Deep Appearance Prefiltering

Physically based rendering of complex scenes can be prohibitively costly with a potentially unbounded and uneven distribution complexity across the rendered image. The goal an ideal level detail (LoD) method is to make costs independent 3D scene complexity, while preserving appearance scene. However, current prefiltering LoD methods are limited in appearances they support due their reliance approximate models other heuristics. We propose first comprehensive multi-scale framework for environments geometry materials (e.g., Disney BRDF), maintaining respect ray-traced reference. Using hierarchy scene, we perform data-driven step obtain phase function directional coverage mask at each scale. At heart our approach novel neural representation that encodes this information into compact latent form easy decode inside physically renderer. Once baked out, requires no original geometry, materials, or textures render time. demonstrate compares favorably state-of-the-art achieves considerable savings memory scenes.