feat(cnn_v3): Phase 4 complete — CNNv3Effect C++ + FiLM uniform upload

- cnn_v3/src/cnn_v3_effect.{h,cc}: full Effect subclass with 5 compute
  passes (enc0→enc1→bottleneck→dec1→dec0), shared weights storage buffer,
  per-pass uniform buffers, set_film_params() API
- Fixed WGSL/C++ struct alignment: vec3u has align=16, so CnnV3Params4ch
  is 64 bytes and CnnV3ParamsEnc1 is 96 bytes (not 48/80)
- Weight offsets computed as explicit formulas (e.g. 20*4*9+4) for clarity
- Registered in CMake, shaders.h/cc, demo_effects.h, test_demo_effects.cc
- 35/35 tests pass

handoff(Gemini): CNN v3 Phase 5 next — parity validation (Python ref vs WGSL)

10 files changed, 623 insertions, 5 deletions

diff --git a/PROJECT_CONTEXT.md b/PROJECT_CONTEXT.md
index f42ccf4..4767185 100644
--- a/PROJECT_CONTEXT.md
+++ b/PROJECT_CONTEXT.md
@@ -36,7 +36,7 @@
 - **Audio:** Sample-accurate sync. Zero heap allocations per frame. Variable tempo. OLA-IDCT synthesis (v2 .spec): Hann analysis window, rectangular synthesis, 50% overlap, click-free. V1 (raw DCT-512) preserved for generated notes. .spec files regenerated as v2.
 - **Shaders:** Parameterized effects (UniformHelper, .seq syntax). Beat-synchronized animation support (`beat_time`, `beat_phase`). Modular WGSL composition with ShaderComposer. 27 shared common shaders (math, render, compute). Reusable snippets: `render/scratch_lines`, `render/ntsc_common` (NTSC signal processing, RGB and YIQ input variants via `sample_ntsc_signal` hook), `math/color` (YIQ/NTSC), `math/color_c64` (C64 palette, Bayer dither, border animation).
 - **3D:** Hybrid SDF/rasterization with BVH. Binary scene loader. Blender pipeline.
-- **Effects:** CNN post-processing: CNNEffect (v1) and CNNv2Effect operational. CNN v2: sigmoid activation, storage buffer weights (~3.2 KB), 7D static features, dynamic layers. Training stable, convergence validated. **CNN v3 Phase 3 complete:** 5 WGSL inference shaders (enc0/enc1/bottleneck/dec1/dec0) + `cnn_v3/common` snippet. Zero-pad convs, AvgPool down, NearestUp, FiLM, skip-concat, sigmoid output. Registered in assets + shaders.cc. See `cnn_v3/docs/HOWTO.md` §7.
+- **Effects:** CNN post-processing: CNNEffect (v1) and CNNv2Effect operational. CNN v2: sigmoid activation, storage buffer weights (~3.2 KB), 7D static features, dynamic layers. Training stable, convergence validated. **CNN v3 Phase 4 complete:** `CNNv3Effect` C++ class (5 compute passes, FiLM uniform upload, identity γ/β defaults). `set_film_params()` modulates all layers via beat/audio. WGSL params struct alignment fix (vec3u align=16 → 64/96-byte C++ mirrors). Registered in CMake, shaders.h/cc, demo_effects.h, tests. See `cnn_v3/docs/HOWTO.md`.
 - **Tools:** CNN test tool operational. Texture readback utility functional. Timeline editor (web-based, beat-aligned, audio playback).
 - **Build:** Asset dependency tracking. Size measurement. Hot-reload (debug-only). WSL (Windows 10) supported: native Linux build and cross-compile to `.exe` via `mingw-w64`.
 - **Sequence:** DAG-based effect routing with explicit node system. Python compiler with topological sort and ping-pong optimization. 12 effects operational (Passthrough, Placeholder, GaussianBlur, Heptagon, Particles, RotatingCube, Hybrid3D, Flash, PeakMeter, Scene1, Scene2, Scratch). Effect times are absolute (seq_compiler adds sequence start offset). See `doc/SEQUENCE.md`.
@@ -46,7 +46,7 @@
 
 ## Next Up
 
-**Active:** CNN v3 Phase 4 (C++ CNNv3Effect + FiLM uniform), Spectral Brush Editor
+**Active:** CNN v3 Phase 5 (parity validation), Spectral Brush Editor
 **Ongoing:** Test infrastructure maintenance (35/35 passing)
 **Future:** Size optimization (64k target), 3D enhancements
 
diff --git a/TODO.md b/TODO.md
index 86c3e37..e33072f 100644
--- a/TODO.md
+++ b/TODO.md
@@ -76,7 +76,7 @@ PyTorch / HTML WebGPU / C++ WebGPU.
