Photon-Driven Manifold Sampling

Abstract

Complex lighting phenomenon such as the caustics often arises from specular chains of varying lengths. Thanks to recent advancement in specular light transport, we are able to address most effects caused by shorter specular chains using root-finding techniques. However, finding longer chains of valid specular interaction still remains problematic due to the curse of dimensionality. This creates a difficulty for methods that depend on random sampling to navigate the path space. In this paper, we present Photon-Driven Manifold Sampling (PMS), an unbiased method designed for sampling multi-bounce caustics, derived from Specular Manifold Sampling (SMS). We achieve efficient caustic rending through the combination of photon mapping and path space manifold exploration. Our method, inspired by recent advancements in path reusing techniques, caches multi-bounce caustic photons with the aim of reusing their vertices to seed local exploration via manifold walk. This eliminates the necessity for additional random sampling in seed path sampling as seen in alternative methods that took the guiding approach. Our results highlight a notable decrease in variance compared to other state-of-the-art techniques and demonstrate the adaptability to intricate specular chains.