Yet Another Box

Yet Another Box is an intentionally difficult scene for path guiding techniques. Like in the Indirectly Lit Cornell Box, the light source is facing upward. However, in this scene the light source is additionally tiny and the ceiling is made out of a highly glossy material. This makes the scene almost impossible to render with unidirectional path tracing. Path guiding techniques have to learn a radiance field which is high-frequency both in the spatial and directional domain. In this tough scene, PPG [Müller et al. 2017] is outperformed by the bidirectionally trained GMMs of Vorba et al. [2014] and our unidirectionally-trained radiance-driven neural path guiding (NPG-Radiance). The latter performs the best among all guiding techniques. Interestingly, our product-driven approach (NPG-Product) performs worse than the radiance-driven approach, suggesting that the larger number of dimensions makes it more difficult to optimize our neural networks. Improving upon this limitation is an interesting avenue for future work. Primary-sample-space path-sampling techniques are able to learn some components of the light transport, but fail to provide a consistent improvement over unidirectional path tracing. Our NPS slightly outperforms PSSPG [Guo et al. 2018].