Path Guiding Using Spatio‐Directional Mixture Models

We propose a learning-based method for light-path construction in path tracing algorithms, which iteratively optimizes and samples from what we refer to as spatio-directional Gaussian mixture models (SDMMs). In particular, approximate incident radiance an online-trained 5D that is accelerated by D-tree. Using the same framework, BSDFs pre-trained D mixtures, where number of BSDF parameters. Such approach addresses two major challenges path-guiding models. First, representation naturally captures correlation between spatial directional dimensions. correlations are present in, example parallax caustics. Second, using tangent-space parameterization Gaussians, our mixtures can perform product sampling with arbitrarily oriented BSDFs. Existing only able do this either foregoing anisotropy components or representing field local (normal aligned) coordinates, both make more difficult learn. An additional benefit each individual mapped solid sphere low distortion near its centre mass. Our performs especially well on scenes small, localized luminaires induce high radiance.