HdRestir 8 — Convergence Across Scenes
The final comparison asks a practical question: when does ReSTIR reach a useful image sooner than conventional path tracing? No single scene answers it, and visual inspection alone cannot distinguish lower variance from a fortunate random seed.
This chapter compares CPU and GPU pipelines across several scenes using independent high-sample HDR references, fixed render seeds, noise measurements, RMSE, and equal-time outputs. It closes the series by identifying where temporal reuse helps, where it does not, and which limitations still dominate.
