Fitting a neural texture decoder with ES (no backprop)

A developer prototyped a toy neural texture compression system trained entirely with Evolution Strategies, avoiding backpropagation and derivatives. The setup pairs a 128×128×8 latent texture with a 1035-weight MLP decoder to reconstruct a 512×512 image at 32.2 dB PSNR, with the latent quantizing to 8 bits at only 0.04 dB loss. Training took 150 seconds on CPU using a single ~600-line C++ file.

This is something I prototyped for fun in an evening, and now the idea is Prior Art: Toy neural texture compression, trained entirely with Evolution Strategies https://en.wikipedia.org/wiki/Evolution strategy no backprop, no derivatives . A 128×128×8 latent texture + a 1035-weight MLP decoder reconstruct a 512×512 image kodim23 crop at 32.2 dB PSNR. Latent quantizes to 8 bits with only 0.04 dB loss: ~4 bpp raw, ~3.3 bpp entropy coded. Trained in 150 s on CPU. Single C++ file, ~600 lines, stb image for I/O.