Image-Space Russian Roulette for Monte Carlo Rendering

Geonu Noh , Chanu Yang , Bochang Moon

Gwangju Institute of Science and Technology

ACM SIGGRAPH ASIA 2026 Conference Proceedings

Image-Space Russian Roulette for Monte Carlo Rendering
Equal-time comparison between baseline path tracing with vertex-level Russian roulette (RR) (a) and our method (b). By integrating our unbiased image-space RR on top of baseline path tracing without modifying its underlying vertex-level RR, our approach skips a subset of sample evaluations prior to camera-ray construction while approximating their contributions, achieving lower error under the same rendering time. We measure efficiency as 1/(MSE × time) and report relative efficiency gains with respect to the equal-time baseline. For the two scenes, our method skips 80% and 68% of sample evaluations, respectively, reducing computation while mitigating RR-induced noise amplification. The per-pixel fraction of evaluated samples (d) visualizes how sample evaluations are spatially reduced. 3D model courtesy of cekuhnen (Coffee Machine) and NewSee2l035 (Staircase).

Abstract

Russian roulette (RR) is a widely used technique for reducing computational cost in Monte Carlo (MC) rendering, which often requires a large number of samples to achieve low-noise results. In standard practice, RR probabilistically terminates suffix paths at path vertices when their expected contribution is low, returning a predefined termination value (e.g., zero) in place of the skipped contribution. In this paper, we introduce an imagespace RR framework that operates directly in image space, enabling early termination of entire light paths before ray traversal. We generalize RR by introducing sample-dependent termination values, allowing skipped contributions to be approximated rather than discarded. To make this formulation practical, we exploit spatial correlation through a seed-shifting strategy that reuses primary samples across neighboring pixels. This enables efficient construction of both termination values and termination probabilities using correlated evaluations. We demonstrate that our method achieves lower error than baseline path tracing under equal-time conditions across a variety of scenes while remaining unbiased and compatible with existing techniques such as vertex-level RR.

Contents

BibTex