1 files changed, 21 insertions, 10 deletions

diff --git a/workspaces/main/shaders/cnn/cnn_layer.wgsl b/workspaces/main/shaders/cnn/cnn_layer.wgsl
index b2bab26..1b1b539 100644
--- a/workspaces/main/shaders/cnn/cnn_layer.wgsl
+++ b/workspaces/main/shaders/cnn/cnn_layer.wgsl
@@ -1,5 +1,6 @@
 // CNN layer shader - uses modular convolution snippets
 // Supports multi-pass rendering with residual connections
+// DO NOT EDIT - Generated by train_cnn.py
 
 @group(0) @binding(0) var smplr: sampler;
 @group(0) @binding(1) var txt: texture_2d<f32>;
@@ -7,16 +8,18 @@
 #include "common_uniforms"
 #include "cnn_activation"
 #include "cnn_conv3x3"
+#include "cnn_conv5x5"
 #include "cnn_weights_generated"
 
 struct CNNLayerParams {
     layer_index: i32,
-    use_residual: i32,
+    blend_amount: f32,
     _pad: vec2<f32>,
 };
 
 @group(0) @binding(2) var<uniform> uniforms: CommonUniforms;
 @group(0) @binding(3) var<uniform> params: CNNLayerParams;
+@group(0) @binding(4) var original_input: texture_2d<f32>;
 
 @vertex fn vs_main(@builtin(vertex_index) i: u32) -> @builtin(position) vec4<f32> {
     var pos = array<vec2<f32>, 3>(
@@ -27,20 +30,28 @@ struct CNNLayerParams {
 
 @fragment fn fs_main(@builtin(position) p: vec4<f32>) -> @location(0) vec4<f32> {
     let uv = p.xy / uniforms.resolution;
+    let original = (textureSample(original_input, smplr, uv) - 0.5) * 2.0;  // Normalize to [-1,1]
     var result = vec4<f32>(0.0);
 
-    // Layer 0 uses coordinate-aware convolution
+    // Layer 0: 7→4 (RGBD output)
     if (params.layer_index == 0) {
-        result = cnn_conv3x3_with_coord(txt, smplr, uv, uniforms.resolution,
-                                        rgba_weights_layer0, coord_weights_layer0, bias_layer0);
-        result = cnn_tanh(result);
+        result = cnn_conv3x3_7to4_src(txt, smplr, uv, uniforms.resolution, weights_layer0);
+        result = cnn_tanh(result);  // Keep in [-1,1]
     }
-
-    // Residual connection
-    if (params.use_residual != 0) {
-        let input = textureSample(txt, smplr, uv);
-        result = input + result * 0.3;
+    else if (params.layer_index == 1) {
+        result = cnn_conv5x5_7to4(txt, smplr, uv, uniforms.resolution,
+                                   original, weights_layer1);
+	result = cnn_tanh(result);  // Keep in [-1,1]
     }
+    else if (params.layer_index == 2) {  // last layer
+        let gray_out = cnn_conv3x3_7to1(txt, smplr, uv, uniforms.resolution,
+                                        original, weights_layer2);
 
+        // At this point here, 'gray_out' is what the training script should have learned.
+        // Below is some extra code for visual output, excluded from training:
+        result = vec4<f32>(gray_out, gray_out, gray_out, 1.0);  // Keep in [-1,1]
+        let blended = mix(original, result, params.blend_amount);
+        return (blended + 1.0) * 0.5;  // Denormalize to [0,1] for display
+    }
     return result;
 }