A weighted reservoir compresses a stream of candidates into one representative sample and a small amount of normalization state. That constant-memory property makes it possible to carry useful lighting information from one frame into the next.

This chapter develops reservoir updates and merges from weighted sampling, then introduces temporal reuse: persistent state, reprojection, validation, rejection rules, and the conditions required to reuse history without silently changing the estimator.

Many-lights scene rendered with temporal reservoir reuse

Read the full reservoirs and temporal reuse article →