to the caster to get the other
+            #offload stream in this corner case
+            return None
+        if cast_buffer is not None and cast_buffer.numel() > 50 * (1024 ** 2):
+            #I want my wrongly sized 50MB+ of VRAM back from the caching allocator right now
+            synchronize()
+            del STREAM_CAST_BUFFERS[offload_stream]
+            del cast_buffer
+            #FIXME: This doesn't work in Aimdo because mempool cant clear cache
+            soft_empty_cache()
+        with wf_context:
+            cast_buffer = torch.empty((size), dtype=torch.int8, device=device)
+            STREAM_CAST_BUFFERS[offload_stream] = cast_buffer
+
+        if  size > LARGEST_CASTED_WEIGHT[1]:
+            LARGEST_CASTED_WEIGHT = (ref, size)
+
+    return cast_buffer
+
+def reset_cast_buffers():
+    global LARGEST_CASTED_WEIGHT
+    LARGEST_CASTED_WEIGHT = (None, 0)
+    for offload_stream in STREAM_CAST_BUFFERS:
+        offload_stream.synchronize()
+    STREAM_CAST_BUFFERS.clear()
+    soft_empty_cache()
+
 def get_offload_stream(device):
     stream_counter = stream_counters.get(device, 0)
     if NUM_STREAMS == 0:
@@ -1093,7 +1182,61 @@ def sync_stream(device, stream):
         return
     current_stream(device).wait_stream(stream)
 
-def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None):
+
+def cast_to_gathered(tensors, r, non_blocking=False, stream=None):
+    wf_context = nullcontext()
+    if stream is not None:
+       wf_context = stream
+       if hasattr(wf_context, "as_context"):
+           wf_context = wf_context.as_context(stream)
+
+    dest_views = comfy.memory_management.interpret_gathered_like(tensors, r)
+    with wf_context:
+        for tensor in tensors:
+            dest_view = dest_views.pop(0)
+            if tensor is None:
+                continue
+            dest_view.copy_(tensor, non_blocking=non_blocking)
+
+
+def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, stream=None, r=None):
+    if hasattr(weight, "_v"):
+        #Unexpected usage patterns. There is no reason these don't work but they
+        #have no testing and no callers do this.
+        assert r is None
+        assert stream is None
+
+        cast_geometry = comfy.memory_management.tensors_to_geometries([ weight ])
+
+        if dtype is None:
+            dtype = weight._model_dtype
+
+        signature = comfy_aimdo.model_vbar.vbar_fault(weight._v)
+        if signature is not None:
+            if comfy_aimdo.model_vbar.vbar_signature_compare(signature, weight._v_signature):
+                v_tensor = weight._v_tensor
+            else:
+                raw_tensor = comfy_aimdo.torch.aimdo_to_tensor(weight._v, device)
+                v_tensor = comfy.memory_management.interpret_gathered_like(cast_geometry, raw_tensor)[0]
+                weight._v_tensor = v_tensor
+                weight._v_signature = signature
+                #Send it over
+                v_tensor.copy_(weight, non_blocking=non_blocking)
+            return v_tensor.to(dtype=dtype)
+
+        r = torch.empty_like(weight, dtype=dtype, device=device)
+
+        if weight.dtype != r.dtype and weight.dtype != weight._model_dtype:
+            #Offloaded casting could skip this, however it would make the quantizations
+            #inconsistent between loaded and offloaded weights. So force the double casting
+            #that would happen in regular flow to make offload deterministic.
+            cast_buffer = torch.empty_like(weight, dtype=weight._model_dtype, device=device)
+            cast_buffer.copy_(weight, non_blocking=non_blocking)
+            weight = cast_buffer
+        r.copy_(weight, non_blocking=non_blocking)
+
+        return r
+
     if device is None or weight.device == device:
         if not copy:
             if dtype is None or weight.dtype == dtype:
@@ -1112,10 +1255,12 @@ def cast_to(weight, dtype=None, device=None, non_blocking=False, copy=False, str
         if hasattr(wf_context, "as_context"):
             wf_context = wf_context.as_context(stream)
         with wf_context:
-            r = torch.empty_like(weight, dtype=dtype, device=device)
+            if r is None:
+                r = torch.empty_like(weight, dtype=dtype, device=device)
             r.copy_(weight, non_blocking=non_blocking)
     else:
-        r = torch.empty_like(weight, dtype=dtype, device=device)
+        if r is None:
+            r = torch.empty_like(weight, dtype=dtype, device=device)
         r.copy_(weight, non_blocking=non_blocking)
     return r
 
@@ -1135,14 +1280,14 @@ if not args.disable_pinned_memory:
             MAX_PINNED_MEMORY = get_total_memory(torch.device("cpu")) * 0.95
         logging.info("Enabled pinned memory {}".format(MAX_PINNED_MEMORY // (1024 * 1024)))
 
-PINNING_ALLOWED_TYPES = set(["Parameter", "QuantizedTensor"])
+PINNING_ALLOWED_TYPES = set(["Tensor", "Parameter", "QuantizedTensor"])
 
 def discard_cuda_async_error():
     try:
         a = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
         b = torch.tensor([1], dtype=torch.uint8, device=get_torch_device())
         _ = a + b
-        torch.cuda.synchronize()
+        synchronize()
     except torch.AcceleratorError:
         #Dump it! We already know about it from the synchronous return
         pass
@@ -1546,6 +1691,12 @@ def lora_compute_dtype(device):
     LORA_COMPUTE_DTYPES[device] = dtype
     return dtype
 
+def synchronize():
+    if is_intel_xpu():
+        torch.xpu.synchronize()
+    elif torch.cuda.is_available():
+        torch.cuda.synchronize()
+
 def soft_empty_cache(force=False):
     global cpu_state
     if cpu_state == CPUState.MPS:
@@ -1557,6 +1708,7 @@ def soft_empty_cache(force=False):
     elif is_mlu():
         torch.mlu.empty_cache()
     elif torch.cuda.is_available():
+        torch.cuda.synchronize()
         torch.cuda.empty_cache()
         torch.cuda.ipc_collect()
 
@@ -1568,9 +1720,6 @@ def debug_memory_summary():
         return torch.cuda.memory.memory_summary()
     return ""
 
-#TODO: might be cleaner to put this somewhere else
-import threading
-
 class InterruptProcessingException(Exception):
     pass
 

comfy/model_patcher.py

@@ -19,7 +19,6 @@
 from __future__ import annotations
 
 import collections
-import copy
 import inspect
 import logging
 import math
@@ -38,19 +37,7 @@ from comfy.comfy_types import UnetWrapperFunction
 from comfy.quant_ops import QuantizedTensor
 from comfy.patcher_extension import CallbacksMP, PatcherInjection, WrappersMP
 
-
-def string_to_seed(data):
-    crc = 0xFFFFFFFF
-    for byte in data:
-        if isinstance(byte, str):
-            byte = ord(byte)
-        crc ^= byte
-        for _ in range(8):
-            if crc & 1:
-                crc = (crc >> 1) ^ 0xEDB88320
-            else:
-                crc >>= 1
-    return crc ^ 0xFFFFFFFF
+import comfy_aimdo.model_vbar
 
 def set_model_options_patch_replace(model_options, patch, name, block_name, number, transformer_index=None):
     to = model_options["transformer_options"].copy()
@@ -123,6 +110,10 @@ def move_weight_functions(m, device):
                 memory += f.move_to(device=device)
     return memory
 
+def string_to_seed(data):
+    logging.warning("WARNING: string_to_seed has moved from comfy.model_patcher to comfy.utils")
+    return comfy.utils.string_to_seed(data)
+
 class LowVramPatch:
     def __init__(self, key, patches, convert_func=None, set_func=None):
         self.key = key
@@ -169,6 +160,11 @@ def get_key_weight(model, key):
 
     return weight, set_func, convert_func
 
+def key_param_name_to_key(key, param):
+    if len(key) == 0:
+        return param
+    return "{}.{}".format(key, param)
+
 class AutoPatcherEjector:
     def __init__(self, model: 'ModelPatcher', skip_and_inject_on_exit_only=False):
         self.model = model
@@ -212,6 +208,27 @@ class MemoryCounter:
     def decrement(self, used: int):
         self.value -= used
 
+CustomTorchDevice = collections.namedtuple("FakeDevice", ["type", "index"])("comfy-lazy-caster", 0)
+
+class LazyCastingParam(torch.nn.Parameter):
+    def __new__(cls, model, key, tensor):
+        return super().__new__(cls, tensor)
+
+    def __init__(self, model, key, tensor):
+        self.model = model
+        self.key = key
+
+    @property
+    def device(self):
+        return CustomTorchDevice
+
+    #safetensors will .to() us to the cpu which we catch here to cast on demand. The returned tensor is
+    #then just a short lived thing in the safetensors serialization logic inside its big for loop over
+    #all weights getting garbage collected per-weight
+    def to(self, *args, **kwargs):
+        return self.model.patch_weight_to_device(self.key, device_to=self.model.load_device, return_weight=True).to("cpu")
+
+
 class ModelPatcher:
     def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False):
         self.size = size
@@ -269,6 +286,9 @@ class ModelPatcher:
         if not hasattr(self.model, 'model_offload_buffer_memory'):
             self.model.model_offload_buffer_memory = 0
 
+    def is_dynamic(self):
+        return False
+
     def model_size(self):
         if self.size > 0:
             return self.size
@@ -284,6 +304,9 @@ class ModelPatcher:
     def lowvram_patch_counter(self):
         return self.model.lowvram_patch_counter
 
+    def get_free_memory(self, device):
+        return comfy.model_management.get_free_memory(device)
+
     def clone(self):
         n = self.__class__(self.model, self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update)
         n.patches = {}
@@ -293,7 +316,7 @@ class ModelPatcher:
 
         n.object_patches = self.object_patches.copy()
         n.weight_wrapper_patches = self.weight_wrapper_patches.copy()
-        n.model_options = copy.deepcopy(self.model_options)
+        n.model_options = comfy.utils.deepcopy_list_dict(self.model_options)
         n.backup = self.backup
         n.object_patches_backup = self.object_patches_backup
         n.parent = self
@@ -611,14 +634,14 @@ class ModelPatcher:
                         sd.pop(k)
             return sd
 
-    def patch_weight_to_device(self, key, device_to=None, inplace_update=False):
-        if key not in self.patches:
-            return
-
+    def patch_weight_to_device(self, key, device_to=None, inplace_update=False, return_weight=False):
         weight, set_func, convert_func = get_key_weight(self.model, key)
+        if key not in self.patches:
+            return weight
+
         inplace_update = self.weight_inplace_update or inplace_update
 
-        if key not in self.backup:
+        if key not in self.backup and not return_weight:
             self.backup[key] = collections.namedtuple('Dimension', ['weight', 'inplace_update'])(weight.to(device=self.offload_device, copy=inplace_update), inplace_update)
 
         temp_dtype = comfy.model_management.lora_compute_dtype(device_to)
@@ -631,13 +654,15 @@ class ModelPatcher:
 
         out_weight = comfy.lora.calculate_weight(self.patches[key], temp_weight, key)
         if set_func is None:
-            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key))
-            if inplace_update:
+            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key))
+            if return_weight:
+                return out_weight
+            elif inplace_update:
                 comfy.utils.copy_to_param(self.model, key, out_weight)
             else:
                 comfy.utils.set_attr_param(self.model, key, out_weight)
         else:
-            set_func(out_weight, inplace_update=inplace_update, seed=string_to_seed(key))
+            return set_func(out_weight, inplace_update=inplace_update, seed=comfy.utils.string_to_seed(key), return_weight=return_weight)
 
     def pin_weight_to_device(self, key):
         weight, set_func, convert_func = get_key_weight(self.model, key)
@@ -654,18 +679,19 @@ class ModelPatcher:
         for key in list(self.pinned):
             self.unpin_weight(key)
 
-    def _load_list(self):
+    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"):
@@ -681,7 +707,8 @@ class ModelPatcher:
                         return 0
                     module_offload_mem += check_module_offload_mem("{}.weight".format(n))
                     module_offload_mem += check_module_offload_mem("{}.bias".format(n))
-                loading.append((module_offload_mem, module_mem, n, m, params))
+                prepend = (not hasattr(m, "comfy_cast_weights"),) if prio_comfy_cast_weights else ()
+                loading.append(prepend + (module_offload_mem, module_mem, n, m, params))
         return loading
 
     def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False):
@@ -773,7 +800,7 @@ class ModelPatcher:
                         continue
 
                 for param in params:
-                    key = "{}.{}".format(n, param)
+                    key = key_param_name_to_key(n, param)
                     self.unpin_weight(key)
                     self.patch_weight_to_device(key, device_to=device_to)
                 if comfy.model_management.is_device_cuda(device_to):
@@ -789,7 +816,7 @@ class ModelPatcher:
                 n = x[1]
                 params = x[3]
                 for param in params:
-                    self.pin_weight_to_device("{}.{}".format(n, param))
+                    self.pin_weight_to_device(key_param_name_to_key(n, param))
 
             usable_stat = "{:.2f} MB usable,".format(lowvram_model_memory / (1024 * 1024)) if lowvram_model_memory < 1e32 else ""
             if lowvram_counter > 0:
@@ -895,7 +922,7 @@ class ModelPatcher:
                 if hasattr(m, "comfy_patched_weights") and m.comfy_patched_weights == True:
                     move_weight = True
                     for param in params:
-                        key = "{}.{}".format(n, param)
+                        key = key_param_name_to_key(n, param)
                         bk = self.backup.get(key, None)
                         if bk is not None:
                             if not lowvram_possible:
@@ -946,7 +973,7 @@ class ModelPatcher:
                         logging.debug("freed {}".format(n))
 
                         for param in params:
-                            self.pin_weight_to_device("{}.{}".format(n, param))
+                            self.pin_weight_to_device(key_param_name_to_key(n, param))
 
 
             self.model.model_lowvram = True
@@ -984,6 +1011,9 @@ class ModelPatcher:
 
             return self.model.model_loaded_weight_memory - current_used
 
+    def partially_unload_ram(self, ram_to_unload):
+        pass
+
     def detach(self, unpatch_all=True):
         self.eject_model()
         self.model_patches_to(self.offload_device)
@@ -1317,10 +1347,10 @@ class ModelPatcher:
                                                  key, original_weights=original_weights)
         del original_weights[key]
         if set_func is None:
-            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=string_to_seed(key))
+            out_weight = comfy.float.stochastic_rounding(out_weight, weight.dtype, seed=comfy.utils.string_to_seed(key))
             comfy.utils.copy_to_param(self.model, key, out_weight)
         else:
-            set_func(out_weight, inplace_update=True, seed=string_to_seed(key))
+            set_func(out_weight, inplace_update=True, seed=comfy.utils.string_to_seed(key))
         if self.hook_mode == comfy.hooks.EnumHookMode.MaxSpeed:
             # TODO: disable caching if not enough system RAM to do so
             target_device = self.offload_device
@@ -1355,7 +1385,256 @@ class ModelPatcher:
         self.unpatch_hooks()
         self.clear_cached_hook_weights()
 
+    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+        unet_state_dict = self.model.diffusion_model.state_dict()
+        for k, v in unet_state_dict.items():
+            op_keys = k.rsplit('.', 1)
+            if (len(op_keys) < 2) or op_keys[1] not in ["weight", "bias"]:
+                continue
+            try:
+                op = comfy.utils.get_attr(self.model.diffusion_model, op_keys[0])
+            except:
+                continue
+            if not op or not hasattr(op, "comfy_cast_weights") or \
+                (hasattr(op, "comfy_patched_weights") and op.comfy_patched_weights == True):
+                continue
+            key = "diffusion_model." + k
+            unet_state_dict[k] = LazyCastingParam(self, key, comfy.utils.get_attr(self.model, key))
+        return self.model.state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
+
     def __del__(self):
         self.unpin_all_weights()
         self.detach(unpatch_all=False)
 
+class ModelPatcherDynamic(ModelPatcher):
+
+    def __new__(cls, model=None, load_device=None, offload_device=None, size=0, weight_inplace_update=False):
+        if load_device is not None and comfy.model_management.is_device_cpu(load_device):
+            #reroute to default MP for CPUs
+            return ModelPatcher(model, load_device, offload_device, size, weight_inplace_update)
+        return super().__new__(cls)
+
+    def __init__(self, model, load_device, offload_device, size=0, weight_inplace_update=False):
+        super().__init__(model, load_device, offload_device, size, weight_inplace_update)
+        #this is now way more dynamic and we dont support the same base model for both Dynamic
+        #and non-dynamic patchers.
+        if hasattr(self.model, "model_loaded_weight_memory"):
+            del self.model.model_loaded_weight_memory
+        if not hasattr(self.model, "dynamic_vbars"):
+            self.model.dynamic_vbars = {}
+        assert load_device is not None
+
+    def is_dynamic(self):
+        return True
+
+    def _vbar_get(self, create=False):
+        if self.load_device == torch.device("cpu"):
+            return None
+        vbar = self.model.dynamic_vbars.get(self.load_device, None)
+        if create and vbar is None:
+            # x10. We dont know what model defined type casts we have in the vbar, but virtual address
+            # space is pretty free. This will cover someone casting an entire model from FP4 to FP32
+            # with some left over.
+            vbar = comfy_aimdo.model_vbar.ModelVBAR(self.model_size() * 10, self.load_device.index)
+            self.model.dynamic_vbars[self.load_device] = vbar
+        return vbar
+
+    def loaded_size(self):
+        vbar = self._vbar_get()
+        if vbar is None:
+            return 0
+        return vbar.loaded_size()
+
+    def get_free_memory(self, device):
+        #NOTE: on high condition / batch counts, estimate should have already vacated
+        #all non-dynamic models so this is safe even if its not 100% true that this
+        #would all be avaiable for inference use.
+        return comfy.model_management.get_total_memory(device) - self.model_size()
+
+    #Pinning is deferred to ops time. Assert against this API to avoid pin leaks.
+
+    def pin_weight_to_device(self, key):
+        raise RuntimeError("pin_weight_to_device invalid for dymamic weight loading")
+
+    def unpin_weight(self, key):
+        raise RuntimeError("unpin_weight invalid for dymamic weight loading")
+
+    def unpin_all_weights(self):
+        self.partially_unload_ram(1e32)
+
+    def memory_required(self, input_shape):
+        #Pad this significantly. We are trying to get away from precise estimates. This
+        #estimate is only used when using the ModelPatcherDynamic after ModelPatcher. If you
+        #use all ModelPatcherDynamic this is ignored and its all done dynamically.
+        return super().memory_required(input_shape=input_shape) * 1.3 + (1024 ** 3)
+
+
+    def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False, full_load=False, dirty=False):
+
+        #Force patching doesn't make sense in Dynamic loading, as you dont know what does and
+        #doesn't need to be forced at this stage. The only thing you could do would be patch
+        #it all on CPU which consumes huge RAM.
+        assert not force_patch_weights
+
+        #Full load doesn't make sense as we dont actually have any loader capability here and
+        #now.
+        assert not full_load
+
+        assert device_to == self.load_device
+
+        num_patches = 0
+        allocated_size = 0
+
+        with self.use_ejected():
+            self.unpatch_hooks()
+
+            vbar = self._vbar_get(create=True)
+            if vbar is not None:
+                vbar.prioritize()
+
+            #We force reserve VRAM for the non comfy-weight so we dont have to deal
+            #with pin and unpin syncrhonization which can be expensive for small weights
+            #with a high layer rate (e.g. autoregressive LLMs).
+            #prioritize the non-comfy weights (note the order reverse).
+            loading = self._load_list(prio_comfy_cast_weights=True, default_device=device_to)
+            loading.sort(reverse=True)
+
+            for x in loading:
+                _, _, _, n, m, params = x
+
+                def set_dirty(item, dirty):
+                    if dirty or not hasattr(item, "_v_signature"):
+                        item._v_signature = None
+
+                def setup_param(self, m, n, param_key):
+                    nonlocal num_patches
+                    key = key_param_name_to_key(n, param_key)
+
+                    weight_function = []
+
+                    weight, _, _ = get_key_weight(self.model, key)
+                    if weight is None:
+                        return 0
+                    if key in self.patches:
+                        setattr(m, param_key + "_lowvram_function", LowVramPatch(key, self.patches))
+                        num_patches += 1
+                    else:
+                        setattr(m, param_key + "_lowvram_function", None)
+
+                    if key in self.weight_wrapper_patches:
+                        weight_function.extend(self.weight_wrapper_patches[key])
+                    setattr(m, param_key + "_function", weight_function)
+                    geometry = weight
+                    if not isinstance(weight, QuantizedTensor):
+                        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)
+
+                if hasattr(m, "comfy_cast_weights"):
+                    m.comfy_cast_weights = True
+                    m.pin_failed = False
+                    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")
+
+                    if vbar is not None and not hasattr(m, "_v"):
+                        m._v = vbar.alloc(v_weight_size)
+                    allocated_size += v_weight_size
+
+                else:
+                    for param in params:
+                        key = key_param_name_to_key(n, param)
+                        weight, _, _ = get_key_weight(self.model, key)
+                        weight.seed_key = key
+                        set_dirty(weight, dirty)
+                        geometry = weight
+                        model_dtype = getattr(m, param + "_comfy_model_dtype", None) or weight.dtype
+                        geometry = comfy.memory_management.TensorGeometry(shape=weight.shape, dtype=model_dtype)
+                        weight_size = geometry.numel() * geometry.element_size()
+                        if vbar is not None and not hasattr(weight, "_v"):
+                            weight._v = vbar.alloc(weight_size)
+                            weight._model_dtype = model_dtype
+                        allocated_size += weight_size
+                    vbar.set_watermark_limit(allocated_size)
+
+                move_weight_functions(m, device_to)
+
+            logging.info(f"Model {self.model.__class__.__name__} prepared for dynamic VRAM loading. {allocated_size // (1024 ** 2)}MB Staged. {num_patches} patches attached.")
+
+            self.model.device = device_to
+            self.model.current_weight_patches_uuid = self.patches_uuid
+
+            for callback in self.get_all_callbacks(CallbacksMP.ON_LOAD):
+                #These are all super dangerous. Who knows what the custom nodes actually do here...
+                callback(self, device_to, lowvram_model_memory, force_patch_weights, full_load)
+
+            self.apply_hooks(self.forced_hooks, force_apply=True)
+
+    def partially_unload(self, device_to, memory_to_free=0, force_patch_weights=False):
+        assert not force_patch_weights #See above
+        assert self.load_device != torch.device("cpu")
+
+        vbar = self._vbar_get()
+        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, default_device=self.offload_device)
+        for x in loading:
+            _, _, _, _, m, _ = x
+            ram_to_unload -= comfy.pinned_memory.unpin_memory(m)
+            if ram_to_unload <= 0:
+                return
+
+    def patch_model(self, device_to=None, lowvram_model_memory=0, load_weights=True, force_patch_weights=False):
+        #This isn't used by the core at all and can only be to load a model out of
+        #the control of proper model_managment. If you are a custom node author reading
+        #this, the correct pattern is to call load_models_gpu() to get a proper
+        #managed load of your model.
+        assert not load_weights
+        return super().patch_model(load_weights=load_weights, force_patch_weights=force_patch_weights)
+
+    def unpatch_model(self, device_to=None, unpatch_weights=True):
+        super().unpatch_model(device_to=None, unpatch_weights=False)
+
+        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)
+
+    def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
+        assert not force_patch_weights #See above
+        with self.use_ejected(skip_and_inject_on_exit_only=True):
+            dirty = self.model.current_weight_patches_uuid is not None and (self.model.current_weight_patches_uuid != self.patches_uuid)
+
+            self.unpatch_model(self.offload_device, unpatch_weights=False)
+            self.patch_model(load_weights=False)
+
+            try:
+                self.load(device_to, dirty=dirty)
+            except Exception as e:
+                self.detach()
+                raise e
+            #ModelPatcher::partially_load returns a number on what got loaded but
+            #nothing in core uses this and we have no data in the Dynamic world. Hit
+            #the custom node devs with a None rather than a 0 that would mislead any
+            #logic they might have.
+            return None
+
+    def patch_cached_hook_weights(self, cached_weights: dict, key: str, memory_counter: MemoryCounter):
+        assert False #Should be unreachable - we dont ever cache in the new implementation
+
+    def patch_hook_weight_to_device(self, hooks: comfy.hooks.HookGroup, combined_patches: dict, key: str, original_weights: dict, memory_counter: MemoryCounter):
+        if key not in combined_patches:
+            return
+
+        raise RuntimeError("Hooks not implemented in ModelPatcherDynamic. Please remove --fast arguments form ComfyUI startup")
+
+    def unpatch_hooks(self, whitelist_keys_set: set[str]=None) -> None:
+        pass
+
+CoreModelPatcher = ModelPatcher

