TensoRF: Tensorial Radiance Fields

We present TensoRF, a novel approach to model and reconstruct radiance fields. Unlike NeRF that purely uses MLPs, we the field of scene as 4D tensor, which represents 3D voxel grid with per-voxel multi-channel features. Our central idea is factorize tensor into multiple compact low-rank components. demonstrate applying traditional CANDECOMP/PARAFAC (CP) decomposition – factorizes tensors rank-one components vectors in our framework leads improvements over vanilla NeRF. To further boost performance, introduce vector-matrix (VM) relaxes constraints for two modes vector matrix factors. Beyond superior rendering quality, models CP VM decompositions lead significantly lower memory footprint comparison previous concurrent works directly optimize Experimentally, TensoRF achieves fast reconstruction ( $$<30$$ min) better quality even smaller size $$<4$$ MB) compared Moreover, boosts outperforms state-of-the-art methods, while reducing time $$<10$$ retaining $$<75$$ MB).