Intel shows Texture Set Neural Compression, claims up to 18x smaller texture sets

[harlan4096]

TSNC Variant A delivers 9x texture compression, lower-quality B variant up to 18x



Intel has shared new details on Texture Set Neural Compression, or TSNC, a neural texture compression technology first shown as an Intel Labs research demo at GDC 2025. The company now says it has rebuilt the project into a standalone SDK, with a new decompression API that can compile to C, C++, or HLSL. Intel says the decoder supports both a fallback fused multiply-add path for CPU and GPU, and a linear algebra path that uses XMX acceleration on supported Intel GPUs.

This is yet another compression algorithm based on neural networks, followed by NVIDIA, who also held a GDC talk and mentioned up to 85% compression. Intel is working on two variants, one offering up to 9x compression and another up to 18x. 

TSNC vs BC1

The basic idea is to go beyond standard block compression by training a neural network to encode and decode a set of textures together. Intel stores the latent data in a four-layer BC1-based feature pyramid, then reconstructs the original texture channels through a three-layer MLP decoder. 

Continue Reading...