Speaker: Dejan Belic
Abstract
Real-time global illumination (GI) remains computationally challenging. Path tracing of full paths per pixel is too slow, so alternative approaches like Instant Radiosity are used. These methods generate Virtual Point Lights (VPLs) from emitters, replacing expensive camera paths with many lights. However, sampling millions of VPLs becomes the new bottleneck. Methods like Stochastic Lightcuts address this by clustering lights hierarchically, but in real-time settings, error bounds cannot be tightened enough, leading to persistent noise.
This work introduces a real-time GI method that combines ReSTIR’s (Reservoir-based Spatio-Temporal Importance Resampling) spatio-temporal importance resampling with dynamically generated VPLs. Instead of building light trees, we use ReSTIR’s reservoirs to efficiently select high-contribution VPLs by reusing promising samples across neighboring pixels and previous frames. We explore two extensions: limited cross-frame VPL reuse to reduce generation cost, and a hybrid approach where a light tree guides ReSTIR’s initial candidate selection. The goal is to achieve lower noise than current real-time many-light methods under the same computational budget, making accurate indirect lighting more practical for dynamic scenes.