    - Howto: `cnn_v3/docs/HOWTO.md`
 2. ✅ Training infrastructure: `blender_export.py`, `pack_blender_sample.py`, `pack_photo_sample.py`
 3. ✅ WGSL shaders: cnn_v3_common (snippet), enc0, enc1, bottleneck, dec1, dec0
-4. C++ CNNv3Effect + FiLM uniform upload
+4. ✅ C++ CNNv3Effect + FiLM uniform upload
 5. Parity validation (test vectors, ≤1/255 per pixel)
 
 ## Future: CNN v2 8-bit Quantization
diff --git a/cmake/DemoSourceLists.cmake b/cmake/DemoSourceLists.cmake
index 04bbb3b..742057a 100644
--- a/cmake/DemoSourceLists.cmake
+++ b/cmake/DemoSourceLists.cmake
@@ -41,6 +41,7 @@ set(COMMON_GPU_EFFECTS
     src/effects/scene1_effect.cc
     src/effects/scene2_effect.cc
     cnn_v3/src/gbuffer_effect.cc
+    cnn_v3/src/cnn_v3_effect.cc
     # TODO: Port CNN effects to v2 (complex v1 dependencies)
     # cnn_v1/src/cnn_v1_effect.cc
     # cnn_v2/src/cnn_v2_effect.cc
diff --git a/cnn_v3/docs/HOWTO.md b/cnn_v3/docs/HOWTO.md
index ad71f1f..22266d3 100644
--- a/cnn_v3/docs/HOWTO.md
+++ b/cnn_v3/docs/HOWTO.md
@@ -201,7 +201,7 @@ The CNN v3 design requires exact parity between PyTorch, WGSL (HTML), and C++.
 | 1 — G-buffer (SDF + shadow passes) | TODO | Placeholder in place |
 | 2 — Training infrastructure | ✅ Done | blender_export.py, pack_*_sample.py |
 | 3 — WGSL U-Net shaders | ✅ Done | 5 compute shaders + cnn_v3/common snippet |
-| 4 — C++ CNNv3Effect | TODO | FiLM uniform upload |
+| 4 — C++ CNNv3Effect | ✅ Done | FiLM uniform upload, 35/35 tests pass |
 | 5 — Parity validation | TODO | Test vectors, ≤1/255 |
 
 ---
diff --git a/cnn_v3/src/cnn_v3_effect.cc b/cnn_v3/src/cnn_v3_effect.cc
new file mode 100644
index 0000000..d13799c
--- /dev/null
+++ b/cnn_v3/src/cnn_v3_effect.cc
@@ -0,0 +1,467 @@
+// CNN v3 Effect — U-Net + FiLM inference (5 compute passes)
+// See cnn_v3/docs/CNN_V3.md for architecture, HOWTO.md §7 for shader details.
+
+#include "cnn_v3_effect.h"
+#include "gpu/gpu.h"
+#include "gpu/shader_composer.h"
+#include "util/fatal_error.h"
+#include <cstdint>
+#include <cstring>
+
+// ---------------------------------------------------------------------------
+// Weight layout constants — explicit formulas matching WGSL shader comments
+// ---------------------------------------------------------------------------
+//
+// Format: Conv(IN→OUT, KxK) has OUT*IN*K*K weights + OUT biases
+// Layout: OIHW order (out × in × kH × kW), biases appended after conv weights
+//
+static const uint32_t kEnc0Weights  = 20 * 4 * 9 + 4;   // Conv(20→4,3×3)+bias
+static const uint32_t kEnc1Weights  = 4  * 8 * 9 + 8;   // Conv(4→8,3×3)+bias
+static const uint32_t kBnWeights    = 8  * 8 * 1 + 8;   // Conv(8→8,1×1)+bias
+static const uint32_t kDec1Weights  = 16 * 4 * 9 + 4;   // Conv(16→4,3×3)+bias
+static const uint32_t kDec0Weights  = 8  * 4 * 9 + 4;   // Conv(8→4,3×3)+bias
+
+static const uint32_t kEnc0Offset  = 0;
+static const uint32_t kEnc1Offset  = kEnc0Offset  + kEnc0Weights;
+static const uint32_t kBnOffset    = kEnc1Offset  + kEnc1Weights;
+static const uint32_t kDec1Offset  = kBnOffset    + kBnWeights;
+static const uint32_t kDec0Offset  = kDec1Offset  + kDec1Weights;
+static const uint32_t kTotalF16    = kDec0Offset  + kDec0Weights;
+
+// Weights buffer size in bytes: f16 values are packed two-per-u32.
+// Round up to u32 boundary.
