Richard Geldreich's Blog: Notes on NNTC vs. neural texturing

NNTC uses a bilinear/degree-2 polynomial decoder fitted per PBR material, in contrast to solutions using full MLP-based neural networks. While MLPs typically achieve slightly better PSNR (1-3 dB), they make GPU texture filtering harder and training is far more expensive. After optimizing the NNTC-specific CUDA encoder with BC4/BC5 awareness, it surpassed the neural network approach on raw PSNR. The author argues that correctly configured latent textures using traditional high-resolution weight planes with supercompression (RDO+LZ, DCT) make expensive MLP decoding unnecessary for PBR textures, and that NNTC's CUDA-based encoding in seconds is hard for backprop-based neural solutions to match.

Source: https://richg42.blogspot.com/2026/09/notes-on-nntc-vs-neural-texturing.html. 8 Sync News only summarizes and links out; content copyright belongs to the authors and original sources.