reze-engine 0.11.0 → 0.11.1

package/src/shaders/eye.ts

@@ -60,23 +60,31 @@ fn smith_g1(ndotx: f32, a2: f32) -> f32 {
 }
 
 fn fresnel_schlick(cosTheta: f32, f0: f32) -> f32 {
-  return f0 + (1.0 - f0) * pow(1.0 - cosTheta, 5.0);
+  let m = 1.0 - cosTheta;
+  let m2 = m * m;
+  return f0 + (1.0 - f0) * (m2 * m2 * m);
 }
 
 fn sampleShadow(worldPos: vec3f, n: vec3f) -> f32 {
+  // Back-facing to key light: direct contribution is zero anyway, skip 9 texture samples.
+  if (dot(n, -light.lights[0].direction.xyz) <= 0.0) { return 0.0; }
   let biasedPos = worldPos + n * 0.08;
   let lclip = lightVP.viewProj * vec4f(biasedPos, 1.0);
   let ndc = lclip.xyz / max(lclip.w, 1e-6);
   let suv = vec2f(ndc.x * 0.5 + 0.5, 0.5 - ndc.y * 0.5);
   let cmpZ = ndc.z - 0.001;
-  let ts = 1.0 / 4096.0;
-  var vis = 0.0;
-  for (var y = -1; y <= 1; y++) {
-    for (var x = -1; x <= 1; x++) {
-      vis += textureSampleCompare(shadowMap, shadowSampler, suv + vec2f(f32(x), f32(y)) * ts, cmpZ);
-    }
-  }
-  return vis / 9.0;
+  let ts = 1.0 / 2048.0;
+  // 3x3 PCF unrolled — Safari's Metal backend doesn't unroll nested shadow loops reliably.
+  let s00 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, -ts), cmpZ);
+  let s10 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0, -ts), cmpZ);
+  let s20 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, -ts), cmpZ);
+  let s01 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, 0.0), cmpZ);
+  let s11 = textureSampleCompareLevel(shadowMap, shadowSampler, suv, cmpZ);
+  let s21 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, 0.0), cmpZ);
+  let s02 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts,  ts), cmpZ);
+  let s12 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0,  ts), cmpZ);
+  let s22 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts,  ts), cmpZ);
+  return (s00 + s10 + s20 + s01 + s11 + s21 + s02 + s12 + s22) * (1.0 / 9.0);
 }
 
 @vertex fn vs(
@@ -100,13 +108,19 @@ fn sampleShadow(worldPos: vec3f, n: vec3f) -> f32 {
     skinnedNrm += (mat3x3f(m[0].xyz, m[1].xyz, m[2].xyz) * normal) * w;
   }
   output.position = camera.projection * camera.view * vec4f(skinnedPos.xyz, 1.0);
-  output.normal = normalize(skinnedNrm);
+  // Skip VS normalize — interpolation denormalizes anyway, and FS always does normalize(input.normal).
+  output.normal = skinnedNrm;
   output.uv = uv;
   output.worldPos = skinnedPos.xyz;
   return output;
 }
 
-@fragment fn fs(input: VertexOutput) -> @location(0) vec4f {
+struct FSOut {
+  @location(0) color: vec4f,
+  @location(1) mask: f32,
+};
+
+@fragment fn fs(input: VertexOutput) -> FSOut {
   let alpha = material.alpha;
   if (alpha < 0.001) { discard; }
 
@@ -137,7 +151,10 @@ fn sampleShadow(worldPos: vec3f, n: vec3f) -> f32 {
   // Principled Emission socket: emissive = emission_color × strength, added on top of shading.
   let emission = albedo * EYE_EMISSION_STRENGTH;
 
-  return vec4f(ambient + direct + emission, alpha);
+  var out: FSOut;
+  out.color = vec4f(ambient + direct + emission, alpha);
+  out.mask = 1.0;
+  return out;
 }
 
 `

package/src/shaders/face.ts

@@ -47,21 +47,27 @@ struct LightVP { viewProj: mat4x4f, };
 @group(2) @binding(0) var diffuseTexture: texture_2d<f32>;
 @group(2) @binding(1) var<uniform> material: MaterialUniforms;
 
