Bilevel Layer-Positioning LoRA for Real Image Dehazing

Yan Zhang; Long Ma; Yuxin Feng; Zhe Huang; Fan Zhou; Zhuo Su

cs.CVMar 11, 2026

Bilevel Layer-Positioning LoRA for Real Image Dehazing

Authors: Yan Zhang, Long Ma, Yuxin Feng, Zhe Huang, Fan Zhou, Zhuo Su

Organizations: School of Computer Science and Engineering, Sun Yat-sen University · Dalian University of Technology · Xidian University · National Engineering Research Center of Digital Life

Abstract

Learning-based real image dehazing methods have achieved notable progress, yet they still face adaptation challenges in diverse real haze scenes. These challenges mainly stem from the lack of effective unsupervised mechanisms for unlabeled data and the heavy cost of full model fine-tuning. To address these challenges, we propose the haze-to-clear text-directed loss that leverages CLIP's cross-modal capabilities to reformulate real image dehazing as a semantic alignment problem in latent space, thereby providing explicit unsupervised cross-modal guidance in the absence of reference images. Furthermore, we introduce the Bilevel Layer-positioning LoRA (BiLaLoRA) strategy, which learns both the LoRA parameters and automatically search the injection layers, enabling targeted adaptation of critical network layers. Extensive experiments demonstrate our superiority against state-of-the-art methods on multiple real-world dehazing benchmarks. The code is publicly available at https://github.com/YanZhang-zy/BiLaLoRA.

Figures & tables

Figure 2: Effectiveness of H2C loss in different real-scenes. The left two rows (daytime scenes) are from URHI and Fattal, and the right two rows (nighttime scenes) are from NHRW [ 44 ] .

Figure 3: Contribution of different network components to domain adaptation .

Figure 4: Quantitative results of cross-model flexibility . We evaluate four baseline dehazing architectures on three real dehazing datasets with four non-reference metrics to ensure the generality of this property.

Figure 5: Quantitative results of cross-domain stability . We leverage DEA pre-trained on four synthetic datasets to verify robustness across different source domains.

Figure 6: Qualitative results of cross-domain stability . BiLaLoRA consistently enhances real-world performance across various source domains, with data from Fattal.

Figure 7: Qualitative results of full fine-tuning vs. BiLaLoRA . The left two rows are from RTTS and the right two rows are from NHRW.

Metric		Fine-Tuning	BiLaLoRA	Rate
Performance	FADE $\downarrow$	0.610	0.638	$\downarrow$ 4.59%
	BIQME $\uparrow$	0.617	0.611	$\downarrow$ 0.97%
	Entropy $\uparrow$	7.569	7.572	$\uparrow$ 0.04%
	MUSIQ $\uparrow$	64.43	64.40	$\downarrow$ 0.05%
Efficiency	Train Time (H)	4.215	0.940	$\downarrow$ 77.70%
	Params. (M)	3.653	3.764	$\uparrow$ 3.03%

Table 1: Quantitative results of full fine-tuning vs. BiLaLoRA .

Method			RTTS [ 25 ]				URHI [ 25 ]				Fattal [ 10 ]				Average
Name		Venue	FADE	BIQME	Entropy	MUSIQ	FADE	BIQME	Entropy	MUSIQ	FADE	BIQME	Entropy	MUSIQ	FADE	BIQME	Entropy	MUSIQ
Synthetic	MSBDN [ 8 ]	CVPR 20	1.483	0.549	7.273	52.93	1.517	0.542	7.264	54.68	0.613	0.555	7.408	63.71	1.204	0.568	7.315	57.11
	DeHamer [ 18 ]	CVPR 22	1.806	0.542	7.215	52.90	1.853	0.537	7.217	54.67	0.756	0.552	7.411	64.31	1.472	0.544	7.281	57.29
	C 2 PNet [ 46 ]	CVPR 23	2.050	0.531	7.168	54.18	2.054	0.524	7.157	56.48	0.720	0.551	7.399	65.00	1.608	0.535	7.241	58.55
	DEA [ 4 ]	TIP 24	1.781	0.541	7.196	53.13	1.891	0.534	7.198	54.89	0.696	0.557	7.423	64.23	1.456	0.544	7.272	57.42
All-In-One	PromptIR [ 33 ]	NeurIPS 23	1.765	0.546	7.189	53.80	1.747	0.544	7.390	55.97	0.668	0.558	7.498	64.96	1.393	0.549	7.359	58.24

Table 2: Quantitative evaluations . All metrics are computed on RTTS, URHI and Fattal.

Figure 8: Visual comparisons on different datasets . Boxes indicate specific areas that highlight differences.

