Di$^2$CycleSB: Towards High-Quality Unsupervised Nighttime Visibility Enhancement via Schrödinger Bridge Transformer

Hanting Li; Xin Sun; Wei Ye; Jungong Han; Liang-jie Zhang

cs.CVAug 29, 2026

Di $^2$ CycleSB: Towards High-Quality Unsupervised Nighttime Visibility Enhancement via Schrödinger Bridge Transformer

Authors: Hanting Li, Xin Sun, Wei Ye, Jungong Han, Liang-jie Zhang

Organizations: Faculty of Data Science, City University of Macau, 999078, SAR Macao, China · College of Electronic and Information Engineering, Shanghai Research Institute for Autonomous Intelligent Systems, State Key Laboratory of Autonomous Intelligent Unmanned Systems, Tongji University, Shanghai, China · Department of Automation, Tsinghua University, Beijing, China · College of Computer Science and Software Engineering, Shenzhen University, Shenzhen, China

Abstract

Light-effect contamination poses a significant challenge to nighttime visibility enhancement. Most methods suppress light effects by estimating and decomposing them through prior-driven regularization, yet they are often limited by hand-crafted priors and ill-posed nature of decomposition. This work proposes Di $^2$ CycleSB, a unsupervised Cycle Schrödinger Bridge Transformer framework guided by dynamic integral image priors, for high-quality unsupervised nighttime visibility enhancement. Specifically, a novel light-effect estimator is introduced to parameterize Gaussian-like adaptive priors by aggregating dynamic integral image representations for non-uniform glow estimation. Then, we propose a prior-informed Generator that exploits light-effect representations to guide long-range dependency modeling within our specific Transformer blocks. We formulate light-effect suppression as a Schrödinger bridge problem and construct forward and backward bridges with cycle consistency constraints to achieve visually pleasing enhancement. Extensive experiments on real-world datasets demonstrate the remarkable effectiveness of our Di $^2$ CycleSB in enhancing nighttime visibility. In particular, it achieves effective end-to-end light-effect suppression without any regularization constraints and image decomposition. The code and models are available at https://github.com/LHTcode/Di2CycleSB.

Figures & tables

Fig. 1: Comparison of Di²CycleSB, UNIE [ 12 ] , NightEnhance [ 13 ] , and DehazeSB [ 14 ] on the Light-Effects [ 15 ] and NTIRE 2025 Challenge [ 16 ] datasets. Our method achieves the best performance across User Study , LMM-IQA , Q-Align [ 17 ] , CLIP-IQA+ [ 18 ] , MANIQA [ 19 ] , and LIQE [ 20 ] . For visualization, each metric is first normalized separately for each dataset and then averaged.

Panel of Figure 2 — Fig. 2: Our results lead to a superior nighttime visibility experience. Di²CycleSB naturally suppresses light effects while preserving visual fidelity. All unsupervised and semi-supervised methods methods are retrained for a fair comparison. Please zoom in for better viewing.

Fig. 3: Overview of the Di²CycleSB framework. (a) illustrates the unpaired training process of Di²CycleSB. During training, we jointly optimize the forward and backward Schrödinger bridges at randomly sampled time steps and establish a cycle path to further regularize the transport process. (b) Di²LE estimates the light-effect map from the image intensity component. (c) The PI Generator performs light-effect suppression under the guidance of the estimated light-effect map. (d) The PIT Block is the core component of the PI Generator.

Fig. 4: Visual qualitative comparisons of state-of-the-art methods on the Light-Effects dataset.

Fig. 5: Visual qualitative comparisons of state-of-the-art methods on the NTIRE 2025 Challenge 2K resolution dataset.

Methods	Venue	User Study $\uparrow$	LMM-IQA $\uparrow$	Q-Align $\uparrow$	CLIP-IQA+ $\uparrow$	MANIQA $\uparrow$	LIQE $\uparrow$
			(sampled/full dataset)
Original Dataset	–	–	–	3.368	0.603	0.368	2.745
EnlightenGAN [ 2 ]	TIP’21	-1.158	-1.635/-1.853	2.552	0.451	0.330	1.991
Dehazeformer [ 52 ]	TIP’23	-1.491	-1.011/-0.849	3.015	0.513	0.321	2.399
UNIE [ 12 ]	ECCV’22	0.356	-0.164/-1.991	3.070	0.480	0.274	1.839
NightEnhance [ 13 ]	MM’23	0.855	0.770 / 0.763	3.469	0.567	0.334	2.574

TABLE I: Quantitative comparison on the Light-Effects dataset via user studies and no-reference IQA metrics.