+static const uint32_t kWeightsBufBytes = ((kTotalF16 + 1) / 2) * 4;
+
+// ---------------------------------------------------------------------------
+// Shader source externs (registered in shaders.cc via InitShaderComposer)
+// ---------------------------------------------------------------------------
+extern const char* cnn_v3_enc0_wgsl;
+extern const char* cnn_v3_enc1_wgsl;
+extern const char* cnn_v3_bottleneck_wgsl;
+extern const char* cnn_v3_dec1_wgsl;
+extern const char* cnn_v3_dec0_wgsl;
+
+// ---------------------------------------------------------------------------
+// Helpers
+// ---------------------------------------------------------------------------
+
+static WGPUShaderModule make_shader(WGPUDevice device, const char* wgsl) {
+  const std::string composed =
+      ShaderComposer::Get().Compose({"cnn_v3/common"}, wgsl);
+
+  WGPUShaderSourceWGSL src = {};
+  src.chain.sType = WGPUSType_ShaderSourceWGSL;
+  src.code        = str_view(composed.c_str());
+
+  WGPUShaderModuleDescriptor desc = {};
+  desc.nextInChain = &src.chain;
+  return wgpuDeviceCreateShaderModule(device, &desc);
+}
+
+static WGPUBindGroupLayout make_bgl(WGPUDevice device,
+                                    const WGPUBindGroupLayoutEntry* entries,
+                                    uint32_t count) {
+  WGPUBindGroupLayoutDescriptor desc = {};
+  desc.entryCount = count;
+  desc.entries    = entries;
+  return wgpuDeviceCreateBindGroupLayout(device, &desc);
+}
+
+static WGPUComputePipeline make_compute_pipeline(WGPUDevice device,
+                                                 WGPUShaderModule shader,
+                                                 const char* entry,
+                                                 WGPUBindGroupLayout bgl) {
+  WGPUPipelineLayoutDescriptor pl_desc = {};
+  pl_desc.bindGroupLayoutCount = 1;
+  pl_desc.bindGroupLayouts     = &bgl;
+  WGPUPipelineLayout pl = wgpuDeviceCreatePipelineLayout(device, &pl_desc);
+
+  WGPUComputePipelineDescriptor pipe_desc = {};
+  pipe_desc.layout                = pl;
+  pipe_desc.compute.module        = shader;
+  pipe_desc.compute.entryPoint    = str_view(entry);
+  WGPUComputePipeline pipe = wgpuDeviceCreateComputePipeline(device, &pipe_desc);
+
+  wgpuPipelineLayoutRelease(pl);
+  return pipe;
+}
+
+// BGL entry helpers
+static WGPUBindGroupLayoutEntry bgl_uint_tex(uint32_t binding) {
+  WGPUBindGroupLayoutEntry e = {};
+  e.binding                    = binding;
+  e.visibility                 = WGPUShaderStage_Compute;
+  e.texture.sampleType         = WGPUTextureSampleType_Uint;
+  e.texture.viewDimension      = WGPUTextureViewDimension_2D;
+  return e;
+}
+static WGPUBindGroupLayoutEntry bgl_float_tex(uint32_t binding) {
+  WGPUBindGroupLayoutEntry e = {};
+  e.binding               = binding;
+  e.visibility            = WGPUShaderStage_Compute;
+  e.texture.sampleType    = WGPUTextureSampleType_Float;
+  e.texture.viewDimension = WGPUTextureViewDimension_2D;
+  return e;
+}
+static WGPUBindGroupLayoutEntry bgl_storage_buf(uint32_t binding) {
+  WGPUBindGroupLayoutEntry e = {};
+  e.binding          = binding;
+  e.visibility       = WGPUShaderStage_Compute;
+  e.buffer.type      = WGPUBufferBindingType_ReadOnlyStorage;
+  return e;
+}
+static WGPUBindGroupLayoutEntry bgl_uniform_buf(uint32_t binding,
+                                                uint64_t min_size) {
+  WGPUBindGroupLayoutEntry e = {};
+  e.binding                 = binding;
+  e.visibility              = WGPUShaderStage_Compute;
+  e.buffer.type             = WGPUBufferBindingType_Uniform;
+  e.buffer.minBindingSize   = min_size;
+  return e;
+}
+static WGPUBindGroupLayoutEntry bgl_storage_tex_write(
+    uint32_t binding, WGPUTextureFormat fmt) {
+  WGPUBindGroupLayoutEntry e = {};
+  e.binding                          = binding;
+  e.visibility                       = WGPUShaderStage_Compute;
+  e.storageTexture.access            = WGPUStorageTextureAccess_WriteOnly;
+  e.storageTexture.format            = fmt;
+  e.storageTexture.viewDimension     = WGPUTextureViewDimension_2D;
+  return e;
+}
+
+// ---------------------------------------------------------------------------
+// Constructor
+// ---------------------------------------------------------------------------
+
+CNNv3Effect::CNNv3Effect(const GpuContext& ctx,
+                         const std::vector<std::string>& inputs,
+                         const std::vector<std::string>& outputs,
+                         float start_time, float end_time)
+    : Effect(ctx, inputs, outputs, start_time, end_time) {
+  HEADLESS_RETURN_IF_NULL(ctx_.device);
+
+  const std::string& prefix =
+      outputs.empty() ? std::string("cnn_v3") : outputs[0];
+  node_enc0_       = prefix + "_enc0";
+  node_enc1_       = prefix + "_enc1";
+  node_bottleneck_ = prefix + "_bottleneck";
+  node_dec1_       = prefix + "_dec1";
+
+  // Allocate zeroed weights buffer (f16 pairs packed as u32).
+  // Weights are zero-initialized; load_weights() can fill from file later.
+  weights_buf_ = gpu_create_buffer(
+      ctx_.device, kWeightsBufBytes,
+      WGPUBufferUsage_Storage | WGPUBufferUsage_CopyDst);
+
+  // Initialize uniform buffers.
+  enc0_params_buf_.init(ctx_.device);
+  enc1_params_buf_.init(ctx_.device);
+  bn_params_buf_.init(ctx_.device);
+  dec1_params_buf_.init(ctx_.device);
+  dec0_params_buf_.init(ctx_.device);
+
+  // Set weight offsets (FiLM γ/β default to identity: γ=1, β=0).
