On Cosine Prior Distributions for Neural Path Guiding
Loading...
Date
2026
Journal Title
Journal ISSN
Volume Title
Publisher
The Eurographics Association
Abstract
Monte Carlo rendering efficiency depends on the quality of importance sampling distributions. Neural path guiding addresses this by learning adaptive distributions using normalizing flows, which transform simple prior distributions into target distributions. We explore how prior distribution choice affects learning efficiency and show that aligning the prior with components of the rendering integral simplifies the learning task, enabling the use of smaller models. Our cosine-distributed prior, matching the cosine-weighted hemisphere term of the rendering equation, achieves faster convergence and lower noise than uniform priors, with particularly strong improvements in scenes with high geometric complexity.
Description
@inproceedings{10.2312:egs.20261012,
booktitle = {Eurographics 2026 - Short Papers},
editor = {Musialski, Przemyslaw and Lim, Isaak},
title = {{On Cosine Prior Distributions for Neural Path Guiding}},
author = {Gutsch, Jan-Luca and Dereviannykh, Mikhail and Hanika, Johannes},
year = {2026},
publisher = {The Eurographics Association},
ISSN = {2309-5059},
ISBN = {978-3-03868-299-8},
DOI = {10.2312/egs.20261012}
}