BSDF importance sampling using a diffusion model

University of California San Diego,

SIGGRAPH Asia 2024 (conference track)

Top Row: Equal-time comparison of our method against NBRDF and NeuSample for sampling the specular metal material COPPER-SHEET from RGL DATASET, under global illumination using BRDF sampling only. Bottom Row: (left) Rendering results for a rough dielectric material using our model and reference analytical sampling method at 2048 samples per pixel (spp). (right) PDF slices for a fixed incoming direction $\omega_i$, that is close to normal (top) and approaching the grazing angle (bottom).

Abstract

Previous neural sampling methods, primarily using analytical lobe mixtures and normalizing flows, often struggle with specular materials, particularly at grazing angles. Furthermore, they are limited to reflection, and do not handle transmission. Our key observation is that previous normalizing flows impose significant restriction in their network architecture for easy computation of the Jacobian. However, for low-dimensional BSDF sampling, the Jacobian computation is not the bottleneck. Therefore, we propose to use diffusion models to importance sample full BSDFs. Our method has two variants, one for most reflective materials that learns a distribution on a disk, and the other for extremely specular reflective materials and full BSDFs, which learns a distribution on a sphere. Our equal-time evaluations show that our method outperforms normalizing flows and significantly surpasses them in certain specular materials.

Video

* Since the demo is mostly noisy, please set the playback quality to 1080p or higher to avoid crazy artifacts caused by low-bitrate video compression. Alternatively, consider downloading the video directly (300MB) for optimal clarity.