+  enc0_params_.weight_offset = kEnc0Offset;
+  for (int i = 0; i < 4; ++i) { enc0_params_.gamma[i] = 1.0f; }
+
+  enc1_params_.weight_offset = kEnc1Offset;
+  for (int i = 0; i < 4; ++i) {
+    enc1_params_.gamma_lo[i] = 1.0f;
+    enc1_params_.gamma_hi[i] = 1.0f;
+  }
+
+  bn_params_.weight_offset = kBnOffset;
+
+  dec1_params_.weight_offset = kDec1Offset;
+  for (int i = 0; i < 4; ++i) { dec1_params_.gamma[i] = 1.0f; }
+
+  dec0_params_.weight_offset = kDec0Offset;
+  for (int i = 0; i < 4; ++i) { dec0_params_.gamma[i] = 1.0f; }
+
+  create_pipelines();
+}
+
+// ---------------------------------------------------------------------------
+// declare_nodes
+// ---------------------------------------------------------------------------
+
+void CNNv3Effect::declare_nodes(NodeRegistry& registry) {
+  // enc0_tex: rgba16float full-res
+  registry.declare_node(node_enc0_,       NodeType::GBUF_ALBEDO,    -1, -1);
+  // enc1_tex: rgba32uint half-res
+  registry.declare_node(node_enc1_,       NodeType::GBUF_RGBA32UINT, -1, -1);
+  // bottleneck_tex: rgba32uint quarter-res — declare at 1/4 resolution
+  registry.declare_node(node_bottleneck_, NodeType::GBUF_RGBA32UINT, -1, -1);
+  // dec1_tex: rgba16float half-res
+  registry.declare_node(node_dec1_,       NodeType::GBUF_ALBEDO,    -1, -1);
+  // output_tex: rgba16float full-res (the declared output_nodes_[0])
+}
+
+// ---------------------------------------------------------------------------
+// set_film_params — simple linear mapping, no MLP yet
+// ---------------------------------------------------------------------------
+
+void CNNv3Effect::set_film_params(const CNNv3FiLMParams& fp) {
+  // Identity + audio/beat modulation.
+  // Replace with FiLM MLP output once training is done.
+  const float a = fp.audio_intensity;
+  const float b = fp.beat_phase;
+
+  for (int i = 0; i < 4; ++i) {
+    enc0_params_.gamma[i] = 1.0f + a * 0.5f;
+    enc0_params_.beta[i]  = b * 0.1f;
+  }
+  for (int i = 0; i < 4; ++i) {
+    enc1_params_.gamma_lo[i] = 1.0f + a * 0.3f;
+    enc1_params_.gamma_hi[i] = 1.0f + a * 0.3f;
+    enc1_params_.beta_lo[i]  = fp.beat_norm * 0.1f;
+    enc1_params_.beta_hi[i]  = fp.beat_norm * 0.1f;
+  }
+  for (int i = 0; i < 4; ++i) {
+    dec1_params_.gamma[i] = 1.0f + fp.style_p0 * 0.5f;
+    dec1_params_.beta[i]  = fp.style_p1 * 0.1f;
+    dec0_params_.gamma[i] = 1.0f + fp.style_p0 * 0.5f;
+    dec0_params_.beta[i]  = fp.style_p1 * 0.1f;
+  }
+}
+
+// ---------------------------------------------------------------------------
+// render
+// ---------------------------------------------------------------------------
+
+void CNNv3Effect::render(WGPUCommandEncoder encoder,
+                         const UniformsSequenceParams& params,
+                         NodeRegistry& nodes) {
+  // Upload params uniforms.
+  enc0_params_buf_.update(ctx_.queue, enc0_params_);
+  enc1_params_buf_.update(ctx_.queue, enc1_params_);
+  bn_params_buf_.update(ctx_.queue, bn_params_);
+  dec1_params_buf_.update(ctx_.queue, dec1_params_);
+  dec0_params_buf_.update(ctx_.queue, dec0_params_);
+
+  update_bind_groups(nodes);
+
+  const int W = (int)params.resolution.x;
+  const int H = (int)params.resolution.y;
+
+  // Dispatch helper: ceil(dim / 8) workgroups
+  auto dispatch = [&](WGPUComputePipeline pipe, WGPUBindGroup bg,
+                      int w, int h) {
+    WGPUComputePassDescriptor pass_desc = {};
+    WGPUComputePassEncoder pass =
+        wgpuCommandEncoderBeginComputePass(encoder, &pass_desc);
+    wgpuComputePassEncoderSetPipeline(pass, pipe);
+    wgpuComputePassEncoderSetBindGroup(pass, 0, bg, 0, nullptr);
+    wgpuComputePassEncoderDispatchWorkgroups(
+        pass,
+        (uint32_t)((w + 7) / 8),
+        (uint32_t)((h + 7) / 8),
+        1);
+    wgpuComputePassEncoderEnd(pass);
+    wgpuComputePassEncoderRelease(pass);
+  };
+
+  dispatch(enc0_pipeline_.get(), enc0_bg_.get(),  W,     H);
+  dispatch(enc1_pipeline_.get(), enc1_bg_.get(),  W / 2, H / 2);
+  dispatch(bn_pipeline_.get(),   bn_bg_.get(),    W / 4, H / 4);