-// 3x3 PCF shadow sampling, 4096 map, normal-bias 0.08, depth-bias 0.001
+// 3x3 PCF shadow sampling, 2048 map, normal-bias 0.08, depth-bias 0.001
 fn sampleShadow(worldPos: vec3f, n: vec3f) -> f32 {
+  // Back-facing to key light: direct contribution is zero anyway, skip 9 texture samples.
+  if (dot(n, -light.lights[0].direction.xyz) <= 0.0) { return 0.0; }
   let biasedPos = worldPos + n * 0.08;
   let lclip = lightVP.viewProj * vec4f(biasedPos, 1.0);
   let ndc = lclip.xyz / max(lclip.w, 1e-6);
   let suv = vec2f(ndc.x * 0.5 + 0.5, 0.5 - ndc.y * 0.5);
   let cmpZ = ndc.z - 0.001;
-  let ts = 1.0 / 4096.0;
-  var vis = 0.0;
-  for (var y = -1; y <= 1; y++) {
-    for (var x = -1; x <= 1; x++) {
-      vis += textureSampleCompare(shadowMap, shadowSampler, suv + vec2f(f32(x), f32(y)) * ts, cmpZ);
-    }
-  }
-  return vis / 9.0;
+  let ts = 1.0 / 2048.0;
+  // 3x3 PCF unrolled — Safari's Metal backend doesn't unroll nested shadow loops reliably.
+  let s00 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, -ts), cmpZ);
+  let s10 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0, -ts), cmpZ);
+  let s20 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, -ts), cmpZ);
+  let s01 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, 0.0), cmpZ);
+  let s11 = textureSampleCompareLevel(shadowMap, shadowSampler, suv, cmpZ);
+  let s21 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, 0.0), cmpZ);
+  let s02 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts,  ts), cmpZ);
+  let s12 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0,  ts), cmpZ);
+  let s22 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts,  ts), cmpZ);
+  return (s00 + s10 + s20 + s01 + s11 + s21 + s02 + s12 + s22) * (1.0 / 9.0);
 }
 
 const PI_F: f32 = 3.141592653589793;
@@ -73,6 +79,10 @@ const FACE_RIM2_BG: vec3f = vec3f(1.0, 0.4684903025627136, 0.3698573112487793);
 const FACE_WARM_AO_MUL: f32 = 0.30000001192092896; // 运算.004 MULTIPLY after invert (was 0.5 in older trace)
 const FACE_BRIGHT_TEX_THRESH: f32 = 0.9300000071525574; // 运算.005 GREATER_THAN Value_001
 const FACE_MIX_NPR: f32 = 0.5; // 混合着色器.001 Fac
+// EEVEE Light Clamp equivalent (Render Props → Sampling → Clamping). Caps direct
+// specular firefly from the noise-bumped normal's NDF aliasing — Blender hides this
+// via TAA, which we don't have. Value mirrors EEVEE's default Clamp Indirect=10.0.
+const FACE_SPEC_CLAMP: f32 = 10.0;
 
 @vertex fn vs(
   @location(0) position: vec3f,
@@ -95,7 +105,8 @@ const FACE_MIX_NPR: f32 = 0.5; // 混合着色器.001 Fac
     skinnedNrm += (mat3x3f(m[0].xyz, m[1].xyz, m[2].xyz) * normal) * w;
   }
   output.position = camera.projection * camera.view * vec4f(skinnedPos.xyz, 1.0);
-  output.normal = normalize(skinnedNrm);
+  // Skip VS normalize — interpolation denormalizes anyway, and FS always does normalize(input.normal).
+  output.normal = skinnedNrm;
   output.uv = uv;
   output.worldPos = skinnedPos.xyz;
   return output;
@@ -105,7 +116,12 @@ const FACE_MIX_NPR: f32 = 0.5; // 混合着色器.001 Fac
 // TEX → HueSat shadow/lit → toon gate → BrightContrast → AO chain → emission stack
 // Fresnel rims, warm AO emission, bright-texture gate, noise-bumped Principled
 // Final = mix(Principled, NPR, 0.5)
-@fragment fn fs(input: VertexOutput) -> @location(0) vec4f {
+struct FSOut {
+  @location(0) color: vec4f,
+  @location(1) mask: f32,
+};
+
+@fragment fn fs(input: VertexOutput) -> FSOut {
   let alpha = material.alpha;
   if (alpha < 0.001) { discard; }
 
