← All resources

Fitting Neural Textures and PBR Material Maps with ES (No Backprop)

Auto-discovered

Blog Advanced

Richard Geldreich

Visit resource ↗

Description

Fitting Neural Textures and PBR Material Maps with ES (No Backprop) is a September 2026 Richard Geldreich blog post that documents a C++ experiment mirrored in public source code. Geldreich explores fitting neural textures and multi-map PBR materials using evolution strategies instead of gradient backpropagation through the decoder.

The approach stores images as low-resolution latent grids decoded by a small MLP, targeting albedo and related PBR channels with quantization-aware training goals. The write-up discusses why ES is attractive when derivatives through rendering or compression are awkward, and shares qualitative comparisons alongside the linked GitHub repository released the same week.

Advanced graphics programmers familiar with texture baking, material authoring, and optimization will get the most from the post. It assumes comfort reading experimental code paths and interpreting loss curves without a polished production toolchain.

Geldreich's long-standing graphics blog and open-source history make his optimization experiments a common reference point for engine programmers experimenting with neural material.