Few TensoRF: Enhance the Few-shot on Tensorial Radiance Fields

Thanh-Hai Le; Hoang-Hau Tran; Trong-Nghia Vu

cs.CVMar 26, 2026

Few TensoRF: Enhance the Few-shot on Tensorial Radiance Fields

Authors: Thanh-Hai Le, Hoang-Hau Tran, Trong-Nghia Vu

Organizations: School of Computer Science and Engineering, The Saigon International University, Ho Chi Minh City 700000, Vietnam · Department of Informatic Technology, FPT University, Ho Chi Minh City 700000, Vietnam

Abstract

This paper presents Few TensoRF, a 3D reconstruction framework that combines TensorRF's efficient tensor based representation with FreeNeRF's frequency driven few shot regularization. Using TensorRF to significantly accelerate rendering speed and introducing frequency and occlusion masks, the method improves stability and reconstruction quality under sparse input views. Experiments on the Synthesis NeRF benchmark show that Few TensoRF method improves the average PSNR from 21.45 dB (TensorRF) to 23.70 dB, with the fine tuned version reaching 24.52 dB, while maintaining TensorRF's fast $\approx10-15$ minute training time. Experiments on the THuman 2.0 dataset further demonstrate competitive performance in human body reconstruction, achieving 27.37 - 34.00 dB with only eight input images. These results highlight Few TensoRF as an efficient and data effective solution for real-time 3D reconstruction across diverse scenes.

Figures & tables

Figure 1: Few TensoRF in the 3D NeRF-like Pipeline. The diagram shows where Few TensoRF fits into the core of the 3D NeRF-like reconstruction pipeline. It all begins with multiple scene images, and either COLMAP or Pose Diffusion calculates precise camera details. These ones, along with the images, train Few TensoRF. The trained Few TensoRF model then takes center stage in the final steps, ensuring a streamlined and accurate 3D model reconstruction within the NeRF-like framework. This visual guide underscores the seamless integration of Few TensoRF at the heart of the 3D NeRF-like process.

Figure 2: The Few TensoRF overview. This illustration was adapted from the original TensoRF paper Chen et al. (2022) . It highlights two significant enhancements as we move from left to right: incorporating density frequency masks and integrating appearance color frequency on each rendered voxel. Additionally, a supplementary frequency mask is applied to refine the positional encoding fed into the neural networks.

Figure 3: Comparing the rendered novel views quality of TensoRF baseline across different setups. To assess TensoRF’s performance on sparse input scenarios, we retrain the default TensoRF setup on different amounts of training image inputs. Following the image order, we have the original testing image (Ground Truth), and the novel view images rendered with TensoRF after training with 10 and 3 images (TensoRF_10img and TensorRF_10img) respectively.

Figure 4: Diagram illustrating Frequency Masking on the appearance grid $G_{c}$ and view direction $d$ .

PSNR	Lego	Chair	Drums	Ficus	Mic	Ship	Materials	Hotdog	Avg.
FreeNeRF_50k iters (repro.)	24.81	27.01	20.46	20.59	25.54	23.79	21.69	29.40	24.16
Tensorf (repro.)	23.83	23.45	18.32	19.34	23.3	19.96	20.78	21.85	21.45
Our Few TensoRF	25.68	27.70	18.88	21.09	24.56	21.64	22.04	27.99	23.70
Our Few TensoRF (Fine tune)	25.73	27.71	18.04	21.36	27.44	23.77	22.36	29.71	24.52

Table 1: PSNR metrics comparison on each Synthesis NeRF’s scene. The best-performing and second-best results from our experiments are denoted by red and blue highlights, respectively.

Metrics	PSNR $\uparrow$	Training time $\downarrow$
FreeNeRF_50k (repro.)	29.40	04:55:00
FreeNeRF_15k (repro.)	28.02	01:25:40
TensoRF (repro.)	21.85	00:15:00
Our Few TensoRF	27.99	00:15:20
Our Few TensoRF (Fine tune)	29.71	00:10:22

Table 2: Training time comparison on the Hotdog scene. The bold markings indicate the most optimal results obtained in our experiments.

Figure 5: Comparing the view rendered of Hotdog object between methods.

Figure 6: Visual comparison between methods.

Figure 7: Comparing the quality of rendered images with different numbers of images between TensoRF and Few TensoRF.

PSNR	0525	0300
TensoRF (50 images)	40.98	45.58
TensoRF(8 images)	28,37	34.28
Few TensoRF (8 images)	27.37	34.00

Table 3: Comparison of PSNR parameters between TensoRF and Few TensoRF with different numbers of input images.

Figure 8: Comparing 3D meshes rendered for 0300 object.

Neural Radiance Fields

Neural radiance fields represent 3D scenes as continuous functions of position and viewing direction to synthesize images from new camera views.

See all 82 papers

Temporal Residual Neural Radiance Fields for Monocular Video Dynamic Human Body Reconstruction
Sep 4, 2026Tianle Du, Jie Wang, Xiaolong Xie +33D Human Reconstruction Dynamic Scene Reconstruction
Nanchang University
Low-Cost Neural Radiance Fields
May 10, 2026Alice Huang, Prathamesh Sonawane, Yashdeep Thorat +1Neural Rendering Sparse-View 3D Reconstruction
University of Illinois Urbana Champaign
OTT3R: Multi-View 3D Reconstruction and Fast Dataset Generation at 1% Compute
Sep 28, 2026Brandon Leblanc, Charalambos PoullisMulti-View 3D Reconstruction Pseudo-Labeling
Immersive and Creative Technologies Lab, Concordia University, Montreal, Canada

Sep 4, 2026J/K move · Enter open · S save

Temporal Residual Neural Radiance Fields for Monocular Video Dynamic Human Body Reconstruction

Tianle Du, Jie Wang, Xiaolong Xie, Wei Li+2

Nanchang University

3D Human Reconstruction Dynamic Scene Reconstruction Neural Radiance Fields Monocular 3D Reconstruction

Open paper AskPDF arXiv

In the field of computer vision and graphics, high-quality reconstruction of the human body in static scenes has been achieved in recent years by a single multilayer perceptron (MLP) in a number of approaches. However, MLPs have capacity limitations, requiring substantial training time and computational resources for dynamic scene reconstruction. And the quality of reconstruction is significantly constrained. This paper proposes a method for effectively processing complex spatiotemporal signals in dynamic scene human 3D modeling. The proposed method uses Temporal Residual Neural Radiance Fields to achieve novel view rendering and new pose synthesis of human bodies.To address the problem of representing temporal signals in video sequences, we construct a temporal residual field which is not related to the MLP architecture. Secondly, to improve reconstruction efficiency, we propose an integrated approach that reduces trainable parameters and accelerates rendering, thereby enhancing the network's feature representation capability. Finally, we design a multi-dimensional loss function to accurately measure the loss between predicted and actual spatial pixel values. The experimental results show that our proposed approach improves the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) accuracy metrics compared to the latest representative methods. It maintains similar accuracy to Anim-NeRF and Neural Body while achieving a nearly 780-fold increase in time efficiency.

Abstract

Neural Radiance Fields

Explore similar work

Temporal Residual Neural Radiance Fields for Monocular Video Dynamic Human Body Reconstruction

Low-Cost Neural Radiance Fields

OTT3R: Multi-View 3D Reconstruction and Fast Dataset Generation at 1% Compute