NRRS: Neural Russian Roulette and Splitting

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Jin, Haojie, Ren, Jierui, Chen, Yisong, Wang, Guoping, Li, Sheng
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866914082441920512
author Jin, Haojie
Ren, Jierui
Chen, Yisong
Wang, Guoping
Li, Sheng
author_facet Jin, Haojie
Ren, Jierui
Chen, Yisong
Wang, Guoping
Li, Sheng
contents We propose a novel framework for Russian Roulette and Splitting (RRS) tailored to wavefront path tracing, a highly parallel rendering architecture that processes path states in batched, stage-wise execution for efficient GPU utilization. Traditional RRS methods, with unpredictable path counts, are fundamentally incompatible with wavefront's preallocated memory and scheduling requirements. To resolve this, we introduce a normalized RRS formulation with a bounded path count, enabling stable and memory-efficient execution. Furthermore, we pioneer the use of neural networks to learn RRS factors, presenting two models: NRRS and AID-NRRS. At a high level, both feature a carefully designed RRSNet that explicitly incorporates RRS normalization, with only subtle differences in their implementation. To balance computational cost and inference accuracy, we introduce Mix-Depth, a path-depth-aware mechanism that adaptively regulates neural evaluation, further improving efficiency. Extensive experiments demonstrate that our method outperforms traditional heuristics and recent RRS techniques in both rendering quality and performance across a variety of complex scenes.
format Preprint
id arxiv_https___arxiv_org_abs_2510_07868
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle NRRS: Neural Russian Roulette and Splitting
Jin, Haojie
Ren, Jierui
Chen, Yisong
Wang, Guoping
Li, Sheng
Graphics
We propose a novel framework for Russian Roulette and Splitting (RRS) tailored to wavefront path tracing, a highly parallel rendering architecture that processes path states in batched, stage-wise execution for efficient GPU utilization. Traditional RRS methods, with unpredictable path counts, are fundamentally incompatible with wavefront's preallocated memory and scheduling requirements. To resolve this, we introduce a normalized RRS formulation with a bounded path count, enabling stable and memory-efficient execution. Furthermore, we pioneer the use of neural networks to learn RRS factors, presenting two models: NRRS and AID-NRRS. At a high level, both feature a carefully designed RRSNet that explicitly incorporates RRS normalization, with only subtle differences in their implementation. To balance computational cost and inference accuracy, we introduce Mix-Depth, a path-depth-aware mechanism that adaptively regulates neural evaluation, further improving efficiency. Extensive experiments demonstrate that our method outperforms traditional heuristics and recent RRS techniques in both rendering quality and performance across a variety of complex scenes.
title NRRS: Neural Russian Roulette and Splitting
topic Graphics
url https://arxiv.org/abs/2510.07868