demo.git - Vide-coded 64k demo system

Age	Commit message (Expand)	Author
10 hours	Replace hard clamp with sigmoid activation in CNN v2	skal
10 hours	Streamline CNN v2 training pipeline output	skal
10 hours	Fix CNN v2 training: always save final checkpoint, derive num_layers	skal
11 hours	Fix --mix option: blend prev layer with static p4-p7, not p0-p3	skal
11 hours	Fix CNN v2 static feature channel mapping (p4-p7 → channels 8-11)	skal
11 hours	gen_identity_weights: Change --mix to 50-50 blend	skal
11 hours	gen_identity_weights: Add --p47 option for static feature visualization	skal
11 hours	gen_identity_weights: Add --mix option for static feature blending	skal
12 hours	CNN v2: Fix Layer 0 visualization scale (was 0.5, now 1.0)	skal
12 hours	CNN v2: Add debugging tools for mismatch investigation	skal
12 hours	CNN v2 training: Fix float64/float32 dtype mismatch in depth feature	skal
12 hours	CNN v2: Alpha channel depth handling and layer visualization	skal
13 hours	CNN v2: Use alpha channel for p3 depth feature + layer visualization	skal
18 hours	CNN v2 training: Add --grayscale-loss option for luminance-based loss computa...	skal
19 hours	CNN v2: Change feature #6 from sin(10x) to sin(20y)	skal
19 hours	CNN v2: Add TODO for flexible feature layout in binary format v3	skal
19 hours	CNN v2: Add mip-level support to runtime effect	skal
19 hours	CNN v2 export: Read and display mip_level from checkpoints	skal
19 hours	CNN v2: Add --mip-level option for parametric features	skal
20 hours	CNN v2: Fix activation function mismatch between training and inference	skal
23 hours	CNN v2 training: Use target image alpha channel	skal
23 hours	CNN v2: Restore per-layer kernel sizes support	skal
23 hours	CNN v2: Refactor to uniform 12D→4D architecture	skal
27 hours	Add weights/ subdirectory to workspaces for CNN training outputs	skal
45 hours	test_demo: Add beat-synchronized CNN post-processing with version selection	skal
47 hours	Refine training script output and validation	skal
48 hours	TODO: 8-bit weight quantization for 2× size reduction	skal
48 hours	CNN v2: Storage buffer complete - real weights exported	skal
48 hours	CNN v2: storage buffer architecture foundation	skal
48 hours	TODO: Add random sampling to patch-based training	skal
48 hours	CNN v2: Patch-based training as default (like CNN v1)	skal
48 hours	Fix: CNN v2 training - handle variable image sizes	skal
2 days	CNN v2: parametric static features - Phases 1-4	skal
2 days	remove more stale files	skal
2 days	feat: implement beat-based timing system	skal
2 days	add trained layers	skal
3 days	docs: Update CNN comments and add bias fix summary	skal
3 days	fix: CNN bias accumulation and output format improvements	skal
3 days	update cnn code	skal
3 days	refactor: Use linspace(-1,1) directly for coords	skal
3 days	fix: Compute gray from [0,1] RGB in CNN shader generator	skal
3 days	fix: Complete auxiliary texture initialization fix	skal
3 days	add --save-intermediates to train.py and cnn_test	skal
3 days	fix: Move sigmoid activation to call site in CNN layer shader	skal
3 days	fix: Replace clamp with sigmoid in CNN final layer	skal
3 days	feat: Add early stopping to CNN training	skal
3 days	fix: CNN training/inference to match WGSL sliding window	skal
3 days	format .wgsl layer code (cosmetics)	skal
3 days	fix: Use patch-based inference to match CNN training distribution	skal
3 days	opt: Move invariant in1 calculation outside CNN convolution loops	skal