comfy/ops.py

@@ -19,10 +19,16 @@
 import torch
 import logging
 import comfy.model_management
-from comfy.cli_args import args, PerformanceFeature
+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
+import comfy.utils
+
+import comfy_aimdo.model_vbar
+import comfy_aimdo.torch
 
 def run_every_op():
     if torch.compiler.is_compiling():
@@ -48,6 +54,8 @@ try:
             SDPA_BACKEND_PRIORITY.insert(0, SDPBackend.CUDNN_ATTENTION)
 
             def scaled_dot_product_attention(q, k, v, *args, **kwargs):
+                if q.nelement() < 1024 * 128:  # arbitrary number, for small inputs cudnn attention seems slower
+                    return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
                 with sdpa_kernel(SDPA_BACKEND_PRIORITY, set_priority=True):
                     return torch.nn.functional.scaled_dot_product_attention(q, k, v, *args, **kwargs)
         else:
@@ -72,7 +80,122 @@ 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(s, input=None, dtype=None, device=None, bias_dtype=None, offloadable=False):
+def cast_bias_weight_with_vbar(s, dtype, device, bias_dtype, non_blocking, compute_dtype):
+    offload_stream = None
+    xfer_dest = None
+
+    signature = comfy_aimdo.model_vbar.vbar_fault(s._v)
+    resident = comfy_aimdo.model_vbar.vbar_signature_compare(signature, s._v_signature)
+    if signature is not None:
+        if resident:
+            weight = s._v_weight
+            bias = s._v_bias
+        else:
+            xfer_dest = comfy_aimdo.torch.aimdo_to_tensor(s._v, device)
+
+    if not resident:
+        cast_geometry = comfy.memory_management.tensors_to_geometries([ s.weight, s.bias ])
+        cast_dest = None
+
+        xfer_source = [ s.weight, s.bias ]
+
+        pin = comfy.pinned_memory.get_pin(s)
+        if pin is not None:
+            xfer_source = [ pin ]
+
+        for data, geometry in zip([ s.weight, s.bias ], cast_geometry):
+            if data is None:
+                continue
+            if data.dtype != geometry.dtype:
+                cast_dest = xfer_dest
+                if cast_dest is None:
+                    cast_dest = torch.empty((comfy.memory_management.vram_aligned_size(cast_geometry),), dtype=torch.uint8, device=device)
+                xfer_dest = None
+                break
+
+        dest_size = comfy.memory_management.vram_aligned_size(xfer_source)
+        offload_stream = comfy.model_management.get_offload_stream(device)
+        if xfer_dest is None and offload_stream is not None:
+                xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
+                if xfer_dest is None:
+                    offload_stream = comfy.model_management.get_offload_stream(device)
+                    xfer_dest = comfy.model_management.get_cast_buffer(offload_stream, device, dest_size, s)
+        if xfer_dest is None:
+            xfer_dest = torch.empty((dest_size,), dtype=torch.uint8, device=device)
+            offload_stream = None
+
+        if signature is None and pin is None:
+            comfy.pinned_memory.pin_memory(s)
+            pin = comfy.pinned_memory.get_pin(s)
+        else:
+            pin = None
+
+        if pin is not None:
+            comfy.model_management.cast_to_gathered(xfer_source, pin)
+            xfer_source = [ pin ]
+        #send it over
+        comfy.model_management.cast_to_gathered(xfer_source, xfer_dest, non_blocking=non_blocking, stream=offload_stream)
+        comfy.model_management.sync_stream(device, offload_stream)
+
+        if cast_dest is not None:
+            for pre_cast, post_cast in zip(comfy.memory_management.interpret_gathered_like([s.weight, s.bias ], xfer_dest),
+                                           comfy.memory_management.interpret_gathered_like(cast_geometry, cast_dest)):
+                if post_cast is not None:
+                    post_cast.copy_(pre_cast)
+            xfer_dest = cast_dest
+
+        params = comfy.memory_management.interpret_gathered_like(cast_geometry, xfer_dest)
+        weight = params[0]
+        bias = params[1]
+        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)
+        fns = getattr(s, param_key + "_function", [])
+
+        orig = x
+
+        def to_dequant(tensor, dtype):
+            tensor = tensor.to(dtype=dtype)
+            if isinstance(tensor, QuantizedTensor):
+                tensor = tensor.dequantize()
+            return tensor
+
+        if orig.dtype != dtype or len(fns) > 0:
+            x = to_dequant(x, dtype)
+        if not resident and lowvram_fn is not None:
+            x = to_dequant(x, dtype if compute_dtype is None else compute_dtype)
+            #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)):
+                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:
+                    #The layer actually wants our freshly saved QT
+                    x = y
+            elif update_weight:
+                y = comfy.float.stochastic_rounding(x, orig.dtype, seed = comfy.utils.string_to_seed(s.seed_key))
+            if update_weight:
+                orig.copy_(y)
+        for f in fns:
+            x = f(x)
+        return x
+
+    update_weight = signature is not None
+
+    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)
+
+    #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):
     # 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.
@@ -87,22 +210,38 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
         if device is None:
             device = input.device
 
+    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)
+
     if offloadable and (device != s.weight.device or
                         (s.bias is not None and device != s.bias.device)):
         offload_stream = comfy.model_management.get_offload_stream(device)
     else:
         offload_stream = None
 
-    non_blocking = comfy.model_management.device_supports_non_blocking(device)
+    bias = None
+    weight = None
+
+    if offload_stream is not None and not args.cuda_malloc:
+        cast_buffer_size = comfy.memory_management.vram_aligned_size([ s.weight, s.bias ])
+        cast_buffer = comfy.model_management.get_cast_buffer(offload_stream, device, cast_buffer_size, s)
+        #The streams can be uneven in buffer capability and reject us. Retry to get the other stream
+        if cast_buffer is None:
+            offload_stream = comfy.model_management.get_offload_stream(device)
+            cast_buffer = comfy.model_management.get_cast_buffer(offload_stream, device, cast_buffer_size, s)
+        params = comfy.memory_management.interpret_gathered_like([ s.weight, s.bias ], cast_buffer)
+        weight = params[0]
+        bias = params[1]
 
     weight_has_function = len(s.weight_function) > 0
     bias_has_function = len(s.bias_function) > 0
 
-    weight = comfy.model_management.cast_to(s.weight, None, device, non_blocking=non_blocking, copy=weight_has_function, stream=offload_stream)
+    weight = comfy.model_management.cast_to(s.weight, None, device, non_blocking=non_blocking, copy=weight_has_function, stream=offload_stream, r=weight)
 
-    bias = None
     if s.bias is not None:
-        bias = comfy.model_management.cast_to(s.bias, bias_dtype, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream)
+        bias = comfy.model_management.cast_to(s.bias, None, device, non_blocking=non_blocking, copy=bias_has_function, stream=offload_stream, r=bias)
 
     comfy.model_management.sync_stream(device, offload_stream)
 
@@ -110,6 +249,7 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of
     weight_a = weight
 
     if s.bias is not None:
+        bias = bias.to(dtype=bias_dtype)
         for f in s.bias_function:
             bias = f(bias)
 
@@ -131,14 +271,20 @@ def uncast_bias_weight(s, weight, bias, offload_stream):
     if offload_stream is None:
         return
     os, weight_a, bias_a = offload_stream
+    device=None
+    #FIXME: This is not good RTTI
+    if not isinstance(weight_a, torch.Tensor):
+        comfy_aimdo.model_vbar.vbar_unpin(s._v)
+        device = weight_a
     if os is None:
         return
-    if weight_a is not None:
-        device = weight_a.device
-    else:
-        if bias_a is None:
-            return
-        device = bias_a.device
+    if device is None:
+        if weight_a is not None:
+            device = weight_a.device
+        else:
+            if bias_a is None:
+                return
+            device = bias_a.device
     os.wait_stream(comfy.model_management.current_stream(device))
 
 
@@ -149,6 +295,57 @@ class CastWeightBiasOp:
 
 class disable_weight_init:
     class Linear(torch.nn.Linear, CastWeightBiasOp):
+
+        def __init__(self, in_features, out_features, bias=True, device=None, dtype=None):
+            if not comfy.model_management.WINDOWS or not enables_dynamic_vram():
+                super().__init__(in_features, out_features, bias, device, dtype)
+                return
+
+            # Issue is with `torch.empty` still reserving the full memory for the layer.
+            # Windows doesn't over-commit memory so without this, We are momentarily commit
+            # charged for the weight even though we might zero-copy it when we load the
+            # state dict. If the commit charge exceeds the ceiling we can destabilize the
+            # system.
+            torch.nn.Module.__init__(self)
+            self.in_features = in_features
+            self.out_features = out_features
+            self.weight = None
+            self.bias = None
+            self.comfy_need_lazy_init_bias=bias
+            self.weight_comfy_model_dtype = dtype
+            self.bias_comfy_model_dtype = dtype
+
+        def _load_from_state_dict(self, state_dict, prefix, local_metadata,
+                                strict, missing_keys, unexpected_keys, error_msgs):
+
+            if not comfy.model_management.WINDOWS or not enables_dynamic_vram():
+                return super()._load_from_state_dict(state_dict, prefix, local_metadata, strict,
+                                                     missing_keys, unexpected_keys, error_msgs)
+            assign_to_params_buffers = local_metadata.get("assign_to_params_buffers", False)
+            prefix_len = len(prefix)
+            for k,v in state_dict.items():
+                if k[prefix_len:] == "weight":
+                    if not assign_to_params_buffers:
+                        v = v.clone()
+                    self.weight = torch.nn.Parameter(v, requires_grad=False)
+                elif k[prefix_len:] == "bias" and v is not None:
+                    if not assign_to_params_buffers:
+                        v = v.clone()
+                    self.bias = torch.nn.Parameter(v, requires_grad=False)
+                else:
+                    unexpected_keys.append(k)
+
+            #Reconcile default construction of the weight if its missing.
+            if self.weight is None:
+                v = torch.zeros(self.in_features, self.out_features)
+                self.weight = torch.nn.Parameter(v, requires_grad=False)
+                missing_keys.append(prefix+"weight")
+            if self.bias is None and self.comfy_need_lazy_init_bias:
+                v = torch.zeros(self.out_features,)
+                self.bias = torch.nn.Parameter(v, requires_grad=False)
+                missing_keys.append(prefix+"bias")
+
+
         def reset_parameters(self):
             return None
 
@@ -655,8 +852,8 @@ 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):
-                weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
+            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)
                 x = self._forward(input, weight, bias)
                 uncast_bias_weight(self, weight, bias, offload_stream)
                 return x
@@ -666,6 +863,8 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 
                 input_shape = input.shape
                 reshaped_3d = False
+                #If cast needs to apply lora, it should be done in the compute dtype
+                compute_dtype = input.dtype
 
                 if (getattr(self, 'layout_type', None) is not None and
                     not isinstance(input, QuantizedTensor) and not self._full_precision_mm and
@@ -684,7 +883,8 @@ 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)
+
+                output = self.forward_comfy_cast_weights(input, compute_dtype)
 
                 # Reshape output back to 3D if input was 3D
                 if reshaped_3d:

comfy/pinned_memory.py

@@ -0,0 +1,29 @@
+import torch
+import comfy.model_management
+import comfy.memory_management
+
+from comfy.cli_args import args
+
+def get_pin(module):
+    return getattr(module, "_pin", None)
+
+def pin_memory(module):
+    if module.pin_failed or args.disable_pinned_memory or get_pin(module) is not None:
+        return
+    #FIXME: This is a RAM cache trigger event
+    size = comfy.memory_management.vram_aligned_size([ module.weight, module.bias ])
+    pin = torch.empty((size,), dtype=torch.uint8)
+    if comfy.model_management.pin_memory(pin):
+        module._pin = pin
+    else:
+        module.pin_failed = True
+        return False
+    return True
+
+def unpin_memory(module):
+    if get_pin(module) is None:
+        return 0
+    size = module._pin.numel() * module._pin.element_size()
+    comfy.model_management.unpin_memory(module._pin)
+    del module._pin
+    return size

comfy/sampler_helpers.py

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

comfy/samplers.py

@@ -9,7 +9,6 @@ if TYPE_CHECKING:
 import torch
 from functools import partial
 import collections
-from comfy import model_management
 import math
 import logging
 import comfy.sampler_helpers
@@ -260,7 +259,7 @@ def _calc_cond_batch(model: BaseModel, conds: list[list[dict]], x_in: torch.Tens
             to_batch_temp.reverse()
             to_batch = to_batch_temp[:1]
 
-            free_memory = model_management.get_free_memory(x_in.device)
+            free_memory = model.current_patcher.get_free_memory(x_in.device)
             for i in range(1, len(to_batch_temp) + 1):
                 batch_amount = to_batch_temp[:len(to_batch_temp)//i]
                 input_shape = [len(batch_amount) * first_shape[0]] + list(first_shape)[1:]

comfy/sd.py

@@ -59,6 +59,7 @@ import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.jina_clip_2
 import comfy.text_encoders.newbie
 import comfy.text_encoders.anima
+import comfy.text_encoders.ace15
 
 import comfy.model_patcher
 import comfy.lora
@@ -228,8 +229,10 @@ class CLIP:
                     self.cond_stage_model.to(offload_device)
                     logging.warning("Had to shift TE back.")
 
+        model_management.archive_model_dtypes(self.cond_stage_model)
+
         self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
-        self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
+        self.patcher = comfy.model_patcher.CoreModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
         #Match torch.float32 hardcode upcast in TE implemention
         self.patcher.set_model_compute_dtype(torch.float32)
         self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram
@@ -389,8 +392,18 @@ class CLIP:
 
     def load_sd(self, sd, full_model=False):
         if full_model:
-            return self.cond_stage_model.load_state_dict(sd, strict=False)
+            return self.cond_stage_model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
         else:
+            can_assign = self.patcher.is_dynamic()
+            self.cond_stage_model.can_assign_sd = can_assign
+
+            # The CLIP models are a pretty complex web of wrappers and its
+            # a bit of an API change to plumb this all the way through.
+            # So spray paint the model with this flag that the loading
+            # nn.Module can then inspect for itself.
+            for m in self.cond_stage_model.modules():
+                m.can_assign_sd = can_assign
+
             return self.cond_stage_model.load_sd(sd)
 
     def get_sd(self):
@@ -440,6 +453,8 @@ class VAE:
         self.extra_1d_channel = None
         self.crop_input = True
 
+        self.audio_sample_rate = 44100
+
         if config is None:
             if "decoder.mid.block_1.mix_factor" in sd:
                 encoder_config = {'double_z': True, 'z_channels': 4, 'resolution': 256, 'in_channels': 3, 'out_ch': 3, 'ch': 128, 'ch_mult': [1, 2, 4, 4], 'num_res_blocks': 2, 'attn_resolutions': [], 'dropout': 0.0}
@@ -537,14 +552,27 @@ class VAE:
                                                                     encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': ddconfig},
                                                                     decoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Decoder", 'params': ddconfig})
             elif "decoder.layers.1.layers.0.beta" in sd:
-                self.first_stage_model = AudioOobleckVAE()
+                config = {}
+                param_key = None
+                self.upscale_ratio = 2048
+                self.downscale_ratio = 2048
+                if "decoder.layers.2.layers.1.weight_v" in sd:
+                    param_key = "decoder.layers.2.layers.1.weight_v"
+                if "decoder.layers.2.layers.1.parametrizations.weight.original1" in sd:
+                    param_key = "decoder.layers.2.layers.1.parametrizations.weight.original1"
+                if param_key is not None:
+                    if sd[param_key].shape[-1] == 12:
+                        config["strides"] = [2, 4, 4, 6, 10]
+                        self.audio_sample_rate = 48000
+                        self.upscale_ratio = 1920
+                        self.downscale_ratio = 1920
+
+                self.first_stage_model = AudioOobleckVAE(**config)
                 self.memory_used_encode = lambda shape, dtype: (1000 * shape[2]) * model_management.dtype_size(dtype)
                 self.memory_used_decode = lambda shape, dtype: (1000 * shape[2] * 2048) * model_management.dtype_size(dtype)
                 self.latent_channels = 64
                 self.output_channels = 2
                 self.pad_channel_value = "replicate"
-                self.upscale_ratio = 2048
-                self.downscale_ratio =  2048
                 self.latent_dim = 1
                 self.process_output = lambda audio: audio
                 self.process_input = lambda audio: audio
@@ -765,13 +793,6 @@ class VAE:
             self.first_stage_model = AutoencoderKL(**(config['params']))
         self.first_stage_model = self.first_stage_model.eval()
 
-        m, u = self.first_stage_model.load_state_dict(sd, strict=False)
-        if len(m) > 0:
-            logging.warning("Missing VAE keys {}".format(m))
-
-        if len(u) > 0:
-            logging.debug("Leftover VAE keys {}".format(u))
-
         if device is None:
             device = model_management.vae_device()
         self.device = device
@@ -780,9 +801,21 @@ class VAE:
             dtype = model_management.vae_dtype(self.device, self.working_dtypes)
         self.vae_dtype = dtype
         self.first_stage_model.to(self.vae_dtype)
+        model_management.archive_model_dtypes(self.first_stage_model)
         self.output_device = model_management.intermediate_device()
 
-        self.patcher = comfy.model_patcher.ModelPatcher(self.first_stage_model, load_device=self.device, offload_device=offload_device)
+        mp = comfy.model_patcher.CoreModelPatcher
+        if self.disable_offload:
+            mp = comfy.model_patcher.ModelPatcher
+        self.patcher = mp(self.first_stage_model, load_device=self.device, offload_device=offload_device)
+
+        m, u = self.first_stage_model.load_state_dict(sd, strict=False, assign=self.patcher.is_dynamic())
+        if len(m) > 0:
+            logging.warning("Missing VAE keys {}".format(m))
+
+        if len(u) > 0:
+            logging.debug("Leftover VAE keys {}".format(u))
+
         logging.info("VAE load device: {}, offload device: {}, dtype: {}".format(self.device, offload_device, self.vae_dtype))
         self.model_size()
 
@@ -838,7 +871,7 @@ class VAE:
             / 3.0)
         return output
 
-    def decode_tiled_1d(self, samples, tile_x=128, overlap=32):
+    def decode_tiled_1d(self, samples, tile_x=256, overlap=32):
         if samples.ndim == 3:
             decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
         else:
@@ -897,7 +930,7 @@ class VAE:
         try:
             memory_used = self.memory_used_decode(samples_in.shape, self.vae_dtype)
             model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
-            free_memory = model_management.get_free_memory(self.device)
+            free_memory = self.patcher.get_free_memory(self.device)
             batch_number = int(free_memory / memory_used)
             batch_number = max(1, batch_number)
 
@@ -942,7 +975,7 @@ class VAE:
         if overlap is not None:
             args["overlap"] = overlap
 
-        if dims == 1:
+        if dims == 1 or self.extra_1d_channel is not None:
             args.pop("tile_y")
             output = self.decode_tiled_1d(samples, **args)
         elif dims == 2:
@@ -971,7 +1004,7 @@ class VAE:
         try:
             memory_used = self.memory_used_encode(pixel_samples.shape, self.vae_dtype)
             model_management.load_models_gpu([self.patcher], memory_required=memory_used, force_full_load=self.disable_offload)
-            free_memory = model_management.get_free_memory(self.device)
+            free_memory = self.patcher.get_free_memory(self.device)
             batch_number = int(free_memory / max(1, memory_used))
             batch_number = max(1, batch_number)
             samples = None
@@ -1409,6 +1442,14 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
                 clip_data_jina = clip_data[0]
             tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None)
             tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None)
+        elif clip_type == CLIPType.ACE:
+            te_models = [detect_te_model(clip_data[0]), detect_te_model(clip_data[1])]
+            if TEModel.QWEN3_4B in te_models:
+                model_type = "qwen3_4b"
+            else:
+                model_type = "qwen3_2b"
+            clip_target.clip = comfy.text_encoders.ace15.te(lm_model=model_type, **llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.ace15.ACE15Tokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer
@@ -1432,7 +1473,7 @@ def load_gligen(ckpt_path):
     model = gligen.load_gligen(data)
     if model_management.should_use_fp16():
         model = model.half()
-    return comfy.model_patcher.ModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
+    return comfy.model_patcher.CoreModelPatcher(model, load_device=model_management.get_torch_device(), offload_device=model_management.unet_offload_device())
 
 def model_detection_error_hint(path, state_dict):
     filename = os.path.basename(path)
@@ -1520,7 +1561,8 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
     if output_model:
         inital_load_device = model_management.unet_inital_load_device(parameters, unet_dtype)
         model = model_config.get_model(sd, diffusion_model_prefix, device=inital_load_device)
-        model.load_model_weights(sd, diffusion_model_prefix)
+        model_patcher = comfy.model_patcher.CoreModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device())
+        model.load_model_weights(sd, diffusion_model_prefix, assign=model_patcher.is_dynamic())
 
     if output_vae:
         vae_sd = comfy.utils.state_dict_prefix_replace(sd, {k: "" for k in model_config.vae_key_prefix}, filter_keys=True)
@@ -1563,7 +1605,6 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
         logging.debug("left over keys: {}".format(left_over))
 
     if output_model:
-        model_patcher = comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=model_management.unet_offload_device())
         if inital_load_device != torch.device("cpu"):
             logging.info("loaded diffusion model directly to GPU")
             model_management.load_models_gpu([model_patcher], force_full_load=True)
@@ -1655,13 +1696,14 @@ def load_diffusion_model_state_dict(sd, model_options={}, metadata=None):
         model_config.optimizations["fp8"] = True
 
     model = model_config.get_model(new_sd, "")
-    model = model.to(offload_device)
-    model.load_model_weights(new_sd, "")
+    model_patcher = comfy.model_patcher.CoreModelPatcher(model, load_device=load_device, offload_device=offload_device)
+    if not model_management.is_device_cpu(offload_device):
+        model.to(offload_device)
+    model.load_model_weights(new_sd, "", assign=model_patcher.is_dynamic())
     left_over = sd.keys()
     if len(left_over) > 0:
         logging.info("left over keys in diffusion model: {}".format(left_over))
-    return comfy.model_patcher.ModelPatcher(model, load_device=load_device, offload_device=offload_device)
-
+    return model_patcher
 
 def load_diffusion_model(unet_path, model_options={}):
     sd, metadata = comfy.utils.load_torch_file(unet_path, return_metadata=True)
@@ -1692,9 +1734,9 @@ def save_checkpoint(output_path, model, clip=None, vae=None, clip_vision=None, m
     if metadata is None:
         metadata = {}
 
-    model_management.load_models_gpu(load_models, force_patch_weights=True)
+    model_management.load_models_gpu(load_models)
     clip_vision_sd = clip_vision.get_sd() if clip_vision is not None else None
-    sd = model.model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd)
+    sd = model.state_dict_for_saving(clip_sd, vae_sd, clip_vision_sd)
     for k in extra_keys:
         sd[k] = extra_keys[k]
 

comfy/sd1_clip.py

@@ -155,6 +155,8 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         self.execution_device = options.get("execution_device", self.execution_device)
         if isinstance(self.layer, list) or self.layer == "all":
             pass
+        elif isinstance(layer_idx, list):
+            self.layer = layer_idx
         elif layer_idx is None or abs(layer_idx) > self.num_layers:
             self.layer = "last"
         else:
@@ -169,8 +171,9 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
 
     def process_tokens(self, tokens, device):
         end_token = self.special_tokens.get("end", None)
+        pad_token = self.special_tokens.get("pad", -1)
         if end_token is None:
-            cmp_token = self.special_tokens.get("pad", -1)
+            cmp_token = pad_token
         else:
             cmp_token = end_token
 
@@ -184,15 +187,21 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
             other_embeds = []
             eos = False
             index = 0
+            left_pad = False
             for y in x:
                 if isinstance(y, numbers.Integral):
-                    if eos:
+                    token = int(y)
+                    if index == 0 and token == pad_token:
+                        left_pad = True
+
+                    if eos or (left_pad and token == pad_token):
                         attention_mask.append(0)
                     else:
                         attention_mask.append(1)
-                    token = int(y)
+                        left_pad = False
+
                     tokens_temp += [token]
-                    if not eos and token == cmp_token:
+                    if not eos and token == cmp_token and not left_pad:
                         if end_token is None:
                             attention_mask[-1] = 0
                         eos = True
@@ -297,7 +306,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         return self(tokens)
 
     def load_sd(self, sd):
-        return self.transformer.load_state_dict(sd, strict=False)
+        return self.transformer.load_state_dict(sd, strict=False, assign=getattr(self, "can_assign_sd", False))
 
 def parse_parentheses(string):
     result = []

comfy/supported_models.py

@@ -24,6 +24,7 @@ import comfy.text_encoders.hunyuan_image
 import comfy.text_encoders.kandinsky5
 import comfy.text_encoders.z_image
 import comfy.text_encoders.anima
+import comfy.text_encoders.ace15
 
 from . import supported_models_base
 from . import latent_formats
@@ -709,6 +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."]
 
