Reproduce “Neural Visibility Cache for Real-Time Light Sampling” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jakub Bokšanský, Daniel Meister
Researchers at AMD developed the Neural Visibility Cache (NVC), an online-trained neural network, to enhance real-time direct light sampling quality. NVC guides an unbiased Monte Carlo sampling process, reducing FLIP error by 20-45% compared to screen-space ReSTIR and handling dynamic scenes with many lights efficiently.
Key insights
- Neural networks can effectively learn the complex visibility function in real-time, providing an efficient guide for unbiased Monte Carlo light sampling rather than approximating radiance directly.
- Online training, using a hybrid of world-space and screen-space samples, enables the NVC to rapidly adapt to dynamic scenes, including moving cameras, animated geometry, and changing lights.
- The NVC can be integrated with existing state-of-the-art techniques like ReSTIR to provide superior initial candidates, boosting convergence rates during disocclusion events and further reducing noise.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.