Comparison between path tracing, ReSTIR GI [OLK*21] and our method with equal time on two scenes (SAN MIGUEL and BATHROOM) with complex indirect lighting. The mean squared errors (MSE) are shown under the images. The image resolution is 1920 x 1080. Our method outperforms the other methods both visually and quantitatively.

Abstract

With the advent of hardware-accelerated ray tracing, more and more real-time rendering applications tend to render images with ray-traced global illumination (GI). However, the low sample counts at real-time framerates bring enormous challenges to existing path sampling methods. Recent work (ReSTIR GI) samples indirect illumination effectively with a dramatic bias reduction. However, as a screen-space based path resampling approach, it can only reuse the path at the first bounce and brings subtle benefits for complex scenes. To this end, we propose a world-space based spatiotemporal path resampling approach. Our approach caches more path samples into a world-space grid, which allows reusing sub-path starting from non-primary path vertices. Furthermore, we introduce a practical normal-aware hash grid construction approach, providing more efficient candidate samples for path resampling. Eventually, our method achieves improvements ranging from 16.6% to 41.9% in terms of mean squared errors (MSE) compared against the previous method with only 4.4% - 8.4% extra time cost.

Downloads

BibTex

@article{Zhang:2023:WorldReSTIR,
  author = {Hangyu Zhang and Beibei Wang},
  title = {World-Space Spatiotemporal Path Resampling for Path Tracing},
  journal ={Computer Graphics Forum (Proceedings of PG 2023)},
  year = {2023},
}

Back to Homepage