Light Source Selection in Primary-Sample-Space Neural Photon Sampling

Abstract

This paper proposes a light source selection for photon mapping combined with recent deep-learning-based importance sampling. Although applying such neural importance sampling (NIS) to photon mapping is not difficult, a straightforward approach can sample inappropriate photons for each light source because NIS relies on the approximation of a smooth continuous probability density function on the primary sample space, whereas the light source selection follows a discrete probability distribution. To alleviate this problem, we introduce a normalizing flow conditioned by a feature vector representing the index for each light source. When the neural network for NIS is trained to sample visible photons, we achieved lower variance with the same sample budgets, compared to a previous photon sampling using Markov chain Monte Carlo.