We compare scenes of different structure and complexity prefiltered using our approach and display their FLIP error [Andersson et al. 2020] when compared to a path traced reference. For the Pandanus tree, we both include the results of using our optimal thresholds and the stochastic threshold variant.
Each LoD corresponds to a different image resolution, with the lowest LoD rendered at the resolution of 512x512 pixels and the lowest resolution rendered at only 8x8 pixels.
For more details on training times, rendering performance, memory footprint and variance reduction, please consult our paper.