Path Guiding with Vertex Triplet Distributions

Good importance sampling strategies are decisive for the quality and robustness of photorealistic image synthesis with Monte Carlo integration. Path guiding approaches use transport paths sampled by an existing base sampler to build refine a distribution. This distribution then guides subsequent in regions that otherwise hard sample. We observe all terms measurement contribution function during path construction depend on at most three consecutive vertices. thus propose 9D over vertex triplets adapts full Gaussian mixture model (GMM). For incremental sampling, we query last two vertices prefix, resulting 3D conditional which sample next along path. To make this approach scalable, partition scene octree learn local GMM each leaf separately. In learning phase, using current collect resample these online keep only fixed-size subset reservoirs. After progression, obtain new GMMs from triplet samples initial clustering followed expectation maximization. Since positions, our naturally extends participating media. addition, symmetry allows us it constructed light tracer. Therefore method can guide both tracer jointly learned