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.

.gitignore

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

app/node_replace_manager.py

@@ -1,105 +0,0 @@
-from __future__ import annotations
-
-from aiohttp import web
-
-from typing import TYPE_CHECKING, TypedDict
-if TYPE_CHECKING:
-    from comfy_api.latest._io_public import NodeReplace
-
-from comfy_execution.graph_utils import is_link
-import nodes
-
-class NodeStruct(TypedDict):
-    inputs: dict[str, str | int | float | bool | tuple[str, int]]
-    class_type: str
-    _meta: dict[str, str]
-
-def copy_node_struct(node_struct: NodeStruct, empty_inputs: bool = False) -> NodeStruct:
-    new_node_struct = node_struct.copy()
-    if empty_inputs:
-        new_node_struct["inputs"] = {}
-    else:
-        new_node_struct["inputs"] = node_struct["inputs"].copy()
-    new_node_struct["_meta"] = node_struct["_meta"].copy()
-    return new_node_struct
-
-
-class NodeReplaceManager:
-    """Manages node replacement registrations."""
-
-    def __init__(self):
-        self._replacements: dict[str, list[NodeReplace]] = {}
-
-    def register(self, node_replace: NodeReplace):
-        """Register a node replacement mapping."""
-        self._replacements.setdefault(node_replace.old_node_id, []).append(node_replace)
-
-    def get_replacement(self, old_node_id: str) -> list[NodeReplace] | None:
-        """Get replacements for an old node ID."""
-        return self._replacements.get(old_node_id)
-
-    def has_replacement(self, old_node_id: str) -> bool:
-        """Check if a replacement exists for an old node ID."""
-        return old_node_id in self._replacements
-
-    def apply_replacements(self, prompt: dict[str, NodeStruct]):
-        connections: dict[str, list[tuple[str, str, int]]] = {}
-        need_replacement: set[str] = set()
-        for node_number, node_struct in prompt.items():
-            class_type = node_struct["class_type"]
-            # need replacement if not in NODE_CLASS_MAPPINGS and has replacement
-            if class_type not in nodes.NODE_CLASS_MAPPINGS.keys() and self.has_replacement(class_type):
-                need_replacement.add(node_number)
-            # keep track of connections
-            for input_id, input_value in node_struct["inputs"].items():
-                if is_link(input_value):
-                    conn_number = input_value[0]
-                    connections.setdefault(conn_number, []).append((node_number, input_id, input_value[1]))
-        for node_number in need_replacement:
-            node_struct = prompt[node_number]
-            class_type = node_struct["class_type"]
-            replacements = self.get_replacement(class_type)
-            if replacements is None:
-                continue
-            # just use the first replacement
-            replacement = replacements[0]
-            new_node_id = replacement.new_node_id
-            # if replacement is not a valid node, skip trying to replace it as will only cause confusion
-            if new_node_id not in nodes.NODE_CLASS_MAPPINGS.keys():
-                continue
-            # first, replace node id (class_type)
-            new_node_struct = copy_node_struct(node_struct, empty_inputs=True)
-            new_node_struct["class_type"] = new_node_id
-            # TODO: consider replacing display_name in _meta as well for error reporting purposes; would need to query node schema
-            # second, replace inputs
-            if replacement.input_mapping is not None:
-                for input_map in replacement.input_mapping:
-                    if "set_value" in input_map:
-                        new_node_struct["inputs"][input_map["new_id"]] = input_map["set_value"]
-                    elif "old_id" in input_map:
-                        new_node_struct["inputs"][input_map["new_id"]] = node_struct["inputs"][input_map["old_id"]]
-            # finalize input replacement
-            prompt[node_number] = new_node_struct
-            # third, replace outputs
-            if replacement.output_mapping is not None:
-                # re-mapping outputs requires changing the input values of nodes that receive connections from this one
-                if node_number in connections:
-                    for conns in connections[node_number]:
-                        conn_node_number, conn_input_id, old_output_idx = conns
-                        for output_map in replacement.output_mapping:
-                            if output_map["old_idx"] == old_output_idx:
-                                new_output_idx = output_map["new_idx"]
-                                previous_input = prompt[conn_node_number]["inputs"][conn_input_id]
-                                previous_input[1] = new_output_idx
-
-    def as_dict(self):
-        """Serialize all replacements to dict."""
-        return {
-            k: [v.as_dict() for v in v_list]
-            for k, v_list in self._replacements.items()
-        }
-
-    def add_routes(self, routes):
-        @routes.get("/node_replacements")
-        async def get_node_replacements(request):
-            return web.json_response(self.as_dict())

blueprints/.glsl/Brightness_and_Contrast_1.frag

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

blueprints/.glsl/Chromatic_Aberration_16.frag

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

blueprints/.glsl/Color_Adjustment_15.frag

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

blueprints/.glsl/Edge-Preserving_Blur_128.frag

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

blueprints/.glsl/Film_Grain_15.frag

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

blueprints/.glsl/Glow_30.frag

@@ -0,0 +1,133 @@
+#version 300 es
+precision mediump float;
+
+uniform sampler2D u_image0;
+uniform vec2 u_resolution;
+uniform int u_int0;      // Blend mode
+uniform int u_int1;      // Color tint
+uniform float u_float0;  // Intensity
+uniform float u_float1;  // Radius
+uniform float u_float2;  // Threshold
+
+in vec2 v_texCoord;
+out vec4 fragColor;
+
+const int BLEND_ADD      = 0;
+const int BLEND_SCREEN   = 1;
+const int BLEND_SOFT     = 2;
+const int BLEND_OVERLAY  = 3;
+const int BLEND_LIGHTEN  = 4;
+
+const float GOLDEN_ANGLE = 2.39996323;
+const int MAX_SAMPLES = 48;
+const vec3 LUMA = vec3(0.299, 0.587, 0.114);
+
+float hash(vec2 p) {
+    p = fract(p * vec2(123.34, 456.21));
+    p += dot(p, p + 45.32);
+    return fract(p.x * p.y);
+}
+
+vec3 hexToRgb(int h) {
+    return vec3(
+        float((h >> 16) & 255),
+        float((h >> 8) & 255),
+        float(h & 255)
+    ) * (1.0 / 255.0);
+}
+
+vec3 blend(vec3 base, vec3 glow, int mode) {
+    if (mode == BLEND_SCREEN) {
+        return 1.0 - (1.0 - base) * (1.0 - glow);
+    }
+    if (mode == BLEND_SOFT) {
+        return mix(
+            base - (1.0 - 2.0 * glow) * base * (1.0 - base),
+            base + (2.0 * glow - 1.0) * (sqrt(base) - base),
+            step(0.5, glow)
+        );
+    }
+    if (mode == BLEND_OVERLAY) {
+        return mix(
+            2.0 * base * glow,
+            1.0 - 2.0 * (1.0 - base) * (1.0 - glow),
+            step(0.5, base)
+        );
+    }
+    if (mode == BLEND_LIGHTEN) {
+        return max(base, glow);
+    }
+    return base + glow;
+}
+
+void main() {
+    vec4 original = texture(u_image0, v_texCoord);
+    
+    float intensity = u_float0 * 0.05;
+    float radius = u_float1 * u_float1 * 0.012;
+    
+    if (intensity < 0.001 || radius < 0.1) {
+        fragColor = original;
+        return;
+    }
+    
+    float threshold = 1.0 - u_float2 * 0.01;
+    float t0 = threshold - 0.15;
+    float t1 = threshold + 0.15;
+    
+    vec2 texelSize = 1.0 / u_resolution;
+    float radius2 = radius * radius;
+    
+    float sampleScale = clamp(radius * 0.75, 0.35, 1.0);
+    int samples = int(float(MAX_SAMPLES) * sampleScale);
+    
+    float noise = hash(gl_FragCoord.xy);
+    float angleOffset = noise * GOLDEN_ANGLE;
+    float radiusJitter = 0.85 + noise * 0.3;
+    
+    float ca = cos(GOLDEN_ANGLE);
+    float sa = sin(GOLDEN_ANGLE);
+    vec2 dir = vec2(cos(angleOffset), sin(angleOffset));
+    
+    vec3 glow = vec3(0.0);
+    float totalWeight = 0.0;
+    
+    // Center tap
+    float centerMask = smoothstep(t0, t1, dot(original.rgb, LUMA));
+    glow += original.rgb * centerMask * 2.0;
+    totalWeight += 2.0;
+    
+    for (int i = 1; i < MAX_SAMPLES; i++) {
+        if (i >= samples) break;
+        
+        float fi = float(i);
+        float dist = sqrt(fi / float(samples)) * radius * radiusJitter;
+        
+        vec2 offset = dir * dist * texelSize;
+        vec3 c = texture(u_image0, v_texCoord + offset).rgb;
+        float mask = smoothstep(t0, t1, dot(c, LUMA));
+        
+        float w = 1.0 - (dist * dist) / (radius2 * 1.5);
+        w = max(w, 0.0);
+        w *= w;
+        
+        glow += c * mask * w;
+        totalWeight += w;
+        
+        dir = vec2(
+            dir.x * ca - dir.y * sa,
+            dir.x * sa + dir.y * ca
+        );
+    }
+    
+    glow *= intensity / max(totalWeight, 0.001);
+    
+    if (u_int1 > 0) {
+        glow *= hexToRgb(u_int1);
+    }
+    
+    vec3 result = blend(original.rgb, glow, u_int0);
+    result += (noise - 0.5) * (1.0 / 255.0);
+    
+    fragColor = vec4(clamp(result, 0.0, 1.0), original.a);
+}

