← All resources

Neural Visibility Cache for Real-Time Light Sampling

Auto-discovered

Paper Advanced

Jakub Bokšanský, Daniel Meister

Visit resource ↗

Description

Neural Visibility Cache for Real-Time Light Sampling is a Journal of Computer Graphics Techniques Vol. 14, No. 2 paper by Jakub Bokšanský and Daniel Meister at Advanced Micro Devices on online neural visibility for many-light real-time direct illumination.

It trains a fully fused multilayer perceptron with multi-resolution hash-grid encoding to predict nonbinary visibility between 3D positions and lights or light clusters, then drives weighted reservoir sampling with optional LTC-based BRDF weights. The cache can seed ReSTIR candidates. Neural DI uses the same weights as a fast approximate direct lighting estimate without casting shadow rays beyond training. Clustering keeps a fixed network width for arbitrary light counts.

It targets advanced real-time rendering engineers working on many-light sampling, ReSTIR, or neural caches. Comfort with Monte Carlo light sampling and GPU ML inference is assumed.

The JCGT publication page provides open access to the paper and PDF downloads for this volume entry.

It is useful when visibility-aware light selection reduces noise in occluded regions where screen-space reuse alone struggles. Link last verified accessible on 2026-09-29. Link last verified accessible on 2026-09-28.