+  dispatch(dec1_pipeline_.get(), dec1_bg_.get(),  W / 2, H / 2);
+  dispatch(dec0_pipeline_.get(), dec0_bg_.get(),  W,     H);
+}
+
+// ---------------------------------------------------------------------------
+// create_pipelines
+// ---------------------------------------------------------------------------
+
+void CNNv3Effect::create_pipelines() {
+  HEADLESS_RETURN_IF_NULL(ctx_.device);
+  WGPUDevice dev = ctx_.device;
+
+  // --- enc0 ---
+  // B0: feat_tex0 (u32), B1: feat_tex1 (u32), B2: weights (storage),
+  // B3: params (uniform), B4: enc0_out (storage_tex rgba16float write)
+  {
+    WGPUBindGroupLayoutEntry e[5] = {
+        bgl_uint_tex(0),
+        bgl_uint_tex(1),
+        bgl_storage_buf(2),
+        bgl_uniform_buf(3, sizeof(CnnV3Params4ch)),  // 64 bytes
+        bgl_storage_tex_write(4, WGPUTextureFormat_RGBA16Float),
+    };
+    WGPUBindGroupLayout bgl = make_bgl(dev, e, 5);
+    WGPUShaderModule sh = make_shader(dev, cnn_v3_enc0_wgsl);
+    enc0_pipeline_.set(make_compute_pipeline(dev, sh, "enc0_main", bgl));
+    wgpuShaderModuleRelease(sh);
+    wgpuBindGroupLayoutRelease(bgl);
+  }
+
+  // --- enc1 ---
+  // B0: enc0_tex (f32), B1: weights (storage),
+  // B2: params (uniform), B3: enc1_out (storage_tex rgba32uint write)
+  {
+    WGPUBindGroupLayoutEntry e[4] = {
+        bgl_float_tex(0),
+        bgl_storage_buf(1),
+        bgl_uniform_buf(2, sizeof(CnnV3ParamsEnc1)),
+        bgl_storage_tex_write(3, WGPUTextureFormat_RGBA32Uint),
+    };
+    WGPUBindGroupLayout bgl = make_bgl(dev, e, 4);
+    WGPUShaderModule sh = make_shader(dev, cnn_v3_enc1_wgsl);
+    enc1_pipeline_.set(make_compute_pipeline(dev, sh, "enc1_main", bgl));
+    wgpuShaderModuleRelease(sh);
+    wgpuBindGroupLayoutRelease(bgl);
+  }
+
+  // --- bottleneck ---
+  // B0: enc1_tex (u32), B1: weights (storage),
+  // B2: params (uniform), B3: bottleneck_out (storage_tex rgba32uint write)
+  {
+    WGPUBindGroupLayoutEntry e[4] = {
+        bgl_uint_tex(0),
+        bgl_storage_buf(1),
+        bgl_uniform_buf(2, sizeof(CnnV3ParamsBn)),
+        bgl_storage_tex_write(3, WGPUTextureFormat_RGBA32Uint),
+    };
+    WGPUBindGroupLayout bgl = make_bgl(dev, e, 4);
+    WGPUShaderModule sh = make_shader(dev, cnn_v3_bottleneck_wgsl);
+    bn_pipeline_.set(make_compute_pipeline(dev, sh, "bottleneck_main", bgl));
+    wgpuShaderModuleRelease(sh);
+    wgpuBindGroupLayoutRelease(bgl);
+  }
+
+  // --- dec1 ---
+  // B0: bottleneck_tex (u32), B1: enc1_tex (u32), B2: weights (storage),
+  // B3: params (uniform), B4: dec1_out (storage_tex rgba16float write)
+  {
+    WGPUBindGroupLayoutEntry e[5] = {
+        bgl_uint_tex(0),
+        bgl_uint_tex(1),
+        bgl_storage_buf(2),
+        bgl_uniform_buf(3, sizeof(CnnV3Params4ch)),  // 64 bytes
+        bgl_storage_tex_write(4, WGPUTextureFormat_RGBA16Float),
+    };
+    WGPUBindGroupLayout bgl = make_bgl(dev, e, 5);
+    WGPUShaderModule sh = make_shader(dev, cnn_v3_dec1_wgsl);
+    dec1_pipeline_.set(make_compute_pipeline(dev, sh, "dec1_main", bgl));
+    wgpuShaderModuleRelease(sh);
+    wgpuBindGroupLayoutRelease(bgl);
+  }
+
+  // --- dec0 ---
+  // B0: dec1_tex (f32), B1: enc0_tex (f32), B2: weights (storage),
+  // B3: params (uniform), B4: output_tex (storage_tex rgba16float write)
+  {
+    WGPUBindGroupLayoutEntry e[5] = {
+        bgl_float_tex(0),
+        bgl_float_tex(1),
+        bgl_storage_buf(2),
+        bgl_uniform_buf(3, sizeof(CnnV3Params4ch)),  // 64 bytes
+        bgl_storage_tex_write(4, WGPUTextureFormat_RGBA16Float),
+    };
+    WGPUBindGroupLayout bgl = make_bgl(dev, e, 5);
+    WGPUShaderModule sh = make_shader(dev, cnn_v3_dec0_wgsl);
+    dec0_pipeline_.set(make_compute_pipeline(dev, sh, "dec0_main", bgl));
+    wgpuShaderModuleRelease(sh);
+    wgpuBindGroupLayoutRelease(bgl);
+  }
+}
+
+// ---------------------------------------------------------------------------
+// update_bind_groups — rebuilt each frame (node views may be recreated)
+// ---------------------------------------------------------------------------
+
+// Helper: set a texture view binding entry.
