In our experience, the quality of machine learning denoisers are greatly influenced by the training dataset, so in all comparisons in this paper we train all network variants at equal number of steps on the same dataset and loss (SMAPE : Equation 9), to focus on architectural differences and network capacity. [20EGSR Layer Embedding]

As a concrete example, the presence of (particularly indirect) shadows cannot be reliably detected from easily-available auxiliary variables, which leads many denoising algorithms to struggle with shadow fidelity. [19TOG_ConvNet for G-PT]

Determining which in-put signal to trust to correlate with coherence is a highly context-dependent task. We solve it with a convolutional neural network ~. [19TOG_ConvNet for G-PT]

We hypothesize that processing individual samples instead of summaries has a fundamentally better outlook in complex transport scenarios. [19TOG_SBMC]

A comprehensive review of Monte Carlo (MC) denoising and generative adversarial network (GAN) is beyond the scope of this paper, with both topics having been extensively studied in the field. Hence, we focus on the most relevant to our work. [19TOG_AdvCond]

We refer the reader to the review by Zwicker et al. [2015] and focus here on the most relevant and recent developments. [18TOG_AsymLoss]

SphereFace

Even in scenes where these paths do not cause dramatic lighting effects, their presence can lead to unusably slow convergence in renderers that attempt to account for all transport paths [Jakob and Marschner].

In light transport simulation, a path is typically represented as a list of vertices (locations of interactions), or by its start vertex and a sequence of directions. In this paper, we propose a different representation, where the relation of the incident and outgoing directions at interactions is specified by the direction of the (generalized) half vector [HSLT].