Compatibility-Guided Neighbor Selection for ReSTIR

Abstract

Reservoir-based spatiotemporal importance resampling (ReSTIR) improves convergence of real-time Monte Carlo rendering by reusing samples between pixels and frames. Recent reservoir splatting and area ReSTIR methods improve ReSTIR’s temporal robustness in challenging scenes, but much less work focuses on improving spatial reuse. We present a novel neighbor selection algorithm for ReSTIR’s spatial reuse. Our simple method works with any pixel-space ReSTIR technique to improve image quality; we reduce SMAPE by 6–29% while also decreasing temporal covariance by 22–49%, all with a 2–5% incremental cost. We ground our work in an analysis of path compatibility between pixels and an empirical evaluation of the fundamental error-correlation tradeoff in ReSTIR; we improve the Pareto frontier of this tradeoff and provide tunable control over the balance between variance and correlation.