+static void bg_tex(WGPUBindGroupEntry& e, uint32_t binding,
+                   WGPUTextureView view) {
+  e = {};
+  e.binding     = binding;
+  e.textureView = view;
+}
+// Helper: set a buffer binding entry.
+static void bg_buf(WGPUBindGroupEntry& e, uint32_t binding, WGPUBuffer buf,
+                   uint64_t size) {
+  e = {};
+  e.binding = binding;
+  e.buffer  = buf;
+  e.size    = size;
+}
+
+void CNNv3Effect::update_bind_groups(NodeRegistry& nodes) {
+  WGPUDevice dev = ctx_.device;
+
+  WGPUTextureView feat0_view = nodes.get_view(input_nodes_[0]);
+  WGPUTextureView feat1_view = nodes.get_view(input_nodes_[1]);
+  WGPUTextureView enc0_view  = nodes.get_view(node_enc0_);
+  WGPUTextureView enc1_view  = nodes.get_view(node_enc1_);
+  WGPUTextureView bn_view    = nodes.get_view(node_bottleneck_);
+  WGPUTextureView dec1_view  = nodes.get_view(node_dec1_);
+  WGPUTextureView out_view   = nodes.get_view(output_nodes_[0]);
+
+  WGPUBuffer wb = weights_buf_.buffer;
+
+  auto make_bg = [&](WGPUComputePipeline pipe, WGPUBindGroupEntry* e,
+                     uint32_t count) -> WGPUBindGroup {
+    WGPUBindGroupLayout bgl =
+        wgpuComputePipelineGetBindGroupLayout(pipe, 0);
+    WGPUBindGroupDescriptor desc = {};
+    desc.layout     = bgl;
+    desc.entryCount = count;
+    desc.entries    = e;
+    WGPUBindGroup bg = wgpuDeviceCreateBindGroup(dev, &desc);
+    wgpuBindGroupLayoutRelease(bgl);
+    return bg;
+  };
+
+  // enc0: feat_tex0(B0), feat_tex1(B1), weights(B2), params(B3), enc0_out(B4)
+  {
+    WGPUBindGroupEntry e[5] = {};
+    bg_tex(e[0], 0, feat0_view);
+    bg_tex(e[1], 1, feat1_view);
+    bg_buf(e[2], 2, wb, kWeightsBufBytes);
+    bg_buf(e[3], 3, enc0_params_buf_.get().buffer, sizeof(CnnV3Params4ch));
+    bg_tex(e[4], 4, enc0_view);
+    enc0_bg_.set(make_bg(enc0_pipeline_.get(), e, 5));
+  }
+
+  // enc1: enc0_tex(B0), weights(B1), params(B2), enc1_out(B3)
+  {
+    WGPUBindGroupEntry e[4] = {};
+    bg_tex(e[0], 0, enc0_view);
+    bg_buf(e[1], 1, wb, kWeightsBufBytes);
+    bg_buf(e[2], 2, enc1_params_buf_.get().buffer, sizeof(CnnV3ParamsEnc1));
+    bg_tex(e[3], 3, enc1_view);
+    enc1_bg_.set(make_bg(enc1_pipeline_.get(), e, 4));
+  }
+
+  // bottleneck: enc1_tex(B0), weights(B1), params(B2), bn_out(B3)
+  {
+    WGPUBindGroupEntry e[4] = {};
+    bg_tex(e[0], 0, enc1_view);
+    bg_buf(e[1], 1, wb, kWeightsBufBytes);
+    bg_buf(e[2], 2, bn_params_buf_.get().buffer, sizeof(CnnV3ParamsBn));
+    bg_tex(e[3], 3, bn_view);
+    bn_bg_.set(make_bg(bn_pipeline_.get(), e, 4));
+  }
+
+  // dec1: bn_tex(B0), enc1_tex(B1), weights(B2), params(B3), dec1_out(B4)
+  {
+    WGPUBindGroupEntry e[5] = {};
+    bg_tex(e[0], 0, bn_view);
+    bg_tex(e[1], 1, enc1_view);
+    bg_buf(e[2], 2, wb, kWeightsBufBytes);
+    bg_buf(e[3], 3, dec1_params_buf_.get().buffer, sizeof(CnnV3Params4ch));
+    bg_tex(e[4], 4, dec1_view);
+    dec1_bg_.set(make_bg(dec1_pipeline_.get(), e, 5));
+  }
+
+  // dec0: dec1_tex(B0), enc0_tex(B1), weights(B2), params(B3), output(B4)
+  {
+    WGPUBindGroupEntry e[5] = {};
+    bg_tex(e[0], 0, dec1_view);
+    bg_tex(e[1], 1, enc0_view);
+    bg_buf(e[2], 2, wb, kWeightsBufBytes);
+    bg_buf(e[3], 3, dec0_params_buf_.get().buffer, sizeof(CnnV3Params4ch));
+    bg_tex(e[4], 4, out_view);
+    dec0_bg_.set(make_bg(dec0_pipeline_.get(), e, 5));
+  }
+}
diff --git a/cnn_v3/src/cnn_v3_effect.h b/cnn_v3/src/cnn_v3_effect.h
new file mode 100644
index 0000000..c358990
--- /dev/null
+++ b/cnn_v3/src/cnn_v3_effect.h
@@ -0,0 +1,132 @@
+// CNN v3 Effect — U-Net + FiLM inference pass
+// Runs 5 compute passes (enc0→enc1→bottleneck→dec1→dec0) on G-buffer feature
+// textures produced by GBufferEffect.
