TensoRF is a novel approach for modeling and reconstructing radiance fields, which represents scenes as 4D tensors rather than using MLPs as in NeRF. The method factors the 4D tensor into compact low-rank components, enabling efficient scene reconstruction with significantly reduced memory footprint. Two decomposition techniques, CP and VM, are introduced. CP decomposition factorizes the tensor into rank-one components, while VM decomposition further reduces the number of components by using vector-matrix factors, leading to faster reconstruction and better rendering quality. TensoRF with CP decomposition achieves fast reconstruction (<30 min) and a smaller model size (<4 MB) compared to NeRF, while VM decomposition further improves rendering quality and reduces reconstruction time (<10 min) with a compact model size (<75 MB). The method supports various types of per-voxel features, including neural and spherical harmonics features, and enables efficient trilinear interpolation for continuous field modeling. TensoRF achieves state-of-the-art rendering quality and is highly efficient in terms of computation and memory, making it suitable for real-time applications. The approach is evaluated on various datasets and outperforms previous methods in terms of rendering quality, reconstruction speed, and model size.TensoRF is a novel approach for modeling and reconstructing radiance fields, which represents scenes as 4D tensors rather than using MLPs as in NeRF. The method factors the 4D tensor into compact low-rank components, enabling efficient scene reconstruction with significantly reduced memory footprint. Two decomposition techniques, CP and VM, are introduced. CP decomposition factorizes the tensor into rank-one components, while VM decomposition further reduces the number of components by using vector-matrix factors, leading to faster reconstruction and better rendering quality. TensoRF with CP decomposition achieves fast reconstruction (<30 min) and a smaller model size (<4 MB) compared to NeRF, while VM decomposition further improves rendering quality and reduces reconstruction time (<10 min) with a compact model size (<75 MB). The method supports various types of per-voxel features, including neural and spherical harmonics features, and enables efficient trilinear interpolation for continuous field modeling. TensoRF achieves state-of-the-art rendering quality and is highly efficient in terms of computation and memory, making it suitable for real-time applications. The approach is evaluated on various datasets and outperforms previous methods in terms of rendering quality, reconstruction speed, and model size.