NVIDIA Introduces Revolutionary Neural Texture Compression for Material Textures

NVIDIA has unveiled a cutting-edge compression algorithm for material textures, dubbed Neural Texture Compression (NTC). This innovative algorithm tackles the escalating memory requirements of high-resolution textures employed in photorealistic rendering.

NVIDIA's research team has developed a new neural compression method specifically tailored for material textures, delivering 16 times more texels than the conventional GPU-based texture compression technique, Block Compression (BC). The NTC algorithm leverages a compact neural network customized for each material, facilitating on-demand, real-time decompression with random access akin to block texture compression on GPUs.

The NTC algorithm enables two additional levels of detail and offers superior image quality compared to sophisticated image compression methods like AVIF and JPEG XL. Unlike standard BCx algorithms that necessitate specialized hardware, NTC employs matrix multiplication techniques, which modern GPUs now accelerate. This renders the NTC algorithm more feasible and versatile due to reduced disk and memory limitations.

NVIDIA researchers are set to reveal further details about this groundbreaking technology at SIGGRAPH 2023 on August 6. According to their paper, neural textures can be rendered in real-time, achieving up to 16x more texel than the BC methodology. The expense of a 4K render using NTC is 1.15ms, marginally higher than the 0.49ms cost observed on the RTX 4090.