The Neural Light Grid: A Scalable Production-Ready Learned Irradiance Volume
Our system that precomputes and stores indirect diffuse lighting into a volume-based data structure of lighting probes for our large game scenes. Our approach stores a compact, expressive weighting function with each probe, which explicitly defines where in space the probe should contribute. We represent these functions in a novel form that was specifically tailored to their characteristics and allows for efficient evaluation even on the most basic hardware. These representations are non-linear, which traditionally requires a non-trivial optimization to produce. Traditional optimization is prohibitively expensive in a fast-paced production environment, where scenes can have multiple square kilometers and be covered by hundreds of thousands of probes. To solve this problem we introduce a generator network-- a meta-learning approach, where we pre-train a single neural network to directly produce our representation. This allows precomputation times faster by almost two orders of magnitude and makes the approach practical in a production setting.
The Software License and Service Agreement will be updated. Please follow this link [https://www.activision.com/legal/ap-eula] in order to see these changes.
