Temporally Accumulated Embeddings for Real-Time Neural Denoising

Abstract

Neural models for Monte Carlo denoising usually target either real-time or offline performance. Offline denoisers are aimed at higher sample-per-pixel counts, since the offline budget allows for longer rendering times. Multiple works transform these samples into embeddings and (temporally) accumulate them. Real-time denoisers work on 1-spp budgets, meaning that it is not necessary to average multiple samples per pixel. But can the temporal accumulation of the embeddings be beneficial for real-time models? This thesis shows that there is little change in performance and quality, and goes over a few key visual differences.