@@ -897,11 +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
 
@@ -992,7 +1004,7 @@ class CosmosT2IPredict2(supported_models_base.BASE):
 
     memory_usage_factor = 1.0
 
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     def __init__(self, unet_config):
         super().__init__(unet_config)
@@ -1022,11 +1034,7 @@ class Anima(supported_models_base.BASE):
 
     memory_usage_factor = 1.0
 
-    supported_inference_dtypes = [torch.bfloat16, torch.float32]
-
-    def __init__(self, unet_config):
-        super().__init__(unet_config)
-        self.memory_usage_factor = (unet_config.get("model_channels", 2048) / 2048) * 0.95
+    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
 
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.Anima(self, device=device)
@@ -1037,6 +1045,12 @@ class Anima(supported_models_base.BASE):
         detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_06b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.anima.AnimaTokenizer, comfy.text_encoders.anima.te(**detect))
 
+    def set_inference_dtype(self, dtype, manual_cast_dtype, **kwargs):
+        self.memory_usage_factor = (self.unet_config.get("model_channels", 2048) / 2048) * 0.95
+        if dtype is torch.float16:
+            self.memory_usage_factor *= 1.4
+        return super().set_inference_dtype(dtype, manual_cast_dtype, **kwargs)
+
 class CosmosI2VPredict2(CosmosT2IPredict2):
     unet_config = {
         "image_model": "cosmos_predict2",
@@ -1261,6 +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)
@@ -1338,6 +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)
@@ -1596,6 +1627,46 @@ class Kandinsky5Image(Kandinsky5):
         return supported_models_base.ClipTarget(comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage, comfy.text_encoders.kandinsky5.te(**hunyuan_detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
+class ACEStep15(supported_models_base.BASE):
+    unet_config = {
+        "audio_model": "ace1.5",
+    }
+
+    unet_extra_config = {
+    }
+
+    sampling_settings = {
+        "multiplier": 1.0,
+        "shift": 3.0,
+    }
+
+    latent_format = comfy.latent_formats.ACEAudio15
+
+    memory_usage_factor = 4.7
+
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    vae_key_prefix = ["vae."]
+    text_encoder_key_prefix = ["text_encoders."]
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.ACEStep15(self, device=device)
+        return out
+
+    def clip_target(self, state_dict={}):
+        pref = self.text_encoder_key_prefix[0]
+        detect_2b = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref))
+        detect_4b = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref))
+        if "dtype_llama" in detect_2b:
+            detect = detect_2b
+            detect["lm_model"] = "qwen3_2b"
+        elif "dtype_llama" in detect_4b:
+            detect = detect_4b
+            detect["lm_model"] = "qwen3_4b"
+
+        return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**detect))
+
+
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
 
 models += [SVD_img2vid]

comfy/text_encoders/ace15.py

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

comfy/text_encoders/anima.py

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

comfy/text_encoders/flux.py

@@ -118,7 +118,7 @@ class MistralTokenizerClass:
 class Mistral3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         self.tekken_data = tokenizer_data.get("tekken_model", None)
-        super().__init__("", pad_with_end=False, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
+        super().__init__("", pad_with_end=False, embedding_directory=embedding_directory, embedding_size=5120, embedding_key='mistral3_24b', tokenizer_class=MistralTokenizerClass, has_end_token=False, pad_to_max_length=False, pad_token=11, start_token=1, max_length=99999999, min_length=1, pad_left=True, tokenizer_args=load_mistral_tokenizer(self.tekken_data), tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"tekken_model": self.tekken_data}
@@ -176,12 +176,12 @@ def flux2_te(dtype_llama=None, llama_quantization_metadata=None, pruned=False):
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class Qwen3Tokenizer8B(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=4096, embedding_key='qwen3_8b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=512, pad_token=151643, tokenizer_data=tokenizer_data)
 
 class KleinTokenizer(sd1_clip.SD1Tokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}, name="qwen3_4b"):

comfy/text_encoders/llama.py

@@ -6,6 +6,7 @@ import math
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
+import comfy.ops
 import comfy.ldm.common_dit
 import comfy.clip_model
 
@@ -103,6 +104,75 @@ class Qwen3_06BConfig:
     final_norm: bool = True
     lm_head: bool = False
 
+@dataclass
+class Qwen3_06B_ACE15_Config:
+    vocab_size: int = 151669
+    hidden_size: int = 1024
+    intermediate_size: int = 3072
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 32768
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
+@dataclass
+class Qwen3_2B_ACE15_lm_Config:
+    vocab_size: int = 217204
+    hidden_size: int = 2048
+    intermediate_size: int = 6144
+    num_hidden_layers: int = 28
+    num_attention_heads: int = 16
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 40960
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
+@dataclass
+class Qwen3_4B_ACE15_lm_Config:
+    vocab_size: int = 217204
+    hidden_size: int = 2560
+    intermediate_size: int = 9728
+    num_hidden_layers: int = 36
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 40960
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
 @dataclass
 class Qwen3_4BConfig:
     vocab_size: int = 151936
@@ -285,13 +355,6 @@ 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]
@@ -320,20 +383,30 @@ 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)
 
 
@@ -581,10 +654,10 @@ class Llama2_(nn.Module):
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1])
-            mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))
+            mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min / 4)
 
         if seq_len > 1:
-            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
+            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min / 4).triu_(1)
             if mask is not None:
                 mask += causal_mask
             else:
@@ -692,6 +765,21 @@ class BaseLlama:
     def forward(self, input_ids, *args, **kwargs):
         return self.model(input_ids, *args, **kwargs)
 
+class BaseQwen3:
+    def logits(self, x):
+        input = x[:, -1:]
+        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
 
 class Llama2(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
@@ -720,7 +808,7 @@ class Qwen25_3B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_06B(BaseLlama, torch.nn.Module):
+class Qwen3_06B(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_06BConfig(**config_dict)
@@ -729,7 +817,25 @@ class Qwen3_06B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_4B(BaseLlama, torch.nn.Module):
+class Qwen3_06B_ACE15(BaseLlama, BaseQwen3, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_06B_ACE15_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+class Qwen3_2B_ACE15_lm(BaseLlama, BaseQwen3, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_2B_ACE15_lm_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+class Qwen3_4B(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_4BConfig(**config_dict)
@@ -738,7 +844,16 @@ class Qwen3_4B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_8B(BaseLlama, torch.nn.Module):
+class Qwen3_4B_ACE15_lm(BaseLlama, BaseQwen3, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_4B_ACE15_lm_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+class Qwen3_8B(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_8BConfig(**config_dict)

comfy/text_encoders/lt.py

@@ -25,7 +25,7 @@ def ltxv_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, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
+        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}, tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}
@@ -97,6 +97,7 @@ 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)
@@ -125,7 +126,7 @@ class LTXAVTEModel(torch.nn.Module):
             for prefix, component in [("text_embedding_projection.", self.text_embedding_projection), ("video_embeddings_connector.", self.video_embeddings_connector), ("audio_embeddings_connector.", self.audio_embeddings_connector)]:
                 component_sd = {k.replace(prefix, ""): v for k, v in sdo.items() if k.startswith(prefix)}
                 if component_sd:
-                    missing, unexpected = component.load_state_dict(component_sd, strict=False)
+                    missing, unexpected = component.load_state_dict(component_sd, strict=False, assign=getattr(self, "can_assign_sd", False))
                     missing_all.extend([f"{prefix}{k}" for k in missing])
                     unexpected_all.extend([f"{prefix}{k}" for k in unexpected])
 
@@ -138,6 +139,7 @@ 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):

comfy/text_encoders/z_image.py

@@ -6,7 +6,7 @@ import os
 class Qwen3Tokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "qwen25_tokenizer")
-        super().__init__(tokenizer_path, pad_with_end=False, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer_path, pad_with_end=False, embedding_directory=embedding_directory, embedding_size=2560, embedding_key='qwen3_4b', tokenizer_class=Qwen2Tokenizer, has_start_token=False, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, pad_token=151643, tokenizer_data=tokenizer_data)
 
 
 class ZImageTokenizer(sd1_clip.SD1Tokenizer):

comfy/utils.py

@@ -20,41 +20,92 @@
 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
+from comfy.cli_args import args, enables_dynamic_vram
 import json
 import time
+import mmap
+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 = {
+    "F64": torch.float64,
+    "F32": torch.float32,
+    "F16": torch.float16,
+    "BF16": torch.bfloat16,
+    "I64": torch.int64,
+    "I32": torch.int32,
+    "I16": torch.int16,
+    "I8": torch.int8,
+    "U8": torch.uint8,
+    "BOOL": torch.bool,
+    "F8_E4M3": torch.float8_e4m3fn,
+    "F8_E5M2": torch.float8_e5m2,
+    "C64": torch.complex64,
+
+    "U64": torch.uint64,
+    "U32": torch.uint32,
+    "U16": torch.uint16,
+}
+
+def load_safetensors(ckpt):
+    f = open(ckpt, "rb")
+    mapping = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
+    mv = memoryview(mapping)
+
+    header_size = struct.unpack("<Q", mapping[:8])[0]
+    header = json.loads(mapping[8:8+header_size].decode("utf-8"))
+
+    mv = mv[8 + header_size:]
+
+    sd = {}
+    for name, info in header.items():
+        if name == "__metadata__":
+            continue
+
+        start, end = info["data_offsets"]
+        if start == end:
+            sd[name] = torch.empty(info["shape"], dtype =_TYPES[info["dtype"]])
+        else:
+            with warnings.catch_warnings():
+                #We are working with read-only RAM by design
+                warnings.filterwarnings("ignore", message="The given buffer is not writable")
+                sd[name] = torch.frombuffer(mv[start:end], dtype=_TYPES[info["dtype"]]).view(info["shape"])
+
+    return sd, header.get("__metadata__", {}),
+
 
 def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
     if device is None:
@@ -62,15 +113,20 @@ def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
     metadata = None
     if ckpt.lower().endswith(".safetensors") or ckpt.lower().endswith(".sft"):
         try:
-            with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
-                sd = {}
-                for k in f.keys():
-                    tensor = f.get_tensor(k)
-                    if DISABLE_MMAP:  # TODO: Not sure if this is the best way to bypass the mmap issues
-                        tensor = tensor.to(device=device, copy=True)
-                    sd[k] = tensor
-                if return_metadata:
-                    metadata = f.metadata()
+            if enables_dynamic_vram():
+                sd, metadata = load_safetensors(ckpt)
+                if not return_metadata:
+                    metadata = None
+            else:
+                with safetensors.safe_open(ckpt, framework="pt", device=device.type) as f:
+                    sd = {}
+                    for k in f.keys():
+                        tensor = f.get_tensor(k)
+                        if DISABLE_MMAP:  # TODO: Not sure if this is the best way to bypass the mmap issues
+                            tensor = tensor.to(device=device, copy=True)
+                        sd[k] = tensor
+                    if return_metadata:
+                        metadata = f.metadata()
         except Exception as e:
             if len(e.args) > 0:
                 message = e.args[0]
@@ -84,11 +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:
@@ -619,10 +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:
@@ -645,8 +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
                     }
@@ -1100,6 +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
@@ -1308,3 +1387,34 @@ def convert_old_quants(state_dict, model_prefix="", metadata={}):
             state_dict["{}.comfy_quant".format(k)] = torch.tensor(list(json.dumps(v).encode('utf-8')), dtype=torch.uint8)
 
     return state_dict, metadata
+
+def string_to_seed(data):
+    crc = 0xFFFFFFFF
+    for byte in data:
+        if isinstance(byte, str):
+            byte = ord(byte)
+        crc ^= byte
+        for _ in range(8):
+            if crc & 1:
+                crc = (crc >> 1) ^ 0xEDB88320
+            else:
+                crc >>= 1
+    return crc ^ 0xFFFFFFFF
+
+def deepcopy_list_dict(obj, memo=None):
+    if memo is None:
+        memo = {}
+
+    obj_id = id(obj)
+    if obj_id in memo:
+        return memo[obj_id]
+
+    if isinstance(obj, dict):
+        res = {deepcopy_list_dict(k, memo): deepcopy_list_dict(v, memo) for k, v in obj.items()}
+    elif isinstance(obj, list):
+        res = [deepcopy_list_dict(i, memo) for i in obj]
+    else:
+        res = obj
+
+    memo[obj_id] = res
+    return res

comfy/weight_adapter/bypass.py

@@ -21,6 +21,7 @@ from typing import Optional, Union
 import torch
 import torch.nn as nn
 
+import comfy.model_management
 from .base import WeightAdapterBase, WeightAdapterTrainBase
 from comfy.patcher_extension import PatcherInjection
 
@@ -181,18 +182,21 @@ class BypassForwardHook:
             )
             return  # Already injected
 
-        # Move adapter weights to module's device to avoid CPU-GPU transfer on every forward
-        device = None
+        # Move adapter weights to compute device (GPU)
+        # Use get_torch_device() instead of module.weight.device because
+        # with offloading, module weights may be on CPU while compute happens on GPU
+        device = comfy.model_management.get_torch_device()
+
+        # Get dtype from module weight if available
         dtype = None
         if hasattr(self.module, "weight") and self.module.weight is not None:
-            device = self.module.weight.device
             dtype = self.module.weight.dtype
-        elif hasattr(self.module, "W_q"):  # Quantized layers might use different attr
-            device = self.module.W_q.device
-            dtype = self.module.W_q.dtype
 
-        if device is not None:
-            self._move_adapter_weights_to_device(device, dtype)
+        # Only use dtype if it's a standard float type, not quantized
+        if dtype is not None and dtype not in (torch.float32, torch.float16, torch.bfloat16):
+            dtype = None
+
+        self._move_adapter_weights_to_device(device, dtype)
 
         self.original_forward = self.module.forward
         self.module.forward = self._bypass_forward

comfy/windows.py

@@ -0,0 +1,52 @@
+import ctypes
+import logging
+import psutil
+from ctypes import wintypes
+
+import comfy_aimdo.control
+
+psapi = ctypes.WinDLL("psapi")
+kernel32 = ctypes.WinDLL("kernel32")
+
+class PERFORMANCE_INFORMATION(ctypes.Structure):
+    _fields_ = [
+        ("cb", wintypes.DWORD),
+        ("CommitTotal", ctypes.c_size_t),
+        ("CommitLimit", ctypes.c_size_t),
+        ("CommitPeak", ctypes.c_size_t),
+        ("PhysicalTotal", ctypes.c_size_t),
+        ("PhysicalAvailable", ctypes.c_size_t),
+        ("SystemCache", ctypes.c_size_t),
+        ("KernelTotal", ctypes.c_size_t),
+        ("KernelPaged", ctypes.c_size_t),
+        ("KernelNonpaged", ctypes.c_size_t),
+        ("PageSize", ctypes.c_size_t),
+        ("HandleCount", wintypes.DWORD),
+        ("ProcessCount", wintypes.DWORD),
+        ("ThreadCount", wintypes.DWORD),
+    ]
+
+def get_free_ram():
+    #Windows is way too conservative and chalks recently used uncommitted model RAM
+    #as "in-use". So, calculate free RAM for the sake of general use as the greater of:
+    #
+    #1: What psutil says
+    #2: Total Memory - (Committed Memory - VRAM in use)
+    #
+    #We have to subtract VRAM in use from the comitted memory as WDDM creates a naked
+    #commit charge for all VRAM used just incase it wants to page it all out. This just
+    #isn't realistic so "overcommit" on our calculations by just subtracting it off.
+
+    pi = PERFORMANCE_INFORMATION()
+    pi.cb = ctypes.sizeof(pi)
+
+    if not psapi.GetPerformanceInfo(ctypes.byref(pi), pi.cb):
+        logging.warning("WARNING: Failed to query windows performance info. RAM usage may be sub optimal")
+        return psutil.virtual_memory().available
+
+    committed = pi.CommitTotal * pi.PageSize
+    total = pi.PhysicalTotal * pi.PageSize
+
+    return max(psutil.virtual_memory().available,
+               total - (committed - comfy_aimdo.control.get_total_vram_usage()))
+

comfy_api/latest/__init__.py

@@ -7,7 +7,7 @@ from comfy_api.internal.singleton import ProxiedSingleton
 from comfy_api.internal.async_to_sync import create_sync_class
 from ._input import ImageInput, AudioInput, MaskInput, LatentInput, VideoInput
 from ._input_impl import VideoFromFile, VideoFromComponents
-from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL
+from ._util import VideoCodec, VideoContainer, VideoComponents, MESH, VOXEL, File3D
 from . import _io_public as io
 from . import _ui_public as ui
 from comfy_execution.utils import get_executing_context
@@ -105,6 +105,7 @@ class Types:
     VideoComponents = VideoComponents
     MESH = MESH
     VOXEL = VOXEL
+    File3D = File3D
 
 ComfyAPI = ComfyAPI_latest
 

comfy_api/latest/_input/video_types.py

@@ -34,6 +34,21 @@ class VideoInput(ABC):
         """
         pass
 
+    @abstractmethod
+    def as_trimmed(
+        self,
+        start_time: float | None = None,
+        duration: float | None = None,
+        strict_duration: bool = False,
+    ) -> VideoInput | None:
+        """
+        Create a new VideoInput which is trimmed to have the corresponding start_time and duration
+
+        Returns:
+            A new VideoInput, or None if the result would have negative duration
+        """
+        pass
+
     def get_stream_source(self) -> Union[str, io.BytesIO]:
         """
         Get a streamable source for the video. This allows processing without

comfy_api/latest/_input_impl/video_types.py

@@ -6,6 +6,7 @@ from typing import Optional
 from .._input import AudioInput, VideoInput
 import av
 import io
+import itertools
 import json
 import numpy as np
 import math
@@ -29,7 +30,6 @@ def container_to_output_format(container_format: str | None) -> str | None:
     formats = container_format.split(",")
     return formats[0]
 
-
 def get_open_write_kwargs(
     dest: str | io.BytesIO, container_format: str, to_format: str | None
 ) -> dict:
@@ -57,12 +57,14 @@ class VideoFromFile(VideoInput):
     Class representing video input from a file.
     """
 
-    def __init__(self, file: str | io.BytesIO):
+    def __init__(self, file: str | io.BytesIO, *, start_time: float=0, duration: float=0):
         """
         Initialize the VideoFromFile object based off of either a path on disk or a BytesIO object
         containing the file contents.
         """
         self.__file = file
+        self.__start_time = start_time
+        self.__duration = duration
 
     def get_stream_source(self) -> str | io.BytesIO:
         """
@@ -96,6 +98,16 @@ class VideoFromFile(VideoInput):
         Returns:
             Duration in seconds
         """
+        raw_duration = self._get_raw_duration()
+        if self.__start_time < 0:
+            duration_from_start = min(raw_duration, -self.__start_time)
+        else:
+            duration_from_start = raw_duration - self.__start_time
+        if self.__duration:
+            return min(self.__duration, duration_from_start)
+        return duration_from_start
+
+    def _get_raw_duration(self) -> float:
         if isinstance(self.__file, io.BytesIO):
             self.__file.seek(0)
         with av.open(self.__file, mode="r") as container:
@@ -113,9 +125,13 @@ class VideoFromFile(VideoInput):
             if video_stream and video_stream.average_rate:
                 frame_count = 0
                 container.seek(0)
-                for packet in container.demux(video_stream):
-                    for _ in packet.decode():
-                        frame_count += 1
+                frame_iterator = (
+                    container.decode(video_stream)
+                    if video_stream.codec.capabilities & 0x100
+                    else container.demux(video_stream)
+                )
+                for packet in frame_iterator:
+                    frame_count += 1
                 if frame_count > 0:
                     return float(frame_count / video_stream.average_rate)
 
@@ -131,36 +147,54 @@ class VideoFromFile(VideoInput):
 
         with av.open(self.__file, mode="r") as container:
             video_stream = self._get_first_video_stream(container)
-            # 1. Prefer the frames field if available
-            if video_stream.frames and video_stream.frames > 0:
+            # 1. Prefer the frames field if available and usable
+            if (
+                video_stream.frames
+                and video_stream.frames > 0
+                and not self.__start_time
+                and not self.__duration
+            ):
                 return int(video_stream.frames)
 
             # 2. Try to estimate from duration and average_rate using only metadata
-            if container.duration is not None and video_stream.average_rate:
-                duration_seconds = float(container.duration / av.time_base)
-                estimated_frames = int(round(duration_seconds * float(video_stream.average_rate)))
-                if estimated_frames > 0:
-                    return estimated_frames
-
             if (
                 getattr(video_stream, "duration", None) is not None
                 and getattr(video_stream, "time_base", None) is not None
                 and video_stream.average_rate
             ):
-                duration_seconds = float(video_stream.duration * video_stream.time_base)
+                raw_duration = float(video_stream.duration * video_stream.time_base)
+                if self.__start_time < 0:
+                    duration_from_start = min(raw_duration, -self.__start_time)
+                else:
+                    duration_from_start = raw_duration - self.__start_time
+                duration_seconds = min(self.__duration, duration_from_start)
                 estimated_frames = int(round(duration_seconds * float(video_stream.average_rate)))
                 if estimated_frames > 0:
                     return estimated_frames
 