@@ -123,7 +139,7 @@ const FACE_MIX_NPR: f32 = 0.5; // 混合着色器.001 Fac
   let ndotl_raw = shader_to_rgb_diffuse(n, l, sun, light.ambientColor.xyz, shadow);
   // ramp.008 CONSTANT — edge AA avoids binary fac shimmer / white specks on terminator (fwidth + smoothstep)
   let toon = ramp_constant_edge_aa(ndotl_raw, 0.2966, vec4f(0,0,0,1), vec4f(1,1,1,1)).r;
-  let ao = ao_fake(n, v);
+  let ao = 1.0; // ao_fake(n, v) — no SSAO yet; inline 1.0 so the ramp/mix chain folds at compile time.
 
   // ═══ TOON COLOR ═══
   let shadow_tint = hue_sat(0.46000000834465027, 2.0, 0.3499999940395355, 1.0, tex_color); // HueSat.002
@@ -163,7 +179,12 @@ const FACE_MIX_NPR: f32 = 0.5; // 混合着色器.001 Fac
   let rim2_mixed = mix(emission3, FACE_RIM2_BG, rim2_fac);
 
   // 转接点.005(tex) → 运算.005 GREATER_THAN Value_001
-  let tex_gate = math_greater_than(tex_color.r, FACE_BRIGHT_TEX_THRESH);
+  // Blender implicitly converts Color → Float via BT.601 grayscale when plugging a
+  // color output into a Math node's Value input. Our earlier trace used tex_color.r,
+  // which fires aggressively on R-dominant skin — single near-white R pixels produced
+  // firefly speckles. color_to_value matches the actual Blender socket semantic and
+  // only fires on genuinely near-white painted features (the author's intent).
+  let tex_gate = math_greater_than(color_to_value(tex_color), FACE_BRIGHT_TEX_THRESH);
   let bright_emit = vec3f(tex_gate) * 3.0; // Emission.002(Strength=3.0)
 
   // ═══ NPR STACK (AddShader chain) ═══
@@ -172,9 +193,8 @@ const FACE_MIX_NPR: f32 = 0.5; // 混合着色器.001 Fac
   let npr_stack = add0 + warm_emission; // AddShader.001
 
   // ═══ PRINCIPLED BSDF ═══
-  // Noise-based bump normal
-  let gen = mapping_point(input.worldPos, vec3f(0.0), vec3f(0.0), vec3f(1.0, 1.0, 1.5));
-  let noise_val = tex_noise(gen, 1.0, 2.0, 0.5, 0.0);
+  // Noise-based bump normal. Mapping loc=rot=0 → plain scale multiply, inline.
+  let noise_val = tex_noise_d2(input.worldPos * vec3f(1.0, 1.0, 1.5), 1.0);
   let noise_ramp = ramp_linear(noise_val, 0.0, vec4f(0,0,0,1), 1.0, vec4f(1,1,1,1)).r;
   let bumped_n = bump_lh(0.324644535779953, noise_ramp, n, input.worldPos); // 凹凸 Strength; LH bump
 
@@ -182,9 +202,9 @@ const FACE_MIX_NPR: f32 = 0.5; // 混合着色器.001 Fac
   let principled_base = mix_blend(noise_ramp, bc, vec3f(0.6832, 0.1947, 0.1373));
   // Emission input from reroute.011 (bc), Strength=0.2
   let p_emission = bc * 0.2;
-  // AO.002 → ramp.005 LINEAR [0.003→black, 1.0→gray] for subsurface approx
-  let ao2 = ao_fake(n, v);
-  let sss = ramp_linear(ao2, 0.003, vec4f(0,0,0,1), 1.0, vec4f(0.0786, 0.0786, 0.0786, 1.0)).r;
+  // AO.002 → ramp.005 LINEAR [0.003→black, 1.0→gray] for subsurface approx.
+  // Reuse 'ao' (ao_fake(n, v) above) — identical inputs, avoid a second procedural AO pass.
+  let sss = ramp_linear(ao, 0.003, vec4f(0,0,0,1), 1.0, vec4f(0.0786, 0.0786, 0.0786, 1.0)).r;
 
