← All resources

Neural Render Proxies for Interactive and Differentiable Lighting

Auto-discovered

Paper Advanced

Sergio Sancho, Alexander Rath, Marco Manzi, Pascal Chang, Amit H. Bermano, Derek Nowrouzezahrai, Markus Gross, Marios Papas

Visit resource ↗

Description

Neural Render Proxies for Interactive and Differentiable Lighting is an EGSR 2026 paper from Disney Research Studios and academic collaborators on accelerating lighting iteration in animation pipelines. It targets static scenes with fixed camera and materials where full path-traced relights would otherwise take minutes per frame.

The method introduces a scene-specific neural render proxy that decouples path sampling from emission evaluation. A light-agnostic pass collects path transport samples, then a lightweight network learns how light from any surface location reaches each pixel, enabling on-the-fly relighting without re-running a production path tracer.

The approach stays compatible with non-differentiable offline renderers, keeps inference memory modest, and scales with resolution and light count rather than full scene complexity. Evaluation reports interactive relighting at roughly 30 - 60 Hz with fidelity close to ground-truth path tracing, plus gradient-based inverse lighting from image-space edits.

Researchers and lighting TDs working on differentiable rendering or fast look-dev loops are the primary audience. The work sits in active EGSR and neural rendering literature on proxy acceleration for production lighting.

Link last verified accessible on 2026-09-28.