             # 3. Last resort: decode frames and count them (streaming)
-            frame_count = 0
-            container.seek(0)
-            for packet in container.demux(video_stream):
-                for _ in packet.decode():
-                    frame_count += 1
-
-            if frame_count == 0:
-                raise ValueError(f"Could not determine frame count for file '{self.__file}'")
+            if self.__start_time < 0:
+                start_time = max(self._get_raw_duration() + self.__start_time, 0)
+            else:
+                start_time = self.__start_time
+            frame_count = 1
+            start_pts = int(start_time / video_stream.time_base)
+            end_pts = int((start_time + self.__duration) / video_stream.time_base)
+            container.seek(start_pts, stream=video_stream)
+            frame_iterator = (
+                container.decode(video_stream)
+                if video_stream.codec.capabilities & 0x100
+                else container.demux(video_stream)
+            )
+            for frame in frame_iterator:
+                if frame.pts >= start_pts:
+                    break
+            else:
+                raise ValueError(f"Could not determine frame count for file '{self.__file}'\nNo frames exist for start_time {self.__start_time}")
+            for frame in frame_iterator:
+                if frame.pts >= end_pts:
+                    break
+                frame_count += 1
             return frame_count
 
     def get_frame_rate(self) -> Fraction:
@@ -199,9 +233,21 @@ class VideoFromFile(VideoInput):
             return container.format.name
 
     def get_components_internal(self, container: InputContainer) -> VideoComponents:
+        video_stream = self._get_first_video_stream(container)
+        if self.__start_time < 0:
+            start_time = max(self._get_raw_duration() + self.__start_time, 0)
+        else:
+            start_time = self.__start_time
         # Get video frames
         frames = []
-        for frame in container.decode(video=0):
+        start_pts = int(start_time / video_stream.time_base)
+        end_pts = int((start_time + self.__duration) / video_stream.time_base)
+        container.seek(start_pts, stream=video_stream)
+        for frame in container.decode(video_stream):
+            if frame.pts < start_pts:
+                continue
+            if self.__duration and frame.pts >= end_pts:
+                break
             img = frame.to_ndarray(format='rgb24')  # shape: (H, W, 3)
             img = torch.from_numpy(img) / 255.0  # shape: (H, W, 3)
             frames.append(img)
@@ -209,31 +255,44 @@ class VideoFromFile(VideoInput):
         images = torch.stack(frames) if len(frames) > 0 else torch.zeros(0, 3, 0, 0)
 
         # Get frame rate
-        video_stream = next(s for s in container.streams if s.type == 'video')
-        frame_rate = Fraction(video_stream.average_rate) if video_stream and video_stream.average_rate else Fraction(1)
+        frame_rate = Fraction(video_stream.average_rate) if video_stream.average_rate else Fraction(1)
 
         # Get audio if available
         audio = None
-        try:
-            container.seek(0)  # Reset the container to the beginning
-            for stream in container.streams:
-                if stream.type != 'audio':
-                    continue
-                assert isinstance(stream, av.AudioStream)
-                audio_frames = []
-                for packet in container.demux(stream):
-                    for frame in packet.decode():
-                        assert isinstance(frame, av.AudioFrame)
-                        audio_frames.append(frame.to_ndarray())  # shape: (channels, samples)
-                if len(audio_frames) > 0:
-                    audio_data = np.concatenate(audio_frames, axis=1)  # shape: (channels, total_samples)
-                    audio_tensor = torch.from_numpy(audio_data).unsqueeze(0)  # shape: (1, channels, total_samples)
-                    audio = AudioInput({
-                        "waveform": audio_tensor,
-                        "sample_rate": int(stream.sample_rate) if stream.sample_rate else 1,
-                    })
-        except StopIteration:
-            pass  # No audio stream
+        container.seek(start_pts, stream=video_stream)
+        # Use last stream for consistency
+        if len(container.streams.audio):
+            audio_stream = container.streams.audio[-1]
+            audio_frames = []
+            resample = av.audio.resampler.AudioResampler(format='fltp').resample
+            frames = itertools.chain.from_iterable(
+                map(resample, container.decode(audio_stream))
+            )
+
+            has_first_frame = False
+            for frame in frames:
+                offset_seconds = start_time - frame.pts * audio_stream.time_base
+                to_skip = int(offset_seconds * audio_stream.sample_rate)
+                if to_skip < frame.samples:
+                    has_first_frame = True
+                    break
+            if has_first_frame:
+                audio_frames.append(frame.to_ndarray()[..., to_skip:])
+
+            for frame in frames:
+                if frame.time > start_time + self.__duration:
+                    break
+                audio_frames.append(frame.to_ndarray())  # shape: (channels, samples)
+            if len(audio_frames) > 0:
+                audio_data = np.concatenate(audio_frames, axis=1)  # shape: (channels, total_samples)
+                if self.__duration:
+                    audio_data = audio_data[..., :int(self.__duration * audio_stream.sample_rate)]
+
+                audio_tensor = torch.from_numpy(audio_data).unsqueeze(0)  # shape: (1, channels, total_samples)
+                audio = AudioInput({
+                    "waveform": audio_tensor,
+                    "sample_rate": int(audio_stream.sample_rate) if audio_stream.sample_rate else 1,
+                })
 
         metadata = container.metadata
         return VideoComponents(images=images, audio=audio, frame_rate=frame_rate, metadata=metadata)
@@ -250,7 +309,7 @@ class VideoFromFile(VideoInput):
         path: str | io.BytesIO,
         format: VideoContainer = VideoContainer.AUTO,
         codec: VideoCodec = VideoCodec.AUTO,
-        metadata: Optional[dict] = None
+        metadata: Optional[dict] = None,
     ):
         if isinstance(self.__file, io.BytesIO):
             self.__file.seek(0)  # Reset the BytesIO object to the beginning
@@ -262,15 +321,14 @@ class VideoFromFile(VideoInput):
                 reuse_streams = False
             if codec != VideoCodec.AUTO and codec != video_encoding and video_encoding is not None:
                 reuse_streams = False
+            if self.__start_time or self.__duration:
+                reuse_streams = False
 
             if not reuse_streams:
                 components = self.get_components_internal(container)
                 video = VideoFromComponents(components)
                 return video.save_to(
-                    path,
-                    format=format,
-                    codec=codec,
-                    metadata=metadata
+                    path, format=format, codec=codec, metadata=metadata
                 )
 
             streams = container.streams
@@ -304,10 +362,21 @@ class VideoFromFile(VideoInput):
                         output_container.mux(packet)
 
     def _get_first_video_stream(self, container: InputContainer):
-        video_stream = next((s for s in container.streams if s.type == "video"), None)
-        if video_stream is None:
-            raise ValueError(f"No video stream found in file '{self.__file}'")
-        return video_stream
+        if len(container.streams.video):
+            return container.streams.video[0]
+        raise ValueError(f"No video stream found in file '{self.__file}'")
+
+    def as_trimmed(
+        self, start_time: float = 0, duration: float = 0, strict_duration: bool = True
+    ) -> VideoInput | None:
+        trimmed = VideoFromFile(
+            self.get_stream_source(),
+            start_time=start_time + self.__start_time,
+            duration=duration,
+        )
+        if trimmed.get_duration() < duration and strict_duration:
+            return None
+        return trimmed
 
 
 class VideoFromComponents(VideoInput):
@@ -322,7 +391,7 @@ class VideoFromComponents(VideoInput):
         return VideoComponents(
             images=self.__components.images,
             audio=self.__components.audio,
-            frame_rate=self.__components.frame_rate
+            frame_rate=self.__components.frame_rate,
         )
 
     def save_to(
@@ -330,7 +399,7 @@ class VideoFromComponents(VideoInput):
         path: str,
         format: VideoContainer = VideoContainer.AUTO,
         codec: VideoCodec = VideoCodec.AUTO,
-        metadata: Optional[dict] = None
+        metadata: Optional[dict] = None,
     ):
         if format != VideoContainer.AUTO and format != VideoContainer.MP4:
             raise ValueError("Only MP4 format is supported for now")
@@ -357,7 +426,10 @@ class VideoFromComponents(VideoInput):
             audio_stream: Optional[av.AudioStream] = None
             if self.__components.audio:
                 audio_sample_rate = int(self.__components.audio['sample_rate'])
-                audio_stream = output.add_stream('aac', rate=audio_sample_rate)
+                waveform = self.__components.audio['waveform']
+                waveform = waveform[0, :, :math.ceil((audio_sample_rate / frame_rate) * self.__components.images.shape[0])]
+                layout = {1: 'mono', 2: 'stereo', 6: '5.1'}.get(waveform.shape[0], 'stereo')
+                audio_stream = output.add_stream('aac', rate=audio_sample_rate, layout=layout)
 
             # Encode video
             for i, frame in enumerate(self.__components.images):
@@ -372,12 +444,21 @@ class VideoFromComponents(VideoInput):
             output.mux(packet)
 
             if audio_stream and self.__components.audio:
-                waveform = self.__components.audio['waveform']
-                waveform = waveform[:, :, :math.ceil((audio_sample_rate / frame_rate) * self.__components.images.shape[0])]
-                frame = av.AudioFrame.from_ndarray(waveform.movedim(2, 1).reshape(1, -1).float().cpu().numpy(), format='flt', layout='mono' if waveform.shape[1] == 1 else 'stereo')
+                frame = av.AudioFrame.from_ndarray(waveform.float().cpu().numpy(), format='fltp', layout=layout)
                 frame.sample_rate = audio_sample_rate
                 frame.pts = 0
                 output.mux(audio_stream.encode(frame))
 
                 # Flush encoder
                 output.mux(audio_stream.encode(None))
+
+    def as_trimmed(
+        self,
+        start_time: float | None = None,
+        duration: float | None = None,
+        strict_duration: bool = True,
+    ) -> VideoInput | None:
+        if self.get_duration() < start_time + duration:
+            return None
+        #TODO Consider tracking duration and trimming at time of save?
+        return VideoFromFile(self.get_stream_source(), start_time=start_time, duration=duration)

comfy_api/latest/_io.py

@@ -27,7 +27,7 @@ if TYPE_CHECKING:
 from comfy_api.internal import (_ComfyNodeInternal, _NodeOutputInternal, classproperty, copy_class, first_real_override, is_class,
     prune_dict, shallow_clone_class)
 from comfy_execution.graph_utils import ExecutionBlocker
-from ._util import MESH, VOXEL, SVG as _SVG
+from ._util import MESH, VOXEL, SVG as _SVG, File3D
 
 
 class FolderType(str, Enum):
@@ -667,6 +667,49 @@ class Voxel(ComfyTypeIO):
 class Mesh(ComfyTypeIO):
     Type = MESH
 
+
+@comfytype(io_type="FILE_3D")
+class File3DAny(ComfyTypeIO):
+    """General 3D file type - accepts any supported 3D format."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_GLB")
+class File3DGLB(ComfyTypeIO):
+    """GLB format 3D file - binary glTF, best for web and cross-platform."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_GLTF")
+class File3DGLTF(ComfyTypeIO):
+    """GLTF format 3D file - JSON-based glTF with external resources."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_FBX")
+class File3DFBX(ComfyTypeIO):
+    """FBX format 3D file - best for game engines and animation."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_OBJ")
+class File3DOBJ(ComfyTypeIO):
+    """OBJ format 3D file - simple geometry format."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_STL")
+class File3DSTL(ComfyTypeIO):
+    """STL format 3D file - best for 3D printing."""
+    Type = File3D
+
+
+@comfytype(io_type="FILE_3D_USDZ")
+class File3DUSDZ(ComfyTypeIO):
+    """USDZ format 3D file - Apple AR format."""
+    Type = File3D
+
+
 @comfytype(io_type="HOOKS")
 class Hooks(ComfyTypeIO):
     if TYPE_CHECKING:
@@ -2037,6 +2080,13 @@ __all__ = [
     "LossMap",
     "Voxel",
     "Mesh",
+    "File3DAny",
+    "File3DGLB",
+    "File3DGLTF",
+    "File3DFBX",
+    "File3DOBJ",
+    "File3DSTL",
+    "File3DUSDZ",
     "Hooks",
     "HookKeyframes",
     "TimestepsRange",

comfy_api/latest/_util/__init__.py

@@ -1,5 +1,5 @@
 from .video_types import VideoContainer, VideoCodec, VideoComponents
-from .geometry_types import VOXEL, MESH
+from .geometry_types import VOXEL, MESH, File3D
 from .image_types import SVG
 
 __all__ = [
@@ -9,5 +9,6 @@ __all__ = [
     "VideoComponents",
     "VOXEL",
     "MESH",
+    "File3D",
     "SVG",
 ]

comfy_api/latest/_util/geometry_types.py

@@ -1,3 +1,8 @@
+import shutil
+from io import BytesIO
+from pathlib import Path
+from typing import IO
+
 import torch
 
 
@@ -10,3 +15,75 @@ class MESH:
     def __init__(self, vertices: torch.Tensor, faces: torch.Tensor):
         self.vertices = vertices
         self.faces = faces
+
+
+class File3D:
+    """Class representing a 3D file from a file path or binary stream.
+
+    Supports both disk-backed (file path) and memory-backed (BytesIO) storage.
+    """
+
+    def __init__(self, source: str | IO[bytes], file_format: str = ""):
+        self._source = source
+        self._format = file_format or self._infer_format()
+
+    def _infer_format(self) -> str:
+        if isinstance(self._source, str):
+            return Path(self._source).suffix.lstrip(".").lower()
+        return ""
+
+    @property
+    def format(self) -> str:
+        return self._format
+
+    @format.setter
+    def format(self, value: str) -> None:
+        self._format = value.lstrip(".").lower() if value else ""
+
+    @property
+    def is_disk_backed(self) -> bool:
+        return isinstance(self._source, str)
+
+    def get_source(self) -> str | IO[bytes]:
+        if isinstance(self._source, str):
+            return self._source
+        if hasattr(self._source, "seek"):
+            self._source.seek(0)
+        return self._source
+
+    def get_data(self) -> BytesIO:
+        if isinstance(self._source, str):
+            with open(self._source, "rb") as f:
+                result = BytesIO(f.read())
+            return result
+        if hasattr(self._source, "seek"):
+            self._source.seek(0)
+        if isinstance(self._source, BytesIO):
+            return self._source
+        return BytesIO(self._source.read())
+
+    def save_to(self, path: str) -> str:
+        dest = Path(path)
+        dest.parent.mkdir(parents=True, exist_ok=True)
+
+        if isinstance(self._source, str):
+            if Path(self._source).resolve() != dest.resolve():
+                shutil.copy2(self._source, dest)
+        else:
+            if hasattr(self._source, "seek"):
+                self._source.seek(0)
+            with open(dest, "wb") as f:
+                f.write(self._source.read())
+        return str(dest)
+
+    def get_bytes(self) -> bytes:
+        if isinstance(self._source, str):
+            return Path(self._source).read_bytes()
+        if hasattr(self._source, "seek"):
+            self._source.seek(0)
+        return self._source.read()
+
+    def __repr__(self) -> str:
+        if isinstance(self._source, str):
+            return f"File3D(source={self._source!r}, format={self._format!r})"
+        return f"File3D(<stream>, format={self._format!r})"

comfy_api_nodes/apis/__init__.py

@@ -1197,12 +1197,6 @@ class KlingImageGenImageReferenceType(str, Enum):
     face = 'face'
 
 
-class KlingImageGenModelName(str, Enum):
-    kling_v1 = 'kling-v1'
-    kling_v1_5 = 'kling-v1-5'
-    kling_v2 = 'kling-v2'
-
-
 class KlingImageGenerationsRequest(BaseModel):
     aspect_ratio: Optional[KlingImageGenAspectRatio] = '16:9'
     callback_url: Optional[AnyUrl] = Field(
@@ -1218,7 +1212,7 @@ class KlingImageGenerationsRequest(BaseModel):
         0.5, description='Reference intensity for user-uploaded images', ge=0.0, le=1.0
     )
     image_reference: Optional[KlingImageGenImageReferenceType] = None
-    model_name: Optional[KlingImageGenModelName] = 'kling-v1'
+    model_name: str = Field(...)
     n: Optional[int] = Field(1, description='Number of generated images', ge=1, le=9)
     negative_prompt: Optional[str] = Field(
         None, description='Negative text prompt', max_length=200

comfy_api_nodes/apis/hitpaw.py

@@ -0,0 +1,51 @@
+from typing import TypedDict
+
+from pydantic import BaseModel, Field
+
+
+class InputVideoModel(TypedDict):
+    model: str
+    resolution: str
+
+
+class ImageEnhanceTaskCreateRequest(BaseModel):
+    model_name: str = Field(...)
+    img_url: str = Field(...)
+    extension: str = Field(".png")
+    exif: bool = Field(False)
+    DPI: int | None = Field(None)
+
+
+class VideoEnhanceTaskCreateRequest(BaseModel):
+    video_url: str = Field(...)
+    extension: str = Field(".mp4")
+    model_name: str | None = Field(...)
+    resolution: list[int] = Field(..., description="Target resolution [width, height]")
+    original_resolution: list[int] = Field(..., description="Original video resolution [width, height]")
+
+
+class TaskCreateDataResponse(BaseModel):
+    job_id: str = Field(...)
+    consume_coins: int | None = Field(None)
+
+
+class TaskStatusPollRequest(BaseModel):
+    job_id: str = Field(...)
+
+
+class TaskCreateResponse(BaseModel):
+    code: int = Field(...)
+    message: str = Field(...)
+    data: TaskCreateDataResponse | None = Field(None)
+
+
+class TaskStatusDataResponse(BaseModel):
+    job_id: str = Field(...)
+    status: str = Field(...)
+    res_url: str = Field("")
+
+
+class TaskStatusResponse(BaseModel):
+    code: int = Field(...)
+    message: str = Field(...)
+    data: TaskStatusDataResponse = Field(...)

comfy_api_nodes/apis/kling.py

@@ -1,12 +1,22 @@
 from pydantic import BaseModel, Field
 
 
+class MultiPromptEntry(BaseModel):
+    index: int = Field(...)
+    prompt: str = Field(...)
+    duration: str = Field(...)
+
+
 class OmniProText2VideoRequest(BaseModel):
     model_name: str = Field(..., description="kling-video-o1")
     aspect_ratio: str = Field(..., description="'16:9', '9:16' or '1:1'")
     duration: str = Field(..., description="'5' or '10'")
     prompt: str = Field(...)
     mode: str = Field("pro")
+    multi_shot: bool | None = Field(None)
+    multi_prompt: list[MultiPromptEntry] | None = Field(None)
+    shot_type: str | None = Field(None)
+    sound: str = Field(..., description="'on' or 'off'")
 
 
 class OmniParamImage(BaseModel):
@@ -26,6 +36,10 @@ class OmniProFirstLastFrameRequest(BaseModel):
     duration: str = Field(..., description="'5' or '10'")
     prompt: str = Field(...)
     mode: str = Field("pro")
+    sound: str | None = Field(None, description="'on' or 'off'")
+    multi_shot: bool | None = Field(None)
+    multi_prompt: list[MultiPromptEntry] | None = Field(None)
+    shot_type: str | None = Field(None)
 
 
 class OmniProReferences2VideoRequest(BaseModel):
@@ -38,6 +52,10 @@ class OmniProReferences2VideoRequest(BaseModel):
     duration: str | None = Field(..., description="From 3 to 10.")
     prompt: str = Field(...)
     mode: str = Field("pro")
+    sound: str | None = Field(None, description="'on' or 'off'")
+    multi_shot: bool | None = Field(None)
+    multi_prompt: list[MultiPromptEntry] | None = Field(None)
+    shot_type: str | None = Field(None)
 
 
 class TaskStatusVideoResult(BaseModel):
@@ -54,6 +72,7 @@ class TaskStatusImageResult(BaseModel):
 class TaskStatusResults(BaseModel):
     videos: list[TaskStatusVideoResult] | None = Field(None)
     images: list[TaskStatusImageResult] | None = Field(None)
+    series_images: list[TaskStatusImageResult] | None = Field(None)
 
 
 class TaskStatusResponseData(BaseModel):
@@ -77,31 +96,42 @@ class OmniImageParamImage(BaseModel):
 
 
 class OmniProImageRequest(BaseModel):
-    model_name: str = Field(..., description="kling-image-o1")
-    resolution: str = Field(..., description="'1k' or '2k'")
+    model_name: str = Field(...)
+    resolution: str = Field(...)
     aspect_ratio: str | None = Field(...)
     prompt: str = Field(...)
     mode: str = Field("pro")
     n: int | None = Field(1, le=9)
     image_list: list[OmniImageParamImage] | None = Field(..., max_length=10)
+    result_type: str | None = Field(None, description="Set to 'series' for series generation")
+    series_amount: int | None = Field(None, ge=2, le=9, description="Number of images in a series")
 
 
 class TextToVideoWithAudioRequest(BaseModel):
-    model_name: str = Field(..., description="kling-v2-6")
+    model_name: str = Field(...)
     aspect_ratio: str = Field(..., description="'16:9', '9:16' or '1:1'")
-    duration: str = Field(..., description="'5' or '10'")
-    prompt: str = Field(...)
+    duration: str = Field(...)
+    prompt: str | None = Field(...)
+    negative_prompt: str | None = Field(None)
     mode: str = Field("pro")
     sound: str = Field(..., description="'on' or 'off'")
+    multi_shot: bool | None = Field(None)
+    multi_prompt: list[MultiPromptEntry] | None = Field(None)
+    shot_type: str | None = Field(None)
 
 
 class ImageToVideoWithAudioRequest(BaseModel):
-    model_name: str = Field(..., description="kling-v2-6")
+    model_name: str = Field(...)
     image: str = Field(...)
-    duration: str = Field(..., description="'5' or '10'")
-    prompt: str = Field(...)
+    image_tail: str | None = Field(None)
+    duration: str = Field(...)
+    prompt: str | None = Field(...)
+    negative_prompt: str | None = Field(None)
     mode: str = Field("pro")
     sound: str = Field(..., description="'on' or 'off'")
+    multi_shot: bool | None = Field(None)
+    multi_prompt: list[MultiPromptEntry] | None = Field(None)
+    shot_type: str | None = Field(None)
 
 
 class MotionControlRequest(BaseModel):

comfy_api_nodes/apis/meshy.py

@@ -109,14 +109,19 @@ class MeshyTextureRequest(BaseModel):
 
 class MeshyModelsUrls(BaseModel):
     glb: str = Field("")
+    fbx: str = Field("")
+    usdz: str = Field("")
+    obj: str = Field("")
 
 
 class MeshyRiggedModelsUrls(BaseModel):
     rigged_character_glb_url: str = Field("")
+    rigged_character_fbx_url: str = Field("")
 
 
 class MeshyAnimatedModelsUrls(BaseModel):
     animation_glb_url: str = Field("")
+    animation_fbx_url: str = Field("")
 
 
 class MeshyResultTextureUrls(BaseModel):

