Neural Global Illumination via HoloMap and AdaptMap

We present a neural rendering framework for real-time global illumination, composed of a Neural Baker and a Neural Renderer. The Neural Baker transforms G-Atlas into a UV-space feature atlas, HoloMap, encoding static geometry and material information. The Neural Renderer takes HoloMap along with view and lighting parameters to learn a dynamic neural field, AdaptMap. Training occurs in two phases: first, both modules are jointly optimized to construct a consistent HoloMap; then, the Neural Baker is detached, HoloMap is fixed, and AdaptMap is initialized from it and refined to capture view- and lighting-dependent effects. At inference, the Neural Renderer uses both HoloMap and AdaptMap to produce global illumination. Experiments show that our staged training design reduces model complexity and memory footprint, while achieving high-quality, real-time rendering with improved efficiency. Furthermore, our optimized training strategy, combining multi-threaded asynchronous rendering with an adaptive sampling policy, reduces the total training time by approximately 50% compared to baseline methods.