Compact Representation of Mipmapped SVBRDFs via Shared Gaussians

Mipmapped SVBRDF stacks are expensive to ship: they combine high-resolution, multi-channel material maps with multiple mip levels, and that quickly turns into a storage problem for games and tools. A new technique called Gaussian Texture Compression (GTC) attacks that cost with a compact 2D Gaussian representation built around shared spatial support.

The key idea is to exploit two common redundancies at once: the same surface structure tends to repeat across mip levels, and the same support also repeats across different material maps. In GTC, the Gaussian footprint is shared while the values attached to it can vary per level and per map, which makes the representation more compact without forcing a neural network into the decode path.

That matters because the current options all have awkward tradeoffs. Block codecs like ASTC are hardware-friendly and support random access, but they only squeeze redundancy locally. Image codecs can compress better, but they are not designed for direct texture access in a renderer. Neural texture compression can go even further on size, but decoding cost is a real concern, especially on mobile and other latency-sensitive targets.

GTC is trained with a progressive optimization pipeline and is designed for non-neural, random-access decoding. In experiments, it reportedly beats ASTC on both reconstruction quality and memory usage while keeping the runtime model simple enough for real-time rendering. For teams shipping lots of material variants, that could translate into smaller texture budgets without sacrificing the access patterns engines...