2 files changed, 11 insertions, 17 deletions
diff --git a/src/gpu/effects/cnn_v2_effect.cc b/src/gpu/effects/cnn_v2_effect.cc
index 566686e..366a232 100644
--- a/src/gpu/effects/cnn_v2_effect.cc
+++ b/src/gpu/effects/cnn_v2_effect.cc
@@ -111,17 +111,21 @@ void CNNv2Effect::load_weights() {
     layer_info_.push_back(info);
   }
 
-  // Create GPU storage buffer for weights
-  // Buffer contains: header + layer info + packed f16 weights (as u32)
+  // Create GPU storage buffer for weights (skip header + layer info, upload only weights)
+  size_t header_size = 20;  // 5 u32
+  size_t layer_info_size = 20 * num_layers;  // 5 u32 per layer
+  size_t weights_offset = header_size + layer_info_size;
+  size_t weights_only_size = weights_size - weights_offset;
+
   WGPUBufferDescriptor buffer_desc = {};
-  buffer_desc.size = weights_size;
+  buffer_desc.size = weights_only_size;
   buffer_desc.usage = WGPUBufferUsage_Storage | WGPUBufferUsage_CopyDst;
   buffer_desc.mappedAtCreation = false;
 
   weights_buffer_ = wgpuDeviceCreateBuffer(ctx_.device, &buffer_desc);
 
-  // Upload weights data
-  wgpuQueueWriteBuffer(ctx_.queue, weights_buffer_, 0, weights_data, weights_size);
+  // Upload only weights (skip header + layer info)
+  wgpuQueueWriteBuffer(ctx_.queue, weights_buffer_, 0, weights_data + weights_offset, weights_only_size);
 
   // Create uniform buffers for layer params (one per layer)
   for (uint32_t i = 0; i < num_layers; ++i) {
@@ -233,18 +237,6 @@ void CNNv2Effect::create_pipelines() {
   WGPUShaderModuleDescriptor shader_desc = {};
   shader_desc.nextInChain = &wgsl_src.chain;
 
-  WGPUShaderModule static_module = wgpuDeviceCreateShaderModule(ctx_.device, &shader_desc);
-  if (!static_module) {
-    return;
-  }
-
-  WGPUComputePipelineDescriptor pipeline_desc = {};
-  pipeline_desc.compute.module = static_module;
-  pipeline_desc.compute.entryPoint = str_view("main");
-
-  static_pipeline_ = wgpuDeviceCreateComputePipeline(ctx_.device, &pipeline_desc);
-  wgpuShaderModuleRelease(static_module);
-
   // Create bind group layout for static features compute
   // Bindings: 0=input_tex, 1=input_mip1, 2=input_mip2, 3=depth_tex, 4=output, 5=params
   WGPUBindGroupLayoutEntry bgl_entries[6] = {};
@@ -542,6 +534,7 @@ void CNNv2Effect::compute(WGPUCommandEncoder encoder,
     params.weight_offset = info.weight_offset;
     params.is_output_layer = (i == layer_info_.size() - 1) ? 1 : 0;
     params.blend_amount = effective_blend;
+    params.is_layer_0 = (i == 0) ? 1 : 0;
 
     wgpuQueueWriteBuffer(ctx_.queue, layer_params_buffers_[i], 0, &params, sizeof(params));
 
diff --git a/src/gpu/effects/cnn_v2_effect.h b/src/gpu/effects/cnn_v2_effect.h
index 47dedf5..8a2e1b6 100644
--- a/src/gpu/effects/cnn_v2_effect.h
+++ b/src/gpu/effects/cnn_v2_effect.h
@@ -45,6 +45,7 @@ private:
     uint32_t weight_offset;
     uint32_t is_output_layer;
     float blend_amount;
+    uint32_t is_layer_0;
   };
 
   struct StaticFeatureParams {