Neural Precomputed Radiance Transfer

Recent advances in neural rendering indicate immense promise for architectures that learn light transport, allowing efficient of global illumination effects once such methods are trained. The training phase these can be seen as a form pre-computation, which has long standing history Computer Graphics. In particular, Pre-computed Radiance Transfer (PRT) achieves real-time by freezing some variables the scene (geometry, materials) and encoding distribution others, interactive at runtime. We adopt same configuration PRT – static scenes under dynamic environment lighting investigate different network architectures, inspired design principles theoretical analysis PRT. introduce four show those based on knowledge transport models PRT-inspired improve quality predictions equal time size, without need high-end ray-tracing hardware.