+//
+// Inputs:  feat_tex0, feat_tex1   (rgba32uint, 20-channel G-buffer)
+// Output:  output_tex             (rgba16float, 4-channel RGBA)
+
+#pragma once
+
+#include <cstdint>
+
+#include "gpu/effect.h"
+#include "gpu/sequence.h"
+#include "gpu/uniform_helper.h"
+#include "gpu/wgpu_resource.h"
+
+// ---------------------------------------------------------------------------
+// Per-pass params uniform layouts (mirror WGSL Params structs exactly)
+// ---------------------------------------------------------------------------
+
+// enc0, dec1, dec0: 4-channel FiLM
+//
+// WGSL layout (vec3u has align=16, so _pad sits at offset 16):
+//   offset  0: weight_offset (u32,    4 bytes)
+//   offset  4: (12 bytes implicit padding before vec3u)
+//   offset 16: _pad          (vec3u, 12 bytes)
+//   offset 28: (4 bytes implicit padding before vec4f)
+//   offset 32: gamma         (vec4f, 16 bytes)
+//   offset 48: beta          (vec4f, 16 bytes)
+//   total: 64 bytes
+struct CnnV3Params4ch {
+  uint32_t weight_offset;  // offset 0
+  uint32_t _pad[7];        // offsets 4-31 (mirrors implicit + vec3u + post-pad)
+  float gamma[4];          // offset 32
+  float beta[4];           // offset 48
+};
+static_assert(sizeof(CnnV3Params4ch) == 64, "CnnV3Params4ch must be 64 bytes");
+
+// enc1: 8-channel FiLM (split into lo/hi vec4 pairs)
+//
+// WGSL layout (same header padding as above):
+//   offset  0: weight_offset (u32,    4 bytes)
+//   offset 16: _pad          (vec3u, 12 bytes)
+//   offset 32: gamma_lo      (vec4f, 16 bytes)
+//   offset 48: gamma_hi      (vec4f, 16 bytes)
+//   offset 64: beta_lo       (vec4f, 16 bytes)
+//   offset 80: beta_hi       (vec4f, 16 bytes)
+//   total: 96 bytes
+struct CnnV3ParamsEnc1 {
+  uint32_t weight_offset;  // offset 0
+  uint32_t _pad[7];        // offsets 4-31
+  float gamma_lo[4];       // offset 32
+  float gamma_hi[4];       // offset 48
+  float beta_lo[4];        // offset 64
+  float beta_hi[4];        // offset 80
+};
+static_assert(sizeof(CnnV3ParamsEnc1) == 96,
+              "CnnV3ParamsEnc1 must be 96 bytes");
+
+// bottleneck: no FiLM — 4 plain u32s, no alignment gap
+struct CnnV3ParamsBn {
+  uint32_t weight_offset;
+  uint32_t _pad[3];
+};
+static_assert(sizeof(CnnV3ParamsBn) == 16, "CnnV3ParamsBn must be 16 bytes");
+
+// ---------------------------------------------------------------------------
+// FiLM conditioning inputs (CPU-side, uploaded via set_film_params each frame)
+// ---------------------------------------------------------------------------
+struct CNNv3FiLMParams {
+  float beat_phase      = 0.0f;  // 0-1 within current beat
+  float beat_norm       = 0.0f;  // beat_time / 8.0, normalized 8-beat cycle
+  float audio_intensity = 0.0f;  // peak audio level 0-1
+  float style_p0        = 0.0f;  // user-defined style param
+  float style_p1        = 0.0f;  // user-defined style param
+};
+
+class CNNv3Effect : public Effect {
+ public:
+  CNNv3Effect(const GpuContext& ctx, const std::vector<std::string>& inputs,
+              const std::vector<std::string>& outputs, float start_time,
+              float end_time);
+
+  void declare_nodes(NodeRegistry& registry) override;
+
+  void render(WGPUCommandEncoder encoder, const UniformsSequenceParams& params,
+              NodeRegistry& nodes) override;
+
+  // Update FiLM conditioning; call before render() each frame.
