Nvidia's neural texture compression promises gpu-friendly visual upscaling

Nvidia has been under fire lately for its use of AI in its new graphical rendering techniques, particularly the upcoming DLSS 5.0 update. Critics argue that while the AI improves graphics, it does so at the cost of altering the artistic vision of game developers.

Enter neural texture compression (ntc): a deterministic solution

Nvidia is looking to redeem itself with a new technique called Neural Texture Compression (NTC), which uses AI but not in a generative way. Instead, it learns from a single frame to improve textures without generating new pixels or altering the artistic intent. The best part? It doesn't require a top-of-the-line GPU to use.

NTC works by adding an extra layer on top of graphical upscaling techniques, using photogrammetry and improvements from DLSS 5. It dynamically reconstructs textures in real-time using neural models, based on what's shown on screen (angle, lighting, shadows, etc.).

In a demo, Nvidia showcased the power of NTC, reducing 6.5GB of VRAM usage with traditional BCn compression to just 970MB while maintaining visual texture quality. This is achieved through