Neural Radiance Fields with Hash-Low-Rank Decomposition
Abstract
HashRF combines low-rank tensor decomposition and multi-resolution hash encoding to create an efficient neural radiance field. It replaces high-rank matrix components with compact 2D hash tables and uses a lightweight adaptive fusion MLP to reconcile the two encodings. The representation improves fine-detail rendering while retaining compact model size and efficient training.
Type
Publication
Applied Sciences, 14(23), 11277
HashRF uses hash-low-rank decomposition to allocate representation capacity efficiently, preserving high-frequency scene details with a compact parameter budget.
