Glossy Kitchen

The Glossy Kitchen consists mostly of rough metal objects, lit indirectly through tiny spherical light sources. This scene is the toughest to render with unidirectional path tracing among all our scenes. Radiance-based path guiding dramatically improves upon unidirectional path tracing. PPG [Müller et al. 2017] and our NPG-Radiance perform on-par, whereas the GMMs by Vorba et al. [2014] can not complete the rendering process due to numerical instabilities caused by the extremely small light sources emitting large densities of radiance. Due to the glossy materials, incorporating the full triple product of the BSDF, incident illumination, and the cosine term yields further large improvements. In this scene, one-blob encoding the inputs to our neural networks is particularly important, as shown by the weak performance of NPG-Product with scalar input. Primary-sample-space path-sampling techniques offer little to no benefit over unidirectional path tracing with our NPS slightly outperforming PSSPG [Guo et al. 2018].