Neural Rendering for Real-Time 3D Lighting Editing

This DisneyResearch|Studios paper introduces a neural render proxy for interactively and differentially relighting static scenes with fixed cameras and materials. It separates path sampling from emiss

In animation lighting workflows, even small changes can require global illumination and complex shaders to be recalculated, meaning that rerendering a single frame can take anywhere from several minutes to several hours. Neural Render Proxies for Interactive and Differentiable Lighting proposes neural render proxies (NRPs) for static scenes with fixed cameras and materials to reduce this iteration time.

The method begins by dividing traditional rendering into path sampling and emission computation. After collecting samples of light-transport paths in a single render pass that does not depend on lighting conditions, a lightweight scene-specific neural network learns how light is transported from positions in the scene to each image pixel. Lighting conditions can then be changed and the results evaluated rapidly.

The researchers explain that the approach is compatible with non-differentiable production renderers while keeping memory requirements low during inference. Computational cost increases with resolution, the number of light sources, and the number of lighting parameters, rather than with the complexity of the scene or appearance. In evaluations, the method demonstrated relighting at approximately 30–60 Hz while maintaining image quality close to ground-truth path tracing. It can also be used for gradient-based inverse tasks that find lighting parameters for image-space edits or target images generated by generative models.

The paper was presented at the 2026 Eurographics Symposium on Rendering (EGSR), with Sergio Sancho, Alexander Rath, Marco Manzi, Pascal Chang, Amit H. Bermano, Derek Nowrouzezahrai, Markus Gross, and Marios Papas as authors.

Maru