comfy_api_nodes/nodes_hitpaw.py

@@ -0,0 +1,342 @@
+import math
+
+from typing_extensions import override
+
+from comfy_api.latest import IO, ComfyExtension, Input
+from comfy_api_nodes.apis.hitpaw import (
+    ImageEnhanceTaskCreateRequest,
+    InputVideoModel,
+    TaskCreateDataResponse,
+    TaskCreateResponse,
+    TaskStatusPollRequest,
+    TaskStatusResponse,
+    VideoEnhanceTaskCreateRequest,
+)
+from comfy_api_nodes.util import (
+    ApiEndpoint,
+    download_url_to_image_tensor,
+    download_url_to_video_output,
+    downscale_image_tensor,
+    get_image_dimensions,
+    poll_op,
+    sync_op,
+    upload_image_to_comfyapi,
+    upload_video_to_comfyapi,
+    validate_video_duration,
+)
+
+VIDEO_MODELS_MODELS_MAP = {
+    "Portrait Restore Model (1x)": "portrait_restore_1x",
+    "Portrait Restore Model (2x)": "portrait_restore_2x",
+    "General Restore Model (1x)": "general_restore_1x",
+    "General Restore Model (2x)": "general_restore_2x",
+    "General Restore Model (4x)": "general_restore_4x",
+    "Ultra HD Model (2x)": "ultrahd_restore_2x",
+    "Generative Model (1x)": "generative_1x",
+}
+
+# Resolution name to target dimension (shorter side) in pixels
+RESOLUTION_TARGET_MAP = {
+    "720p": 720,
+    "1080p": 1080,
+    "2K/QHD": 1440,
+    "4K/UHD": 2160,
+    "8K": 4320,
+}
+
+# Square (1:1) resolutions use standard square dimensions
+RESOLUTION_SQUARE_MAP = {
+    "720p": 720,
+    "1080p": 1080,
+    "2K/QHD": 1440,
+    "4K/UHD": 2048,  # DCI 4K square
+    "8K": 4096,  # DCI 8K square
+}
+
+# Models with limited resolution support (no 8K)
+LIMITED_RESOLUTION_MODELS = {"Generative Model (1x)"}
+
+# Resolution options for different model types
+RESOLUTIONS_LIMITED = ["original", "720p", "1080p", "2K/QHD", "4K/UHD"]
+RESOLUTIONS_FULL = ["original", "720p", "1080p", "2K/QHD", "4K/UHD", "8K"]
+
+# Maximum output resolution in pixels
+MAX_PIXELS_GENERATIVE = 32_000_000
+MAX_MP_GENERATIVE = MAX_PIXELS_GENERATIVE // 1_000_000
+
+
+class HitPawGeneralImageEnhance(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="HitPawGeneralImageEnhance",
+            display_name="HitPaw General Image Enhance",
+            category="api node/image/HitPaw",
+            description="Upscale low-resolution images to super-resolution, eliminate artifacts and noise. "
+            f"Maximum output: {MAX_MP_GENERATIVE} megapixels.",
+            inputs=[
+                IO.Combo.Input("model", options=["generative_portrait", "generative"]),
+                IO.Image.Input("image"),
+                IO.Combo.Input("upscale_factor", options=[1, 2, 4]),
+                IO.Boolean.Input(
+                    "auto_downscale",
+                    default=False,
+                    tooltip="Automatically downscale input image if output would exceed the limit.",
+                ),
+            ],
+            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"]),
+                expr="""
+                (
+                  $prices := {
+                    "generative_portrait": {"min": 0.02, "max": 0.06},
+                    "generative": {"min": 0.05, "max": 0.15}
+                  };
+                  $price := $lookup($prices, widgets.model);
+                  {
+                    "type": "range_usd",
+                    "min_usd": $price.min,
+                    "max_usd": $price.max
+                  }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: str,
+        image: Input.Image,
+        upscale_factor: int,
+        auto_downscale: bool,
+    ) -> IO.NodeOutput:
+        height, width = get_image_dimensions(image)
+        requested_scale = upscale_factor
+        output_pixels = height * width * requested_scale * requested_scale
+        if output_pixels > MAX_PIXELS_GENERATIVE:
+            if auto_downscale:
+                input_pixels = width * height
+                scale = 1
+                max_input_pixels = MAX_PIXELS_GENERATIVE
+
+                for candidate in [4, 2, 1]:
+                    if candidate > requested_scale:
+                        continue
+                    scale_output_pixels = input_pixels * candidate * candidate
+                    if scale_output_pixels <= MAX_PIXELS_GENERATIVE:
+                        scale = candidate
+                        max_input_pixels = None
+                        break
+                    # Check if we can downscale input by at most 2x to fit
+                    downscale_ratio = math.sqrt(scale_output_pixels / MAX_PIXELS_GENERATIVE)
+                    if downscale_ratio <= 2.0:
+                        scale = candidate
+                        max_input_pixels = MAX_PIXELS_GENERATIVE // (candidate * candidate)
+                        break
+
+                if max_input_pixels is not None:
+                    image = downscale_image_tensor(image, total_pixels=max_input_pixels)
+                upscale_factor = scale
+            else:
+                output_width = width * requested_scale
+                output_height = height * requested_scale
+                raise ValueError(
+                    f"Output size ({output_width}x{output_height} = {output_pixels:,} pixels) "
+                    f"exceeds maximum allowed size of {MAX_PIXELS_GENERATIVE:,} pixels ({MAX_MP_GENERATIVE}MP). "
+                    f"Enable auto_downscale or use a smaller input image or a lower upscale factor."
+                )
+
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/photo-enhancer", method="POST"),
+            response_model=TaskCreateResponse,
+            data=ImageEnhanceTaskCreateRequest(
+                model_name=f"{model}_{upscale_factor}x",
+                img_url=await upload_image_to_comfyapi(cls, image, total_pixels=None),
+            ),
+            wait_label="Creating task",
+            final_label_on_success="Task created",
+        )
+        if initial_res.code != 200:
+            raise ValueError(f"Task creation failed with code {initial_res.code}: {initial_res.message}")
+        request_price = initial_res.data.consume_coins / 1000
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/task-status", method="POST"),
+            data=TaskCreateDataResponse(job_id=initial_res.data.job_id),
+            response_model=TaskStatusResponse,
+            status_extractor=lambda x: x.data.status,
+            price_extractor=lambda x: request_price,
+            poll_interval=10.0,
+            max_poll_attempts=480,
+        )
+        return IO.NodeOutput(await download_url_to_image_tensor(final_response.data.res_url))
+
+
+class HitPawVideoEnhance(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        model_options = []
+        for model_name in VIDEO_MODELS_MODELS_MAP:
+            if model_name in LIMITED_RESOLUTION_MODELS:
+                resolutions = RESOLUTIONS_LIMITED
+            else:
+                resolutions = RESOLUTIONS_FULL
+            model_options.append(
+                IO.DynamicCombo.Option(
+                    model_name,
+                    [IO.Combo.Input("resolution", options=resolutions)],
+                )
+            )
+
+        return IO.Schema(
+            node_id="HitPawVideoEnhance",
+            display_name="HitPaw Video Enhance",
+            category="api node/video/HitPaw",
+            description="Upscale low-resolution videos to high resolution, eliminate artifacts and noise. "
+            "Prices shown are per second of video.",
+            inputs=[
+                IO.DynamicCombo.Input("model", options=model_options),
+                IO.Video.Input("video"),
+            ],
+            outputs=[
+                IO.Video.Output(),
+            ],
+            hidden=[
+                IO.Hidden.auth_token_comfy_org,
+                IO.Hidden.api_key_comfy_org,
+                IO.Hidden.unique_id,
+            ],
+            is_api_node=True,
+            price_badge=IO.PriceBadge(
+                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.resolution"]),
+                expr="""
+                (
+                  $m := $lookup(widgets, "model");
+                  $res := $lookup(widgets, "model.resolution");
+                  $standard_model_prices := {
+                    "original": {"min": 0.01, "max": 0.198},
+                    "720p": {"min": 0.01, "max": 0.06},
+                    "1080p": {"min": 0.015, "max": 0.09},
+                    "2k/qhd": {"min": 0.02, "max": 0.117},
+                    "4k/uhd": {"min": 0.025, "max": 0.152},
+                    "8k": {"min": 0.033, "max": 0.198}
+                  };
+                  $ultra_hd_model_prices := {
+                    "original": {"min": 0.015, "max": 0.264},
+                    "720p": {"min": 0.015, "max": 0.092},
+                    "1080p": {"min": 0.02, "max": 0.12},
+                    "2k/qhd": {"min": 0.026, "max": 0.156},
+                    "4k/uhd": {"min": 0.034, "max": 0.203},
+                    "8k": {"min": 0.044, "max": 0.264}
+                  };
+                  $generative_model_prices := {
+                    "original": {"min": 0.015, "max": 0.338},
+                    "720p": {"min": 0.008, "max": 0.090},
+                    "1080p": {"min": 0.05, "max": 0.15},
+                    "2k/qhd": {"min": 0.038, "max": 0.225},
+                    "4k/uhd": {"min": 0.056, "max": 0.338}
+                  };
+                  $prices := $contains($m, "ultra hd") ? $ultra_hd_model_prices :
+                             $contains($m, "generative") ? $generative_model_prices :
+                             $standard_model_prices;
+                  $price := $lookup($prices, $res);
+                  {
+                    "type": "range_usd",
+                    "min_usd": $price.min,
+                    "max_usd": $price.max,
+                    "format": {"approximate": true, "suffix": "/second"}
+                  }
+                )
+                """,
+            ),
+        )
+
+    @classmethod
+    async def execute(
+        cls,
+        model: InputVideoModel,
+        video: Input.Video,
+    ) -> IO.NodeOutput:
+        validate_video_duration(video, min_duration=0.5, max_duration=60 * 60)
+        resolution = model["resolution"]
+        src_width, src_height = video.get_dimensions()
+
+        if resolution == "original":
+            output_width = src_width
+            output_height = src_height
+        else:
+            if src_width == src_height:
+                target_size = RESOLUTION_SQUARE_MAP[resolution]
+                if target_size < src_width:
+                    raise ValueError(
+                        f"Selected resolution {resolution} ({target_size}x{target_size}) is smaller than "
+                        f"the input video ({src_width}x{src_height}). Please select a higher resolution or 'original'."
+                    )
+                output_width = target_size
+                output_height = target_size
+            else:
+                min_dimension = min(src_width, src_height)
+                target_size = RESOLUTION_TARGET_MAP[resolution]
+                if target_size < min_dimension:
+                    raise ValueError(
+                        f"Selected resolution {resolution} ({target_size}p) is smaller than "
+                        f"the input video's shorter dimension ({min_dimension}p). "
+                        f"Please select a higher resolution or 'original'."
+                    )
+                if src_width > src_height:
+                    output_height = target_size
+                    output_width = int(target_size * (src_width / src_height))
+                else:
+                    output_width = target_size
+                    output_height = int(target_size * (src_height / src_width))
+        initial_res = await sync_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/video-enhancer", method="POST"),
+            response_model=TaskCreateResponse,
+            data=VideoEnhanceTaskCreateRequest(
+                video_url=await upload_video_to_comfyapi(cls, video),
+                resolution=[output_width, output_height],
+                original_resolution=[src_width, src_height],
+                model_name=VIDEO_MODELS_MODELS_MAP[model["model"]],
+            ),
+            wait_label="Creating task",
+            final_label_on_success="Task created",
+        )
+        request_price = initial_res.data.consume_coins / 1000
+        if initial_res.code != 200:
+            raise ValueError(f"Task creation failed with code {initial_res.code}: {initial_res.message}")
+        final_response = await poll_op(
+            cls,
+            ApiEndpoint(path="/proxy/hitpaw/api/task-status", method="POST"),
+            data=TaskStatusPollRequest(job_id=initial_res.data.job_id),
+            response_model=TaskStatusResponse,
+            status_extractor=lambda x: x.data.status,
+            price_extractor=lambda x: request_price,
+            poll_interval=10.0,
+            max_poll_attempts=320,
+        )
+        return IO.NodeOutput(await download_url_to_video_output(final_response.data.res_url))
+
+
+class HitPawExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[IO.ComfyNode]]:
+        return [
+            HitPawGeneralImageEnhance,
+            HitPawVideoEnhance,
+        ]
+
+
+async def comfy_entrypoint() -> HitPawExtension:
+    return HitPawExtension()

comfy_api_nodes/nodes_hunyuan3d.py

@@ -1,5 +1,3 @@
-import os
-
 from typing_extensions import override
 
 from comfy_api.latest import IO, ComfyExtension, Input
@@ -14,7 +12,7 @@ from comfy_api_nodes.apis.hunyuan3d import (
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
-    download_url_to_bytesio,
+    download_url_to_file_3d,
     downscale_image_tensor_by_max_side,
     poll_op,
     sync_op,
@@ -22,14 +20,13 @@ from comfy_api_nodes.util import (
     validate_image_dimensions,
     validate_string,
 )
-from folder_paths import get_output_directory
 
 
-def get_glb_obj_from_response(response_objs: list[ResultFile3D]) -> ResultFile3D:
+def get_file_from_response(response_objs: list[ResultFile3D], file_type: str) -> ResultFile3D | None:
     for i in response_objs:
-        if i.Type.lower() == "glb":
+        if i.Type.lower() == file_type.lower():
             return i
-    raise ValueError("No GLB file found in response. Please report this to the developers.")
+    return None
 
 
 class TencentTextToModelNode(IO.ComfyNode):
@@ -74,7 +71,9 @@ class TencentTextToModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DOBJ.Output(display_name="OBJ"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -124,19 +123,20 @@ class TencentTextToModelNode(IO.ComfyNode):
         )
         if response.Error:
             raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
+        task_id = response.JobId
         result = await poll_op(
             cls,
             ApiEndpoint(path="/proxy/tencent/hunyuan/3d-pro/query", method="POST"),
-            data=To3DProTaskQueryRequest(JobId=response.JobId),
+            data=To3DProTaskQueryRequest(JobId=task_id),
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
-        model_file = f"hunyuan_model_{response.JobId}.glb"
-        await download_url_to_bytesio(
-            get_glb_obj_from_response(result.ResultFile3Ds).Url,
-            os.path.join(get_output_directory(), model_file),
+        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
         )
-        return IO.NodeOutput(model_file)
 
 
 class TencentImageToModelNode(IO.ComfyNode):
@@ -184,7 +184,9 @@ class TencentImageToModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DOBJ.Output(display_name="OBJ"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -269,19 +271,20 @@ class TencentImageToModelNode(IO.ComfyNode):
         )
         if response.Error:
             raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
+        task_id = response.JobId
         result = await poll_op(
             cls,
             ApiEndpoint(path="/proxy/tencent/hunyuan/3d-pro/query", method="POST"),
-            data=To3DProTaskQueryRequest(JobId=response.JobId),
+            data=To3DProTaskQueryRequest(JobId=task_id),
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
-        model_file = f"hunyuan_model_{response.JobId}.glb"
-        await download_url_to_bytesio(
-            get_glb_obj_from_response(result.ResultFile3Ds).Url,
-            os.path.join(get_output_directory(), model_file),
+        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
         )
-        return IO.NodeOutput(model_file)
 
 
 class TencentHunyuan3DExtension(ComfyExtension):

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
@@ -103,11 +127,20 @@ 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 +201,10 @@ 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 +226,7 @@ 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,
         )
@@ -257,8 +295,14 @@ 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 +365,9 @@ 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 +386,7 @@ 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,
         )
@@ -877,8 +925,8 @@ 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

@@ -1,5 +1,3 @@
-import os
-
 from typing_extensions import override
 
 from comfy_api.latest import IO, ComfyExtension, Input
@@ -20,13 +18,12 @@ from comfy_api_nodes.apis.meshy import (
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
-    download_url_to_bytesio,
+    download_url_to_file_3d,
     poll_op,
     sync_op,
     upload_images_to_comfyapi,
     validate_string,
 )
-from folder_paths import get_output_directory
 
 
 class MeshyTextToModelNode(IO.ComfyNode):
@@ -79,8 +76,10 @@ class MeshyTextToModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -122,16 +121,20 @@ class MeshyTextToModelNode(IO.ComfyNode):
                 seed=seed,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{task_id}"),
             response_model=MeshyModelResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id),
+        )
 
 
 class MeshyRefineNode(IO.ComfyNode):
@@ -167,8 +170,10 @@ class MeshyRefineNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -210,16 +215,20 @@ class MeshyRefineNode(IO.ComfyNode):
                 ai_model=model,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v2/text-to-3d/{task_id}"),
             response_model=MeshyModelResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id),
+        )
 
 
 class MeshyImageToModelNode(IO.ComfyNode):
@@ -303,8 +312,10 @@ class MeshyImageToModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -368,16 +379,20 @@ class MeshyImageToModelNode(IO.ComfyNode):
                 seed=seed,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/image-to-3d/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/image-to-3d/{task_id}"),
             response_model=MeshyModelResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id),
+        )
 
 
 class MeshyMultiImageToModelNode(IO.ComfyNode):
@@ -464,8 +479,10 @@ class MeshyMultiImageToModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MESHY_TASK_ID").Output(display_name="meshy_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -531,16 +548,20 @@ class MeshyMultiImageToModelNode(IO.ComfyNode):
                 seed=seed,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/multi-image-to-3d/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/multi-image-to-3d/{task_id}"),
             response_model=MeshyModelResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id),
+        )
 
 
 class MeshyRigModelNode(IO.ComfyNode):
@@ -571,8 +592,10 @@ class MeshyRigModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MESHY_RIGGED_TASK_ID").Output(display_name="rig_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -606,18 +629,20 @@ class MeshyRigModelNode(IO.ComfyNode):
                 texture_image_url=texture_image_url,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/rigging/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/rigging/{task_id}"),
             response_model=MeshyRiggedResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(
-            result.result.rigged_character_glb_url, os.path.join(get_output_directory(), model_file)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.result.rigged_character_glb_url, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.result.rigged_character_fbx_url, "fbx", task_id=task_id),
         )
-        return IO.NodeOutput(model_file, response.result)
 
 
 class MeshyAnimateModelNode(IO.ComfyNode):
@@ -640,7 +665,9 @@ class MeshyAnimateModelNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -669,16 +696,19 @@ class MeshyAnimateModelNode(IO.ComfyNode):
                 action_id=action_id,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/animations/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/animations/{task_id}"),
             response_model=MeshyAnimationResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.result.animation_glb_url, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            await download_url_to_file_3d(result.result.animation_glb_url, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.result.animation_fbx_url, "fbx", task_id=task_id),
+        )
 
 
 class MeshyTextureNode(IO.ComfyNode):
@@ -715,8 +745,10 @@ class MeshyTextureNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="meshy_task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
+                IO.File3DFBX.Output(display_name="FBX"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -760,16 +792,20 @@ class MeshyTextureNode(IO.ComfyNode):
                 image_style_url=image_style_url,
             ),
         )
+        task_id = response.result
         result = await poll_op(
             cls,
-            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/retexture/{response.result}"),
+            ApiEndpoint(path=f"/proxy/meshy/openapi/v1/retexture/{task_id}"),
             response_model=MeshyModelResult,
             status_extractor=lambda r: r.status,
             progress_extractor=lambda r: r.progress,
         )
-        model_file = f"meshy_model_{response.result}.glb"
-        await download_url_to_bytesio(result.model_urls.glb, os.path.join(get_output_directory(), model_file))
-        return IO.NodeOutput(model_file, response.result)
+        return IO.NodeOutput(
+            f"{task_id}.glb",
+            task_id,
+            await download_url_to_file_3d(result.model_urls.glb, "glb", task_id=task_id),
+            await download_url_to_file_3d(result.model_urls.fbx, "fbx", task_id=task_id),
+        )
 
 
 class MeshyExtension(ComfyExtension):

comfy_api_nodes/nodes_moonvalley.py

@@ -219,8 +219,8 @@ class MoonvalleyImg2VideoNode(IO.ComfyNode):
                 ),
                 IO.Int.Input(
                     "steps",
-                    default=33,
-                    min=1,
+                    default=80,
+                    min=75,  # steps should be greater or equal to cooldown_steps(75) + warmup_steps(0)
                     max=100,
                     step=1,
                     tooltip="Number of denoising steps",
@@ -340,8 +340,8 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode):
                 ),
                 IO.Int.Input(
                     "steps",
-                    default=33,
-                    min=1,
+                    default=60,
+                    min=60,  # steps should be greater or equal to cooldown_steps(36) + warmup_steps(24)
                     max=100,
                     step=1,
                     display_mode=IO.NumberDisplay.number,
@@ -370,7 +370,7 @@ class MoonvalleyVideo2VideoNode(IO.ComfyNode):
         video: Input.Video | None = None,
         control_type: str = "Motion Transfer",
         motion_intensity: int | None = 100,
-        steps=33,
+        steps=60,
         prompt_adherence=4.5,
     ) -> IO.NodeOutput:
         validated_video = validate_video_to_video_input(video)
@@ -465,8 +465,8 @@ class MoonvalleyTxt2VideoNode(IO.ComfyNode):
                 ),
                 IO.Int.Input(
                     "steps",
-                    default=33,
-                    min=1,
+                    default=80,
+                    min=75,  # steps should be greater or equal to cooldown_steps(75) + warmup_steps(0)
                     max=100,
                     step=1,
                     tooltip="Inference steps",

comfy_api_nodes/nodes_rodin.py

@@ -10,7 +10,6 @@ import folder_paths as comfy_paths
 import os
 import logging
 import math
-from typing import Optional
 from io import BytesIO
 from typing_extensions import override
 from PIL import Image
@@ -28,8 +27,9 @@ from comfy_api_nodes.util import (
     poll_op,
     ApiEndpoint,
     download_url_to_bytesio,
+    download_url_to_file_3d,
 )
-from comfy_api.latest import ComfyExtension, IO
+from comfy_api.latest import ComfyExtension, IO, Types
 
 
 COMMON_PARAMETERS = [
@@ -177,7 +177,7 @@ def check_rodin_status(response: Rodin3DCheckStatusResponse) -> str:
         return "DONE"
     return "Generating"
 
-def extract_progress(response: Rodin3DCheckStatusResponse) -> Optional[int]:
+def extract_progress(response: Rodin3DCheckStatusResponse) -> int | None:
     if not response.jobs:
         return None
     completed_count = sum(1 for job in response.jobs if job.status == JobStatus.Done)
@@ -207,17 +207,25 @@ async def get_rodin_download_list(uuid: str, cls: type[IO.ComfyNode]) -> Rodin3D
     )
 
 
-async def download_files(url_list, task_uuid: str):
+async def download_files(url_list, task_uuid: str) -> tuple[str | None, Types.File3D | None]:
     result_folder_name = f"Rodin3D_{task_uuid}"
     save_path = os.path.join(comfy_paths.get_output_directory(), result_folder_name)
     os.makedirs(save_path, exist_ok=True)
     model_file_path = None
+    file_3d = None
+
     for i in url_list.list:
         file_path = os.path.join(save_path, i.name)
-        if file_path.endswith(".glb"):
+        if i.name.lower().endswith(".glb"):
             model_file_path = os.path.join(result_folder_name, i.name)
-        await download_url_to_bytesio(i.url, file_path)
-    return model_file_path
+            file_3d = await download_url_to_file_3d(i.url, "glb")
+            # Save to disk for backward compatibility
+            with open(file_path, "wb") as f:
+                f.write(file_3d.get_bytes())
+        else:
+            await download_url_to_bytesio(i.url, file_path)
+
+    return model_file_path, file_3d
 
 
 class Rodin3D_Regular(IO.ComfyNode):
@@ -234,7 +242,10 @@ class Rodin3D_Regular(IO.ComfyNode):
                 IO.Image.Input("Images"),
                 *COMMON_PARAMETERS,
             ],
-            outputs=[IO.String.Output(display_name="3D Model Path")],
+            outputs=[
+                IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
                 IO.Hidden.api_key_comfy_org,
@@ -271,9 +282,9 @@ class Rodin3D_Regular(IO.ComfyNode):
         )
         await poll_for_task_status(subscription_key, cls)
         download_list = await get_rodin_download_list(task_uuid, cls)
-        model = await download_files(download_list, task_uuid)
+        model_path, file_3d = await download_files(download_list, task_uuid)
 
-        return IO.NodeOutput(model)
+        return IO.NodeOutput(model_path, file_3d)
 
 
 class Rodin3D_Detail(IO.ComfyNode):
@@ -290,7 +301,10 @@ class Rodin3D_Detail(IO.ComfyNode):
                 IO.Image.Input("Images"),
                 *COMMON_PARAMETERS,
             ],
-            outputs=[IO.String.Output(display_name="3D Model Path")],
+            outputs=[
+                IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
                 IO.Hidden.api_key_comfy_org,
@@ -327,9 +341,9 @@ class Rodin3D_Detail(IO.ComfyNode):
         )
         await poll_for_task_status(subscription_key, cls)
         download_list = await get_rodin_download_list(task_uuid, cls)
-        model = await download_files(download_list, task_uuid)
+        model_path, file_3d = await download_files(download_list, task_uuid)
 
-        return IO.NodeOutput(model)
+        return IO.NodeOutput(model_path, file_3d)
 
 
 class Rodin3D_Smooth(IO.ComfyNode):
@@ -346,7 +360,10 @@ class Rodin3D_Smooth(IO.ComfyNode):
                 IO.Image.Input("Images"),
                 *COMMON_PARAMETERS,
             ],
-            outputs=[IO.String.Output(display_name="3D Model Path")],
+            outputs=[
+                IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
                 IO.Hidden.api_key_comfy_org,
@@ -382,9 +399,9 @@ class Rodin3D_Smooth(IO.ComfyNode):
         )
         await poll_for_task_status(subscription_key, cls)
         download_list = await get_rodin_download_list(task_uuid, cls)
-        model = await download_files(download_list, task_uuid)
+        model_path, file_3d = await download_files(download_list, task_uuid)
 
