Conditional Resampled Importance Sampling and ReSTIR

SIGGRAPH Asia (Conference Track), 2023

Markus Kettunen*Daqi Lin*Ravi Ramamoorthi
Thomas Bashford-RogersChris Wyman
University of WarwickNVIDIA
*Joint First Authors
[Paper (Preprint)]
[Supplemental Document]
[Paper Video] [Source Code]
[Interative Comparisons]


Recent work on generalized resampled importance sampling (GRIS) enables importance-sampled Monte Carlo integration with random variable weights replacing the usual division by probability density. This enables very flexible spatiotemporal sample reuse, even if neighboring samples (e.g., light paths) have intractable probability densities. Unlike typical Monte Carlo integration, which samples according to some PDF, GRIS instead resamples existing samples. But resampling with GRIS assumes samples have tractable marginal contribution weights, which is problematic if reusing, for example, light subpaths from unidirectionally-sampled paths. Reusing such subpaths requires conditioning by (non-reused) segments of the path prefixes.

In this paper, we extend GRIS to conditional probability spaces, showing correctness given certain conditional independence between integration variables and their unbiased contribution weights. We show proper conditioning when using GRIS over randomized conditional domains and how to formulate a joint unbiased contribution weight for unbiased integration.

To show our theory has practical impact, we prototype a modified ReSTIR PT with a final gather pass. This reuses subpaths, postponing reuse at least one bounce along each light path. As in photon mapping, such a final gather reduces blotchy artifacts from sample correlation and reduced correlation improves the behavior of modern denoisers on ReSTIR PT signals.

cris-fig Our new conditional RIS theory enables new types of unbiased subpath reuse by resampling in conditional probability spaces. To show the theory has practical use, we prototype an algorithm resampling multiple ReSTIR-driven path suffixes in a photon map like final gather. While our proof-of-concept is unoptimized, we compare with two state-of-the-art methods without conditional resampling, including a final gather using West et al.’s [2022] marginal multiple importance sampling (MMIS) and full-path resampling using Lin et al.’s [2022] ReSTIR PT sample code. The Tower Bridge [Pobursky 2021] is lit by the Shanghai Bund probe; the camera sees an almost entirely indirectly lit region. ReSTIR PT is very fast, but complex lighting plus specular surfaces can cause large spatiotemporal correlations, boiling, and color shifts (left inset). While currently more expensive, our subpath resampling gives spatiotemporally stable results without visible correlations. Compared to an MMIS gather, our prototype improves quality given a similar ray budget. (Bottom) Below each image we show (𝑥, 𝑡) plots taken from videos (without movement); rows come from sequential video frames, so temporal correlations appear as vertical blobs and spatial correlations show up as horizontal blobs. All techniques are unbiased, converging to reference in time, but results here use only one full path per-pixel for integration.



  title         = {Conditional Resampled Importance Sampling and ReSTIR},
  author        = {Kettunen, Markus and Lin, Daqi and Ramamoorthi, Ravi and Bashford-Rogers, Thomas and Wyman, Chris},
  month         = {December},
  booktitle     = {SIGGRAPH Asia (Conference Track)},
  year          = {2023},
  doi           = {10.1145/3610548.3618245},