LITA-GS: Illumination-Agnostic Novel View Synthesis via Reference-Free 3D Gaussian Splatting and Physical Priors

TL;DR

The paper tackles the problem of novel view synthesis under adverse illumination, where traditional NeRF-based methods and even 3D Gaussian Splatting (3DGS) struggle due to limited SfM cues and noisy, poorly exposed imagery. It introduces LITA-GS, an unsupervised, reference-free framework that combines an illumination-invariant physical prior, a lighting-agnostic structure rendering pipeline, and a progressive denoising module to robustly optimize 3D Gaussians. Key contributions include a Kubelka-Munk–based illumination-invariant prior, depth- and structure-guided rendering with per-Gaussian attributes, and a multi-stage denoising strategy, all trained without ground-truth references. Experiments on the LOM dataset show that LITA-GS outperforms state-of-the-art NeRF-based approaches with faster convergence and rendering, while offering improved multi-view consistency over exposure-correction pipelines. Overall, the work provides a practical, fast, and unsupervised solution for high-quality 3D scene reconstruction under challenging lighting conditions.

Abstract

Directly employing 3D Gaussian Splatting (3DGS) on images with adverse illumination conditions exhibits considerable difficulty in achieving high-quality, normally-exposed representations due to: (1) The limited Structure from Motion (SfM) points estimated in adverse illumination scenarios fail to capture sufficient scene details; (2) Without ground-truth references, the intensive information loss, significant noise, and color distortion pose substantial challenges for 3DGS to produce high-quality results; (3) Combining existing exposure correction methods with 3DGS does not achieve satisfactory performance due to their individual enhancement processes, which lead to the illumination inconsistency between enhanced images from different viewpoints. To address these issues, we propose LITA-GS, a novel illumination-agnostic novel view synthesis method via reference-free 3DGS and physical priors. Firstly, we introduce an illumination-invariant physical prior extraction pipeline. Secondly, based on the extracted robust spatial structure prior, we develop the lighting-agnostic structure rendering strategy, which facilitates the optimization of the scene structure and object appearance. Moreover, a progressive denoising module is introduced to effectively mitigate the noise within the light-invariant representation. We adopt the unsupervised strategy for the training of LITA-GS and extensive experiments demonstrate that LITA-GS surpasses the state-of-the-art (SOTA) NeRF-based method while enjoying faster inference speed and costing reduced training time. The code is released at https://github.com/LowLevelAI/LITA-GS.

Figures (4)

• Figure 1: First row: (a) Input image; (b) Our rendered illumination-invariant structure prior; (c) Our rendered depth map; (d) SfM point cloud estimated from low-light scenes; (e) Our optimized point cloud. The second row provides visual comparisons between our method and other SOTA approaches. With the rendering of illumination-invariant structure prior and depth map, our method effectively represents the structure and spatial geometry of the scene, thereby achieving superior performance compared to current SOTA approaches.
• Figure 2: The framework of our LITA-GS. We introduce a physical prior extraction pipeline to capture structural information $\mathbf{P}$ from images with low or high exposure. Then, the extracted $\mathbf{P}$ is integrated into our developed lighting-agnostic structure rendering process. Furthermore, we employ a progressive denoising module (PDM) for noise reduction and optimize our LITA-GS without GT references.
• Figure 3: Our extracted $\mathbf{P}$ is invariant to varying illumination.
• Figure 4: Novel view synthesis comparison in low-light and over-exposure conditions. Compared to other methods, our LITA-GS achieves more vivid enhancement results and preserves more details in novel view synthesis.