Richard Geldreich's Blog: Neural GPU block compression

An experimental approach to neural GPU texture block compression is described, using a small MLP (7→36→36→3, 1,731 weights) trained with Evolution Strategies over 3,000 iterations to decode block textures from latent data. The method uses two latent textures (a full-res 3-bpp texture and a quarter-res 32-bpp control texture), achieving roughly 5.0 bpp amortized across a texture stack, with support for PBR material stacks of up to 4 correlated textures. A variant adds in-loop deblocking with a 5-tap filter, boosting quality by about 0.9 dB in one test. Because Evolution Strategies is derivative-free, the GPU texture transcoder (e.g., a BC7 encoder with discrete mode decisions) can be placed directly inside the optimization loop without needing differentiability. Some visible BC1-like chroma block artifacts remain, with ideas proposed to address them, and a prior art disclosure repo is mentioned as forthcoming.

Source: https://richg42.blogspot.com/2026/09/neural-gpu-block-compression.html. 8 Sync News only summarizes and links out; content copyright belongs to the authors and original sources.