-        return IO.NodeOutput(model)
+        return IO.NodeOutput(model_path, file_3d)
 
 
 class Rodin3D_Sketch(IO.ComfyNode):
@@ -408,7 +425,10 @@ class Rodin3D_Sketch(IO.ComfyNode):
                     optional=True,
                 ),
             ],
-            outputs=[IO.String.Output(display_name="3D Model Path")],
+            outputs=[
+                IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
                 IO.Hidden.api_key_comfy_org,
@@ -441,9 +461,9 @@ class Rodin3D_Sketch(IO.ComfyNode):
         )
         await poll_for_task_status(subscription_key, cls)
         download_list = await get_rodin_download_list(task_uuid, cls)
-        model = await download_files(download_list, task_uuid)
+        model_path, file_3d = await download_files(download_list, task_uuid)
 
-        return IO.NodeOutput(model)
+        return IO.NodeOutput(model_path, file_3d)
 
 
 class Rodin3D_Gen2(IO.ComfyNode):
@@ -475,7 +495,10 @@ class Rodin3D_Gen2(IO.ComfyNode):
                 ),
                 IO.Boolean.Input("TAPose", default=False),
             ],
-            outputs=[IO.String.Output(display_name="3D Model Path")],
+            outputs=[
+                IO.String.Output(display_name="3D Model Path"),  # for backward compatibility only
+                IO.File3DGLB.Output(display_name="GLB"),
+            ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
                 IO.Hidden.api_key_comfy_org,
@@ -511,9 +534,9 @@ class Rodin3D_Gen2(IO.ComfyNode):
         )
         await poll_for_task_status(subscription_key, cls)
         download_list = await get_rodin_download_list(task_uuid, cls)
-        model = await download_files(download_list, task_uuid)
+        model_path, file_3d = await download_files(download_list, task_uuid)
 
-        return IO.NodeOutput(model)
+        return IO.NodeOutput(model_path, file_3d)
 
 
 class Rodin3DExtension(ComfyExtension):

comfy_api_nodes/nodes_tripo.py

@@ -1,10 +1,6 @@
-import os
-from typing import Optional
-
-import torch
 from typing_extensions import override
 
-from comfy_api.latest import IO, ComfyExtension
+from comfy_api.latest import IO, ComfyExtension, Input
 from comfy_api_nodes.apis.tripo import (
     TripoAnimateRetargetRequest,
     TripoAnimateRigRequest,
@@ -26,12 +22,11 @@ from comfy_api_nodes.apis.tripo import (
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
-    download_url_as_bytesio,
+    download_url_to_file_3d,
     poll_op,
     sync_op,
     upload_images_to_comfyapi,
 )
-from folder_paths import get_output_directory
 
 
 def get_model_url_from_response(response: TripoTaskResponse) -> str:
@@ -45,7 +40,7 @@ def get_model_url_from_response(response: TripoTaskResponse) -> str:
 async def poll_until_finished(
     node_cls: type[IO.ComfyNode],
     response: TripoTaskResponse,
-    average_duration: Optional[int] = None,
+    average_duration: int | None = None,
 ) -> IO.NodeOutput:
     """Polls the Tripo API endpoint until the task reaches a terminal state, then returns the response."""
     if response.code != 0:
@@ -69,12 +64,8 @@ async def poll_until_finished(
     )
     if response_poll.data.status == TripoTaskStatus.SUCCESS:
         url = get_model_url_from_response(response_poll)
-        bytesio = await download_url_as_bytesio(url)
-        # Save the downloaded model file
-        model_file = f"tripo_model_{task_id}.glb"
-        with open(os.path.join(get_output_directory(), model_file), "wb") as f:
-            f.write(bytesio.getvalue())
-        return IO.NodeOutput(model_file, task_id)
+        file_glb = await download_url_to_file_3d(url, "glb", task_id=task_id)
+        return IO.NodeOutput(f"{task_id}.glb", task_id, file_glb)
     raise RuntimeError(f"Failed to generate mesh: {response_poll}")
 
 
@@ -107,8 +98,9 @@ class TripoTextToModelNode(IO.ComfyNode):
                 IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -155,18 +147,18 @@ class TripoTextToModelNode(IO.ComfyNode):
     async def execute(
         cls,
         prompt: str,
-        negative_prompt: Optional[str] = None,
+        negative_prompt: str | None = None,
         model_version=None,
-        style: Optional[str] = None,
-        texture: Optional[bool] = None,
-        pbr: Optional[bool] = None,
-        image_seed: Optional[int] = None,
-        model_seed: Optional[int] = None,
-        texture_seed: Optional[int] = None,
-        texture_quality: Optional[str] = None,
-        geometry_quality: Optional[str] = None,
-        face_limit: Optional[int] = None,
-        quad: Optional[bool] = None,
+        style: str | None = None,
+        texture: bool | None = None,
+        pbr: bool | None = None,
+        image_seed: int | None = None,
+        model_seed: int | None = None,
+        texture_seed: int | None = None,
+        texture_quality: str | None = None,
+        geometry_quality: str | None = None,
+        face_limit: int | None = None,
+        quad: bool | None = None,
     ) -> IO.NodeOutput:
         style_enum = None if style == "None" else style
         if not prompt:
@@ -232,8 +224,9 @@ class TripoImageToModelNode(IO.ComfyNode):
                 IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -279,19 +272,19 @@ class TripoImageToModelNode(IO.ComfyNode):
     @classmethod
     async def execute(
         cls,
-        image: torch.Tensor,
-        model_version: Optional[str] = None,
-        style: Optional[str] = None,
-        texture: Optional[bool] = None,
-        pbr: Optional[bool] = None,
-        model_seed: Optional[int] = None,
+        image: Input.Image,
+        model_version: str | None = None,
+        style: str | None = None,
+        texture: bool | None = None,
+        pbr: bool | None = None,
+        model_seed: int | None = None,
         orientation=None,
-        texture_seed: Optional[int] = None,
-        texture_quality: Optional[str] = None,
-        geometry_quality: Optional[str] = None,
-        texture_alignment: Optional[str] = None,
-        face_limit: Optional[int] = None,
-        quad: Optional[bool] = None,
+        texture_seed: int | None = None,
+        texture_quality: str | None = None,
+        geometry_quality: str | None = None,
+        texture_alignment: str | None = None,
+        face_limit: int | None = None,
+        quad: bool | None = None,
     ) -> IO.NodeOutput:
         style_enum = None if style == "None" else style
         if image is None:
@@ -368,8 +361,9 @@ class TripoMultiviewToModelNode(IO.ComfyNode):
                 IO.Combo.Input("geometry_quality", default="standard", options=["standard", "detailed"], optional=True),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -411,21 +405,21 @@ class TripoMultiviewToModelNode(IO.ComfyNode):
     @classmethod
     async def execute(
         cls,
-        image: torch.Tensor,
-        image_left: Optional[torch.Tensor] = None,
-        image_back: Optional[torch.Tensor] = None,
-        image_right: Optional[torch.Tensor] = None,
-        model_version: Optional[str] = None,
-        orientation: Optional[str] = None,
-        texture: Optional[bool] = None,
-        pbr: Optional[bool] = None,
-        model_seed: Optional[int] = None,
-        texture_seed: Optional[int] = None,
-        texture_quality: Optional[str] = None,
-        geometry_quality: Optional[str] = None,
-        texture_alignment: Optional[str] = None,
-        face_limit: Optional[int] = None,
-        quad: Optional[bool] = None,
+        image: Input.Image,
+        image_left: Input.Image | None = None,
+        image_back: Input.Image | None = None,
+        image_right: Input.Image | None = None,
+        model_version: str | None = None,
+        orientation: str | None = None,
+        texture: bool | None = None,
+        pbr: bool | None = None,
+        model_seed: int | None = None,
+        texture_seed: int | None = None,
+        texture_quality: str | None = None,
+        geometry_quality: str | None = None,
+        texture_alignment: str | None = None,
+        face_limit: int | None = None,
+        quad: bool | None = None,
     ) -> IO.NodeOutput:
         if image is None:
             raise RuntimeError("front image for multiview is required")
@@ -487,8 +481,9 @@ class TripoTextureNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -512,11 +507,11 @@ class TripoTextureNode(IO.ComfyNode):
     async def execute(
         cls,
         model_task_id,
-        texture: Optional[bool] = None,
-        pbr: Optional[bool] = None,
-        texture_seed: Optional[int] = None,
-        texture_quality: Optional[str] = None,
-        texture_alignment: Optional[str] = None,
+        texture: bool | None = None,
+        pbr: bool | None = None,
+        texture_seed: int | None = None,
+        texture_quality: str | None = None,
+        texture_alignment: str | None = None,
     ) -> IO.NodeOutput:
         response = await sync_op(
             cls,
@@ -547,8 +542,9 @@ class TripoRefineNode(IO.ComfyNode):
                 IO.Custom("MODEL_TASK_ID").Input("model_task_id", tooltip="Must be a v1.4 Tripo model"),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("MODEL_TASK_ID").Output(display_name="model task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -583,8 +579,9 @@ class TripoRigNode(IO.ComfyNode):
             category="api node/3d/Tripo",
             inputs=[IO.Custom("MODEL_TASK_ID").Input("original_model_task_id")],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("RIG_TASK_ID").Output(display_name="rig task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,
@@ -642,8 +639,9 @@ class TripoRetargetNode(IO.ComfyNode):
                 ),
             ],
             outputs=[
-                IO.String.Output(display_name="model_file"),
+                IO.String.Output(display_name="model_file"),  # for backward compatibility only
                 IO.Custom("RETARGET_TASK_ID").Output(display_name="retarget task_id"),
+                IO.File3DGLB.Output(display_name="GLB"),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,

comfy_api_nodes/util/__init__.py

@@ -28,6 +28,7 @@ from .conversions import (
 from .download_helpers import (
     download_url_as_bytesio,
     download_url_to_bytesio,
+    download_url_to_file_3d,
     download_url_to_image_tensor,
     download_url_to_video_output,
 )
@@ -69,6 +70,7 @@ __all__ = [
     # Download helpers
     "download_url_as_bytesio",
     "download_url_to_bytesio",
+    "download_url_to_file_3d",
     "download_url_to_image_tensor",
     "download_url_to_video_output",
     # Conversions

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/download_helpers.py

@@ -11,7 +11,8 @@ import torch
 from aiohttp.client_exceptions import ClientError, ContentTypeError
 
 from comfy_api.latest import IO as COMFY_IO
-from comfy_api.latest import InputImpl
+from comfy_api.latest import InputImpl, Types
+from folder_paths import get_output_directory
 
 from . import request_logger
 from ._helpers import (
@@ -166,27 +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
@@ -261,3 +260,38 @@ def _generate_operation_id(method: str, url: str, attempt: int) -> str:
     except Exception:
         slug = "download"
     return f"{method}_{slug}_try{attempt}_{uuid.uuid4().hex[:8]}"
+
+
+async def download_url_to_file_3d(
+    url: str,
+    file_format: str,
+    *,
+    task_id: str | None = None,
+    timeout: float | None = None,
+    max_retries: int = 5,
+    cls: type[COMFY_IO.ComfyNode] = None,
+) -> Types.File3D:
+    """Downloads a 3D model file from a URL into memory as BytesIO.
+
+    If task_id is provided, also writes the file to disk in the output directory
+    for backward compatibility with the old save-to-disk behavior.
+    """
+    file_format = file_format.lstrip(".").lower()
+    data = BytesIO()
+    await download_url_to_bytesio(
+        url,
+        data,
+        timeout=timeout,
+        max_retries=max_retries,
+        cls=cls,
+    )
+
+    if task_id is not None:
+        # This is only for backward compatability with current behavior when every 3D node is output node
+        # All new API nodes should not use "task_id" and instead users should use "SaveGLB" node to save results
+        output_dir = Path(get_output_directory())
+        output_path = output_dir / f"{task_id}.{file_format}"
+        output_path.write_bytes(data.getvalue())
+        data.seek(0)
+
+    return Types.File3D(source=data, file_format=file_format)

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

@@ -94,7 +94,7 @@ async def upload_image_to_comfyapi(
     *,
     mime_type: str | None = None,
     wait_label: str | None = "Uploading",
-    total_pixels: int = 2048 * 2048,
+    total_pixels: int | None = 2048 * 2048,
 ) -> str:
     """Uploads a single image to ComfyUI API and returns its download URL."""
     return (
@@ -255,17 +255,14 @@ 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 +308,27 @@ 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_execution/jobs.py

@@ -20,10 +20,60 @@ class JobStatus:
 
 
 # Media types that can be previewed in the frontend
-PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio'})
+PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio', '3d'})
 
 # 3D file extensions for preview fallback (no dedicated media_type exists)
-THREE_D_EXTENSIONS = frozenset({'.obj', '.fbx', '.gltf', '.glb'})
+THREE_D_EXTENSIONS = frozenset({'.obj', '.fbx', '.gltf', '.glb', '.usdz'})
+
+
+def has_3d_extension(filename: str) -> bool:
+    lower = filename.lower()
+    return any(lower.endswith(ext) for ext in THREE_D_EXTENSIONS)
+
+
+def normalize_output_item(item):
+    """Normalize a single output list item for the jobs API.
+
+    Returns the normalized item, or None to exclude it.
+    String items with 3D extensions become {filename, type, subfolder} dicts.
+    """
+    if item is None:
+        return None
+    if isinstance(item, str):
+        if has_3d_extension(item):
+            return {'filename': item, 'type': 'output', 'subfolder': '', 'mediaType': '3d'}
+        return None
+    if isinstance(item, dict):
+        return item
+    return None
+
+
+def normalize_outputs(outputs: dict) -> dict:
+    """Normalize raw node outputs for the jobs API.
+
+    Transforms string 3D filenames into file output dicts and removes
+    None items. All other items (non-3D strings, dicts, etc.) are
+    preserved as-is.
+    """
+    normalized = {}
+    for node_id, node_outputs in outputs.items():
+        if not isinstance(node_outputs, dict):
+            normalized[node_id] = node_outputs
+            continue
+        normalized_node = {}
+        for media_type, items in node_outputs.items():
+            if media_type == 'animated' or not isinstance(items, list):
+                normalized_node[media_type] = items
+                continue
+            normalized_items = []
+            for item in items:
+                if item is None:
+                    continue
+                norm = normalize_output_item(item)
+                normalized_items.append(norm if norm is not None else item)
+            normalized_node[media_type] = normalized_items
+        normalized[node_id] = normalized_node
+    return normalized
 
 
 def _extract_job_metadata(extra_data: dict) -> tuple[Optional[int], Optional[str]]:
@@ -45,9 +95,9 @@ def is_previewable(media_type: str, item: dict) -> bool:
     Maintains backwards compatibility with existing logic.
 
     Priority:
-    1. media_type is 'images', 'video', or 'audio'
+    1. media_type is 'images', 'video', 'audio', or '3d'
     2. format field starts with 'video/' or 'audio/'
-    3. filename has a 3D extension (.obj, .fbx, .gltf, .glb)
+    3. filename has a 3D extension (.obj, .fbx, .gltf, .glb, .usdz)
     """
     if media_type in PREVIEWABLE_MEDIA_TYPES:
         return True
@@ -139,7 +189,7 @@ def normalize_history_item(prompt_id: str, history_item: dict, include_outputs:
     })
 
     if include_outputs:
-        job['outputs'] = outputs
+        job['outputs'] = normalize_outputs(outputs)
         job['execution_status'] = status_info
         job['workflow'] = {
             'prompt': prompt,
@@ -171,18 +221,23 @@ def get_outputs_summary(outputs: dict) -> tuple[int, Optional[dict]]:
                 continue
 
             for item in items:
-                count += 1
-
-                if not isinstance(item, dict):
+                normalized = normalize_output_item(item)
+                if normalized is None:
                     continue
 
-                if preview_output is None and is_previewable(media_type, item):
+                count += 1
+
+                if preview_output is not None:
+                    continue
+
+                if isinstance(normalized, dict) and is_previewable(media_type, normalized):
                     enriched = {
-                        **item,
+                        **normalized,
                         'nodeId': node_id,
-                        'mediaType': media_type
                     }
-                    if item.get('type') == 'output':
+                    if 'mediaType' not in normalized:
+                        enriched['mediaType'] = media_type
+                    if normalized.get('type') == 'output':
                         preview_output = enriched
                     elif fallback_preview is None:
                         fallback_preview = enriched

comfy_extras/nodes_ace.py

@@ -28,12 +28,45 @@ class TextEncodeAceStepAudio(io.ComfyNode):
         conditioning = node_helpers.conditioning_set_values(conditioning, {"lyrics_strength": lyrics_strength})
         return io.NodeOutput(conditioning)
 
+class TextEncodeAceStepAudio15(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="TextEncodeAceStepAudio1.5",
+            category="conditioning",
+            inputs=[
+                io.Clip.Input("clip"),
+                io.String.Input("tags", multiline=True, dynamic_prompts=True),
+                io.String.Input("lyrics", multiline=True, dynamic_prompts=True),
+                io.Int.Input("seed", default=0, min=0, max=0xffffffffffffffff, control_after_generate=True),
+                io.Int.Input("bpm", default=120, min=10, max=300),
+                io.Float.Input("duration", default=120.0, min=0.0, max=2000.0, step=0.1),
+                io.Combo.Input("timesignature", options=['2', '3', '4', '6']),
+                io.Combo.Input("language", options=["en", "ja", "zh", "es", "de", "fr", "pt", "ru", "it", "nl", "pl", "tr", "vi", "cs", "fa", "id", "ko", "uk", "hu", "ar", "sv", "ro", "el"]),
+                io.Combo.Input("keyscale", options=[f"{root} {quality}" for quality in ["major", "minor"] for root in ["C", "C#", "Db", "D", "D#", "Eb", "E", "F", "F#", "Gb", "G", "G#", "Ab", "A", "A#", "Bb", "B"]]),
+                io.Boolean.Input("generate_audio_codes", default=True, tooltip="Enable the LLM that generates audio codes. This can be slow but will increase the quality of the generated audio. Turn this off if you are giving the model an audio reference.", advanced=True),
+                io.Float.Input("cfg_scale", default=2.0, min=0.0, max=100.0, step=0.1, advanced=True),
+                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, 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)
+
 
 class EmptyAceStepLatentAudio(io.ComfyNode):
     @classmethod
     def define_schema(cls):
         return io.Schema(
             node_id="EmptyAceStepLatentAudio",
+            display_name="Empty Ace Step 1.0 Latent Audio",
             category="latent/audio",
             inputs=[
                 io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.1),
@@ -51,12 +84,61 @@ class EmptyAceStepLatentAudio(io.ComfyNode):
         return io.NodeOutput({"samples": latent, "type": "audio"})
 
 
+class EmptyAceStep15LatentAudio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="EmptyAceStep1.5LatentAudio",
+            display_name="Empty Ace Step 1.5 Latent Audio",
+            category="latent/audio",
+            inputs=[
+                io.Float.Input("seconds", default=120.0, min=1.0, max=1000.0, step=0.01),
+                io.Int.Input(
+                    "batch_size", default=1, min=1, max=4096, tooltip="The number of latent images in the batch."
+                ),
+            ],
+            outputs=[io.Latent.Output()],
+        )
+
+    @classmethod
+    def execute(cls, seconds, batch_size) -> io.NodeOutput:
+        length = round((seconds * 48000 / 1920))
+        latent = torch.zeros([batch_size, 64, length], device=comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": latent, "type": "audio"})
+
+class ReferenceAudio(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="ReferenceTimbreAudio",
+            display_name="Reference Audio",
+            category="advanced/conditioning/audio",
+            is_experimental=True,
+            description="This node sets the reference audio for ace step 1.5",
+            inputs=[
+                io.Conditioning.Input("conditioning"),
+                io.Latent.Input("latent", optional=True),
+            ],
+            outputs=[
+                io.Conditioning.Output(),
+            ]
+        )
+
+    @classmethod
+    def execute(cls, conditioning, latent=None) -> io.NodeOutput:
+        if latent is not None:
+            conditioning = node_helpers.conditioning_set_values(conditioning, {"reference_audio_timbre_latents": [latent["samples"]]}, append=True)
+        return io.NodeOutput(conditioning)
+
 class AceExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
         return [
             TextEncodeAceStepAudio,
             EmptyAceStepLatentAudio,
+            TextEncodeAceStepAudio15,
+            EmptyAceStep15LatentAudio,
+            ReferenceAudio,
         ]
 
 async def comfy_entrypoint() -> AceExtension:

comfy_extras/nodes_audio.py

@@ -82,17 +82,31 @@ class VAEEncodeAudio(IO.ComfyNode):
     @classmethod
     def execute(cls, vae, audio) -> IO.NodeOutput:
         sample_rate = audio["sample_rate"]
-        if 44100 != sample_rate:
-            waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, 44100)
+        vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+        if vae_sample_rate != sample_rate:
+            waveform = torchaudio.functional.resample(audio["waveform"], sample_rate, vae_sample_rate)
         else:
             waveform = audio["waveform"]
 
         t = vae.encode(waveform.movedim(1, -1))
-        return IO.NodeOutput({"samples":t})
+        return IO.NodeOutput({"samples": t})
 
     encode = execute  # TODO: remove
 
 
+def vae_decode_audio(vae, samples, tile=None, overlap=None):
+    if tile is not None:
+        audio = vae.decode_tiled(samples["samples"], tile_y=tile, overlap=overlap).movedim(-1, 1)
+    else:
+        audio = vae.decode(samples["samples"]).movedim(-1, 1)
+
+    std = torch.std(audio, dim=[1, 2], keepdim=True) * 5.0
+    std[std < 1.0] = 1.0
+    audio /= std
+    vae_sample_rate = getattr(vae, "audio_sample_rate", 44100)
+    return {"waveform": audio, "sample_rate": vae_sample_rate if "sample_rate" not in samples else samples["sample_rate"]}
+
+
 class VAEDecodeAudio(IO.ComfyNode):
     @classmethod
     def define_schema(cls):
@@ -110,15 +124,33 @@ class VAEDecodeAudio(IO.ComfyNode):
 
