Richard Geldreich's Blog: Neural block texture compression trained with Evolution Strategies (ES) on GitHub

A self-contained C++/CUDA research testbed trains neural block texture compression using Evolution Strategies (ES) instead of backpropagation or autodiff. Textures are encoded as a shared low-resolution latent plus a tiny MLP decoder, both optimized via antithetic ES with footprint attribution for variance reduction, plus optional finite-difference polishing. The write-up documents extensive experiments: nearest-sampled block latents, in-loop deblocking, a learned per-pixel selector level chosen by exhaustive search (forming a BC-style block format), multi-channel selectors, and a GPU implementation running 20-30x faster than 32 CPU threads on an RTX 5090. Results span single images (kodim23, mario), multi-layer PBR materials, and game textures, with detailed PSNR/bitrate tables comparing configurations against formats like BC1/BC7. The code and a timestamped disclosure of the technique are published on GitHub as prior art.

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