Methods	Venue	User Study $\uparrow$	LMM-IQA $\uparrow$	Q-Align $\uparrow$	CLIP-IQA+ $\uparrow$	MANIQA $\uparrow$	LIQE $\uparrow$
			(sampled/full dataset)
Original Dataset	–	–	–	2.241	0.293	0.175	1.186
EnlightenGAN [ 2 ]	TIP’21	-0.678	-1.640/-1.700	1.603	0.301	0.156	1.164
Dehazeformer [ 52 ]	TIP’23	-0.416	-0.695/-0.667	1.960	0.306	0.153	1.154
UNIE [ 12 ]	ECCV’22	0.845	1.014 / 0.878	2.609	0.315	0.188	1.706
NightEnhance [ 13 ]	MM’23	0.206	0.222/0.093	1.942	0.325	0.152	1.168

TABLE II: Quantitative comparison on the NTIRE 2025 Challenge unpaired dataset via user studies and no-reference IQA metrics.

Experiments	GPT-5.6 Sol $\uparrow$	Q-Align $\uparrow$	CLIP-IQA+ $\uparrow$	MANIQA $\uparrow$	LIQE $\uparrow$
w/o Cycle	0.556	3.516	0.581	0.338	2.854
w/o SB	-1.157	3.391	0.561	0.342	2.719
w/o Di²LE	0.467	3.463	0.629	0.341	2.752
w/o PI generator	-1.601	2.722	0.478	0.190	1.373
w/o LLIE	0.681	3.547	0.634	0.348	2.899
Di²CycleSB	1.053	3.567	0.644	0.360	2.929

TABLE III: Breakdown ablation study of the proposed framework.

Fig. 6: Qualitative visual results of the ablation study. Please zoom in for better viewing.

Experiments	GPT-5.6 Sol $\uparrow$
Max-RGB input	-0.295
Gaussian kernel ( $5\times 5$ )	-0.372
Gaussian kernel ( $11\times 11$ )	0.795
Gaussian kernel ( $21\times 21$ )	0.184
Gaussian kernel ( $31\times 31$ )	-0.011
Gaussian kernel ( $43\times 43$ )	-1.892

TABLE IV: Ablation study of the Di²LE module.

Fig. 7: Normalized heatmap visualizations of the $K=5$ dynamic integral regions (top row) and the corresponding dynamic weighting coefficients (bottom row). The top row shows that pixels closer to light-effect regions activate larger low-pass filtering windows, whereas pixels farther away from light effects tend to activate smaller ones.

Panel of Figure 12 — Fig. 8: Visualization of the light-effect map estimated by the Di²LE module.

Explore similar work

UBLLIE: Unified Backlight and Low-Light Image Enhancement
Aug 5, 2026Yasmin Yasin, Muhammad Usman, Ibrahim Radwan +1Image Enhancement Low-Light Image Enhancement
King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia · Ontario Tech University, Oshawa, Canada · University of Western Australia, Perth, Australia +1
SIMI: Self-information Mining Network for Low-light Image Enhancement
May 8, 2026Xuanshuo Fu, Lei Kang, Javier Vazquez-CorralLow-Light Image Enhancement
Computer Vision Center, Barcelona, Spain · Computer Science Dept., Universitat Aut`onoma de Barcelona, Barcelona, Spain
ControlLight: Towards Controllable, Consistent, and Generalizable Low-Light Enhancement
May 25, 2026Yufeng Yang, Jianzhuang Liu, Jisheng Chu +4Image Enhancement Low-Light Image Enhancement
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences · Zhejiang University

Aug 5, 2026J/K move · Enter open · S save

UBLLIE: Unified Backlight and Low-Light Image Enhancement

Yasmin Yasin, Muhammad Usman, Ibrahim Radwan, Saeed Anwar

King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia · Ontario Tech University, Oshawa, Canada · University of Western Australia, Perth, Australia+1

Image Enhancement Low-Light Image Enhancement

Open paper AskPDF arXiv

Backlit and low-light images often suffer from severe exposure imbalance or global underexposure, presenting significant challenges for both visual perception and downstream computer vision tasks. In this paper, we propose a unified, unsupervised enhancement framework that addresses both types of degradation without relying on paired ground-truth data. Our approach builds on CLIP-guided prompt learning to semantically supervise enhancement using learned positive and negative textual prompts. To improve the quality of our improvements over prior work, we design a symmetric residual U-Net backbone augmented with an Atrous Spatial Pyramid Pooling module. This architecture captures multi-scale contextual information, enabling adaptive correction under spatially heterogeneous illumination. During training, the enhancement network is guided by CLIP-based semantic similarity losses and refined via an iterative prompt optimization mechanism. Extensive experiments on both paired and unpaired datasets, including BAID, Backlit300, LOL, and VE-LOL-L, demonstrate that our framework consistently outperforms state-of-the-art supervised and unsupervised methods in terms of fidelity, perceptual quality, and generalization. Furthermore, our work emphasizes the need for stronger benchmarking protocols for backlit enhancement, a relatively underexplored area. The proposed framework provides a robust, scalable solution for real-world illumination enhancement across diverse lighting conditions.

Abstract

Explore similar work

UBLLIE: Unified Backlight and Low-Light Image Enhancement

SIMI: Self-information Mining Network for Low-light Image Enhancement

ControlLight: Towards Controllable, Consistent, and Generalizable Low-Light Enhancement