     @classmethod
     def execute(cls, vae, samples) -> IO.NodeOutput:
-        audio = vae.decode(samples["samples"]).movedim(-1, 1)
-        std = torch.std(audio, dim=[1,2], keepdim=True) * 5.0
-        std[std < 1.0] = 1.0
-        audio /= std
-        return IO.NodeOutput({"waveform": audio, "sample_rate": 44100 if "sample_rate" not in samples else samples["sample_rate"]})
+        return IO.NodeOutput(vae_decode_audio(vae, samples))
 
     decode = execute  # TODO: remove
 
 
+class VAEDecodeAudioTiled(IO.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return IO.Schema(
+            node_id="VAEDecodeAudioTiled",
+            search_aliases=["latent to audio"],
+            display_name="VAE Decode Audio (Tiled)",
+            category="latent/audio",
+            inputs=[
+                IO.Latent.Input("samples"),
+                IO.Vae.Input("vae"),
+                IO.Int.Input("tile_size", default=512, min=32, max=8192, step=8),
+                IO.Int.Input("overlap", default=64, min=0, max=1024, step=8),
+            ],
+            outputs=[IO.Audio.Output()],
+        )
+
+    @classmethod
+    def execute(cls, vae, samples, tile_size, overlap) -> IO.NodeOutput:
+        return IO.NodeOutput(vae_decode_audio(vae, samples, tile_size, overlap))
+
+
 class SaveAudio(IO.ComfyNode):
     @classmethod
     def define_schema(cls):
@@ -673,6 +705,7 @@ class AudioExtension(ComfyExtension):
             EmptyLatentAudio,
             VAEEncodeAudio,
             VAEDecodeAudio,
+            VAEDecodeAudioTiled,
             SaveAudio,
             SaveAudioMP3,
             SaveAudioOpus,

comfy_extras/nodes_custom_sampler.py

@@ -622,6 +622,7 @@ class SamplerSASolver(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerSASolver",
+            search_aliases=["sde"],
             category="sampling/custom_sampling/samplers",
             inputs=[
                 io.Model.Input("model"),
@@ -666,6 +667,7 @@ class SamplerSEEDS2(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="SamplerSEEDS2",
+            search_aliases=["sde", "exp heun"],
             category="sampling/custom_sampling/samplers",
             inputs=[
                 io.Combo.Input("solver_type", options=["phi_1", "phi_2"]),

comfy_extras/nodes_easycache.py

@@ -9,6 +9,14 @@ if TYPE_CHECKING:
     from uuid import UUID
 
 
+def _extract_tensor(data, output_channels):
+    """Extract tensor from data, handling both single tensors and lists."""
+    if isinstance(data, list):
+        # LTX2 AV tensors: [video, audio]
+        return data[0][:, :output_channels], data[1][:, :output_channels]
+    return data[:, :output_channels], None
+
+
 def easycache_forward_wrapper(executor, *args, **kwargs):
     # get values from args
     transformer_options: dict[str] = args[-1]
@@ -17,7 +25,7 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
         if not transformer_options:
             transformer_options = args[-2]
     easycache: EasyCacheHolder = transformer_options["easycache"]
-    x: torch.Tensor = args[0][:, :easycache.output_channels]
+    x, ax = _extract_tensor(args[0], easycache.output_channels)
     sigmas = transformer_options["sigmas"]
     uuids = transformer_options["uuids"]
     if sigmas is not None and easycache.is_past_end_timestep(sigmas):
@@ -35,7 +43,11 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
         if easycache.skip_current_step and can_apply_cache_diff:
             if easycache.verbose:
                 logging.info(f"EasyCache [verbose] - was marked to skip this step by {easycache.first_cond_uuid}. Present uuids: {uuids}")
-            return easycache.apply_cache_diff(x, uuids)
+            result = easycache.apply_cache_diff(x, uuids)
+            if ax is not None:
+                result_audio = easycache.apply_cache_diff(ax, uuids, is_audio=True)
+                return [result, result_audio]
+            return result
         if easycache.initial_step:
             easycache.first_cond_uuid = uuids[0]
             has_first_cond_uuid = easycache.has_first_cond_uuid(uuids)
@@ -51,13 +63,18 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
                         logging.info(f"EasyCache [verbose] - skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
                     # other conds should also skip this step, and instead use their cached values
                     easycache.skip_current_step = True
-                    return easycache.apply_cache_diff(x, uuids)
+                    result = easycache.apply_cache_diff(x, uuids)
+                    if ax is not None:
+                        result_audio = easycache.apply_cache_diff(ax, uuids, is_audio=True)
+                        return [result, result_audio]
+                    return result
                 else:
                     if easycache.verbose:
                         logging.info(f"EasyCache [verbose] - NOT skipping step; cumulative_change_rate: {easycache.cumulative_change_rate}, reuse_threshold: {easycache.reuse_threshold}")
                     easycache.cumulative_change_rate = 0.0
 
-    output: torch.Tensor = executor(*args, **kwargs)
+    full_output: torch.Tensor = executor(*args, **kwargs)
+    output, audio_output = _extract_tensor(full_output, easycache.output_channels)
     if has_first_cond_uuid and easycache.has_output_prev_norm():
         output_change = (easycache.subsample(output, uuids, clone=False) - easycache.output_prev_subsampled).flatten().abs().mean()
         if easycache.verbose:
@@ -74,13 +91,15 @@ def easycache_forward_wrapper(executor, *args, **kwargs):
             logging.info(f"EasyCache [verbose] - output_change_rate: {output_change_rate}")
     # TODO: allow cache_diff to be offloaded
     easycache.update_cache_diff(output, next_x_prev, uuids)
+    if audio_output is not None:
+        easycache.update_cache_diff(audio_output, ax, uuids, is_audio=True)
     if has_first_cond_uuid:
         easycache.x_prev_subsampled = easycache.subsample(next_x_prev, uuids)
         easycache.output_prev_subsampled = easycache.subsample(output, uuids)
         easycache.output_prev_norm = output.flatten().abs().mean()
         if easycache.verbose:
             logging.info(f"EasyCache [verbose] - x_prev_subsampled: {easycache.x_prev_subsampled.shape}")
-    return output
+    return full_output
 
 def lazycache_predict_noise_wrapper(executor, *args, **kwargs):
     # get values from args
@@ -89,8 +108,8 @@ def lazycache_predict_noise_wrapper(executor, *args, **kwargs):
     easycache: LazyCacheHolder = model_options["transformer_options"]["easycache"]
     if easycache.is_past_end_timestep(timestep):
         return executor(*args, **kwargs)
-    # prepare next x_prev
     x: torch.Tensor = args[0][:, :easycache.output_channels]
+    # prepare next x_prev
     next_x_prev = x
     input_change = None
     do_easycache = easycache.should_do_easycache(timestep)
@@ -197,6 +216,7 @@ class EasyCacheHolder:
         self.output_prev_subsampled: torch.Tensor = None
         self.output_prev_norm: torch.Tensor = None
         self.uuid_cache_diffs: dict[UUID, torch.Tensor] = {}
+        self.uuid_cache_diffs_audio: dict[UUID, torch.Tensor] = {}
         self.output_change_rates = []
         self.approx_output_change_rates = []
         self.total_steps_skipped = 0
@@ -245,20 +265,21 @@ class EasyCacheHolder:
     def can_apply_cache_diff(self, uuids: list[UUID]) -> bool:
         return all(uuid in self.uuid_cache_diffs for uuid in uuids)
 
-    def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID]):
-        if self.first_cond_uuid in uuids:
+    def apply_cache_diff(self, x: torch.Tensor, uuids: list[UUID], is_audio: bool = False):
+        if self.first_cond_uuid in uuids and not is_audio:
             self.total_steps_skipped += 1
+        cache_diffs = self.uuid_cache_diffs_audio if is_audio else self.uuid_cache_diffs
         batch_offset = x.shape[0] // len(uuids)
         for i, uuid in enumerate(uuids):
             # slice out only what is relevant to this cond
             batch_slice = [slice(i*batch_offset,(i+1)*batch_offset)]
             # if cached dims don't match x dims, cut off excess and hope for the best (cosmos world2video)
-            if x.shape[1:] != self.uuid_cache_diffs[uuid].shape[1:]:
+            if x.shape[1:] != cache_diffs[uuid].shape[1:]:
                 if not self.allow_mismatch:
                     raise ValueError(f"Cached dims {self.uuid_cache_diffs[uuid].shape} don't match x dims {x.shape} - this is no good")
                 slicing = []
                 skip_this_dim = True
-                for dim_u, dim_x in zip(self.uuid_cache_diffs[uuid].shape, x.shape):
+                for dim_u, dim_x in zip(cache_diffs[uuid].shape, x.shape):
                     if skip_this_dim:
                         skip_this_dim = False
                         continue
@@ -270,10 +291,11 @@ class EasyCacheHolder:
                     else:
                         slicing.append(slice(None))
                 batch_slice = batch_slice + slicing
-            x[tuple(batch_slice)] += self.uuid_cache_diffs[uuid].to(x.device)
+            x[tuple(batch_slice)] += cache_diffs[uuid].to(x.device)
         return x
 
-    def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID]):
+    def update_cache_diff(self, output: torch.Tensor, x: torch.Tensor, uuids: list[UUID], is_audio: bool = False):
+        cache_diffs = self.uuid_cache_diffs_audio if is_audio else self.uuid_cache_diffs
         # if output dims don't match x dims, cut off excess and hope for the best (cosmos world2video)
         if output.shape[1:] != x.shape[1:]:
             if not self.allow_mismatch:
@@ -293,7 +315,7 @@ class EasyCacheHolder:
         diff = output - x
         batch_offset = diff.shape[0] // len(uuids)
         for i, uuid in enumerate(uuids):
-            self.uuid_cache_diffs[uuid] = diff[i*batch_offset:(i+1)*batch_offset, ...]
+            cache_diffs[uuid] = diff[i*batch_offset:(i+1)*batch_offset, ...]
 
     def has_first_cond_uuid(self, uuids: list[UUID]) -> bool:
         return self.first_cond_uuid in uuids
@@ -324,6 +346,8 @@ class EasyCacheHolder:
         self.output_prev_norm = None
         del self.uuid_cache_diffs
         self.uuid_cache_diffs = {}
+        del self.uuid_cache_diffs_audio
+        self.uuid_cache_diffs_audio = {}
         self.total_steps_skipped = 0
         self.state_metadata = None
         return self