blueprints/.glsl/Hue_and_Saturation_1.frag

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

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

comfy/ldm/flux/layers.py

@@ -196,9 +196,6 @@ class DoubleStreamBlock(nn.Module):
         else:
             (img_mod1, img_mod2), (txt_mod1, txt_mod2) = vec
 
-        transformer_patches = transformer_options.get("patches", {})
-        extra_options = transformer_options.copy()
-
         # prepare image for attention
         img_modulated = self.img_norm1(img)
         img_modulated = apply_mod(img_modulated, (1 + img_mod1.scale), img_mod1.shift, modulation_dims_img)
@@ -227,12 +224,6 @@ class DoubleStreamBlock(nn.Module):
         attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
         del q, k, v
 
-        if "attn1_output_patch" in transformer_patches:
-            extra_options["img_slice"] = [txt.shape[1], attn.shape[1]]
-            patch = transformer_patches["attn1_output_patch"]
-            for p in patch:
-                attn = p(attn, extra_options)
-
         txt_attn, img_attn = attn[:, : txt.shape[1]], attn[:, txt.shape[1]:]
 
         # calculate the img bloks
@@ -312,9 +303,6 @@ class SingleStreamBlock(nn.Module):
         else:
             mod = vec
 
-        transformer_patches = transformer_options.get("patches", {})
-        extra_options = transformer_options.copy()
-
         qkv, mlp = torch.split(self.linear1(apply_mod(self.pre_norm(x), (1 + mod.scale), mod.shift, modulation_dims)), [3 * self.hidden_size, self.mlp_hidden_dim_first], dim=-1)
 
         q, k, v = qkv.view(qkv.shape[0], qkv.shape[1], 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
@@ -324,12 +312,6 @@ class SingleStreamBlock(nn.Module):
         # compute attention
         attn = attention(q, k, v, pe=pe, mask=attn_mask, transformer_options=transformer_options)
         del q, k, v
-
-        if "attn1_output_patch" in transformer_patches:
-            patch = transformer_patches["attn1_output_patch"]
-            for p in patch:
-                attn = p(attn, extra_options)
-
         # compute activation in mlp stream, cat again and run second linear layer
         if self.yak_mlp:
             mlp = self.mlp_act(mlp[..., self.mlp_hidden_dim_first // 2:]) * mlp[..., :self.mlp_hidden_dim_first // 2]

comfy/ldm/flux/model.py

@@ -142,7 +142,6 @@ class Flux(nn.Module):
         attn_mask: Tensor = None,
     ) -> Tensor:
 
-        transformer_options = transformer_options.copy()
         patches = transformer_options.get("patches", {})
         patches_replace = transformer_options.get("patches_replace", {})
         if img.ndim != 3 or txt.ndim != 3:
@@ -232,7 +231,6 @@ class Flux(nn.Module):
 
         transformer_options["total_blocks"] = len(self.single_blocks)
         transformer_options["block_type"] = "single"
-        transformer_options["img_slice"] = [txt.shape[1], img.shape[1]]
         for i, block in enumerate(self.single_blocks):
             transformer_options["block_index"] = i
             if ("single_block", i) in blocks_replace:

comfy/ldm/hunyuan_video/model.py

@@ -304,7 +304,6 @@ class HunyuanVideo(nn.Module):
         control=None,
         transformer_options={},
     ) -> Tensor:
-        transformer_options = transformer_options.copy()
         patches_replace = transformer_options.get("patches_replace", {})
 
         initial_shape = list(img.shape)
@@ -417,7 +416,6 @@ class HunyuanVideo(nn.Module):
 
         transformer_options["total_blocks"] = len(self.single_blocks)
         transformer_options["block_type"] = "single"
-        transformer_options["img_slice"] = [txt.shape[1], img.shape[1]]
         for i, block in enumerate(self.single_blocks):
             transformer_options["block_index"] = i
             if ("single_block", i) in blocks_replace:

comfy/ldm/modules/diffusionmodules/model.py

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

comfy/lora.py

@@ -374,31 +374,6 @@ def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Ten
 
     return padded_tensor
 
-def calculate_shape(patches, weight, key, original_weights=None):
-    current_shape = weight.shape
-
-    for p in patches:
-        v = p[1]
-        offset = p[3]
-
-        # Offsets restore the old shape; lists force a diff without metadata
-        if offset is not None or isinstance(v, list):
-            continue
-
-        if isinstance(v, weight_adapter.WeightAdapterBase):
-            adapter_shape = v.calculate_shape(key)
-            if adapter_shape is not None:
-                current_shape = adapter_shape
-            continue
-
-        # Standard diff logic with padding
-        if len(v) == 2:
-            patch_type, patch_data = v[0], v[1]
-            if patch_type == "diff" and len(patch_data) > 1 and patch_data[1]['pad_weight']:
-                current_shape = patch_data[0].shape
-
-    return current_shape
-
 def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, original_weights=None):
     for p in patches:
         strength = p[0]

comfy/model_patcher.py

@@ -406,16 +406,13 @@ class ModelPatcher:
     def memory_required(self, input_shape):
         return self.model.memory_required(input_shape=input_shape)
 
-    def disable_model_cfg1_optimization(self):
-        self.model_options["disable_cfg1_optimization"] = True
-
     def set_model_sampler_cfg_function(self, sampler_cfg_function, disable_cfg1_optimization=False):
         if len(inspect.signature(sampler_cfg_function).parameters) == 3:
             self.model_options["sampler_cfg_function"] = lambda args: sampler_cfg_function(args["cond"], args["uncond"], args["cond_scale"]) #Old way
         else:
             self.model_options["sampler_cfg_function"] = sampler_cfg_function
         if disable_cfg1_optimization:
-            self.disable_model_cfg1_optimization()
+            self.model_options["disable_cfg1_optimization"] = True
 
     def set_model_sampler_post_cfg_function(self, post_cfg_function, disable_cfg1_optimization=False):
         self.model_options = set_model_options_post_cfg_function(self.model_options, post_cfg_function, disable_cfg1_optimization)
@@ -1517,10 +1514,8 @@ class ModelPatcherDynamic(ModelPatcher):
 
                     weight, _, _ = get_key_weight(self.model, key)
                     if weight is None:
-                        return (False, 0)
+                        return 0
                     if key in self.patches:
-                        if comfy.lora.calculate_shape(self.patches[key], weight, key) != weight.shape:
-                            return (True, 0)
                         setattr(m, param_key + "_lowvram_function", LowVramPatch(key, self.patches))
                         num_patches += 1
                     else:
@@ -1534,13 +1529,7 @@ class ModelPatcherDynamic(ModelPatcher):
                         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 (False, comfy.memory_management.vram_aligned_size(geometry))
-
-                def force_load_param(self, param_key, device_to):
-                    key = key_param_name_to_key(n, param_key)
-                    if key in self.backup:
-                        comfy.utils.set_attr_param(self.model, key, self.backup[key].weight)
-                    self.patch_weight_to_device(key, device_to=device_to)
+                    return comfy.memory_management.vram_aligned_size(geometry)
 
                 if hasattr(m, "comfy_cast_weights"):
                     m.comfy_cast_weights = True
@@ -1548,19 +1537,13 @@ class ModelPatcherDynamic(ModelPatcher):
                     m.seed_key = n
                     set_dirty(m, dirty)
 
-                    force_load, v_weight_size = setup_param(self, m, n, "weight")
-                    force_load_bias, v_weight_bias = setup_param(self, m, n, "bias")
-                    force_load = force_load or force_load_bias
-                    v_weight_size += v_weight_bias
+                    v_weight_size = 0
+                    v_weight_size += setup_param(self, m, n, "weight")
+                    v_weight_size += setup_param(self, m, n, "bias")
 
-                    if force_load:
-                        logging.info(f"Module {n} has resizing Lora - force loading")
-                        force_load_param(self, "weight", device_to)
-                        force_load_param(self, "bias", device_to)
-                    else:
-                        if vbar is not None and not hasattr(m, "_v"):
-                            m._v = vbar.alloc(v_weight_size)
-                        allocated_size += v_weight_size
+                    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:
@@ -1623,11 +1606,6 @@ class ModelPatcherDynamic(ModelPatcher):
             for m in self.model.modules():
                 move_weight_functions(m, device_to)
 
-            keys = list(self.backup.keys())
-            for k in keys:
-                bk = self.backup[k]
-                comfy.utils.set_attr_param(self.model, k, bk.weight)
-
     def partially_load(self, device_to, extra_memory=0, force_patch_weights=False):
         assert not force_patch_weights #See above
         with self.use_ejected(skip_and_inject_on_exit_only=True):

comfy/ops.py

@@ -21,6 +21,7 @@ import logging
 import comfy.model_management
 from comfy.cli_args import args, PerformanceFeature, enables_dynamic_vram
 import comfy.float
+import comfy.rmsnorm
 import json
 import comfy.memory_management
 import comfy.pinned_memory
@@ -462,7 +463,7 @@ class disable_weight_init:
             else:
                 return super().forward(*args, **kwargs)
 
-    class RMSNorm(torch.nn.RMSNorm, CastWeightBiasOp):
+    class RMSNorm(comfy.rmsnorm.RMSNorm, CastWeightBiasOp):
         def reset_parameters(self):
             self.bias = None
             return None
@@ -474,7 +475,8 @@ class disable_weight_init:
                 weight = None
                 bias = None
                 offload_stream = None
-            x = torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
+            x = comfy.rmsnorm.rms_norm(input, weight, self.eps)  # TODO: switch to commented out line when old torch is deprecated
+            # x = torch.nn.functional.rms_norm(input, self.normalized_shape, weight, self.eps)
             uncast_bias_weight(self, weight, bias, offload_stream)
             return x
 

comfy/rmsnorm.py

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

comfy/weight_adapter/base.py

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

comfy/weight_adapter/lora.py

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

comfy_api/feature_flags.py

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

comfy_api/latest/__init__.py

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

comfy_api/latest/_io.py

@@ -75,12 +75,6 @@ class NumberDisplay(str, Enum):
     slider = "slider"
 
 
-class ControlAfterGenerate(str, Enum):
-    fixed = "fixed"
-    increment = "increment"
-    decrement = "decrement"
-    randomize = "randomize"
-
 class _ComfyType(ABC):
     Type = Any
     io_type: str = None
@@ -269,7 +263,7 @@ class Int(ComfyTypeIO):
     class Input(WidgetInput):
         '''Integer input.'''
         def __init__(self, id: str, display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
-                    default: int=None, min: int=None, max: int=None, step: int=None, control_after_generate: bool | ControlAfterGenerate=None,
+                    default: int=None, min: int=None, max: int=None, step: int=None, control_after_generate: bool=None,
                     display_mode: NumberDisplay=None, socketless: bool=None, force_input: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
             super().__init__(id, display_name, optional, tooltip, lazy, default, socketless, None, force_input, extra_dict, raw_link, advanced)
             self.min = min
@@ -351,7 +345,7 @@ class Combo(ComfyTypeIO):
             tooltip: str=None,
             lazy: bool=None,
             default: str | int | Enum = None,
-            control_after_generate: bool | ControlAfterGenerate=None,
+            control_after_generate: bool=None,
             upload: UploadType=None,
             image_folder: FolderType=None,
             remote: RemoteOptions=None,
@@ -395,7 +389,7 @@ class MultiCombo(ComfyTypeI):
     Type = list[str]
     class Input(Combo.Input):
         def __init__(self, id: str, options: list[str], display_name: str=None, optional=False, tooltip: str=None, lazy: bool=None,
-                    default: list[str]=None, placeholder: str=None, chip: bool=None, control_after_generate: bool | ControlAfterGenerate=None,
+                    default: list[str]=None, placeholder: str=None, chip: bool=None, control_after_generate: bool=None,
                     socketless: bool=None, extra_dict=None, raw_link: bool=None, advanced: bool=None):
             super().__init__(id, options, display_name, optional, tooltip, lazy, default, control_after_generate, socketless=socketless, extra_dict=extra_dict, raw_link=raw_link, advanced=advanced)
             self.multiselect = True
@@ -2036,74 +2030,11 @@ class _UIOutput(ABC):
         ...
 
 
-class InputMapOldId(TypedDict):
-    """Map an old node input to a new node input by ID."""
-    new_id: str
-    old_id: str
-
-class InputMapSetValue(TypedDict):
-    """Set a specific value for a new node input."""
-    new_id: str
-    set_value: Any
-
-InputMap = InputMapOldId | InputMapSetValue
-"""
-Input mapping for node replacement. Type is inferred by dictionary keys:
-- {"new_id": str, "old_id": str} - maps old input to new input
-- {"new_id": str, "set_value": Any} - sets a specific value for new input
-"""
-
-class OutputMap(TypedDict):
-    """Map outputs of node replacement via indexes."""
-    new_idx: int
-    old_idx: int
-
-class NodeReplace:
-    """
-    Defines a possible node replacement, mapping inputs and outputs of the old node to the new node.
-
-    Also supports assigning specific values to the input widgets of the new node.
-
-    Args:
-        new_node_id: The class name of the new replacement node.
-        old_node_id: The class name of the deprecated node.
-        old_widget_ids: Ordered list of input IDs for widgets that may not have an input slot
-            connected. The workflow JSON stores widget values by their relative position index,
-            not by ID. This list maps those positional indexes to input IDs, enabling the
-            replacement system to correctly identify widget values during node migration.
-        input_mapping: List of input mappings from old node to new node.
-        output_mapping: List of output mappings from old node to new node.
-    """
-    def __init__(self,
-        new_node_id: str,
-        old_node_id: str,
-        old_widget_ids: list[str] | None=None,
-        input_mapping: list[InputMap] | None=None,
-        output_mapping: list[OutputMap] | None=None,
-    ):
-        self.new_node_id = new_node_id
-        self.old_node_id = old_node_id
-        self.old_widget_ids = old_widget_ids
-        self.input_mapping = input_mapping
-        self.output_mapping = output_mapping
-
-    def as_dict(self):
-        """Create serializable representation of the node replacement."""
-        return {
-            "new_node_id": self.new_node_id,
-            "old_node_id": self.old_node_id,
-            "old_widget_ids": self.old_widget_ids,
-            "input_mapping": list(self.input_mapping) if self.input_mapping else None,
-            "output_mapping": list(self.output_mapping) if self.output_mapping else None,
-        }
-
-
 __all__ = [
     "FolderType",
     "UploadType",
     "RemoteOptions",
     "NumberDisplay",
-    "ControlAfterGenerate",
 
     "comfytype",
     "Custom",
@@ -2190,5 +2121,4 @@ __all__ = [
     "ImageCompare",
     "PriceBadgeDepends",
     "PriceBadge",
-    "NodeReplace",
 ]

comfy_api_nodes/apis/bria.py

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

comfy_api_nodes/apis/hunyuan3d.py

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

comfy_api_nodes/nodes_bria.py

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

comfy_api_nodes/nodes_hunyuan3d.py

@@ -1,48 +1,31 @@
 from typing_extensions import override
 
-from comfy_api.latest import IO, ComfyExtension, Input, Types
+from comfy_api.latest import IO, ComfyExtension, Input
 from comfy_api_nodes.apis.hunyuan3d import (
     Hunyuan3DViewImage,
     InputGenerateType,
     ResultFile3D,
-    TextureEditTaskRequest,
     To3DProTaskCreateResponse,
     To3DProTaskQueryRequest,
     To3DProTaskRequest,
     To3DProTaskResultResponse,
-    To3DUVFileInput,
-    To3DUVTaskRequest,
 )
 from comfy_api_nodes.util import (
     ApiEndpoint,
     download_url_to_file_3d,
-    download_url_to_image_tensor,
     downscale_image_tensor_by_max_side,
     poll_op,
     sync_op,
-    upload_3d_model_to_comfyapi,
     upload_image_to_comfyapi,
     validate_image_dimensions,
     validate_string,
 )
 
 
-def _is_tencent_rate_limited(status: int, body: object) -> bool:
-    return (
-        status == 400
-        and isinstance(body, dict)
-        and "RequestLimitExceeded" in str(body.get("Response", {}).get("Error", {}).get("Code", ""))
-    )
-
-
-def get_file_from_response(
-    response_objs: list[ResultFile3D], file_type: str, raise_if_not_found: bool = True
-) -> ResultFile3D | None:
+def get_file_from_response(response_objs: list[ResultFile3D], file_type: str) -> ResultFile3D | None:
     for i in response_objs:
         if i.Type.lower() == file_type.lower():
             return i
-    if raise_if_not_found:
-        raise ValueError(f"'{file_type}' file type is not found in the response.")
     return None
 
 
@@ -52,7 +35,7 @@ class TencentTextToModelNode(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="TencentTextToModelNode",
-            display_name="Hunyuan3D: Text to Model",
+            display_name="Hunyuan3D: Text to Model (Pro)",
             category="api node/3d/Tencent",
             inputs=[
                 IO.Combo.Input(
@@ -137,7 +120,6 @@ class TencentTextToModelNode(IO.ComfyNode):
                 EnablePBR=generate_type.get("pbr", None),
                 PolygonType=generate_type.get("polygon_type", None),
             ),
-            is_rate_limited=_is_tencent_rate_limited,
         )
         if response.Error:
             raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
@@ -149,14 +131,11 @@ class TencentTextToModelNode(IO.ComfyNode):
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
+        glb_result = get_file_from_response(result.ResultFile3Ds, "glb")
+        obj_result = get_file_from_response(result.ResultFile3Ds, "obj")
+        file_glb = await download_url_to_file_3d(glb_result.Url, "glb", task_id=task_id) if glb_result else None
         return IO.NodeOutput(
-            f"{task_id}.glb",
-            await download_url_to_file_3d(
-                get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb", task_id=task_id
-            ),
-            await download_url_to_file_3d(
-                get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj", task_id=task_id
-            ),
+            file_glb, file_glb, await download_url_to_file_3d(obj_result.Url, "obj", task_id=task_id) if obj_result else None
         )
 
 
@@ -166,7 +145,7 @@ class TencentImageToModelNode(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="TencentImageToModelNode",
-            display_name="Hunyuan3D: Image(s) to Model",
+            display_name="Hunyuan3D: Image(s) to Model (Pro)",
             category="api node/3d/Tencent",
             inputs=[
                 IO.Combo.Input(
@@ -289,7 +268,6 @@ class TencentImageToModelNode(IO.ComfyNode):
                 EnablePBR=generate_type.get("pbr", None),
                 PolygonType=generate_type.get("polygon_type", None),
             ),
-            is_rate_limited=_is_tencent_rate_limited,
         )
         if response.Error:
             raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
@@ -301,257 +279,11 @@ class TencentImageToModelNode(IO.ComfyNode):
             response_model=To3DProTaskResultResponse,
             status_extractor=lambda r: r.Status,
         )
+        glb_result = get_file_from_response(result.ResultFile3Ds, "glb")
+        obj_result = get_file_from_response(result.ResultFile3Ds, "obj")
+        file_glb = await download_url_to_file_3d(glb_result.Url, "glb", task_id=task_id) if glb_result else None
         return IO.NodeOutput(
-            f"{task_id}.glb",
-            await download_url_to_file_3d(
-                get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb", task_id=task_id
-            ),
-            await download_url_to_file_3d(
-                get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj", task_id=task_id
-            ),
-        )
-
-
-class TencentModelTo3DUVNode(IO.ComfyNode):
-
-    @classmethod
-    def define_schema(cls):
-        return IO.Schema(
-            node_id="TencentModelTo3DUVNode",
-            display_name="Hunyuan3D: Model to UV",
-            category="api node/3d/Tencent",
-            description="Perform UV unfolding on a 3D model to generate UV texture. "
-            "Input model must have less than 30000 faces.",
-            inputs=[
-                IO.MultiType.Input(
-                    "model_3d",
-                    types=[IO.File3DGLB, IO.File3DOBJ, IO.File3DFBX, IO.File3DAny],
-                    tooltip="Input 3D model (GLB, OBJ, or FBX)",
-                ),
-                IO.Int.Input(
-                    "seed",
-                    default=1,
-                    min=0,
-                    max=2147483647,
-                    display_mode=IO.NumberDisplay.number,
-                    control_after_generate=True,
-                    tooltip="Seed controls whether the node should re-run; "
-                    "results are non-deterministic regardless of seed.",
-                ),
-            ],
-            outputs=[
-                IO.File3DOBJ.Output(display_name="OBJ"),
-                IO.File3DFBX.Output(display_name="FBX"),
-                IO.Image.Output(),
-            ],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
-            is_api_node=True,
-            price_badge=IO.PriceBadge(expr='{"type":"usd","usd":0.2}'),
-        )
-
-    SUPPORTED_FORMATS = {"glb", "obj", "fbx"}
-
-    @classmethod
-    async def execute(
-        cls,
-        model_3d: Types.File3D,
-        seed: int,
-    ) -> IO.NodeOutput:
-        _ = seed
-        file_format = model_3d.format.lower()
-        if file_format not in cls.SUPPORTED_FORMATS:
-            raise ValueError(
-                f"Unsupported file format: '{file_format}'. "
-                f"Supported formats: {', '.join(sorted(cls.SUPPORTED_FORMATS))}."
-            )
-        response = await sync_op(
-            cls,
-            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-uv", method="POST"),
-            response_model=To3DProTaskCreateResponse,
-            data=To3DUVTaskRequest(
-                File=To3DUVFileInput(
-                    Type=file_format.upper(),
-                    Url=await upload_3d_model_to_comfyapi(cls, model_3d, file_format),
-                )
-            ),
-            is_rate_limited=_is_tencent_rate_limited,
-        )
-        if response.Error:
-            raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
-        result = await poll_op(
-            cls,
-            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-uv/query", method="POST"),
-            data=To3DProTaskQueryRequest(JobId=response.JobId),
-            response_model=To3DProTaskResultResponse,
-            status_extractor=lambda r: r.Status,
-        )
-        return IO.NodeOutput(
-            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "obj").Url, "obj"),
-            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"),
-            await download_url_to_image_tensor(get_file_from_response(result.ResultFile3Ds, "image").Url),
-        )
-
-
-class Tencent3DTextureEditNode(IO.ComfyNode):
-
-    @classmethod
-    def define_schema(cls):
-        return IO.Schema(
-            node_id="Tencent3DTextureEditNode",
-            display_name="Hunyuan3D: 3D Texture Edit",
-            category="api node/3d/Tencent",
-            description="After inputting the 3D model, perform 3D model texture redrawing.",
-            inputs=[
-                IO.MultiType.Input(
-                    "model_3d",
-                    types=[IO.File3DFBX, IO.File3DAny],
-                    tooltip="3D model in FBX format. Model should have less than 100000 faces.",
-                ),
-                IO.String.Input(
-                    "prompt",
-                    multiline=True,
-                    default="",
-                    tooltip="Describes texture editing. Supports up to 1024 UTF-8 characters.",
-                ),
-                IO.Int.Input(
-                    "seed",
-                    default=0,
-                    min=0,
-                    max=2147483647,
-                    display_mode=IO.NumberDisplay.number,
-                    control_after_generate=True,
-                    tooltip="Seed controls whether the node should re-run; "
-                    "results are non-deterministic regardless of seed.",
-                ),
-            ],
-            outputs=[
-                IO.File3DGLB.Output(display_name="GLB"),
-                IO.File3DFBX.Output(display_name="FBX"),
-            ],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
-            is_api_node=True,
-            price_badge=IO.PriceBadge(
-                expr="""{"type":"usd","usd": 0.6}""",
-            ),
-        )
-
-    @classmethod
-    async def execute(
-        cls,
-        model_3d: Types.File3D,
-        prompt: str,
-        seed: int,
-    ) -> IO.NodeOutput:
-        _ = seed
-        file_format = model_3d.format.lower()
-        if file_format != "fbx":
-            raise ValueError(f"Unsupported file format: '{file_format}'. Only FBX format is supported.")
-        validate_string(prompt, field_name="prompt", min_length=1, max_length=1024)
-        model_url = await upload_3d_model_to_comfyapi(cls, model_3d, file_format)
-        response = await sync_op(
-            cls,
-            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-texture-edit", method="POST"),
-            response_model=To3DProTaskCreateResponse,
-            data=TextureEditTaskRequest(
-                File3D=To3DUVFileInput(Type=file_format.upper(), Url=model_url),
-                Prompt=prompt,
-                EnablePBR=True,
-            ),
-            is_rate_limited=_is_tencent_rate_limited,
-        )
-        if response.Error:
-            raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
-
-        result = await poll_op(
-            cls,
-            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-texture-edit/query", method="POST"),
-            data=To3DProTaskQueryRequest(JobId=response.JobId),
-            response_model=To3DProTaskResultResponse,
-            status_extractor=lambda r: r.Status,
-        )
-        return IO.NodeOutput(
-            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "glb").Url, "glb"),
-            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"),
-        )
-
-
-class Tencent3DPartNode(IO.ComfyNode):
-
-    @classmethod
-    def define_schema(cls):
-        return IO.Schema(
-            node_id="Tencent3DPartNode",
-            display_name="Hunyuan3D: 3D Part",
-            category="api node/3d/Tencent",
-            description="Automatically perform component identification and generation based on the model structure.",
-            inputs=[
-                IO.MultiType.Input(
-                    "model_3d",
-                    types=[IO.File3DFBX, IO.File3DAny],
-                    tooltip="3D model in FBX format. Model should have less than 30000 faces.",
-                ),
-                IO.Int.Input(
-                    "seed",
-                    default=0,
-                    min=0,
-                    max=2147483647,
-                    display_mode=IO.NumberDisplay.number,
-                    control_after_generate=True,
-                    tooltip="Seed controls whether the node should re-run; "
-                    "results are non-deterministic regardless of seed.",
-                ),
-            ],
-            outputs=[
-                IO.File3DFBX.Output(display_name="FBX"),
-            ],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
-            is_api_node=True,
-            price_badge=IO.PriceBadge(expr='{"type":"usd","usd":0.6}'),
-        )
-
-    @classmethod
-    async def execute(
-        cls,
-        model_3d: Types.File3D,
-        seed: int,
-    ) -> IO.NodeOutput:
-        _ = seed
-        file_format = model_3d.format.lower()
-        if file_format != "fbx":
-            raise ValueError(f"Unsupported file format: '{file_format}'. Only FBX format is supported.")
-        model_url = await upload_3d_model_to_comfyapi(cls, model_3d, file_format)
-        response = await sync_op(
-            cls,
-            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-part", method="POST"),
-            response_model=To3DProTaskCreateResponse,
-            data=To3DUVTaskRequest(
-                File=To3DUVFileInput(Type=file_format.upper(), Url=model_url),
-            ),
-            is_rate_limited=_is_tencent_rate_limited,
-        )
-        if response.Error:
-            raise ValueError(f"Task creation failed with code {response.Error.Code}: {response.Error.Message}")
-        result = await poll_op(
-            cls,
-            ApiEndpoint(path="/proxy/tencent/hunyuan/3d-part/query", method="POST"),
-            data=To3DProTaskQueryRequest(JobId=response.JobId),
-            response_model=To3DProTaskResultResponse,
-            status_extractor=lambda r: r.Status,
-        )
-        return IO.NodeOutput(
-            await download_url_to_file_3d(get_file_from_response(result.ResultFile3Ds, "fbx").Url, "fbx"),
+            file_glb, file_glb, await download_url_to_file_3d(obj_result.Url, "obj", task_id=task_id) if obj_result else None
         )
 
 
@@ -561,9 +293,6 @@ class TencentHunyuan3DExtension(ComfyExtension):
         return [
             TencentTextToModelNode,
             TencentImageToModelNode,
-            # TencentModelTo3DUVNode,
-            # Tencent3DTextureEditNode,
-            Tencent3DPartNode,
         ]
 
 

comfy_api_nodes/nodes_openai.py

@@ -43,6 +43,7 @@ class SupportedOpenAIModel(str, Enum):
     o1 = "o1"
     o3 = "o3"
     o1_pro = "o1-pro"
+    gpt_4o = "gpt-4o"
     gpt_4_1 = "gpt-4.1"
     gpt_4_1_mini = "gpt-4.1-mini"
     gpt_4_1_nano = "gpt-4.1-nano"
@@ -648,6 +649,11 @@ class OpenAIChatNode(IO.ComfyNode):
                     "usd": [0.01, 0.04],
                     "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" }
                   }
+                  : $contains($m, "gpt-4o") ? {
+                    "type": "list_usd",
+                    "usd": [0.0025, 0.01],
+                    "format": { "approximate": true, "separator": "-", "suffix": " per 1K tokens" }
+                  }
                   : $contains($m, "gpt-4.1-nano") ? {
                     "type": "list_usd",
                     "usd": [0.0001, 0.0004],

comfy_api_nodes/nodes_vidu.py

@@ -54,7 +54,6 @@ async def execute_task(
         response_model=TaskStatusResponse,
         status_extractor=lambda r: r.state,
         progress_extractor=lambda r: r.progress,
-        price_extractor=lambda r: r.credits * 0.005 if r.credits is not None else None,
         max_poll_attempts=max_poll_attempts,
     )
     if not response.creations:
@@ -1307,36 +1306,6 @@ class Vidu3TextToVideoNode(IO.ComfyNode):
                                 ),
                             ],
                         ),
-                        IO.DynamicCombo.Option(
-                            "viduq3-turbo",
-                            [
-                                IO.Combo.Input(
-                                    "aspect_ratio",
-                                    options=["16:9", "9:16", "3:4", "4:3", "1:1"],
-                                    tooltip="The aspect ratio of the output video.",
-                                ),
-                                IO.Combo.Input(
-                                    "resolution",
-                                    options=["720p", "1080p"],
-                                    tooltip="Resolution of the output video.",
-                                ),
-                                IO.Int.Input(
-                                    "duration",
-                                    default=5,
-                                    min=1,
-                                    max=16,
-                                    step=1,
-                                    display_mode=IO.NumberDisplay.slider,
-                                    tooltip="Duration of the output video in seconds.",
-                                ),
-                                IO.Boolean.Input(
-                                    "audio",
-                                    default=False,
-                                    tooltip="When enabled, outputs video with sound "
-                                    "(including dialogue and sound effects).",
-                                ),
-                            ],
-                        ),
                     ],
                     tooltip="Model to use for video generation.",
                 ),
@@ -1365,20 +1334,13 @@ class Vidu3TextToVideoNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.duration", "model.resolution"]),
+                depends_on=IO.PriceBadgeDepends(widgets=["model.duration", "model.resolution"]),
                 expr="""
                 (
                   $res := $lookup(widgets, "model.resolution");
-                  $d := $lookup(widgets, "model.duration");
-                  $contains(widgets.model, "turbo")
-                    ? (
-                        $rate := $lookup({"720p": 0.06, "1080p": 0.08}, $res);
-                        {"type":"usd","usd": $rate * $d}
-                      )
-                    : (
-                        $rate := $lookup({"720p": 0.15, "1080p": 0.16}, $res);
-                        {"type":"usd","usd": $rate * $d}
-                      )
+                  $base := $lookup({"720p": 0.075, "1080p": 0.1}, $res);
+                  $perSec := $lookup({"720p": 0.025, "1080p": 0.05}, $res);
+                  {"type":"usd","usd": $base + $perSec * ($lookup(widgets, "model.duration") - 1)}
                 )
                 """,
             ),
@@ -1447,31 +1409,6 @@ class Vidu3ImageToVideoNode(IO.ComfyNode):
                                 ),
                             ],
                         ),
-                        IO.DynamicCombo.Option(
-                            "viduq3-turbo",
-                            [
-                                IO.Combo.Input(
-                                    "resolution",
-                                    options=["720p", "1080p"],
-                                    tooltip="Resolution of the output video.",
-                                ),
-                                IO.Int.Input(
-                                    "duration",
-                                    default=5,
-                                    min=1,
-                                    max=16,
-                                    step=1,
-                                    display_mode=IO.NumberDisplay.slider,
-                                    tooltip="Duration of the output video in seconds.",
-                                ),
-                                IO.Boolean.Input(
-                                    "audio",
-                                    default=False,
-                                    tooltip="When enabled, outputs video with sound "
-                                    "(including dialogue and sound effects).",
-                                ),
-                            ],
-                        ),
                     ],
                     tooltip="Model to use for video generation.",
                 ),
@@ -1505,20 +1442,13 @@ class Vidu3ImageToVideoNode(IO.ComfyNode):
             ],
             is_api_node=True,
             price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.duration", "model.resolution"]),
+                depends_on=IO.PriceBadgeDepends(widgets=["model.duration", "model.resolution"]),
                 expr="""
                 (
                   $res := $lookup(widgets, "model.resolution");
-                  $d := $lookup(widgets, "model.duration");
-                  $contains(widgets.model, "turbo")
-                    ? (
-                        $rate := $lookup({"720p": 0.06, "1080p": 0.08}, $res);
-                        {"type":"usd","usd": $rate * $d}
-                      )
-                    : (
-                        $rate := $lookup({"720p": 0.15, "1080p": 0.16, "2k": 0.2}, $res);
-                        {"type":"usd","usd": $rate * $d}
-                      )
+                  $base := $lookup({"720p": 0.075, "1080p": 0.275, "2k": 0.35}, $res);
+                  $perSec := $lookup({"720p": 0.05, "1080p": 0.075, "2k": 0.075}, $res);
+                  {"type":"usd","usd": $base + $perSec * ($lookup(widgets, "model.duration") - 1)}
                 )
                 """,
             ),
@@ -1551,145 +1481,6 @@ class Vidu3ImageToVideoNode(IO.ComfyNode):
         return IO.NodeOutput(await download_url_to_video_output(results[0].url))
 
 
-class Vidu3StartEndToVideoNode(IO.ComfyNode):
-
-    @classmethod
-    def define_schema(cls):
-        return IO.Schema(
-            node_id="Vidu3StartEndToVideoNode",
-            display_name="Vidu Q3 Start/End Frame-to-Video Generation",
-            category="api node/video/Vidu",
-            description="Generate a video from a start frame, an end frame, and a prompt.",
-            inputs=[
-                IO.DynamicCombo.Input(
-                    "model",
-                    options=[
-                        IO.DynamicCombo.Option(
-                            "viduq3-pro",
-                            [
-                                IO.Combo.Input(
-                                    "resolution",
-                                    options=["720p", "1080p"],
-                                    tooltip="Resolution of the output video.",
-                                ),
-                                IO.Int.Input(
-                                    "duration",
-                                    default=5,
-                                    min=1,
-                                    max=16,
-                                    step=1,
-                                    display_mode=IO.NumberDisplay.slider,
-                                    tooltip="Duration of the output video in seconds.",
-                                ),
-                                IO.Boolean.Input(
-                                    "audio",
-                                    default=False,
-                                    tooltip="When enabled, outputs video with sound "
-                                    "(including dialogue and sound effects).",
-                                ),
-                            ],
-                        ),
-                        IO.DynamicCombo.Option(
-                            "viduq3-turbo",
-                            [
-                                IO.Combo.Input(
-                                    "resolution",
-                                    options=["720p", "1080p"],
-                                    tooltip="Resolution of the output video.",
-                                ),
-                                IO.Int.Input(
-                                    "duration",
-                                    default=5,
-                                    min=1,
-                                    max=16,
-                                    step=1,
-                                    display_mode=IO.NumberDisplay.slider,
-                                    tooltip="Duration of the output video in seconds.",
-                                ),
-                                IO.Boolean.Input(
-                                    "audio",
-                                    default=False,
-                                    tooltip="When enabled, outputs video with sound "
-                                    "(including dialogue and sound effects).",
-                                ),
-                            ],
-                        ),
-                    ],
-                    tooltip="Model to use for video generation.",
-                ),
-                IO.Image.Input("first_frame"),
-                IO.Image.Input("end_frame"),
-                IO.String.Input(
-                    "prompt",
-                    multiline=True,
-                    tooltip="Prompt description (max 2000 characters).",
-                ),
-                IO.Int.Input(
-                    "seed",
-                    default=1,
-                    min=0,
-                    max=2147483647,
-                    step=1,
-                    display_mode=IO.NumberDisplay.number,
-                    control_after_generate=True,
-                ),
-            ],
-            outputs=[
-                IO.Video.Output(),
-            ],
-            hidden=[
-                IO.Hidden.auth_token_comfy_org,
-                IO.Hidden.api_key_comfy_org,
-                IO.Hidden.unique_id,
-            ],
-            is_api_node=True,
-            price_badge=IO.PriceBadge(
-                depends_on=IO.PriceBadgeDepends(widgets=["model", "model.duration", "model.resolution"]),
-                expr="""
-                (
-                  $res := $lookup(widgets, "model.resolution");
-                  $d := $lookup(widgets, "model.duration");
-                  $contains(widgets.model, "turbo")
-                    ? (
-                        $rate := $lookup({"720p": 0.06, "1080p": 0.08}, $res);
-                        {"type":"usd","usd": $rate * $d}
-                      )
-                    : (
-                        $rate := $lookup({"720p": 0.15, "1080p": 0.16}, $res);
-                        {"type":"usd","usd": $rate * $d}
-                      )
-                )
-                """,
-            ),
-        )
-
-    @classmethod
-    async def execute(
-        cls,
-        model: dict,
-        first_frame: Input.Image,
-        end_frame: Input.Image,
-        prompt: str,
-        seed: int,
-    ) -> IO.NodeOutput:
-        validate_string(prompt, max_length=2000)
-        validate_images_aspect_ratio_closeness(first_frame, end_frame, min_rel=0.8, max_rel=1.25, strict=False)
-        payload = TaskCreationRequest(
-            model=model["model"],
-            prompt=prompt,
-            duration=model["duration"],
-            seed=seed,
-            resolution=model["resolution"],
-            audio=model["audio"],
-            images=[
-                (await upload_images_to_comfyapi(cls, frame, max_images=1, mime_type="image/png"))[0]
-                for frame in (first_frame, end_frame)
-            ],
-        )
-        results = await execute_task(cls, VIDU_START_END_VIDEO, payload)
-        return IO.NodeOutput(await download_url_to_video_output(results[0].url))
-
-
 class ViduExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[IO.ComfyNode]]:
@@ -1706,7 +1497,6 @@ class ViduExtension(ComfyExtension):
             ViduMultiFrameVideoNode,
             Vidu3TextToVideoNode,
             Vidu3ImageToVideoNode,
-            Vidu3StartEndToVideoNode,
         ]
 
