Fast Implicit Neural Light Field Representation via Geometric Decomposition and Multi-Resolution Low-Rank Features

Yao Guo; Ligen Shi; Shuchen Sun; Jun Qiu; Chang Liu

cs.CVAug 14, 2026

Fast Implicit Neural Light Field Representation via Geometric Decomposition and Multi-Resolution Low-Rank Features

Authors: Yao Guo, Ligen Shi, Shuchen Sun, Jun Qiu, Chang Liu

Organizations: Institute of Computational Imaging, Beijing Information Science and Technology University, Beijing 102206, China · College of Computer Science (College of Software), Inner Mongolia University, Hohhot 010021, China

Abstract

Implicit neural representations provide a compact and continuous way to reconstruct dense light fields from sampled ray coordinates. However, fast light field reconstruction remains challenging because a light field is a high-dimensional signal with strong spatial-angular redundancy and structured disparity variations. Directly fitting 4D ray coordinates with a neural network often requires considerable optimization time to recover both view appearance and cross-view consistency. To address this issue, this paper proposes a fast implicit light field representation based on geometric decomposition and multi-resolution low-rank features. The proposed method decomposes a 4D light field into a horizontal disparity plane, a spatial texture plane, and a vertical disparity plane. Each plane is represented by a low-rank structure that combines a low-resolution 2D grid with the element-wise product of two high-resolution 1D line features at multiple resolution levels. The fused features are decoded by a lightweight multilayer perceptron to predict RGB values. Experiments on public light field datasets show that the proposed method achieves competitive reconstruction quality while providing a better trade-off among model parameters, training time, and inference efficiency.

Figures & tables

Fig. 1: Overview of the proposed framework. A queried ray $\mathbf{r}=(u,v,x,y)$ is mapped onto three feature planes. Each plane uses a multi-resolution low-rank representation that combines a low-resolution 2D grid $\mathbf{G}_{ab}^{\ell}$ with the element-wise product of two high-resolution 1D line features $\mathbf{L}_{a}^{\ell}$ and $\mathbf{L}_{b}^{\ell}$ . The concatenated features are decoded by a lightweight MLP to predict the ray color.

Dataset	Method	PSNR (dB) $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
EPFL	SIGNET	33.96 $\pm$ 2.67	0.93 $\pm$ 0.04	0.09 $\pm$ 0.05
	RSEN	30.10 $\pm$ 1.64	0.87 $\pm$ 0.05	0.19 $\pm$ 0.06
	RDF	31.85 $\pm$ 1.73	0.94 $\pm$ 0.02	0.07 $\pm$ 0.05
	Ours	35.45 $\pm$ 2.55	0.93 $\pm$ 0.03	0.02 $\pm$ 0.02
INRIA	SIGNET	35.43 $\pm$ 3.07	0.91 $\pm$ 0.06	0.12 $\pm$ 0.05
	RSEN	30.99 $\pm$ 1.81	0.85 $\pm$ 0.09	0.24 $\pm$ 0.07

TABLE I: Quantitative comparison on public LF datasets. Best results are marked in bold. The average is computed over datasets.

Fig. 2: Qualitative visual comparison of different methods on representative images selected from the EPFL, INRIA Lytro, and Stanford Gantry datasets, including the challenging Tarot scene. The results are shown for the center views, including reconstructed images and corresponding error maps.

Method	Params. (M)	Train. (min)	Infer. (s)
SIGNET	2.37	360	20
RSEN	1.40	180	9
RDF	0.66	72	12
Ours	1.55	21	8

TABLE II: Efficiency comparison of different methods on the INRIA Lytro dataset.

Scene	Target PSNR (dB)	Train.(min)
Scene	Target PSNR (dB)	SIGNET	RSEN	RDF	Ours
Bee	25.33	11	29	75	0.8
Building	29.93	37	168	70	2.0
Hublai	29.17	45	155	73	1.5
MessyDesk	33.44	72	173	62	2.0
Sculpture	32.19	60	180	68	2.0

TABLE III: Training time required to reach the same target PSNR on the INRIA Lytro dataset.

Setting	PSNR (dB) $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
w/o multi-res.	34.67 $\pm$ 3.21	0.90 $\pm$ 0.06	0.05 $\pm$ 0.03
w/o line prod.	35.46 $\pm$ 3.21	0.91 $\pm$ 0.05	0.04 $\pm$ 0.02
w/o 2D grid.	33.60 $\pm$ 3.26	0.87 $\pm$ 0.07	0.09 $\pm$ 0.05
Full model	37.61 $\pm$ 3.61	0.92 $\pm$ 0.05	0.04 $\pm$ 0.03

TABLE IV: Ablation study results on the INRIA Lytro dataset.

Levels	PSNR (dB) $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Train. (min)
4 levels	36.10 $\pm$ 5.61	0.96 $\pm$ 0.02	0.02 $\pm$ 0.01	13
6 levels	36.91 $\pm$ 5.36	0.96 $\pm$ 0.01	0.02 $\pm$ 0.01	17
8 levels	37.30 $\pm$ 5.22	0.97 $\pm$ 0.01	0.01 $\pm$ 0.01	21

TABLE V: Impact of multi-resolution levels on the Stanford Gantry dataset.

Explore similar work

Light Field Primitive for Novel View Synthesis
Sep 25, 2026Liang Chen, Jiahui Ning, Xun Jiang +4Novel View Synthesis Neural Rendering
Tongji University · Hong Kong Metropolitan University · City University of Hong Kong
Floating Radiance Networks
Aug 6, 2026Krzysztof Byrski, Rafał Tobiasz, Grzegorz Wilczyński +5Neural Rendering Novel View Synthesis
Jagiellonian University, Faculty of Mathematics and Computer Science · IDEAS Research Institute · Poznań University of Technology, Institute of Robotics and Machine Intelligence
Frequency-Structured Field Learning for Light-Field Disparity Estimation
Jul 16, 2026Sara Monji-Azad, Yulin Liu, Jürgen HesserDepth Estimation Frequency-Domain Feature Learning
Mannheim Institute for Intelligent Systems in Medicine (MIISM) · Medical Faculty Mannheim, Heidelberg University, Mannheim, Germany · Interdisciplinary Center for Scientific Computing (IWR), Heidelberg University +2

Sep 25, 2026J/K move · Enter open · S save

Light Field Primitive for Novel View Synthesis

Liang Chen, Jiahui Ning, Xun Jiang, Xing Xu+3

Tongji University · Hong Kong Metropolitan University · City University of Hong Kong

Novel View Synthesis Neural Rendering Differentiable Rendering

Open paper AskPDF arXiv

We present Light Field Primitives (LFP), a formulation for novel view synthesis that replaces the dense ray database with a compact set of differentiable primitives in the classical two-plane parameterization. Each primitive condenses a group of rays into one learned record, and its response to a query is governed by how closely that query belongs to the group. Rendering a camera ray then reduces to compositing all responses it elicits, and a scene can be optimized directly from posed images and rendered in real time with rays. Beyond its competitive performance on standard benchmarks, the main advantage of LFP is structural: its primitives reside directly in the 4D ray space, so optical and appearance effects that are already operations on the light field become behaviors of a single shared renderer. With minimal changes to that renderer, LFP supports multi-scale anti-aliasing, defocus deblurring with refocusing, rendering for fisheye cameras, and even transparent object reconstruction with ray refraction, matching specialized frameworks that devote substantial machinery to these effects.

Abstract

Explore similar work

Light Field Primitive for Novel View Synthesis

Floating Radiance Networks

Frequency-Structured Field Learning for Light-Field Disparity Estimation