← All resources

Neural Texture Compression in Metal 4

Auto-discovered

Blog Advanced

Kiriakos Gavras

Visit resource ↗

Description

Neural Texture Compression in Metal 4 is Kyriakos Gavras' September 2026 article documenting MetalNTC, an Apple Silicon implementation inspired by NVIDIA random-access neural material compression research. The post documents learned texture codecs for custom Metal renderers.

It covers ground-truth preparation, MLP training, quantization, Metal inference, latent grids, per-pixel decode, and quality presets, with notes on M-series neural accelerators. Figures compare latent visualizations, mip selection, and inferred textures against baseline maps.

Advanced Metal and ML engineers should understand texture sampling and basic training loops.

Published on Gavras' blog with PDF export and linked source.

Discussed in neural rendering and Metal communities as a practical port of research-grade NTC to Apple's GPU stack.