demo.git/cnn_v3/shaders, branch main

fix(cnn_v3): L1 loss + depth-grad tanh normalization to reduce flat convergence

2026-03-27T07:41:05Z

- Switch MSELoss → L1Loss in train_cnn_v3.py (median-seeking, avoids gray-blob) - Normalize depth_grad channels with tanh(10x) in cnn_v3_utils.py (bounds ±∞ signed values) - Match normalization in gbuf_pack.wgsl: tanh((right-left)*5.0) == tanh(10*central_diff) handoff(Gemini): training pipeline only; no C++ or test changes needed.

fix(cnn_v3): remove dec0 ReLU, load FiLM MLP at runtime

2026-03-27T06:59:00Z

Two bugs blocking training convergence: 1. dec0 ReLU before sigmoid constrained output to [0.5,1.0] — network could never produce dark pixels. Removed F.relu in train_cnn_v3.py and max(0,…) in cnn_v3_dec0.wgsl. Test vectors regenerated. 2. set_film_params() used hardcoded heuristics instead of the trained MLP. Added CNNv3FilmMlp struct + load_film_mlp() to cnn_v3_effect.h/.cc. MLP auto-loaded from ASSET_WEIGHTS_CNN_V3_FILM_MLP at construction; Linear(5→16)→ReLU→Linear(16→72) runs CPU-side each frame. 36/36 tests pass. Parity max_err=4.88e-4 unchanged. handoff(Gemini): retrain from scratch — needs ≥50 samples (currently 11). See cnn_v3/docs/HOWTO.md §2-3.

feat(cnn_v3): upgrade architecture to enc_channels=[8,16]

2026-03-26T06:03:01Z

Double encoder capacity: enc0 4→8ch, enc1 8→16ch, bottleneck 16→16ch, dec1 32→8ch, dec0 16→4ch. Total weights 2476→7828 f16 (~15.3 KB). FiLM MLP output 40→72 params (L1: 16×40→16×72). 16-ch textures split into _lo/_hi rgba32uint pairs (enc1, bottleneck). enc0 and dec1 textures changed from rgba16float to rgba32uint (8ch). GBUF_RGBA32UINT node gains CopySrc for parity test readback. - WGSL shaders: all 5 passes rewritten for new channel counts - C++ CNNv3Effect: new weight offsets/sizes, 8ch uniform structs - Web tool (shaders.js + tester.js): matching texture formats and bindings - Parity test: readback_rgba32uint_8ch helper, updated vector counts - Training scripts: default enc_channels=[8,16], updated docstrings - Docs + architecture PNG regenerated handoff(Gemini): CNN v3 [8,16] upgrade complete. All code, tests, web tool, training scripts, and docs updated. Next: run training pass.

feat(cnn_v3): 3×3 dilated bottleneck + Sobel loss + FiLM warmup + architecture PNG

2026-03-25T09:05:42Z

- Replace 1×1 pointwise bottleneck with Conv(8→8, 3×3, dilation=2): effective RF grows from ~13px to ~29px at ¼res (~+1 KB weights) - Add Sobel edge loss in training (--edge-loss-weight, default 0.1) - Add FiLM 2-phase training: freeze MLP for warmup epochs then unfreeze at lr×0.1 (--film-warmup-epochs, default 50) - Update weight layout: BN 72→584 f16, total 1964→2476 f16 (4952 B) - Cascade offsets in C++ effect, JS tool, export/gen_test_vectors scripts - Regenerate test_vectors.h (1238 u32); parity max_err=9.77e-04 - Generate dark-theme U-Net+FiLM architecture PNG (gen_architecture_png.py) - Replace ASCII art in CNN_V3.md and HOW_TO_CNN.md with PNG embed handoff(Gemini): bottleneck dilation + Sobel loss + FiLM warmup landed. Next: run first real training pass (see cnn_v3/docs/HOWTO.md §3).

feat(cnn_v3): shadow→dif migration complete (ch18)

