NRRS: Neural Russian Roulette and Splitting.

Area of Science:

• Computer Graphics
• Rendering Algorithms
• Parallel Computing

Background:

• Wavefront path tracing utilizes batched, stage-wise execution for GPU efficiency.
• Traditional Russian Roulette and Splitting (RRS) methods are incompatible with wavefront's memory and scheduling due to unpredictable path counts.

Purpose of the Study:

• To develop a novel framework for RRS compatible with wavefront path tracing.
• To enable stable, memory-efficient, and high-performance rendering on parallel architectures.

Main Methods:

• Introduced a normalized RRS formulation with a bounded path count.
• Pioneered neural networks (NRRS, AID-NRRS) to learn RRS factors using RRSNet.
• Implemented Mix-Depth, a path-depth-aware mechanism for adaptive neural evaluation.

Main Results:

• The proposed normalized RRS enables stable and memory-efficient execution on wavefront architectures.
• Neural network models NRRS and AID-NRRS, with Mix-Depth, effectively learn RRS factors.
• Achieved superior rendering quality and performance compared to traditional heuristics and existing RRS techniques across complex scenes.

Conclusions:

• The novel RRS framework is well-suited for wavefront path tracing, addressing limitations of traditional methods.
• Neural network integration offers adaptive and efficient control over rendering complexity.
• The approach significantly advances parallel rendering efficiency and visual fidelity.