   // 原理化BSDF (EEVEE port): metallic=0, specular=0.5, roughness=0.3, specular_tint=0.
   let NL = max(dot(bumped_n, l), 0.0);
@@ -192,10 +212,11 @@ const FACE_MIX_NPR: f32 = 0.5; // 混合着色器.001 Fac
 
   let f0 = vec3f(0.08 * FACE_SPECULAR);
   let f90 = mix(f0, vec3f(1.0), sqrt(FACE_SPECULAR));
-  let split_sum = brdf_lut_baked(NV, FACE_ROUGHNESS);
-  let reflection_color = F_brdf_multi_scatter(f0, f90, split_sum);
+  let brdf_lut = brdf_lut_sample(NV, FACE_ROUGHNESS);
+  let reflection_color = F_brdf_multi_scatter(f0, f90, brdf_lut.xy);
 
-  let spec_direct = bsdf_ggx(bumped_n, l, v, FACE_ROUGHNESS) * sun * shadow * ltc_brdf_scale(NV, FACE_ROUGHNESS);
+  let spec_direct_raw = bsdf_ggx(bumped_n, l, v, NL, NV, FACE_ROUGHNESS) * sun * shadow * ltc_brdf_scale_from_lut(brdf_lut);
+  let spec_direct = min(spec_direct_raw, vec3f(FACE_SPEC_CLAMP));
   let spec_indirect = light.ambientColor.xyz;
   let spec_radiance = (spec_direct + spec_indirect) * reflection_color;
 
@@ -207,7 +228,10 @@ const FACE_MIX_NPR: f32 = 0.5; // 混合着色器.001 Fac
   // 混合着色器.001: Shader=相加着色器.001, Shader_001=原理化BSDF — Fac blends toward second
   let final_color = mix(npr_stack, principled, FACE_MIX_NPR);
 
-  return vec4f(final_color, alpha);
+  var out: FSOut;
+  out.color = vec4f(final_color, alpha);
+  out.mask = 1.0;
+  return out;
 }
 
 `

package/src/shaders/hair.ts

@@ -49,19 +49,25 @@ struct LightVP { viewProj: mat4x4f, };
 @group(2) @binding(1) var<uniform> material: MaterialUniforms;
 
 fn sampleShadow(worldPos: vec3f, n: vec3f) -> f32 {
+  // Back-facing to key light: direct contribution is zero anyway, skip 9 texture samples.
+  if (dot(n, -light.lights[0].direction.xyz) <= 0.0) { return 0.0; }
   let biasedPos = worldPos + n * 0.08;
   let lclip = lightVP.viewProj * vec4f(biasedPos, 1.0);
   let ndc = lclip.xyz / max(lclip.w, 1e-6);
   let suv = vec2f(ndc.x * 0.5 + 0.5, 0.5 - ndc.y * 0.5);
   let cmpZ = ndc.z - 0.001;
-  let ts = 1.0 / 4096.0;
-  var vis = 0.0;
-  for (var y = -1; y <= 1; y++) {
-    for (var x = -1; x <= 1; x++) {
-      vis += textureSampleCompare(shadowMap, shadowSampler, suv + vec2f(f32(x), f32(y)) * ts, cmpZ);
-    }
-  }
-  return vis / 9.0;
+  let ts = 1.0 / 2048.0;
+  // 3x3 PCF unrolled — Safari's Metal backend doesn't unroll nested shadow loops reliably.
+  let s00 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, -ts), cmpZ);
+  let s10 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0, -ts), cmpZ);
+  let s20 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, -ts), cmpZ);
+  let s01 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, 0.0), cmpZ);
+  let s11 = textureSampleCompareLevel(shadowMap, shadowSampler, suv, cmpZ);
+  let s21 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, 0.0), cmpZ);
+  let s02 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts,  ts), cmpZ);
+  let s12 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0,  ts), cmpZ);
+  let s22 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts,  ts), cmpZ);
+  return (s00 + s10 + s20 + s01 + s11 + s21 + s02 + s12 + s22) * (1.0 / 9.0);
 }
 
