Physics-based forward rendering is a central topic in computer graphics and focuses on synthesizing physically accurate images of fully described virtual scenes. Differentiable rendering, on the contrary, is concerned with differentiating forward renderings with respect to parameters (e.g., the pose or shape of an object, and the reflectance of a surface). Due to its applications in a wide array of areas including computer graphics, vision, computational imaging, and virtual/augmented reality (VR/AR), differentiable rendering has recently been an active research topic.

A grand challenge in differentiable rendering is to develop efficient systems that can scale out to high-resolution images and large numbers (e.g., billions) of scene parameters. To overcome this challenge, we introduce PSDR-Enzyme: a new differentiable renderer developed in C++ based on the Enzyme AD framework. This system implements state-of-the-art algorithms (published in SIGGRAPH conferences) and offers a similar level of performance as many GPU-based systems. In this talk, we will go over the design of PSDR-Enzyme, demonstrate its applications, and discuss future challenges.