Figure 9: Generalization evaluation on HazyDet dataset . All models were evaluated on the testing dataset without retraining.

Figure 10: Generalization evaluation on other real datasets . Dense-Haze (Top) and O-Haze (Bottom). All models were evaluated on the testing dataset without retraining.

Model	H2C Loss		Layer-Positioning		Averaged Performance
Model	Positive	Negative	Naïve	Bilevel	FADE	BIQME	Entropy	MUSIQ
$\text{Q}_{\mathtt{a}}$	$\times$	$\times$	$\times$	$\times$	1.018	0.582	7.438	62.05
$\text{Q}_{\mathtt{b}}$	✓	$\times$	$\times$	$\times$	0.862	0.589	7.544	62.23
$\text{Q}_{\mathtt{c}}$	✓	$\times$	✓	$\times$	0.705	0.592	7.559	62.35
$\text{Q}_{\mathtt{d}}$	✓	$\times$	$\times$	✓	0.680	0.601	7.563	62.57
$\text{Q}_{\mathtt{e}}$	$\times$	✓	$\times$	$\times$	0.774	0.561	7.533	60.47

Table 3: Quantitative evaluation of ablation study .

Figure 11: Qualitative results of ablation study . Visual comparison of different combinations in Table 3 , with data from URHI.

Figure 12: Convergence behaviors . H2C loss convergence (Top) and MUSIQ improvement (Bottom) over training epochs.

Figure 13: Parameter analysis . The black dashed circle indicates the optimal selection.

Image Dehazing

Image dehazing restores clear scene details and colors from images obscured by haze, fog, or smoke.

See all 35 papers

Hihaze: One-Step Diffusion for High-Resolution Real-World Image Dehazing
Oct 19, 2026Zheng Chen, Yuchen Huang, Xiaoyang Liu +3Image Dehazing Diffusion Models for Image Restoration
Shanghai Jiao Tong University
Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling
Jun 24, 2026Yifei Qu, Ru Li, Junjie Chen +1Physics-Guided Image Restoration Image Restoration
Faculty of Computing, Harbin Institute of Technology, Harbin 150006, China
HazeSpikeMamba: Coupling Spiking-Inspired and State-Space Features for Self-Supervised Real-World Dehazing
Aug 7, 2026Haoran Liu, Huibin Li, Mingzhe Liu +2Unsupervised Domain Adaptation Spiking Neural Networks
School of Artificial Intelligence and Electronic Engineering, Sichuan Technology and Business University, Chengdu 611745, China · Chengdu University of Technology, College of Nuclear Technology and Automation Engineering, Chengdu 610059, China · Sigray, Inc., Concord, CA 94520, USA

Oct 19, 2026J/K move · Enter open · S save

Hihaze: One-Step Diffusion for High-Resolution Real-World Image Dehazing

Zheng Chen, Yuchen Huang, Xiaoyang Liu, Jue Gong+2

Shanghai Jiao Tong University

Image Dehazing Diffusion Models for Image Restoration Degradation-Aware Image Restoration Diffusion Models

Open paper AskPDF arXiv

Diffusion models have recently achieved encouraging results in real-world image dehazing. However, restoring high-resolution images remains challenging due to spatially non-uniform haze and inconsistent predictions across local patches. To address these issues, we propose Hihaze, a one-step diffusion framework for high-resolution real-world dehazing. Hihaze adopts two haze-aware modulation modules: attentive skip compensation (ASC) and a spatial degradation adapter (SDA). ASC selectively transfers reliable encoder information to the variational autoencoder (VAE) decoder, preserving structural details while suppressing haze-degraded features. SDA extracts multi-scale spatial degradation representations and injects them into the U-Net to guide restoration according to local haze variations. Furthermore, we design a training-free global-guided patch fusion (GPF) method. It uses a global prediction to align patch appearance and downweight unreliable predictions, followed by structure-guided detail compensation. Notably, this training-free design also makes GPF a plug-and-play module for other dehazing models. Experiments across four real-world benchmarks show that Hihaze surpasses state-of-the-art approaches in both visual quality and inference speed. Specifically, it delivers PSNR improvements of 3.62~~dB on Dense-Haze and 3.69~~dB on NH-HAZE compared with the strongest baselines. Code: https://github.com/zhengchen1999/Hihaze.

Abstract

Image Dehazing

Explore similar work

Hihaze: One-Step Diffusion for High-Resolution Real-World Image Dehazing

Efficient Real-World Dehazing via Physics-Inspired Global-Local Decoupling

HazeSpikeMamba: Coupling Spiking-Inspired and State-Space Features for Self-Supervised Real-World Dehazing