{"slug": "neural-block-texture-compression-with-cuda", "title": "Neural block texture compression with CUDA", "summary": "A developer built a neural block texture compression system that runs entirely on the GPU using CUDA 13.1, combining per-pixel selectors, per-4×4-block latents, and a tiny 1,767-weight MLP decoder trained with evolution strategies rather than backpropagation. The method reaches 40.78 dB in fp32 and 40.71 dB with the block latent quantized to 8 bits, at 6.03 bpp raw and 5.54 bpp entropy-coded, completing 3,000 iterations in 55 seconds on an RTX 5090.", "body_md": "The 2 latent layers (containing the per-block parameters) are sampled with nearest sampling.\n- Format: per-pixel selectors + one latent per 4×4 block + tiny MLP block decoder\n- Selectors: 1024×1024×2, 3 bits + 1 bit per pixel, picked by exact exhaustive search each iteration\n- Block latent: 256×256×4, 8 bits/value, trained by ES with footprint attribution\n- Decoder: 8→36→36→3 MLP, 1767 weights; inputs = 2 selectors + 4 block values + 2 block-local coords\n- Decoder training: antithetic ES (64 pairs, 4096-pixel minibatches) then central finite differences for the last quarter\n- Latent training: 4 antithetic pairs/iteration on the full image, no backprop anywhere\n- 3000 iterations, lr annealed 1× → 0.05× over the second half\n- 40.78 dB fp32, 40.71 dB with the block latent quantized to 8 bits\n- 6.03 bpp raw (4 bpp selectors + 2 bpp block latent + 0.05 bpp MLP), 5.54 bpp entropy-coded\n- 96 bits per 4×4 block: 64 selector bits + 32 latent bits\n- 55 s total on an RTX 5090, everything on the GPU (CUDA 13.1)  \n\nLeft: original, right: reconstruction \n\nThe block artifacts are there if you look closely:", "url": "https://wpnews.pro/news/neural-block-texture-compression-with-cuda", "canonical_source": "https://richg42.blogspot.com/2026/09/neural-block-texture-compression-with.html", "published_at": "2026-09-05 16:29:34+00:00", "updated_at": "2026-09-14 17:25:14.716107+00:00", "lang": "en", "topics": ["machine-learning", "neural-networks", "computer-vision", "ai-research", "ai-infrastructure"], "entities": ["CUDA", "RTX 5090", "NVIDIA"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/neural-block-texture-compression-with-cuda", "markdown": "https://wpnews.pro/news/neural-block-texture-compression-with-cuda.md", "text": "https://wpnews.pro/news/neural-block-texture-compression-with-cuda.txt", "jsonld": "https://wpnews.pro/news/neural-block-texture-compression-with-cuda.jsonld"}}