Many tasks in graphics and vision require computing derivatives of integrals of discontinuous functions, which have previously either required specialized routines or suffered from high variance. We introduce a program transform and boundary sampling technique that computes accurate derivatives for arbitrary shader programs, enabling a range of applications such as painterly rendering, CSG, rasterization, discontinuous textures, and more.

Recent methods have been developed to reconstruct 3D hair strand geometry from images. We introduce an inverse hair grooming pipeline to transform these unstructured hair strands into procedural hair grooms controlled by a small set of guide strands and artist-friendly grooming operators, enabling easy editing of hair shape and style.

Variance reduction techniques for Monte Carlo integration are typically designed for scalar-valued integrands, even though many rendering and inverse rendering tasks actually involve vector-valued integrands. We show that ratio control variates, compared to conventional difference control variates, can significantly reduce the error of vector-valued integration with minimal overhead.

We introduce a spatiotemporal optimizer for inverse rendering which combines the temporal filtering of Adam with spatial cross-bilateral filtering to enable higher quality reconstructions in texture, volume, and geometry recovery.

We propose a real-time path guiding method, Voxel Path Guiding (VXPG), that significantly improves fitting efficiency under limited sampling budget. We show that our method can outperform other real-time path guiding and virtual point light methods, particularly in handling complex dynamic scenes.

Physically-based inverse rendering algorithms that utilize differentiable rendering typically use gradient descent to optimize for scene parameters such as materials and lighting. We observe that the scene often changes slowly from frame to frame during optimization, similar to animation, and therefore adapt temporal reuse from ReSTIR to reuse samples across gradient iterations and accelerate inverse rendering.