Neural Render Proxies for Interactive and Differentiable Lighting (studios.disneyresearch.com)

A team of researchers from ETH Zurich and Disney Research has unveiled a groundbreaking Neural Render Proxy (NRP) that significantly accelerates the relighting process in computer graphics. Traditionally, lighting adjustments in CGI can take considerable time due to the complexity of rendering large scenes with global illumination and intricate shaders. The NRP addresses this bottleneck by decoupling rendering into path sampling and emission computation, allowing for the collection of light transport data from a single, light-agnostic render pass. This innovation enables artists to achieve interactive relighting of static scenes at frame rates of 30–60 Hz, while maintaining high visual fidelity comparable to traditional methods. The significance of this development for the AI and machine learning community lies in its potential to streamline workflows in CG animation and other related fields. The NRP is built on a lightweight neural network that learns specific light transport characteristics for each scene, yielding rapid adjustments without heavy computational overhead. Moreover, it facilitates a differentiable workflow that allows for efficient optimization of lighting parameters based on image-space edits. This capability not only enhances the artistic flexibility but also opens doors for advanced generative workflows, thus bridging the gap between machine learning techniques and practical applications in the creative industries.