Neural Luminaire Light-Field Compression

Speaker: David Köppl

Abstract

This master’s thesis investigates how the near-field emission of a luminaire can be turned into a compact representation for real-time rendering. The emitted light is modeled as a neural surface emission field, a small neural network on a simplified geometry of the luminaire that predicts the HDR radiance for each surface point and viewing direction. An end-to-end pipeline was developed that generates training data from goniophotometric views, trains the network, and evaluates it in real time in a GPU renderer. The results show that the model reproduces the appearance of complex luminaires from a small fraction of the original data, while light sources hidden behind a cover remain the main challenge.