 

comfy_api_nodes/util/__init__.py

@@ -33,7 +33,6 @@ from .download_helpers import (
     download_url_to_video_output,
 )
 from .upload_helpers import (
-    upload_3d_model_to_comfyapi,
     upload_audio_to_comfyapi,
     upload_file_to_comfyapi,
     upload_image_to_comfyapi,
@@ -63,7 +62,6 @@ __all__ = [
     "sync_op",
     "sync_op_raw",
     # Upload helpers
-    "upload_3d_model_to_comfyapi",
     "upload_audio_to_comfyapi",
     "upload_file_to_comfyapi",
     "upload_image_to_comfyapi",

comfy_api_nodes/util/conversions.py

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

comfy_api_nodes/util/upload_helpers.py

@@ -164,27 +164,6 @@ async def upload_video_to_comfyapi(
     return await upload_file_to_comfyapi(cls, video_bytes_io, filename, upload_mime_type, wait_label)
 
 
-_3D_MIME_TYPES = {
-    "glb": "model/gltf-binary",
-    "obj": "model/obj",
-    "fbx": "application/octet-stream",
-}
-
-
-async def upload_3d_model_to_comfyapi(
-    cls: type[IO.ComfyNode],
-    model_3d: Types.File3D,
-    file_format: str,
-) -> str:
-    """Uploads a 3D model file to ComfyUI API and returns its download URL."""
-    return await upload_file_to_comfyapi(
-        cls,
-        model_3d.get_data(),
-        f"{uuid.uuid4()}.{file_format}",
-        _3D_MIME_TYPES.get(file_format, "application/octet-stream"),
-    )
-
-
 async def upload_file_to_comfyapi(
     cls: type[IO.ComfyNode],
     file_bytes_io: BytesIO,

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_lora_extract.py

@@ -7,7 +7,6 @@ import logging
 from enum import Enum
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, io
-from tqdm.auto import trange
 
 CLAMP_QUANTILE = 0.99
 
@@ -50,22 +49,12 @@ LORA_TYPES = {"standard": LORAType.STANDARD,
               "full_diff": LORAType.FULL_DIFF}
 
 def calc_lora_model(model_diff, rank, prefix_model, prefix_lora, output_sd, lora_type, bias_diff=False):
-    comfy.model_management.load_models_gpu([model_diff])
+    comfy.model_management.load_models_gpu([model_diff], force_patch_weights=True)
     sd = model_diff.model_state_dict(filter_prefix=prefix_model)
 
-    sd_keys = list(sd.keys())
-    for index in trange(len(sd_keys), unit="weight"):
-        k = sd_keys[index]
-        op_keys = sd_keys[index].rsplit('.', 1)
-        if len(op_keys) < 2 or op_keys[1] not in ["weight", "bias"] or (op_keys[1] == "bias" and not bias_diff):
-            continue
-        op = comfy.utils.get_attr(model_diff.model, op_keys[0])
-        if hasattr(op, "comfy_cast_weights") and not getattr(op, "comfy_patched_weights", False):
-            weight_diff = model_diff.patch_weight_to_device(k, model_diff.load_device, return_weight=True)
-        else:
+    for k in sd:
+        if k.endswith(".weight"):
             weight_diff = sd[k]
-
-        if op_keys[1] == "weight":
             if lora_type == LORAType.STANDARD:
                 if weight_diff.ndim < 2:
                     if bias_diff:
@@ -80,8 +69,8 @@ def calc_lora_model(model_diff, rank, prefix_model, prefix_lora, output_sd, lora
             elif lora_type == LORAType.FULL_DIFF:
                 output_sd["{}{}.diff".format(prefix_lora, k[len(prefix_model):-7])] = weight_diff.contiguous().half().cpu()
 
-        elif bias_diff and op_keys[1] == "bias":
-            output_sd["{}{}.diff_b".format(prefix_lora, k[len(prefix_model):-5])] = weight_diff.contiguous().half().cpu()
+        elif bias_diff and k.endswith(".bias"):
+            output_sd["{}{}.diff_b".format(prefix_lora, k[len(prefix_model):-5])] = sd[k].contiguous().half().cpu()
     return output_sd
 
 class LoraSave(io.ComfyNode):

comfy_extras/nodes_nag.py

@@ -1,99 +0,0 @@
-import torch
-from comfy_api.latest import ComfyExtension, io
-from typing_extensions import override
-
-
-class NAGuidance(io.ComfyNode):
-    @classmethod
-    def define_schema(cls) -> io.Schema:
-        return io.Schema(
-            node_id="NAGuidance",
-            display_name="Normalized Attention Guidance",
-            description="Applies Normalized Attention Guidance to models, enabling negative prompts on distilled/schnell models.",
-            category="",
-            is_experimental=True,
-            inputs=[
-                io.Model.Input("model", tooltip="The model to apply NAG to."),
-                io.Float.Input("nag_scale", min=0.0, default=5.0, max=50.0, step=0.1, tooltip="The guidance scale factor. Higher values push further from the negative prompt."),
-                io.Float.Input("nag_alpha", min=0.0, default=0.5, max=1.0, step=0.01, tooltip="Blending factor for the normalized attention. 1.0 is full replacement, 0.0 is no effect."),
-                io.Float.Input("nag_tau", min=1.0, default=1.5, max=10.0, step=0.01),
-                # io.Float.Input("start_percent", min=0.0, default=0.0, max=1.0, step=0.01, tooltip="The relative sampling step to begin applying NAG."),
-                # io.Float.Input("end_percent", min=0.0, default=1.0, max=1.0, step=0.01, tooltip="The relative sampling step to stop applying NAG."),
-            ],
-            outputs=[
-                io.Model.Output(tooltip="The patched model with NAG enabled."),
-            ],
-        )
-
-    @classmethod
-    def execute(cls, model: io.Model.Type, nag_scale: float, nag_alpha: float, nag_tau: float) -> io.NodeOutput:
-        m = model.clone()
-
-        # sigma_start = m.get_model_object("model_sampling").percent_to_sigma(start_percent)
-        # sigma_end = m.get_model_object("model_sampling").percent_to_sigma(end_percent)
-
-        def nag_attention_output_patch(out, extra_options):
-            cond_or_uncond = extra_options.get("cond_or_uncond", None)
-            if cond_or_uncond is None:
-                return out
-
-            if not (1 in cond_or_uncond and 0 in cond_or_uncond):
-                return out
-
-            # sigma = extra_options.get("sigmas", None)
-            # if sigma is not None and len(sigma) > 0:
-            #     sigma = sigma[0].item()
-            #     if sigma > sigma_start or sigma < sigma_end:
-            #         return out
-
-            img_slice = extra_options.get("img_slice", None)
-
-            if img_slice is not None:
-                orig_out = out
-                out = out[:, img_slice[0]:img_slice[1]]  # only apply on img part
-
-            batch_size = out.shape[0]
-            half_size = batch_size // len(cond_or_uncond)
-
-            ind_neg = cond_or_uncond.index(1)
-            ind_pos = cond_or_uncond.index(0)
-            z_pos = out[half_size * ind_pos:half_size * (ind_pos + 1)]
-            z_neg = out[half_size * ind_neg:half_size * (ind_neg + 1)]
-
-            guided = z_pos * nag_scale - z_neg * (nag_scale - 1.0)
-
-            eps = 1e-6
-            norm_pos = torch.norm(z_pos, p=1, dim=-1, keepdim=True).clamp_min(eps)
-            norm_guided = torch.norm(guided, p=1, dim=-1, keepdim=True).clamp_min(eps)
-
-            ratio = norm_guided / norm_pos
-            scale_factor = torch.minimum(ratio, torch.full_like(ratio, nag_tau)) / ratio
-
-            guided_normalized = guided * scale_factor
-
-            z_final = guided_normalized * nag_alpha + z_pos * (1.0 - nag_alpha)
-
-            if img_slice is not None:
-                orig_out[half_size * ind_neg:half_size * (ind_neg + 1), img_slice[0]:img_slice[1]] = z_final
-                orig_out[half_size * ind_pos:half_size * (ind_pos + 1), img_slice[0]:img_slice[1]] = z_final
-                return orig_out
-            else:
-                out[half_size * ind_pos:half_size * (ind_pos + 1)] = z_final
-            return out
-
-        m.set_model_attn1_output_patch(nag_attention_output_patch)
-        m.disable_model_cfg1_optimization()
-
-        return io.NodeOutput(m)
-
-
-class NagExtension(ComfyExtension):
-    @override
-    async def get_node_list(self) -> list[type[io.ComfyNode]]:
-        return [
-            NAGuidance,
-        ]
-
-
-async def comfy_entrypoint() -> NagExtension:
-    return NagExtension()

comfy_extras/nodes_post_processing.py

@@ -655,7 +655,6 @@ class BatchImagesMasksLatentsNode(io.ComfyNode):
             batched = batch_masks(values)
         return io.NodeOutput(batched)
 
