Authors
Bing Xu, Liwen Wu, Milos Hasan, Fujun Luan, Iliyan Georgiev, Zexiang Xu, Ravi Ramamoorthi
University of California, San Diego; Adobe Research
Portals
Abstract
Neural material representations have recently been proposed to augment the material appearance toolbox used in realistic rendering. These models are successful at tasks ranging from measured BTF compression, through efficient rendering of synthetic displaced materials with occlusions, to BSDF layering. However, importance sampling has been an after-thought in most neural material approaches, and has been handled by inefficient cosine-hemisphere sampling or mixing it with an additional simple analytic lobe. In this paper we fill that gap, by evaluating and comparing various pdf-learning approaches for sampling spatially varying neural materials, and proposing new variations of these approaches. We investigate three sampling approaches: analytic-lobe mixtures, normalizing flows, and histogram prediction. Within each type, we introduce improvements beyond previous work, and we extensively evaluate and compare these approaches in terms of sampling rate, wall-clock time, and final visual quality. Our versions of normalizing flows and histogram mixtures perform well and can be used in practical rendering systems, potentially facilitating the broader adoption of neural material models in production.
Contribution
- We study three sampling approaches: analytic mixtures, normalizing flows, and histogram prediction (Section 3)
- Within each approach, we improve upon previous work and extensively evaluate and compare these approaches, in terms of sampling rate, wall-clock time, and final visual quality
Related Works
Neural materials; Analytic approximations; Normalizing flows; Histogram prediction