comfy_extras/nodes_glsl.py

@@ -0,0 +1,897 @@
+import os
+import sys
+import re
+import logging
+import ctypes.util
+import importlib.util
+from typing import TypedDict
+
+import numpy as np
+import torch
+
+import nodes
+from comfy_api.latest import ComfyExtension, io, ui
+from typing_extensions import override
+from utils.install_util import get_missing_requirements_message
+
+logger = logging.getLogger(__name__)
+
+
+def _check_opengl_availability():
+    """Early check for OpenGL availability. Raises RuntimeError if unlikely to work."""
+    logger.debug("_check_opengl_availability: starting")
+    missing = []
+
+    # Check Python packages (using find_spec to avoid importing)
+    logger.debug("_check_opengl_availability: checking for glfw package")
+    if importlib.util.find_spec("glfw") is None:
+        missing.append("glfw")
+
+    logger.debug("_check_opengl_availability: checking for OpenGL package")
+    if importlib.util.find_spec("OpenGL") is None:
+        missing.append("PyOpenGL")
+
+    if missing:
+        raise RuntimeError(
+            f"OpenGL dependencies not available.\n{get_missing_requirements_message()}\n"
+        )
+
+    # On Linux without display, check if headless backends are available
+    logger.debug(f"_check_opengl_availability: platform={sys.platform}")
+    if sys.platform.startswith("linux"):
+        has_display = os.environ.get("DISPLAY") or os.environ.get("WAYLAND_DISPLAY")
+        logger.debug(f"_check_opengl_availability: has_display={bool(has_display)}")
+        if not has_display:
+            # Check for EGL or OSMesa libraries
+            logger.debug("_check_opengl_availability: checking for EGL library")
+            has_egl = ctypes.util.find_library("EGL")
+            logger.debug("_check_opengl_availability: checking for OSMesa library")
+            has_osmesa = ctypes.util.find_library("OSMesa")
+
+            # Error disabled for CI as it fails this check
+            # if not has_egl and not has_osmesa:
+            #     raise RuntimeError(
+            #         "GLSL Shader node: No display and no headless backend (EGL/OSMesa) found.\n"
+            #         "See error below for installation instructions."
+            #     )
+            logger.debug(f"Headless mode: EGL={'yes' if has_egl else 'no'}, OSMesa={'yes' if has_osmesa else 'no'}")
+
+    logger.debug("_check_opengl_availability: completed")
+
+
+# Run early check at import time
+logger.debug("nodes_glsl: running _check_opengl_availability at import time")
+_check_opengl_availability()
+
+# OpenGL modules - initialized lazily when context is created
+gl = None
+glfw = None
+EGL = None
+
+
+def _import_opengl():
+    """Import OpenGL module. Called after context is created."""
+    global gl
+    if gl is None:
+        logger.debug("_import_opengl: importing OpenGL.GL")
+        import OpenGL.GL as _gl
+        gl = _gl
+        logger.debug("_import_opengl: import completed")
+    return gl
+
+
+class SizeModeInput(TypedDict):
+    size_mode: str
+    width: int
+    height: int
+
+
+MAX_IMAGES = 5      # u_image0-4
+MAX_UNIFORMS = 5    # u_float0-4, u_int0-4
+MAX_OUTPUTS = 4     # fragColor0-3 (MRT)
+
+# Vertex shader using gl_VertexID trick - no VBO needed.
+# Draws a single triangle that covers the entire screen:
+#
+#     (-1,3)
+#       /|
+#      / |  <- visible area is the unit square from (-1,-1) to (1,1)
+#     /  |     parts outside get clipped away
+# (-1,-1)---(3,-1)
+#
+# v_texCoord is computed from clip space: * 0.5 + 0.5 maps (-1,1) -> (0,1)
+VERTEX_SHADER = """#version 330 core
+out vec2 v_texCoord;
+void main() {
+    vec2 verts[3] = vec2[](vec2(-1, -1), vec2(3, -1), vec2(-1, 3));
+    v_texCoord = verts[gl_VertexID] * 0.5 + 0.5;
+    gl_Position = vec4(verts[gl_VertexID], 0, 1);
+}
+"""
+
+DEFAULT_FRAGMENT_SHADER = """#version 300 es
+precision highp float;
+
+uniform sampler2D u_image0;
+uniform vec2 u_resolution;
+
+in vec2 v_texCoord;
+layout(location = 0) out vec4 fragColor0;
+
+void main() {
+    fragColor0 = texture(u_image0, v_texCoord);
+}
+"""
+
+
+def _convert_es_to_desktop(source: str) -> str:
+    """Convert GLSL ES (WebGL) shader source to desktop GLSL 330 core."""
+    # Remove any existing #version directive
+    source = re.sub(r"#version\s+\d+(\s+es)?\s*\n?", "", source, flags=re.IGNORECASE)
+    # Remove precision qualifiers (not needed in desktop GLSL)
+    source = re.sub(r"precision\s+(lowp|mediump|highp)\s+\w+\s*;\s*\n?", "", source)
+    # Prepend desktop GLSL version
+    return "#version 330 core\n" + source
+
+
+def _detect_output_count(source: str) -> int:
+    """Detect how many fragColor outputs are used in the shader.
+
+    Returns the count of outputs needed (1 to MAX_OUTPUTS).
+    """
+    matches = re.findall(r"fragColor(\d+)", source)
+    if not matches:
+        return 1  # Default to 1 output if none found
+    max_index = max(int(m) for m in matches)
+    return min(max_index + 1, MAX_OUTPUTS)
+
+
+def _detect_pass_count(source: str) -> int:
+    """Detect multi-pass rendering from #pragma passes N directive.
+
+    Returns the number of passes (1 if not specified).
+    """
+    match = re.search(r'#pragma\s+passes\s+(\d+)', source)
+    if match:
+        return max(1, int(match.group(1)))
+    return 1
+
+
+def _init_glfw():
+    """Initialize GLFW. Returns (window, glfw_module). Raises RuntimeError on failure."""
+    logger.debug("_init_glfw: starting")
+    # On macOS, glfw.init() must be called from main thread or it hangs forever
+    if sys.platform == "darwin":
+        logger.debug("_init_glfw: skipping on macOS")
+        raise RuntimeError("GLFW backend not supported on macOS")
+
+    logger.debug("_init_glfw: importing glfw module")
+    import glfw as _glfw
+
+    logger.debug("_init_glfw: calling glfw.init()")
+    if not _glfw.init():
+        raise RuntimeError("glfw.init() failed")
+
+    try:
+        logger.debug("_init_glfw: setting window hints")
+        _glfw.window_hint(_glfw.VISIBLE, _glfw.FALSE)
+        _glfw.window_hint(_glfw.CONTEXT_VERSION_MAJOR, 3)
+        _glfw.window_hint(_glfw.CONTEXT_VERSION_MINOR, 3)
+        _glfw.window_hint(_glfw.OPENGL_PROFILE, _glfw.OPENGL_CORE_PROFILE)
+
+        logger.debug("_init_glfw: calling create_window()")
+        window = _glfw.create_window(64, 64, "ComfyUI GLSL", None, None)
+        if not window:
+            raise RuntimeError("glfw.create_window() failed")
+
+        logger.debug("_init_glfw: calling make_context_current()")
+        _glfw.make_context_current(window)
+        logger.debug("_init_glfw: completed successfully")
+        return window, _glfw
+    except Exception:
+        logger.debug("_init_glfw: failed, terminating glfw")
+        _glfw.terminate()
+        raise
+
+
+def _init_egl():
+    """Initialize EGL for headless rendering. Returns (display, context, surface, EGL_module). Raises RuntimeError on failure."""
+    logger.debug("_init_egl: starting")
+    from OpenGL import EGL as _EGL
+    from OpenGL.EGL import (
+        eglGetDisplay, eglInitialize, eglChooseConfig, eglCreateContext,
+        eglMakeCurrent, eglCreatePbufferSurface, eglBindAPI,
+        eglTerminate, eglDestroyContext, eglDestroySurface,
+        EGL_DEFAULT_DISPLAY, EGL_NO_CONTEXT, EGL_NONE,
+        EGL_SURFACE_TYPE, EGL_PBUFFER_BIT, EGL_RENDERABLE_TYPE, EGL_OPENGL_BIT,
+        EGL_RED_SIZE, EGL_GREEN_SIZE, EGL_BLUE_SIZE, EGL_ALPHA_SIZE, EGL_DEPTH_SIZE,
+        EGL_WIDTH, EGL_HEIGHT, EGL_OPENGL_API,
+    )
+    logger.debug("_init_egl: imports completed")
+
+    display = None
+    context = None
+    surface = None
+
+    try:
+        logger.debug("_init_egl: calling eglGetDisplay()")
+        display = eglGetDisplay(EGL_DEFAULT_DISPLAY)
+        if display == _EGL.EGL_NO_DISPLAY:
+            raise RuntimeError("eglGetDisplay() failed")
+
+        logger.debug("_init_egl: calling eglInitialize()")
+        major, minor = _EGL.EGLint(), _EGL.EGLint()
+        if not eglInitialize(display, major, minor):
+            display = None  # Not initialized, don't terminate
+            raise RuntimeError("eglInitialize() failed")
+        logger.debug(f"_init_egl: EGL version {major.value}.{minor.value}")
+
+        config_attribs = [
+            EGL_SURFACE_TYPE, EGL_PBUFFER_BIT,
+            EGL_RENDERABLE_TYPE, EGL_OPENGL_BIT,
+            EGL_RED_SIZE, 8, EGL_GREEN_SIZE, 8, EGL_BLUE_SIZE, 8, EGL_ALPHA_SIZE, 8,
+            EGL_DEPTH_SIZE, 0, EGL_NONE
+        ]
+        configs = (_EGL.EGLConfig * 1)()
+        num_configs = _EGL.EGLint()
+        if not eglChooseConfig(display, config_attribs, configs, 1, num_configs) or num_configs.value == 0:
+            raise RuntimeError("eglChooseConfig() failed")
+        config = configs[0]
+        logger.debug(f"_init_egl: config chosen, num_configs={num_configs.value}")
+
+        if not eglBindAPI(EGL_OPENGL_API):
+            raise RuntimeError("eglBindAPI() failed")
+
+        logger.debug("_init_egl: calling eglCreateContext()")
+        context_attribs = [
+            _EGL.EGL_CONTEXT_MAJOR_VERSION, 3,
+            _EGL.EGL_CONTEXT_MINOR_VERSION, 3,
+            _EGL.EGL_CONTEXT_OPENGL_PROFILE_MASK, _EGL.EGL_CONTEXT_OPENGL_CORE_PROFILE_BIT,
+            EGL_NONE
+        ]
+        context = eglCreateContext(display, config, EGL_NO_CONTEXT, context_attribs)
+        if context == EGL_NO_CONTEXT:
+            raise RuntimeError("eglCreateContext() failed")
+
+        logger.debug("_init_egl: calling eglCreatePbufferSurface()")
+        pbuffer_attribs = [EGL_WIDTH, 64, EGL_HEIGHT, 64, EGL_NONE]
+        surface = eglCreatePbufferSurface(display, config, pbuffer_attribs)
+        if surface == _EGL.EGL_NO_SURFACE:
+            raise RuntimeError("eglCreatePbufferSurface() failed")
+
+        logger.debug("_init_egl: calling eglMakeCurrent()")
+        if not eglMakeCurrent(display, surface, surface, context):
+            raise RuntimeError("eglMakeCurrent() failed")
+
+        logger.debug("_init_egl: completed successfully")
+        return display, context, surface, _EGL
+
+    except Exception:
+        logger.debug("_init_egl: failed, cleaning up")
+        # Clean up any resources on failure
+        if surface is not None:
+            eglDestroySurface(display, surface)
+        if context is not None:
+            eglDestroyContext(display, context)
+        if display is not None:
+            eglTerminate(display)
+        raise
+
+
+def _init_osmesa():
+    """Initialize OSMesa for software rendering. Returns (context, buffer). Raises RuntimeError on failure."""
+    import ctypes
+
+    logger.debug("_init_osmesa: starting")
+    os.environ["PYOPENGL_PLATFORM"] = "osmesa"
+
+    logger.debug("_init_osmesa: importing OpenGL.osmesa")
+    from OpenGL import GL as _gl
+    from OpenGL.osmesa import (
+        OSMesaCreateContextExt, OSMesaMakeCurrent, OSMesaDestroyContext,
+        OSMESA_RGBA,
+    )
+    logger.debug("_init_osmesa: imports completed")
+
+    ctx = OSMesaCreateContextExt(OSMESA_RGBA, 24, 0, 0, None)
+    if not ctx:
+        raise RuntimeError("OSMesaCreateContextExt() failed")
+
+    width, height = 64, 64
+    buffer = (ctypes.c_ubyte * (width * height * 4))()
+
+    logger.debug("_init_osmesa: calling OSMesaMakeCurrent()")
+    if not OSMesaMakeCurrent(ctx, buffer, _gl.GL_UNSIGNED_BYTE, width, height):
+        OSMesaDestroyContext(ctx)
+        raise RuntimeError("OSMesaMakeCurrent() failed")
+
+    logger.debug("_init_osmesa: completed successfully")
+    return ctx, buffer
+
+
+class GLContext:
+    """Manages OpenGL context and resources for shader execution.
+
+    Tries backends in order: GLFW (desktop) → EGL (headless GPU) → OSMesa (software).
+    """
+
+    _instance = None
+    _initialized = False
+
+    def __new__(cls):
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+        return cls._instance
+
+    def __init__(self):
+        if GLContext._initialized:
+            logger.debug("GLContext.__init__: already initialized, skipping")
+            return
+
+        logger.debug("GLContext.__init__: starting initialization")
+
+        global glfw, EGL
+
+        import time
+        start = time.perf_counter()
+
+        self._backend = None
+        self._window = None
+        self._egl_display = None
+        self._egl_context = None
+        self._egl_surface = None
+        self._osmesa_ctx = None
+        self._osmesa_buffer = None
+        self._vao = None
+
+        # Try backends in order: GLFW → EGL → OSMesa
+        errors = []
+
+        logger.debug("GLContext.__init__: trying GLFW backend")
+        try:
+            self._window, glfw = _init_glfw()
+            self._backend = "glfw"
+            logger.debug("GLContext.__init__: GLFW backend succeeded")
+        except Exception as e:
+            logger.debug(f"GLContext.__init__: GLFW backend failed: {e}")
+            errors.append(("GLFW", e))
+
+        if self._backend is None:
+            logger.debug("GLContext.__init__: trying EGL backend")
+            try:
+                self._egl_display, self._egl_context, self._egl_surface, EGL = _init_egl()
+                self._backend = "egl"
+                logger.debug("GLContext.__init__: EGL backend succeeded")
+            except Exception as e:
+                logger.debug(f"GLContext.__init__: EGL backend failed: {e}")
+                errors.append(("EGL", e))
+
+        if self._backend is None:
+            logger.debug("GLContext.__init__: trying OSMesa backend")
+            try:
+                self._osmesa_ctx, self._osmesa_buffer = _init_osmesa()
+                self._backend = "osmesa"
+                logger.debug("GLContext.__init__: OSMesa backend succeeded")
+            except Exception as e:
+                logger.debug(f"GLContext.__init__: OSMesa backend failed: {e}")
+                errors.append(("OSMesa", e))
+
+        if self._backend is None:
+            if sys.platform == "win32":
+                platform_help = (
+                    "Windows: Ensure GPU drivers are installed and display is available.\n"
+                    "         CPU-only/headless mode is not supported on Windows."
+                )
+            elif sys.platform == "darwin":
+                platform_help = (
+                    "macOS: GLFW is not supported.\n"
+                    "  Install OSMesa via Homebrew: brew install mesa\n"
+                    "  Then: pip install PyOpenGL PyOpenGL-accelerate"
+                )
+            else:
+                platform_help = (
+                    "Linux: Install one of these backends:\n"
+                    "  Desktop:           sudo apt install libgl1-mesa-glx libglfw3\n"
+                    "  Headless with GPU: sudo apt install libegl1-mesa libgl1-mesa-dri\n"
+                    "  Headless (CPU):    sudo apt install libosmesa6"
+                )
+
+            error_details = "\n".join(f"  {name}: {err}" for name, err in errors)
+            raise RuntimeError(
+                f"Failed to create OpenGL context.\n\n"
+                f"Backend errors:\n{error_details}\n\n"
+                f"{platform_help}"
+            )
+
+        # Now import OpenGL.GL (after context is current)
+        logger.debug("GLContext.__init__: importing OpenGL.GL")
+        _import_opengl()
+
+        # Create VAO (required for core profile, but OSMesa may use compat profile)
+        logger.debug("GLContext.__init__: creating VAO")
+        try:
+            vao = gl.glGenVertexArrays(1)
+            gl.glBindVertexArray(vao)
+            self._vao = vao  # Only store after successful bind
+            logger.debug("GLContext.__init__: VAO created successfully")
+        except Exception as e:
+            logger.debug(f"GLContext.__init__: VAO creation failed (may be expected for OSMesa): {e}")
+            # OSMesa with older Mesa may not support VAOs
+            # Clean up if we created but couldn't bind
+            if vao:
+                try:
+                    gl.glDeleteVertexArrays(1, [vao])
+                except Exception:
+                    pass
+
+        elapsed = (time.perf_counter() - start) * 1000
+
+        # Log device info
+        renderer = gl.glGetString(gl.GL_RENDERER)
+        vendor = gl.glGetString(gl.GL_VENDOR)
+        version = gl.glGetString(gl.GL_VERSION)
+        renderer = renderer.decode() if renderer else "Unknown"
+        vendor = vendor.decode() if vendor else "Unknown"
+        version = version.decode() if version else "Unknown"
+
+        GLContext._initialized = True
+        logger.info(f"GLSL context initialized in {elapsed:.1f}ms ({self._backend}) - {renderer} ({vendor}), GL {version}")
+
+    def make_current(self):
+        if self._backend == "glfw":
+            glfw.make_context_current(self._window)
+        elif self._backend == "egl":
+            from OpenGL.EGL import eglMakeCurrent
+            eglMakeCurrent(self._egl_display, self._egl_surface, self._egl_surface, self._egl_context)
+        elif self._backend == "osmesa":
+            from OpenGL.osmesa import OSMesaMakeCurrent
+            OSMesaMakeCurrent(self._osmesa_ctx, self._osmesa_buffer, gl.GL_UNSIGNED_BYTE, 64, 64)
+
+        if self._vao is not None:
+            gl.glBindVertexArray(self._vao)
+
+
+def _compile_shader(source: str, shader_type: int) -> int:
+    """Compile a shader and return its ID."""
+    shader = gl.glCreateShader(shader_type)
+    gl.glShaderSource(shader, source)
+    gl.glCompileShader(shader)
+
+    if gl.glGetShaderiv(shader, gl.GL_COMPILE_STATUS) != gl.GL_TRUE:
+        error = gl.glGetShaderInfoLog(shader).decode()
+        gl.glDeleteShader(shader)
+        raise RuntimeError(f"Shader compilation failed:\n{error}")
+
+    return shader
+
+
+def _create_program(vertex_source: str, fragment_source: str) -> int:
+    """Create and link a shader program."""
+    vertex_shader = _compile_shader(vertex_source, gl.GL_VERTEX_SHADER)
+    try:
+        fragment_shader = _compile_shader(fragment_source, gl.GL_FRAGMENT_SHADER)
+    except RuntimeError:
+        gl.glDeleteShader(vertex_shader)
+        raise
+
+    program = gl.glCreateProgram()
+    gl.glAttachShader(program, vertex_shader)
+    gl.glAttachShader(program, fragment_shader)
+    gl.glLinkProgram(program)
+
+    gl.glDeleteShader(vertex_shader)
+    gl.glDeleteShader(fragment_shader)
+
+    if gl.glGetProgramiv(program, gl.GL_LINK_STATUS) != gl.GL_TRUE:
+        error = gl.glGetProgramInfoLog(program).decode()
+        gl.glDeleteProgram(program)
+        raise RuntimeError(f"Program linking failed:\n{error}")
+
+    return program
+
+
+def _render_shader_batch(
+    fragment_code: str,
+    width: int,
+    height: int,
+    image_batches: list[list[np.ndarray]],
+    floats: list[float],
+    ints: list[int],
+) -> list[list[np.ndarray]]:
+    """
+    Render a fragment shader for multiple batches efficiently.
+
+    Compiles shader once, reuses framebuffer/textures across batches.
+    Supports multi-pass rendering via #pragma passes N directive.
+
+    Args:
+        fragment_code: User's fragment shader code
+        width: Output width
+        height: Output height
+        image_batches: List of batches, each batch is a list of input images (H, W, C) float32 [0,1]
+        floats: List of float uniforms
+        ints: List of int uniforms
+
+    Returns:
+        List of batch outputs, each is a list of output images (H, W, 4) float32 [0,1]
+    """
+    import time
+    start_time = time.perf_counter()
+
+    if not image_batches:
+        return []
+
+    ctx = GLContext()
+    ctx.make_current()
+
+    # Convert from GLSL ES to desktop GLSL 330
+    fragment_source = _convert_es_to_desktop(fragment_code)
+
+    # Detect how many outputs the shader actually uses
+    num_outputs = _detect_output_count(fragment_code)
+
+    # Detect multi-pass rendering
+    num_passes = _detect_pass_count(fragment_code)
+
+    # Track resources for cleanup
+    program = None
+    fbo = None
+    output_textures = []
+    input_textures = []
+    ping_pong_textures = []
+    ping_pong_fbos = []
+
+    num_inputs = len(image_batches[0])
+
+    try:
+        # Compile shaders (once for all batches)
+        try:
+            program = _create_program(VERTEX_SHADER, fragment_source)
+        except RuntimeError:
+            logger.error(f"Fragment shader:\n{fragment_source}")
+            raise
+
+        gl.glUseProgram(program)
+
+        # Create framebuffer with only the needed color attachments
+        fbo = gl.glGenFramebuffers(1)
+        gl.glBindFramebuffer(gl.GL_FRAMEBUFFER, fbo)
+
+        draw_buffers = []
+        for i in range(num_outputs):
+            tex = gl.glGenTextures(1)
+            output_textures.append(tex)
+            gl.glBindTexture(gl.GL_TEXTURE_2D, tex)
+            gl.glTexImage2D(gl.GL_TEXTURE_2D, 0, gl.GL_RGBA32F, width, height, 0, gl.GL_RGBA, gl.GL_FLOAT, None)
+            gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_MIN_FILTER, gl.GL_LINEAR)
+            gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_MAG_FILTER, gl.GL_LINEAR)
+            gl.glFramebufferTexture2D(gl.GL_FRAMEBUFFER, gl.GL_COLOR_ATTACHMENT0 + i, gl.GL_TEXTURE_2D, tex, 0)
+            draw_buffers.append(gl.GL_COLOR_ATTACHMENT0 + i)
+
+        gl.glDrawBuffers(num_outputs, draw_buffers)
+
+        if gl.glCheckFramebufferStatus(gl.GL_FRAMEBUFFER) != gl.GL_FRAMEBUFFER_COMPLETE:
+            raise RuntimeError("Framebuffer is not complete")
+
+        # Create ping-pong resources for multi-pass rendering
+        if num_passes > 1:
+            for _ in range(2):
+                pp_tex = gl.glGenTextures(1)
+                ping_pong_textures.append(pp_tex)
+                gl.glBindTexture(gl.GL_TEXTURE_2D, pp_tex)
+                gl.glTexImage2D(gl.GL_TEXTURE_2D, 0, gl.GL_RGBA32F, width, height, 0, gl.GL_RGBA, gl.GL_FLOAT, None)
+                gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_MIN_FILTER, gl.GL_LINEAR)
+                gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_MAG_FILTER, gl.GL_LINEAR)
+                gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_WRAP_S, gl.GL_CLAMP_TO_EDGE)
+                gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_WRAP_T, gl.GL_CLAMP_TO_EDGE)
+
+                pp_fbo = gl.glGenFramebuffers(1)
+                ping_pong_fbos.append(pp_fbo)
+                gl.glBindFramebuffer(gl.GL_FRAMEBUFFER, pp_fbo)
+                gl.glFramebufferTexture2D(gl.GL_FRAMEBUFFER, gl.GL_COLOR_ATTACHMENT0, gl.GL_TEXTURE_2D, pp_tex, 0)
+                gl.glDrawBuffers(1, [gl.GL_COLOR_ATTACHMENT0])
+
+                if gl.glCheckFramebufferStatus(gl.GL_FRAMEBUFFER) != gl.GL_FRAMEBUFFER_COMPLETE:
+                    raise RuntimeError("Ping-pong framebuffer is not complete")
+
+        # Create input textures (reused for all batches)
+        for i in range(num_inputs):
+            tex = gl.glGenTextures(1)
+            input_textures.append(tex)
+            gl.glActiveTexture(gl.GL_TEXTURE0 + i)
+            gl.glBindTexture(gl.GL_TEXTURE_2D, tex)
+            gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_MIN_FILTER, gl.GL_LINEAR)
+            gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_MAG_FILTER, gl.GL_LINEAR)
+            gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_WRAP_S, gl.GL_CLAMP_TO_EDGE)
+            gl.glTexParameteri(gl.GL_TEXTURE_2D, gl.GL_TEXTURE_WRAP_T, gl.GL_CLAMP_TO_EDGE)
+
+            loc = gl.glGetUniformLocation(program, f"u_image{i}")
+            if loc >= 0:
+                gl.glUniform1i(loc, i)
+
+        # Set static uniforms (once for all batches)
+        loc = gl.glGetUniformLocation(program, "u_resolution")
+        if loc >= 0:
+            gl.glUniform2f(loc, float(width), float(height))
+
+        for i, v in enumerate(floats):
+            loc = gl.glGetUniformLocation(program, f"u_float{i}")
+            if loc >= 0:
+                gl.glUniform1f(loc, v)
+
+        for i, v in enumerate(ints):
+            loc = gl.glGetUniformLocation(program, f"u_int{i}")
+            if loc >= 0:
+                gl.glUniform1i(loc, v)
+
+        # Get u_pass uniform location for multi-pass
+        pass_loc = gl.glGetUniformLocation(program, "u_pass")
+
+        gl.glViewport(0, 0, width, height)
+        gl.glDisable(gl.GL_BLEND)  # Ensure no alpha blending - write output directly
+
+        # Process each batch
+        all_batch_outputs = []
+        for images in image_batches:
+            # Update input textures with this batch's images
+            for i, img in enumerate(images):
+                gl.glActiveTexture(gl.GL_TEXTURE0 + i)
+                gl.glBindTexture(gl.GL_TEXTURE_2D, input_textures[i])
+
+                # Flip vertically for GL coordinates, ensure RGBA
+                h, w, c = img.shape
+                if c == 3:
+                    img_upload = np.empty((h, w, 4), dtype=np.float32)
+                    img_upload[:, :, :3] = img[::-1, :, :]
+                    img_upload[:, :, 3] = 1.0
+                else:
+                    img_upload = np.ascontiguousarray(img[::-1, :, :])
+
+                gl.glTexImage2D(gl.GL_TEXTURE_2D, 0, gl.GL_RGBA32F, w, h, 0, gl.GL_RGBA, gl.GL_FLOAT, img_upload)
+
+            if num_passes == 1:
+                # Single pass - render directly to output FBO
+                gl.glBindFramebuffer(gl.GL_FRAMEBUFFER, fbo)
+                if pass_loc >= 0:
+                    gl.glUniform1i(pass_loc, 0)
+                gl.glClearColor(0, 0, 0, 0)
+                gl.glClear(gl.GL_COLOR_BUFFER_BIT)
+                gl.glDrawArrays(gl.GL_TRIANGLES, 0, 3)
+            else:
+                # Multi-pass rendering with ping-pong
+                for p in range(num_passes):
+                    is_last_pass = (p == num_passes - 1)
+
+                    # Set pass uniform
+                    if pass_loc >= 0:
+                        gl.glUniform1i(pass_loc, p)
+
+                    if is_last_pass:
+                        # Last pass renders to the main output FBO
+                        gl.glBindFramebuffer(gl.GL_FRAMEBUFFER, fbo)
+                    else:
+                        # Intermediate passes render to ping-pong FBO
+                        target_fbo = ping_pong_fbos[p % 2]
+                        gl.glBindFramebuffer(gl.GL_FRAMEBUFFER, target_fbo)
+
+                    # Set input texture for this pass
+                    gl.glActiveTexture(gl.GL_TEXTURE0)
+                    if p == 0:
+                        # First pass reads from original input
+                        gl.glBindTexture(gl.GL_TEXTURE_2D, input_textures[0])
+                    else:
+                        # Subsequent passes read from previous pass output
+                        source_tex = ping_pong_textures[(p - 1) % 2]
+                        gl.glBindTexture(gl.GL_TEXTURE_2D, source_tex)
+
+                    gl.glClearColor(0, 0, 0, 0)
+                    gl.glClear(gl.GL_COLOR_BUFFER_BIT)
+                    gl.glDrawArrays(gl.GL_TRIANGLES, 0, 3)
+
+            # Read back outputs for this batch
+            # (glGetTexImage is synchronous, implicitly waits for rendering)
+            batch_outputs = []
+            for tex in output_textures:
+                gl.glBindTexture(gl.GL_TEXTURE_2D, tex)
+                data = gl.glGetTexImage(gl.GL_TEXTURE_2D, 0, gl.GL_RGBA, gl.GL_FLOAT)
+                img = np.frombuffer(data, dtype=np.float32).reshape(height, width, 4)
+                batch_outputs.append(np.ascontiguousarray(img[::-1, :, :]))
+
+            # Pad with black images for unused outputs
+            black_img = np.zeros((height, width, 4), dtype=np.float32)
+            for _ in range(num_outputs, MAX_OUTPUTS):
+                batch_outputs.append(black_img)
+
+            all_batch_outputs.append(batch_outputs)
+
+        elapsed = (time.perf_counter() - start_time) * 1000
+        num_batches = len(image_batches)
+        pass_info = f", {num_passes} passes" if num_passes > 1 else ""
+        logger.info(f"GLSL shader executed in {elapsed:.1f}ms ({num_batches} batch{'es' if num_batches != 1 else ''}, {width}x{height}{pass_info})")
+
+        return all_batch_outputs
+
+    finally:
+        # Unbind before deleting
+        gl.glBindFramebuffer(gl.GL_FRAMEBUFFER, 0)
+        gl.glUseProgram(0)
+
+        if input_textures:
+            gl.glDeleteTextures(len(input_textures), input_textures)
+        if output_textures:
+            gl.glDeleteTextures(len(output_textures), output_textures)
+        if ping_pong_textures:
+            gl.glDeleteTextures(len(ping_pong_textures), ping_pong_textures)
+        if fbo is not None:
+            gl.glDeleteFramebuffers(1, [fbo])
+        for pp_fbo in ping_pong_fbos:
+            gl.glDeleteFramebuffers(1, [pp_fbo])
+        if program is not None:
+            gl.glDeleteProgram(program)
+
+class GLSLShader(io.ComfyNode):
+
+    @classmethod
+    def define_schema(cls) -> io.Schema:
+        image_template = io.Autogrow.TemplatePrefix(
+            io.Image.Input("image"),
+            prefix="image",
+            min=1,
+            max=MAX_IMAGES,
+        )
+
+        float_template = io.Autogrow.TemplatePrefix(
+            io.Float.Input("float", default=0.0),
+            prefix="u_float",
+            min=0,
+            max=MAX_UNIFORMS,
+        )
+
+        int_template = io.Autogrow.TemplatePrefix(
+            io.Int.Input("int", default=0),
+            prefix="u_int",
+            min=0,
+            max=MAX_UNIFORMS,
+        )
+
+        return io.Schema(
+            node_id="GLSLShader",
+            display_name="GLSL Shader",
+            category="image/shader",
+            description=(
+                f"Apply GLSL fragment shaders to images. "
+                f"Inputs: u_image0-{MAX_IMAGES-1} (sampler2D), u_resolution (vec2), "
+                f"u_float0-{MAX_UNIFORMS-1}, u_int0-{MAX_UNIFORMS-1}. "
+                f"Outputs: layout(location = 0-{MAX_OUTPUTS-1}) out vec4 fragColor0-{MAX_OUTPUTS-1}."
+            ),
+            inputs=[
+                io.String.Input(
+                    "fragment_shader",
+                    default=DEFAULT_FRAGMENT_SHADER,
+                    multiline=True,
+                    tooltip="GLSL fragment shader source code (GLSL ES 3.00 / WebGL 2.0 compatible)",
+                ),
+                io.DynamicCombo.Input(
+                    "size_mode",
+                    options=[
+                        io.DynamicCombo.Option("from_input", []),
+                        io.DynamicCombo.Option(
+                            "custom",
+                            [
+                                io.Int.Input(
+                                    "width",
+                                    default=512,
+                                    min=1,
+                                    max=nodes.MAX_RESOLUTION,
+                                ),
+                                io.Int.Input(
+                                    "height",
+                                    default=512,
+                                    min=1,
+                                    max=nodes.MAX_RESOLUTION,
+                                ),
+                            ],
+                        ),
+                    ],
+                    tooltip="Output size: 'from_input' uses first input image dimensions, 'custom' allows manual size",
+                ),
+                io.Autogrow.Input("images", template=image_template),
+                io.Autogrow.Input("floats", template=float_template),
+                io.Autogrow.Input("ints", template=int_template),
+            ],
+            outputs=[
+                io.Image.Output(display_name="IMAGE0"),
+                io.Image.Output(display_name="IMAGE1"),
+                io.Image.Output(display_name="IMAGE2"),
+                io.Image.Output(display_name="IMAGE3"),
+            ],
+        )
+
+    @classmethod
+    def execute(
+        cls,
+        fragment_shader: str,
+        size_mode: SizeModeInput,
+        images: io.Autogrow.Type,
+        floats: io.Autogrow.Type = None,
+        ints: io.Autogrow.Type = None,
+        **kwargs,
+    ) -> io.NodeOutput:
+        image_list = [v for v in images.values() if v is not None]
+        float_list = (
+            [v if v is not None else 0.0 for v in floats.values()] if floats else []
+        )
+        int_list = [v if v is not None else 0 for v in ints.values()] if ints else []
+
+        if not image_list:
+            raise ValueError("At least one input image is required")
+
+        # Determine output dimensions
+        if size_mode["size_mode"] == "custom":
+            out_width = size_mode["width"]
+            out_height = size_mode["height"]
+        else:
+            out_height, out_width = image_list[0].shape[1:3]
+
+        batch_size = image_list[0].shape[0]
+
+        # Prepare batches
+        image_batches = []
+        for batch_idx in range(batch_size):
+            batch_images = [img_tensor[batch_idx].cpu().numpy().astype(np.float32) for img_tensor in image_list]
+            image_batches.append(batch_images)
+
+        all_batch_outputs = _render_shader_batch(
+            fragment_shader,
+            out_width,
+            out_height,
+            image_batches,
+            float_list,
+            int_list,
+        )
+
+        # Collect outputs into tensors
+        all_outputs = [[] for _ in range(MAX_OUTPUTS)]
+        for batch_outputs in all_batch_outputs:
+            for i, out_img in enumerate(batch_outputs):
+                all_outputs[i].append(torch.from_numpy(out_img))
+
+        output_tensors = [torch.stack(all_outputs[i], dim=0) for i in range(MAX_OUTPUTS)]
+        return io.NodeOutput(
+            *output_tensors,
+            ui=cls._build_ui_output(image_list, output_tensors[0]),
+        )
+
+    @classmethod
+    def _build_ui_output(
+        cls, image_list: list[torch.Tensor], output_batch: torch.Tensor
+    ) -> dict[str, list]:
+        """Build UI output with input and output images for client-side shader execution."""
+        combined_inputs = torch.cat(image_list, dim=0)
+        input_images_ui = ui.ImageSaveHelper.save_images(
+            combined_inputs,
+            filename_prefix="GLSLShader_input",
+            folder_type=io.FolderType.temp,
+            cls=None,
+            compress_level=1,
+        )
+
+        output_images_ui = ui.ImageSaveHelper.save_images(
+            output_batch,
+            filename_prefix="GLSLShader_output",
+            folder_type=io.FolderType.temp,
+            cls=None,
+            compress_level=1,
+        )
+
+        return {"input_images": input_images_ui, "images": output_images_ui}
+
+
+class GLSLExtension(ComfyExtension):
+    @override
+    async def get_node_list(self) -> list[type[io.ComfyNode]]:
+        return [GLSLShader]
+
+
+async def comfy_entrypoint() -> GLSLExtension:
+    return GLSLExtension()

comfy_extras/nodes_hunyuan3d.py

@@ -618,18 +618,31 @@ class SaveGLB(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="SaveGLB",
+            display_name="Save 3D Model",
             search_aliases=["export 3d model", "save mesh"],
             category="3d",
             is_output_node=True,
             inputs=[
-                IO.Mesh.Input("mesh"),
+                IO.MultiType.Input(
+                    IO.Mesh.Input("mesh"),
+                    types=[
+                        IO.File3DGLB,
+                        IO.File3DGLTF,
+                        IO.File3DOBJ,
+                        IO.File3DFBX,
+                        IO.File3DSTL,
+                        IO.File3DUSDZ,
+                        IO.File3DAny,
+                    ],
+                    tooltip="Mesh or 3D file to save",
+                ),
                 IO.String.Input("filename_prefix", default="mesh/ComfyUI"),
             ],
             hidden=[IO.Hidden.prompt, IO.Hidden.extra_pnginfo]
         )
 
     @classmethod
-    def execute(cls, mesh, filename_prefix) -> IO.NodeOutput:
+    def execute(cls, mesh: Types.MESH | Types.File3D, filename_prefix: str) -> IO.NodeOutput:
         full_output_folder, filename, counter, subfolder, filename_prefix = folder_paths.get_save_image_path(filename_prefix, folder_paths.get_output_directory())
         results = []
 
@@ -641,15 +654,27 @@ class SaveGLB(IO.ComfyNode):
                 for x in cls.hidden.extra_pnginfo:
                     metadata[x] = json.dumps(cls.hidden.extra_pnginfo[x])
 
-        for i in range(mesh.vertices.shape[0]):
-            f = f"{filename}_{counter:05}_.glb"
-            save_glb(mesh.vertices[i], mesh.faces[i], os.path.join(full_output_folder, f), metadata)
+        if isinstance(mesh, Types.File3D):
+            # Handle File3D input - save BytesIO data to output folder
+            ext = mesh.format or "glb"
+            f = f"{filename}_{counter:05}_.{ext}"
+            mesh.save_to(os.path.join(full_output_folder, f))
             results.append({
                 "filename": f,
                 "subfolder": subfolder,
                 "type": "output"
             })
-            counter += 1
+        else:
+            # Handle Mesh input - save vertices and faces as GLB
+            for i in range(mesh.vertices.shape[0]):
+                f = f"{filename}_{counter:05}_.glb"
+                save_glb(mesh.vertices[i], mesh.faces[i], os.path.join(full_output_folder, f), metadata)
+                results.append({
+                    "filename": f,
+                    "subfolder": subfolder,
+                    "type": "output"
+                })
+                counter += 1
         return IO.NodeOutput(ui={"3d": results})