 const PI_H: f32 = 3.141592653589793;
@@ -93,13 +99,19 @@ const HAIR_MIX_BG: vec3f = vec3f(0.1673291176557541);
     skinnedNrm += (mat3x3f(m[0].xyz, m[1].xyz, m[2].xyz) * normal) * w;
   }
   output.position = camera.projection * camera.view * vec4f(skinnedPos.xyz, 1.0);
-  output.normal = normalize(skinnedNrm);
+  // Skip VS normalize — interpolation denormalizes anyway, and FS always does normalize(input.normal).
+  output.normal = skinnedNrm;
   output.uv = uv;
   output.worldPos = skinnedPos.xyz;
   return output;
 }
 
-@fragment fn fs(input: VertexOutput) -> @location(0) vec4f {
+struct FSOut {
+  @location(0) color: vec4f,
+  @location(1) mask: f32,
+};
+
+@fragment fn fs(input: VertexOutput) -> FSOut {
   let alpha = material.alpha;
   if (alpha < 0.001) { discard; }
 
@@ -113,11 +125,11 @@ const HAIR_MIX_BG: vec3f = vec3f(0.1673291176557541);
   let shadow = sampleShadow(input.worldPos, n);
 
   // 色相/饱和度/明度 (Hue=0.5 Sat=1.2 Val=0.5 Fac=1) ← reroute from image
-  let hue_sat_shadow = hue_sat(0.5, 1.2, 0.5, 1.0, tex_color);
+  let hue_sat_shadow = hue_sat_id(1.2, 0.5, 1.0, tex_color);
   // 色相/饱和度/明度.002 (0.48, 1.2, 0.7, 1) ← previous
   let hue_sat_002 = hue_sat(0.48, 1.2, 0.7, 1.0, hue_sat_shadow);
   // 色相/饱和度/明度.001 (0.5, 1.5, 1.0, 1) ← image reroute (lit path)
-  let hue_sat_001 = hue_sat(0.5, 1.5, 1.0, 1.0, tex_color);
+  let hue_sat_001 = hue_sat_id(1.5, 1.0, 1.0, tex_color);
 
   // 漫射 BSDF.002 → Shader --> RGB → 颜色渐变.008 CONSTANT [0→0, 0.2966→1]
   let ndotl_raw = shader_to_rgb_diffuse(n, l, sun, light.ambientColor.xyz, shadow);
@@ -133,19 +145,10 @@ const HAIR_MIX_BG: vec3f = vec3f(0.1673291176557541);
   let bevel_z = clamp(n.y, 0.0, 1.0);
   let mix_003 = mix_blend(bevel_z, bc, hue_sat_002);
 
-  // 环境光遮蔽 (AO).001 → 颜色渐变.001 CONSTANT [0→1, 0.3756→0] → 混合.001 → ao_factor
-  let ao = ao_fake(n, v);
-  let ramp_001 = ramp_constant(ao, 0.0, vec4f(1,1,1,1), 0.3756, vec4f(0,0,0,1)).r;
-  let ao_factor = mix(1.0, 0.0, ramp_001);
-
-  // 色相/饱和度/明度.004 (0.5, 0.8, 0.1, 1) ← mix_003
-  let hue_sat_004 = hue_sat(0.5, 0.8, 0.1, 1.0, mix_003);
-
-  // 混合.002 MIX Fac=ao_factor, A=hue_sat_004, B=mix_003
-  let mix_002 = mix_blend(ao_factor, hue_sat_004, mix_003);
-
-  // 自发光(发射).003 Strength=1.0 ← mix_002
-  let emission3 = mix_002 * 1.0;
+  // 环境光遮蔽 (AO).001 → 颜色渐变.001 → 混合.001 → 混合.002 chain collapses with fake AO=1:
+  //   ramp_constant(1, 0→white, 0.3756→black).r = 0 → ao_factor = mix(1,0,0) = 1 → mix_002 = mix_003.
+  //   hue_sat_004 becomes unreachable. When real SSAO lands, restore the original 5-line port.
+  let emission3 = mix_003; // Emission.003 Strength=1.0 (×1 omitted)
 
