NRRS: Neural Russian Roulette and Splitting

TL;DR

This work tackles high-variance, costly global illumination rendering by introducing a normalization-based RRS framework tailored to wavefront path tracing, ensuring a fixed, bounded sampling budget per iteration. ItThen proposes two neural RRS models, NRRS and AID-NRRS, which learn normalized RRS factors under the wavefront constraints, complemented by Mix-Depth to combine multiple RRS strategies with minimal overhead. Empirical results show that NRRS and AID-NRRS consistently outperform traditional heuristics (e.g., EARS, ADRRS) in rendering quality and ray efficiency across diverse scenes and depths, while maintaining stable GPU scheduling and memory usage. The work demonstrates that neural modeling of RRS factors within a normalization framework yields tangible gains in both accuracy and performance, and outlines practical avenues for integration with other neural rendering components. Overall, the approach offers a principled, scalable path to accelerating realistic rendering on GPU architectures with robust budget control.

Abstract

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.

Figures (12)

• Figure 1: Prefix and Suffix Paths. The entire path $\mathbf{z}$ is divided into two parts at point $\mathbf{p}$. The prefix path $\mathbf{x}$ extends from the the camera to point $\mathbf{p}$. And the suffix path $\mathbf{y}$ continues from $\mathbf{p}$ to the light source or infinity.
• Figure 2: Wavefront Architecture with RRS module. The essence of the wavefront path tracing architecture lies in decomposing the control flow of path tracing into a staged dataflow. Leveraging a work-queue-based mechanism, it is expected to achieve highly parallel, low-divergence, and load-balanced execution. Our RRS module regulates the generation and termination of paths, maintaining stable work queue sizes and providing a predictable workload scheduling mechanism for the GPU.
• Figure 3: (A) NRRS consists of two neural networks: StatNet and RRSNet. StatNet predicts the local radiance statistics $L_\text{o}(\mathbf{x}, \omega_o)$, while RRSNet predicts the RRS factors $q_{\text{nrrs}}$. StatNet takes as input the position $\mathbf{p}$, the outgoing direction $\omega_o$, and the surface roughness $\alpha_{\mathbf{p}}$, and outputs both the mean and variance of the local radiance. These predicted statistics serve both as direct inputs to RRSNet (d) and as supervisory signals for training its output. (a) The position $\mathbf{p}$ is encoded using a learnable multi-resolution hash grid TCNN. (b) The direction $\omega_o$ and roughness $\alpha_{\mathbf{p}}$ are encoded using one-blob encoding NIS19. The predicted radiance statistics (after Box–Cox transformation) (e), together with throughput $t_{\mathbf{x}}$, pixel radiance $I_{\text{pixel}}$, and remapped roughness (f), are fed into RRSNet to predict the resampling factor $q_{\text{nrrs}}$ for point $\mathbf{p}$. (B) AID-NRRS is similar to NRRS, except that RRSNet omits the StatNet outputs and directly takes the position $\mathbf{p}$ and outgoing direction $\omega_\text{o}$ as inputs. Note that StatNet and RRSNet each have their own Inputs module with separate trainable parameters.
• Figure 4: Loss Function of RRSNet. The loss function of RRSNet includes three components: Minimize Error Loss, Average Error Loss, and Bound Loss. Solid arrows indicate the forward process—from RRS factor prediction and normalization to scattered path generation and pixel error computation. Dashed arrows represent the backward process, where gradients propagate through Minimize Error Loss and Average Error Loss, while Bound Loss directly constrains the original RRS factors.
• Figure 5: Methods Evaluation and Comparisons on diverse scenes. Each scene is trained for 60 seconds and then inferred for 60 seconds across all methods. All images are rendered at a resolution of $1280\times720$. Procedural results generated during training are not included due to excessive noise. The first row shows the RelMSE error ($\times 10^{-3}$) over the entire image, where lower values indicate better performance. The second row presents RayEffInv, as defined in Sec. \ref{['sec:results_comparisons']}. Ray count is calculated as the total number of rays traced during the 60-seconds inference period, divided by the number of pixels. We color code the RGB]255, 153, 153first, RGB]255, 204, 153second, and RGB]255, 248, 173third lowest numbers. Our approach consistently shows the best results using fewer rays, indicating that each traced ray was utilized as effectively as possible.