diff --git a/piet-gpu/shader/kernel4.comp b/piet-gpu/shader/kernel4.comp
index b0c2a60..b7834cf 100644
--- a/piet-gpu/shader/kernel4.comp
+++ b/piet-gpu/shader/kernel4.comp
@@ -97,12 +97,12 @@ void main() {
 
     uvec2 xy_uint = uvec2(gl_LocalInvocationID.x + TILE_WIDTH_PX * gl_WorkGroupID.x, gl_LocalInvocationID.y + TILE_HEIGHT_PX * gl_WorkGroupID.y);
     vec2 xy = vec2(xy_uint);
-    vec3 rgb[CHUNK];
+    vec4 rgba[CHUNK];
     for (uint i = 0; i < CHUNK; i++) {
-        rgb[i] = vec3(0.5);
+        rgba[i] = vec4(0.0);
 #ifdef ENABLE_IMAGE_INDICES
         if (xy_uint.x < 1024 && xy_uint.y < 1024) {
-            rgb[i] = imageLoad(images[gl_WorkGroupID.x / 64], ivec2(xy_uint + chunk_offset(i))/4).rgb;
+            rgba[i] = imageLoad(images[gl_WorkGroupID.x / 64], ivec2(xy_uint + chunk_offset(i))/4);
         }
 #endif
     }
@@ -187,7 +187,7 @@ void main() {
             vec4 fg = unpacksRGB(color.rgba_color);
             for (uint k = 0; k < CHUNK; k++) {
                 vec4 fg_k = fg * area[k];
-                rgb[k] = rgb[k] * (1.0 - fg_k.a) + fg_k.rgb;
+                rgba[k] = rgba[k] * (1.0 - fg_k.a) + fg_k;
             }
             cmd_ref.offset += 4 + CmdColor_size;
             break;
@@ -196,7 +196,7 @@ void main() {
             vec4 img[CHUNK] = fillImage(xy_uint, fill_img);
             for (uint k = 0; k < CHUNK; k++) {
                 vec4 fg_k = img[k] * area[k];
-                rgb[k] = rgb[k] * (1.0 - fg_k.a) + fg_k.rgb;
+                rgba[k] = rgba[k] * (1.0 - fg_k.a) + fg_k;
             }
             cmd_ref.offset += 4 + CmdImage_size;
             break;
@@ -205,10 +205,11 @@ void main() {
                 gl_LocalInvocationID.x + TILE_WIDTH_PX * gl_LocalInvocationID.y);
             for (uint k = 0; k < CHUNK; k++) {
                 uvec2 offset = chunk_offset(k);
-                uint srgb = packsRGB(vec4(rgb[k], 1.0));
+                uint srgb = packsRGB(vec4(rgba[k]));
                 float alpha = clamp(abs(area[k]), 0.0, 1.0);
                 write_mem(scratch_alloc, base_ix + 0 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX), srgb);
                 write_mem(scratch_alloc, base_ix + 1 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX), floatBitsToUint(alpha));
+                rgba[k] = vec4(0.0);
             }
             clip_depth++;
             cmd_ref.offset += 4;
@@ -221,8 +222,9 @@ void main() {
                 uvec2 offset = chunk_offset(k);
                 uint srgb = read_mem(scratch_alloc, base_ix + 0 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX));
                 uint alpha = read_mem(scratch_alloc, base_ix + 1 + CLIP_STATE_SIZE * (offset.x + offset.y * TILE_WIDTH_PX));
-                vec4 rgba = unpacksRGB(srgb);
-                rgb[k] = mix(rgba.rgb, rgb[k], area[k] * uintBitsToFloat(alpha));
+                vec4 bg = unpacksRGB(srgb);
+                vec4 fg = rgba[k] * area[k] * uintBitsToFloat(alpha);
+                rgba[k] = bg * (1.0 - fg.a) + fg;
             }
             cmd_ref.offset += 4;
             break;
@@ -234,6 +236,6 @@ void main() {
     }
 
     for (uint i = 0; i < CHUNK; i++) {
-        imageStore(image, ivec2(xy_uint + chunk_offset(i)), vec4(tosRGB(rgb[i]), 1.0));
+        imageStore(image, ivec2(xy_uint + chunk_offset(i)), vec4(tosRGB(rgba[i].rgb), rgba[i].a));
     }
 }
diff --git a/piet-gpu/shader/kernel4.spv b/piet-gpu/shader/kernel4.spv
index 71b3193..af7fe18 100644
Binary files a/piet-gpu/shader/kernel4.spv and b/piet-gpu/shader/kernel4.spv differ