Sponza Atrium

The Sponza Atrium is illuminated by extremely narrow directional illumination. In this scene, radiance-driven path-guiding approaches shine, with the bidirectionally trained GMMs by Vorba et al. [2014] performing the best. The SDTree by Müller et al. [2017] struggles with the extremely sparse unidirectional samples in this scene, performing the worst among the radiance-based path-guiding approaches. Training with the same sparse unidirectional samples, our neural path guiding (NPG-Radiance) performs surprisingly well at equal sample counts; almost as well as the bidirectionally trained GMMs. Incorporating the product (NPG-Product) only offers minor per-sample improvements over the GMMs. Primary-sample-space path-sampling techniques offer little to no benefit over unidirectional path tracing with our NPS slightly outperforming PSSPS [Guo et al. 2018] with an equal number of samples.