+  void set_film_params(const CNNv3FiLMParams& fp);
+
+ private:
+  // Intermediate node names (prefixed from output[0])
+  std::string node_enc0_;
+  std::string node_enc1_;
+  std::string node_bottleneck_;
+  std::string node_dec1_;
+
+  // 5 compute pipelines
+  ComputePipeline enc0_pipeline_;
+  ComputePipeline enc1_pipeline_;
+  ComputePipeline bn_pipeline_;
+  ComputePipeline dec1_pipeline_;
+  ComputePipeline dec0_pipeline_;
+
+  // 5 bind groups (rebuilt each render since node views may change)
+  BindGroup enc0_bg_;
+  BindGroup enc1_bg_;
+  BindGroup bn_bg_;
+  BindGroup dec1_bg_;
+  BindGroup dec0_bg_;
+
+  // Params uniform buffers (one per pass)
+  UniformBuffer<CnnV3Params4ch>  enc0_params_buf_;
+  UniformBuffer<CnnV3ParamsEnc1> enc1_params_buf_;
+  UniformBuffer<CnnV3ParamsBn>   bn_params_buf_;
+  UniformBuffer<CnnV3Params4ch>  dec1_params_buf_;
+  UniformBuffer<CnnV3Params4ch>  dec0_params_buf_;
+
+  // Shared packed-f16 weights (storage buffer, read-only in all shaders)
+  GpuBuffer weights_buf_;
+
+  // Per-pass params shadow (updated by set_film_params, uploaded in render)
+  CnnV3Params4ch  enc0_params_{};
+  CnnV3ParamsEnc1 enc1_params_{};
+  CnnV3ParamsBn   bn_params_{};
+  CnnV3Params4ch  dec1_params_{};
+  CnnV3Params4ch  dec0_params_{};
+
+  void create_pipelines();
+  void update_bind_groups(NodeRegistry& nodes);
+};
diff --git a/src/effects/shaders.cc b/src/effects/shaders.cc
index 22c6a6d..f64e135 100644
--- a/src/effects/shaders.cc
+++ b/src/effects/shaders.cc
@@ -117,6 +117,11 @@ const char* ntsc_rgb_shader_wgsl = SafeGetAsset(AssetId::ASSET_SHADER_NTSC_RGB);
 const char* ntsc_yiq_shader_wgsl = SafeGetAsset(AssetId::ASSET_SHADER_NTSC_YIQ);
 const char* gbuf_raster_wgsl     = SafeGetAsset(AssetId::ASSET_SHADER_GBUF_RASTER);
 const char* gbuf_pack_wgsl       = SafeGetAsset(AssetId::ASSET_SHADER_GBUF_PACK);
+const char* cnn_v3_enc0_wgsl        = SafeGetAsset(AssetId::ASSET_SHADER_CNN_V3_ENC0);
+const char* cnn_v3_enc1_wgsl        = SafeGetAsset(AssetId::ASSET_SHADER_CNN_V3_ENC1);
+const char* cnn_v3_bottleneck_wgsl  = SafeGetAsset(AssetId::ASSET_SHADER_CNN_V3_BOTTLENECK);
+const char* cnn_v3_dec1_wgsl        = SafeGetAsset(AssetId::ASSET_SHADER_CNN_V3_DEC1);
+const char* cnn_v3_dec0_wgsl        = SafeGetAsset(AssetId::ASSET_SHADER_CNN_V3_DEC0);
 
 // Compute shaders
 const char* gen_noise_compute_wgsl =
diff --git a/src/effects/shaders.h b/src/effects/shaders.h
index cf095fb..4a77597 100644
--- a/src/effects/shaders.h
+++ b/src/effects/shaders.h
@@ -24,6 +24,13 @@ extern const char* ntsc_yiq_shader_wgsl;
 extern const char* gbuf_raster_wgsl;
 extern const char* gbuf_pack_wgsl;
 
+// CNN v3 inference shaders
+extern const char* cnn_v3_enc0_wgsl;
+extern const char* cnn_v3_enc1_wgsl;
+extern const char* cnn_v3_bottleneck_wgsl;
+extern const char* cnn_v3_dec1_wgsl;
+extern const char* cnn_v3_dec0_wgsl;
+
 // Compute shaders
 extern const char* gen_noise_compute_wgsl;
 extern const char* gen_perlin_compute_wgsl;
diff --git a/src/gpu/demo_effects.h b/src/gpu/demo_effects.h
index 91ab6f2..66b920c 100644
--- a/src/gpu/demo_effects.h
+++ b/src/gpu/demo_effects.h
@@ -32,8 +32,9 @@
 #include "effects/scratch_effect.h"
 #include "effects/ntsc_effect.h"
 
-// CNN v3 G-buffer
+// CNN v3 G-buffer + inference
 #include "../../cnn_v3/src/gbuffer_effect.h"
+#include "../../cnn_v3/src/cnn_v3_effect.h"
 
 // TODO: Port CNN effects
 // #include "../../cnn_v1/src/cnn_v1_effect.h"
diff --git a/src/tests/gpu/test_demo_effects.cc b/src/tests/gpu/test_demo_effects.cc
index c2588f2..f5af5a9 100644
--- a/src/tests/gpu/test_demo_effects.cc
+++ b/src/tests/gpu/test_demo_effects.cc
@@ -84,6 +84,11 @@ static void test_effects() {
            fixture.ctx(), std::vector<std::string>{"source"},
            std::vector<std::string>{"gbuf_feat0", "gbuf_feat1"}, 0.0f,
            1000.0f)},
+      {"CNNv3Effect",
+       std::make_shared<CNNv3Effect>(
+           fixture.ctx(),
+           std::vector<std::string>{"gbuf_feat0", "gbuf_feat1"},
+           std::vector<std::string>{"cnn_v3_output"}, 0.0f, 1000.0f)},
   };
 
   int passed = 0;