Reliability-Weighted Averaging for Real-Time Neural Incident Radiance Caching

Neural radiance caches remain challenging to use in highly dynamic scenes, such as when geometry or light
sources move. The cache must adapt quickly to reduce bias after changes while preserving temporal stability
in static regions to avoid flicker. We introduce a reliability metric based on exponentially weighted correlation
between the cache prediction and the Monte Carlo target. The metric identifies regions where adaptation is needed
and regions where a low-variance prediction is preferable. It does not require tracing additional rays. Using this
signal, we construct a reliability-weighted scheme for real-time neural radiance caching. For each path-traced
cache query, the method averages two global illumination predictors. One predictor adapts rapidly, while the
other is temporally smoothed using exponentially averaged weights. Because the reliability weights are resolved
spatially, the method responds quickly in changing regions while remaining stable elsewhere. At equal samples
per pixel, our method reduces flicker in dynamic scenes, as measured by SMAPE and FLIP. Compared to the
exponentially averaged neural incident radiance cache baseline, it also reduces temporal lag. ... mehrThis is achieved
through data-driven per-query blending between the current predictor and the temporally smoothed predictor.