RGB ReSTIR: Decorrelating Spatiotemporal Importance Resampling with Per-Channel Reservoirs
2026, Mäkitalo, M., Haikio, S., Ikkala, J., Foi, A. and Jääskeläinen, P., In Proceedings of the 21st International Conference on Computer Graphics, Interaction and Visualization Theory and Applications (GRIVAPP).
Abstract: ReSTIR is a family of state-of-the-art spatiotemporal resampling algorithms utilized for improving the efficiency of photorealistic rendering. In particular, ReSTIR PT is commonly used for accelerating path tracing, a method that enables a high level of photorealism through a Monte Carlo based approximation of the global illumination. However, ReSTIR PT produces correlated samples due to its prominent reuse of spatially and temporally close pixels in the sample reservoirs, which typically manifests as visible color noise. ReSTIR-based algorithms typically only use luminance data for estimating their resampling target function, which means that in general, they cannot converge to a fully decorrelated image even with large reservoir sizes. In this paper, we present RGB ReSTIR, a multichannel variant of ReSTIR PT that maintains separate reservoirs and estimates separate target functions for each color channel. This approach allows the resampling to produce images with significantly less color noise than ReSTIR PT, especially for scenes with complex colored lighting. We demonstrate that RGB ReSTIR is able to converge towards a fully decorrelated image as the maximum confidence is increased (i.e., with longer temporal reservoir history), typically reaching an order of magnitude lower average sample autocovariance than ReSTIR PT.