   // 菲涅尔.001 × 层权重.002 → 运算.003 MULTIPLY → 运算.007 POWER(exponent Value_001) → MixShader.002 Fac
   let rim2_raw = fresnel(1.45, n, v) * layer_weight_fresnel(0.61, n, v);
@@ -153,24 +156,28 @@ const HAIR_MIX_BG: vec3f = vec3f(0.1673291176557541);
   // MixShader.002: Shader=Emission.003, Shader_001=背景 — (1-Fac)*emission + Fac*bg
   let mix_shader_002 = mix(emission3, HAIR_MIX_BG, rim2_fac);
 
-  // 运算.004 GREATER_THAN: 图像→Value, threshold Value_001 (R when Color plugs float socket)
-  let tex_gate = math_greater_than(tex_color.r, HAIR_TEX_GATE_THRESH);
+  // 运算.004 GREATER_THAN: 图像→Value, threshold Value_001. Blender converts Color→Float
+  // via BT.601 luminance, not raw R — same socket-semantic fix as M_Face.
+  let tex_gate = math_greater_than(color_to_value(tex_color), HAIR_TEX_GATE_THRESH);
   let gate_emit = vec3f(tex_gate) * 0.1;
 
   // 相加着色器: MixShader.002 + gate emission (color sum in linear space)
   let add_shader = mix_shader_002 + gate_emit;
 
   // 原理化BSDF (EEVEE port): metallic=0, specular=1.0, roughness=0.3, specular_tint=0.
-  // Graph's 噪波→法线贴图 Strength=0.1 is near-identity; plain n matches visually.
+  // Blender graph has 噪波→法线贴图 Strength=0.1 on Principled.Normal, but MixShader.001
+  // weights Principled at only 0.2; spec contribution × that weight is imperceptible in
+  // A/B with the noise-bump port enabled, so we drop it and keep plain n — saves a full
+  // tex_noise + bump_lh per hair fragment.
   let NL = max(dot(n, l), 0.0);
   let NV = max(dot(n, v), 1e-4);
 
   let f0 = vec3f(0.08 * HAIR_SPECULAR);
   let f90 = mix(f0, vec3f(1.0), sqrt(HAIR_SPECULAR));
-  let split_sum = brdf_lut_baked(NV, HAIR_ROUGHNESS);
-  let reflection_color = F_brdf_multi_scatter(f0, f90, split_sum);
+  let brdf_lut = brdf_lut_sample(NV, HAIR_ROUGHNESS);
+  let reflection_color = F_brdf_multi_scatter(f0, f90, brdf_lut.xy);
 
-  let spec_direct = bsdf_ggx(n, l, v, HAIR_ROUGHNESS) * sun * shadow * ltc_brdf_scale(NV, HAIR_ROUGHNESS);
+  let spec_direct = bsdf_ggx(n, l, v, NL, NV, HAIR_ROUGHNESS) * sun * shadow * ltc_brdf_scale_from_lut(brdf_lut);
   let spec_indirect = light.ambientColor.xyz;
   let spec_radiance = (spec_direct + spec_indirect) * reflection_color;
 
@@ -182,7 +189,10 @@ const HAIR_MIX_BG: vec3f = vec3f(0.1673291176557541);
   // 混合着色器.001 Fac=0.2: first socket=相加着色器, second=原理化BSDF
   let final_color = mix(add_shader, principled, 0.2);
 
-  return vec4f(final_color, alpha);
+  var out: FSOut;
+  out.color = vec4f(final_color, alpha);
+  out.mask = 1.0;
+  return out;
 }
 
