Integrating variance-aware multi-lobe BSDF decision sampling into gradient-based 3D Gaussian photon guiding

Develop a practical implementation of a variance-aware distribution for BSDF lobe decision sampling, such as the approach of Variance-Aware Path Guiding, and integrate it into the gradient-based learning framework used to guide photon emission with per-light 3D Gaussian mixtures, so that lobe decisions are sampled according to their actual image contribution rather than BSDF weights.

Background

In caustics rendering, many photon paths traverse materials with multi-lobe BSDFs (e.g., glass exhibiting both reflection and refraction). The current implementation samples lobes by their BSDF weights, which may be suboptimal for variance reduction because the actual image contribution of each lobe can differ significantly from its BSDF weight.

The authors note that adopting a variance-aware distribution for lobe decision sampling—such as the methodology proposed by Variance-Aware Path Guiding—could reduce variance by allocating samples according to estimated contribution rather than local BSDF weights.

However, while the paper presents a gradient-based framework that learns per-light 3D Gaussian mixtures to guide photon emission, incorporating a variance-aware lobe selection mechanism into this learning framework has not been realized and remains an open technical task.