Developer posts neural texture compression demo

Kyriakos Gavras posted details on September 8 of a six-month project implementing neural texture compression with Apple’s Metal stack, describing a system that trains and runs a small decoder network on-device for real-time rendering. Gavras said the work includes Metal-based training, real-time inference, a demo renderer and notes on how to integrate the technique into a graphics pipeline. The post follows a line of research that uses compact latent textures plus a small neural network to reconstruct material data at sample time, rather than storing each texture map in conventional form. NVIDIA researchers described that approach in a 2023 SIGGRAPH paper on random-access neural compression of material textures. ### What did the developer actually build? Kyriakos Gavras described the project as an implementation of neural texture compression in Metal 4, based on NVIDIA’s “Random-Access Neural Compression of Material Textures.” Gavras wrote that he wanted texture compression for a custom engine and was drawn to the quality and compression trade-offs shown in the research. The project page says the implementation covers both the offline training side and the runtime path needed to decode textures during rendering. (x.com) The September 8 social post said the work took six months and packaged the result as more than a paper reproduction. Gavras said the release includes a demo renderer and integration instructions, which turns the project into a working reference for developers evaluating the technique in their own engines. ### How does neural texture compression differ from ordinary texture compression? NVIDIA’s 2023 paper says the method compresses multiple material textures and their mip chains together, then reconstructs texels on demand with a small neural network optimized for each material. (syllogi-graphikon.vercel.app) That differs from standard GPU block compression formats, which typically treat channels more independently and store directly sampleable compressed texels. NVIDIA said its method was designed to preserve random access, so decompression can happen in real time in a way that fits rendering workloads. (x.com) The NVIDIA research page says the technique can unlock “two more levels of detail,” or 16 times more texels, while using less memory than a higher-quality BC baseline in the example shown. The same page says the method targets both disk and memory compression, which is central to why texture compression matters for graphics workloads on memory-constrained devices. (research.nvidia.com) ### Why does Apple’s Metal matter here? Apple said in its WWDC26 session on Metal 4 that developers can train and deploy neural models in real-time rendering pipelines, including with an ML command encoder and shader-level TensorOps. Apple also said Metal 4 lets developers build and evaluate “small, specialized neural networks directly within your shaders,” which matches the kind of lightweight decoder neural texture compression uses. (research.nvidia.com) Apple’s session described neural textures as one of several rendering techniques that can use machine learning to improve quality, performance or memory footprint. That makes Gavras’s project notable as a concrete implementation on Apple hardware rather than a paper-only concept or a Windows-first SDK. ### What does the demo suggest about real-time use? Gavras said the project includes real-time inference and a demo renderer, indicating the decoder is fast enough to sit inside an interactive rendering loop rather than only an offline transcoding stage. (developer.apple.com) NVIDIA’s broader SDK materials describe multiple deployment modes, including decoding in shaders and using sample renderer applications to show compressed materials in 3D scenes. The emphasis on integration notes also matters because neural compression systems usually need more than a trained model: they need asset preprocessing, runtime scheduling, shader hooks and hardware-aware optimization. Gavras’s project page presents the work in that engine-facing form. ### Where can readers find the next step? Gavras’s project write-up is published on his site, and the September 8 post links readers to the demo and implementation notes. (x.com) Apple’s WWDC26 Metal 4 session and NVIDIA’s 2023 research page provide the closest primary references for the APIs and compression method the project builds on. (syllogi-graphikon.vercel.app)