← All resources

Neural Texture Compression using Hypernetworks

Auto-discovered

Paper Advanced

L. Belcour

Visit resource ↗

Description

Neural Texture Compression using Hypernetworks is an EGSR 2026 deep-learning track paper by Laurent Belcour on compressing material textures with a shared hypernetwork instead of per-asset optimization loops. The arXiv preprint summarizes the EGSR contribution.

The method regresses both latent BC1 block features and MLP decoder weights from a single hypernetwork, avoiding per-material gradient descent at training time. Reported quality matches reference neural texture compressors while extending to multi-decoder setups and super-resolution variants useful for mip chains.

The audience is researchers and engine programmers exploring neural material storage for real-time rendering. Familiarity with BC block formats and small MLP decoders on GPU is assumed.

The paper PDF is linked from arXiv as open access. Belcour's prior neural compression work at Adobe and EGSR is well known in real-time shading research circles.

Link last verified accessible on 2026-09-28.