-
 class PostProcessingExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:

comfy_extras/nodes_replacements.py

@@ -1,103 +0,0 @@
-from comfy_api.latest import ComfyExtension, io, ComfyAPI
-
-api = ComfyAPI()
-
-
-async def register_replacements():
-    """Register all built-in node replacements."""
-    await register_replacements_longeredge()
-    await register_replacements_batchimages()
-    await register_replacements_upscaleimage()
-    await register_replacements_controlnet()
-    await register_replacements_load3d()
-    await register_replacements_preview3d()
-    await register_replacements_svdimg2vid()
-    await register_replacements_conditioningavg()
-
-async def register_replacements_longeredge():
-    # No dynamic inputs here
-    await api.node_replacement.register(io.NodeReplace(
-            new_node_id="ImageScaleToMaxDimension",
-            old_node_id="ResizeImagesByLongerEdge",
-            old_widget_ids=["longer_edge"],
-            input_mapping=[
-                {"new_id": "image", "old_id": "images"},
-                {"new_id": "largest_size", "old_id": "longer_edge"},
-                {"new_id": "upscale_method", "set_value": "lanczos"},
-            ],
-            # just to test the frontend output_mapping code, does nothing really here
-            output_mapping=[{"new_idx": 0, "old_idx": 0}],
-        ))
-
-async def register_replacements_batchimages():
-    # BatchImages node uses Autogrow
-    await api.node_replacement.register(io.NodeReplace(
-            new_node_id="BatchImagesNode",
-            old_node_id="ImageBatch",
-            input_mapping=[
-                {"new_id": "images.image0", "old_id": "image1"},
-                {"new_id": "images.image1", "old_id": "image2"},
-            ],
-        ))
-
-async def register_replacements_upscaleimage():
-    # ResizeImageMaskNode uses DynamicCombo
-    await api.node_replacement.register(io.NodeReplace(
-            new_node_id="ResizeImageMaskNode",
-            old_node_id="ImageScaleBy",
-            old_widget_ids=["upscale_method", "scale_by"],
-            input_mapping=[
-                {"new_id": "input", "old_id": "image"},
-                {"new_id": "resize_type", "set_value": "scale by multiplier"},
-                {"new_id": "resize_type.multiplier", "old_id": "scale_by"},
-                {"new_id": "scale_method", "old_id": "upscale_method"},
-            ],
-        ))
-
-async def register_replacements_controlnet():
-    # T2IAdapterLoader → ControlNetLoader
-    await api.node_replacement.register(io.NodeReplace(
-            new_node_id="ControlNetLoader",
-            old_node_id="T2IAdapterLoader",
-            input_mapping=[
-                {"new_id": "control_net_name", "old_id": "t2i_adapter_name"},
-            ],
-        ))
-
-async def register_replacements_load3d():
-    # Load3DAnimation merged into Load3D
-    await api.node_replacement.register(io.NodeReplace(
-            new_node_id="Load3D",
-            old_node_id="Load3DAnimation",
-        ))
-
-async def register_replacements_preview3d():
-    # Preview3DAnimation merged into Preview3D
-    await api.node_replacement.register(io.NodeReplace(
-            new_node_id="Preview3D",
-            old_node_id="Preview3DAnimation",
-        ))
-
-async def register_replacements_svdimg2vid():
-    # Typo fix: SDV → SVD
-    await api.node_replacement.register(io.NodeReplace(
-            new_node_id="SVD_img2vid_Conditioning",
-            old_node_id="SDV_img2vid_Conditioning",
-        ))
-
-async def register_replacements_conditioningavg():
-    # Typo fix: trailing space in node name
-    await api.node_replacement.register(io.NodeReplace(
-            new_node_id="ConditioningAverage",
-            old_node_id="ConditioningAverage ",
-        ))
-
-class NodeReplacementsExtension(ComfyExtension):
-    async def on_load(self) -> None:
-        await register_replacements()
-
-    async def get_node_list(self) -> list[type[io.ComfyNode]]:
-        return []
-
-async def comfy_entrypoint() -> NodeReplacementsExtension:
-    return NodeReplacementsExtension()

nodes.py

@@ -2264,7 +2264,6 @@ async def load_custom_node(module_path: str, ignore=set(), module_parent="custom
                 if not isinstance(extension, ComfyExtension):
                     logging.warning(f"comfy_entrypoint in {module_path} did not return a ComfyExtension, skipping.")
                     return False
-                await extension.on_load()
                 node_list = await extension.get_node_list()
                 if not isinstance(node_list, list):
                     logging.warning(f"comfy_entrypoint in {module_path} did not return a list of nodes, skipping.")
@@ -2433,11 +2432,10 @@ async def init_builtin_extra_nodes():
         "nodes_wanmove.py",
         "nodes_image_compare.py",
         "nodes_zimage.py",
+        "nodes_glsl.py",
         "nodes_lora_debug.py",
         "nodes_color.py",
         "nodes_toolkit.py",
-        "nodes_replacements.py",
-        "nodes_nag.py",
     ]
 
     import_failed = []

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.38.14
-comfyui-workflow-templates==0.8.42
+comfyui-workflow-templates==0.8.38
 comfyui-embedded-docs==0.4.1
 torch
 torchsde
@@ -30,3 +30,6 @@ kornia>=0.7.1
 spandrel
 pydantic~=2.0
 pydantic-settings~=2.0
+PyOpenGL
+PyOpenGL-accelerate
+glfw

server.py

@@ -40,7 +40,6 @@ from app.user_manager import UserManager
 from app.model_manager import ModelFileManager
 from app.custom_node_manager import CustomNodeManager
 from app.subgraph_manager import SubgraphManager
-from app.node_replace_manager import NodeReplaceManager
 from typing import Optional, Union
 from api_server.routes.internal.internal_routes import InternalRoutes
 from protocol import BinaryEventTypes
@@ -205,7 +204,6 @@ class PromptServer():
         self.model_file_manager = ModelFileManager()
         self.custom_node_manager = CustomNodeManager()
         self.subgraph_manager = SubgraphManager()
-        self.node_replace_manager = NodeReplaceManager()
         self.internal_routes = InternalRoutes(self)
         self.supports = ["custom_nodes_from_web"]
         self.prompt_queue = execution.PromptQueue(self)
@@ -889,8 +887,6 @@ class PromptServer():
                 if "partial_execution_targets" in json_data:
                     partial_execution_targets = json_data["partial_execution_targets"]
 
-                self.node_replace_manager.apply_replacements(prompt)
-
                 valid = await execution.validate_prompt(prompt_id, prompt, partial_execution_targets)
                 extra_data = {}
                 if "extra_data" in json_data:
@@ -999,7 +995,6 @@ class PromptServer():
         self.model_file_manager.add_routes(self.routes)
         self.custom_node_manager.add_routes(self.routes, self.app, nodes.LOADED_MODULE_DIRS.items())
         self.subgraph_manager.add_routes(self.routes, nodes.LOADED_MODULE_DIRS.items())
-        self.node_replace_manager.add_routes(self.routes)
         self.app.add_subapp('/internal', self.internal_routes.get_app())
 
         # Prefix every route with /api for easier matching for delegation.