The "compression" process produces a set of latent parameters covering the texture space and a lightweight multi-layer perceptron (2x64 nodes) which is small enough to inference at runtime - either on load or within the sampling pixel shader itself using new Cooperative Vector instructions.