Random-Access Neural Compression of Material Textures

A new Siggrapg 2023 research paper from NVIDIA that improves the GPU compression by 10-16x while keeping random-access property. It exploits correlation in modern graphics texturing practices (multiple textures describing a single physical material) and at the lowest bitrates, it is competitive with entropy-coded JPEG XL and AVIF formats.

Despite using neural networks, compression time is approximately 1 minute for a 4K 9 channel texture, and the decompression overhead ~0.6ms at 2160p (4K) resolution.

This technique can substantially reduce the download and disk sizes of modern games, reduce the memory usage, or significantly improve the details and quality.