UC Irvine

Physics-based differentiable rendering---which is concerned with estimating derivatives of photorealistic rendering with respect to arbitrary scene parameters---has a diverse array of applications from solving inverse-rendering (aka. analysis-by-synthesis) problems to incorporating forward rendering into probabilistic-inference and machine-learning pipelines. Recently, great progress has been made in physics-based differentiable rendering. Unfortunately, most existing techniques lack the generality to support volumetric light transport and the efficiency to handle complex geometries and light transport effects. To address these problems, we introduce in this dissertation a fundamentally new path-space differentiable rendering framework. Specifically, by differentiating the forward-rendering path integrals with respect to arbitrary scene parameters, we establish the mathematical formulation of differential path integrals that capture both interfacial and volumetric light transport. Based on this formulation, we develop new unbiased and consistent differentiable rendering algorithms capable of efficiently handling challenging geometric discontinuities and light-transport phenomena such as soft shadows, interreflection, and caustics. To further improve the robustness of our techniques, we leverage antithetic sampling to efficiently differentiate glossy BSDFs and pixel reconstruction filters.

Lastly, we present the formulation of differential image-loss path integrals that expresses gradients of image losses as another form of differential path integrals. Based on this formulation, we develop a new approach that allows reverse-mode automatic differentiation to be integrated efficiently into our path-space differentiable rendering algorithms.