Richard Geldreich's Blog: Fitting Neural Textures and PBR Material Maps with ES (No Backprop)

A C++ research toy encodes RGB images or multi-layer PBR materials as a low-resolution latent texture plus a tiny MLP decoder, trained entirely with Evolution Strategies (no backprop, no autodiff). The key technique is 'footprint attribution': each pixel's loss difference is credited only to the latent texels its bilinear sampling actually reads, letting a 131k-value latent be updated with just 4 perturbation pairs per step. Results include PSNR/bitrate tables for single images and a 4-layer cobblestone material (normal, roughness, albedo, AO), plus ablations on learning-rate annealing, finite-difference decoder polishing, and two-level latent pyramids. The repository discloses these techniques as public prior art and lists several unimplemented ideas such as quantization-aware training and searching directly in block-compressed GPU texture formats.

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