Progressive Volume Distillation with Active Learning for Efficient NeRF Architecture Conversion
摘要
Neural Radiance Fields (NeRF) have been widely adopted as practical and versatile representations for 3D scenes. However, architectures such as plain MLPs, Tensors, low-rank Tensors, and Hashtables each involve distinct trade-offs, making it difficult for a single model to meet diverse task requirements. For example, Hashtables offer fast rendering but lack clear geometric meaning, complicating spatial-relation-aware editing. To overcome these limitations and maximize the potential of each architecture, we propose Progressive Volume Distillation with Active Learning (PVD-AL), a systematic distillation method that enables any-to-any conversion between diverse architectures, which shifts the relationship between these architectures from competition to collaboration. PVD-AL decomposes each structure into two parts and progressively performs distillation from shallower to deeper volume representation, leveraging effective information retrieved from the rendering process. Additionally, a three-level active learning technique provides continuous feedback from teacher to student, achieving high-performance outcomes. Experimental evidence showcases the effectiveness of our method across benchmarks. For instance, PVD-AL distills an MLP-based model from a Hashtables-based model 10