 `

package/src/shaders/metal.ts

@@ -51,19 +51,25 @@ struct LightVP { viewProj: mat4x4f, };
 @group(2) @binding(1) var<uniform> material: MaterialUniforms;
 
 fn sampleShadow(worldPos: vec3f, n: vec3f) -> f32 {
+  // Back-facing to key light: direct contribution is zero anyway, skip 9 texture samples.
+  if (dot(n, -light.lights[0].direction.xyz) <= 0.0) { return 0.0; }
   let biasedPos = worldPos + n * 0.08;
   let lclip = lightVP.viewProj * vec4f(biasedPos, 1.0);
   let ndc = lclip.xyz / max(lclip.w, 1e-6);
   let suv = vec2f(ndc.x * 0.5 + 0.5, 0.5 - ndc.y * 0.5);
   let cmpZ = ndc.z - 0.001;
-  let ts = 1.0 / 4096.0;
-  var vis = 0.0;
-  for (var y = -1; y <= 1; y++) {
-    for (var x = -1; x <= 1; x++) {
-      vis += textureSampleCompare(shadowMap, shadowSampler, suv + vec2f(f32(x), f32(y)) * ts, cmpZ);
-    }
-  }
-  return vis / 9.0;
+  let ts = 1.0 / 2048.0;
+  // 3x3 PCF unrolled — Safari's Metal backend doesn't unroll nested shadow loops reliably.
+  let s00 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, -ts), cmpZ);
+  let s10 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0, -ts), cmpZ);
+  let s20 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, -ts), cmpZ);
+  let s01 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts, 0.0), cmpZ);
+  let s11 = textureSampleCompareLevel(shadowMap, shadowSampler, suv, cmpZ);
+  let s21 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts, 0.0), cmpZ);
+  let s02 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(-ts,  ts), cmpZ);
+  let s12 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f(0.0,  ts), cmpZ);
+  let s22 = textureSampleCompareLevel(shadowMap, shadowSampler, suv + vec2f( ts,  ts), cmpZ);
+  return (s00 + s10 + s20 + s01 + s11 + s21 + s02 + s12 + s22) * (1.0 / 9.0);
 }
 
 const PI_M: f32 = 3.141592653589793;
@@ -98,13 +104,19 @@ const METAL_VORONOI_SCALE: f32 = 4.3;
     skinnedNrm += (mat3x3f(m[0].xyz, m[1].xyz, m[2].xyz) * normal) * w;
   }
   output.position = camera.projection * camera.view * vec4f(skinnedPos.xyz, 1.0);
-  output.normal = normalize(skinnedNrm);
+  // Skip VS normalize — interpolation denormalizes anyway, and FS always does normalize(input.normal).
+  output.normal = skinnedNrm;
   output.uv = uv;
   output.worldPos = skinnedPos.xyz;
   return output;
 }
 
-@fragment fn fs(input: VertexOutput) -> @location(0) vec4f {
+struct FSOut {
+  @location(0) color: vec4f,
+  @location(1) mask: f32,
+};
+
+@fragment fn fs(input: VertexOutput) -> FSOut {
   let n = normalize(input.normal);
   let v = normalize(camera.viewPos - input.worldPos);
   let l = -light.lights[0].direction.xyz;
@@ -118,23 +130,23 @@ const METAL_VORONOI_SCALE: f32 = 4.3;
   if (out_alpha < 0.001) { discard; }
 
   // ═══ NPR toon stack (图像 → HSV.007 Val=0.8 → 转接点.001) ═══
-  let tex_tint = hue_sat(0.5, 1.0, 0.800000011920929, 1.0, tex_rgb);
+  let tex_tint = hue_sat_id(1.0, 0.800000011920929, 1.0, tex_rgb);
   let lum_shade = shader_to_rgb_diffuse(n, l, sun, amb, shadow);
   let ramp008 = ramp_constant_edge_aa(lum_shade, METAL_TOON_EDGE, vec4f(0,0,0,1), vec4f(1,1,1,1));
   let mix04_fac = math_multiply(ramp008.r, METAL_MIX04_MUL);
 
