Title: PureLight: Learning Complex Luminaires with Light Tracing

Abstract: This paper introduces a neural approach for approximating the visual characteristics of intricate lighting fixtures. Our work specifically targets difficult cases involving complex light transport phenomena, such as small light sources surrounded by multiple specular layers, which pose significant challenges for traditional (bidirectional) path tracing methods. To address this, we employ light tracing to generate paths originating from emitters and terminating at exit surfaces, thereby framing the appearance estimation task as a distribution learning problem. We utilize a large normalizing flow network to model the probability density function (pdf) of the outgoing radiance at these exit surfaces. The outgoing radiance is subsequently recovered by multiplying the estimated pdf by the flux. For efficient inference, we distill the learned appearance into a compact Multi-Layer Perceptron (MLP) capable of directly estimating radiance on the exit surfaces. Furthermore, we train a dedicated sampling network to facilitate effective direct illumination calculations from the luminaire, alongside a blending network designed to integrate the luminaire seamlessly into the scene. This methodology enables the rendering of challenging luminaires with low sample counts within arbitrary environments.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC