BRDF Importance Baking: A Lightweight Neural Solution to Importance Sampling General Parametric BRDFs

Abstract Parametric bidirectional reflectance distribution functions (BRDFs) are widely used due to their flexibility in representing diverse material appearances by simply adjusting parameters. However, high‐quality importance sampling techniques for parametric BRDFs remain limited. Existing methods often rely on coarse approximations or limited subsets of BRDFs, resulting in high variance, and are frequently tailored to specific BRDF types. In this paper, we propose an efficient, reliable, and general method that produces high‐quality approximations for importance sampling of parametric BRDFs. We observe that importance sampling fundamentally involves mapping from a uniform distribution to the target distribution. Specifically, when BRDF parameters are provided, the mapping required for importance sampling on a BRDF slice can be represented as a 2D image, which we term an importance warping map (IWM). Leveraging this insight, we accurately precompute importance maps using optimal transport, then compress them with a lightweight neural network to streamline storage and access. This approach enables parametric BRDF importance sampling to be moved to the precomputation stage, eliminating heavy runtime computation. Since this process resembles light baking—where a set of images is precomputed—we refer to our method as importance baking. Together with neural approximations of both BRDF evaluation and probability density function (PDF) querying, our method supports complete multiple importance sampling (MIS) without requiring modifications to the rendering pipeline.