   // 混合.004: A=HSV.002(Val=0.2 dark), B=tex_tint
-  let dark_tex = hue_sat(0.5, 1.0, 0.19999998807907104, 1.0, tex_tint);
+  let dark_tex = hue_sat_id(1.0, 0.19999998807907104, 1.0, tex_tint);
   let mix04 = mix_blend(mix04_fac, dark_tex, tex_tint);
 
   // AO white/black ramp → 混合.002 factor
-  let ao = ao_fake(n, v);
+  let ao = 1.0; // ao_fake(n, v) — no SSAO yet; inline 1.0 so the ramp/mix chain folds at compile time.
   let ao_ramp_c = ramp_linear(ao, 0.0, vec4f(1,1,1,1), 0.8808, vec4f(0,0,0,1));
   let overlay_fac = mix(1.0, 0.0, ao_ramp_c.r);
 
   // 混合.002 OVERLAY: A=HSV.008(Val=1.0 identity) ← mix04, B=HSV.004(Val=2.0 bright) ← mix04
   let hue008 = mix04; // identity HSV
-  let hue004 = hue_sat(0.5, 1.0, 2.0, 1.0, mix04);
+  let hue004 = hue_sat_id(1.0, 2.0, 1.0, mix04);
   let npr_rgb = mix_overlay(overlay_fac, hue008, hue004);
   let npr_emission = npr_rgb * METAL_EMIT_STR;
 
@@ -145,7 +157,7 @@ const METAL_VORONOI_SCALE: f32 = 4.3;
   let voro = tex_voronoi_f1(refl_dir, METAL_VORONOI_SCALE);
   let voro_ramp = ramp_linear(voro, 0.0, vec4f(0,0,0,1), 1.0, vec4f(1,1,1,1)).r;
   // 混合.005: Fac=voro_ramp, A=voro_color(grayscale), B=HSV.006(Hue=0.5 Sat=1.5 Val=1.3)
-  let hue006 = hue_sat(0.5, 1.5, 1.2999999523162842, 1.0, tex_tint);
+  let hue006 = hue_sat_id(1.5, 1.2999999523162842, 1.0, tex_tint);
   let albedo = mix_blend(voro_ramp, vec3f(voro_ramp), hue006);
 
   // 原理化BSDF (EEVEE port): metallic=1.0, specular=1.0, roughness=0.3.
@@ -153,11 +165,12 @@ const METAL_VORONOI_SCALE: f32 = 4.3;
   // metallic=1 this is just albedo (specular_tint is dielectric-only and ignored here).
   let f0 = albedo;
   let f90 = mix(f0, vec3f(1.0), sqrt(METAL_SPECULAR));
+  let NL = max(dot(n, l), 0.0);
   let NV = max(dot(n, v), 1e-4);
-  let split_sum = brdf_lut_baked(NV, METAL_ROUGHNESS);
-  let reflection_color = F_brdf_multi_scatter(f0, f90, split_sum);
+  let brdf_lut = brdf_lut_sample(NV, METAL_ROUGHNESS);
+  let reflection_color = F_brdf_multi_scatter(f0, f90, brdf_lut.xy);
 
-  let spec_direct = bsdf_ggx(n, l, v, METAL_ROUGHNESS) * sun * shadow * ltc_brdf_scale(NV, METAL_ROUGHNESS);
+  let spec_direct = bsdf_ggx(n, l, v, NL, NV, METAL_ROUGHNESS) * sun * shadow * ltc_brdf_scale_from_lut(brdf_lut);
   let spec_indirect = amb;
   let spec_radiance = (spec_direct + spec_indirect) * reflection_color;
 
@@ -167,7 +180,10 @@ const METAL_VORONOI_SCALE: f32 = 4.3;
   // 混合着色器.001 Fac=0.6967: Shader=npr_emission, Shader_001=principled
   let final_color = mix(npr_emission, principled, METAL_MIX_SHADER_FAC);
 
-  return vec4f(final_color, out_alpha);
+  var out: FSOut;
+  out.color = vec4f(final_color, out_alpha);
+  out.mask = 1.0;
+  return out;
 }
 
 `