Random-access Neural Compression of Material Textures
DescriptionThis work introduces a neural compression technique for mipmapped material texture sets, offering significantly better compression than BCx at comparable quality and even surpassing entropy-coded AVIF and JPEG XL at low bitrates. Our method uses small, optimized neural networks for efficient compression, real-time decompression, and random access on GPUs.
TimeSunday, 6 August 20236pm - 8:30pm PDT
Production & Animation
Research & Education
Artificial Intelligence/Machine Learning