2026-03-22T23:43:20Z

Replace raw shadow (ch18) with dif = max(0,dot(normal,KEY_LIGHT))*shadow across all layers. Channel count stays 20, weight shapes unchanged. - gbuf_pack.wgsl: t1.z = pack4x8unorm(mip2.g, mip2.b, dif, transp); t1.w = 0u - gbuf_deferred.wgsl: read dif from unpack4x8unorm(t1.z).z - gbuf_view.wgsl: revert to 4×5 grid, ch18=dif label, ch19=trns label - tools/shaders.js: FULL_PACK_SHADER adds oct_decode + computes dif - cnn_v3_utils.py: assemble_features() computes dif on-the-fly via oct_decode - docs: CNN_V3.md, HOWTO.md, HOW_TO_CNN.md, GBUF_DIF_MIGRATION.md updated handoff(Gemini): shadow→dif migration done, ready for first training pass

wip(cnn_v3): shadow→dif intermediate + scene tweaks + migration plan

2026-03-22T23:26:52Z

- gbuf_shadow.wgsl: normal bias 0.05→0.02 - gbuf_pack.wgsl: compute dif=diffuse*shadow, drop shadow from t1.z, store dif in t1.w (INTERMEDIATE — incorrect packing, see migration plan) - gbuf_deferred.wgsl: read dif from t1.w.x (matches intermediate packing) - gbuf_view.wgsl: expand to 4×6 grid, show dif.r/g/b in row 5 (INTERMEDIATE — to be reverted to 4×5 with ch18=dif) - gbuffer_effect.cc: add small hovering sphere (r=0.6) above scene; swap cube/sphere positions; both spheres pulsate - docs/GBUF_DIF_MIGRATION.md: full migration plan with checklist handoff(Claude): intermediate commit — GBUF_DIF_MIGRATION.md §Current State describes what is wrong and the full implementation checklist (5 steps).

refactor(cnn_v3): simplify sphere SDF in shadow pass, remove per-frame alloc

2026-03-22T22:51:40Z

gbuf_shadow.wgsl — dfWithID(): - Sphere: replace inv_model local-space transform with direct world-space formula (length(p - center) - radius). Exact, no matrix multiply, no floating-point error from matrix inversion that can corrupt soft-shadow penumbra over 64 march steps. - lp/scale now computed only inside the cases that need them (box/torus/plane) instead of eagerly for every object. gbuffer_effect.cc — upload_scene_data(): - Replace per-frame std::vector heap allocation with a file-static staging buffer s_obj_staging[256]: zero alloc per frame. handoff(Gemini): sphere SDF now exact; shadow march should be cleaner.

fix(cnn_v3): shadow pass — 5 bugs fixed, labels in gbuf_view

2026-03-22T22:17:50Z

1. Camera Y-inversion: proj.m[5] = -proj.m[5] in upload_scene_data + WGPUFrontFace_CCW on raster pipeline. 2. Shadow formula: replace shadowWithStoredDistance with 64-step IQ soft shadow (8*d/t, unbounded). 3. Local→world SDF scale: d *= length(obj.model[0].xyz). 4. Shadow bias: use rasterized normal from normal_mat_tex (binding 4) instead of light direction — fixes terminator self-shadow on spheres. 5. ShaderComposer: GBufViewEffect now resolves #include via ShaderComposer::Get().Compose(). Also: per-tile channel labels in gbuf_view.wgsl via debug_str. Scene simplified to 1 cube + 1 sphere for debugging (restore TODO). Scale propagation for pulsating sphere confirmed correct end-to-end. handoff(Gemini): shadow validated. Next: restore full scene in GBufferEffect::set_scene() (20 cubes + 4 spheres, 2 lights), then run training pass per cnn_v3/docs/HOWTO.md §3.

docs+feat(cnn_v3): compact context, re-enable shadow in GBufDeferredEffect

2026-03-22T19:31:45Z

- TODO/PROJECT_CONTEXT updated to reflect operational pipeline state - GBufDeferredEffect: shadow re-enabled (albedo * (ambient + diffuse * shadow)) feat_tex1 binding restored for shadow channel debugging handoff(Gemini): shadow pass live again — investigate why shadow looks broken.

feat(shaders): add ray_sphere snippet, use in gbuf_raster impostor

2026-03-22T19:28:48Z