From b8b2707d7b3f2eaabd896395f2625e13405a24e2 Mon Sep 17 00:00:00 2001
From: skal <pascal.massimino@gmail.com>
Date: Sat, 21 Mar 2026 08:55:20 +0100
Subject: docs: session handoff — CNN v3 Phase 4 complete
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- TODO.md: mark Phase 4 done, add FiLM MLP training details (blocked on
  train_cnn_v3.py), clarify what 'real' set_film_params() requires
- COMPLETED.md: archive Phase 4 with alignment fix note (vec3u→64/96 bytes)

handoff(Gemini): next up CNN v3 Phase 5 (parity validation) or train_cnn_v3.py
---
 TODO.md | 19 +++++++++++++------
 1 file changed, 13 insertions(+), 6 deletions(-)

(limited to 'TODO.md')

diff --git a/TODO.md b/TODO.md
index e33072f..11c629e 100644
--- a/TODO.md
+++ b/TODO.md
@@ -69,16 +69,23 @@ PyTorch / HTML WebGPU / C++ WebGPU.
 **Design:** `cnn_v3/docs/CNN_V3.md`
 
 **Phases:**
-1. ✅ G-buffer: `GBufferEffect` integrated. Assets, CMake, demo_effects.h, test all wired. 35/35 tests pass.
-   - NodeTypes: `GBUF_ALBEDO`, `GBUF_DEPTH32`, `GBUF_R8`, `GBUF_RGBA32UINT`
-   - Shaders: `cnn_v3/shaders/gbuf_raster.wgsl` (ShaderComposer), `gbuf_pack.wgsl`
-   - SDF/shadow passes TODO (placeholder: shadow=1, transp=0)
-   - Howto: `cnn_v3/docs/HOWTO.md`
+1. ✅ G-buffer: `GBufferEffect` integrated. SDF/shadow placeholder (shadow=1, transp=0).
 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`: 5 compute passes, FiLM uniform upload, `set_film_params()` API
+   - Params alignment fix: WGSL `vec3u` align=16 → C++ structs 64/96 bytes
+   - Weight offsets as explicit formulas (e.g. `20*4*9+4`)
+   - FiLM γ/β: identity defaults; real values require trained MLP (see below)
 5. Parity validation (test vectors, ≤1/255 per pixel)
 
+**FiLM MLP training** (blocks meaningful Phase 4 output):
+- Needs `cnn_v3/training/train_cnn_v3.py` — not yet written
+- MLP: `Linear(5→16) → ReLU → Linear(16→48)` trained jointly with U-Net
+- Input: `[beat_phase, beat_time/8, audio_intensity, style_p0, style_p1]`
+- Output: γ/β for enc0(4ch) + enc1(8ch) + dec1(4ch) + dec0(4ch) = 40 floats
+- Trained weights (~3 KB f16) stored in `.bin` after conv weights; loaded at runtime
+- See `cnn_v3/docs/CNN_V3.md` §5 for full MLP spec and §11 for training pipeline plan
+
 ## Future: CNN v2 8-bit Quantization
 
 Reduce weights from f16 (~3.2 KB) to i8 (~1.6 KB).
-- 
cgit v1.2.3