Bidirectional Cross-frequency Domain Interaction Network for UHD Low-Light Image Enhancement

Bo Fu; Hai Chi; Liyan Wang; Shilin Fu; Wei Wang; Kin-Man Lam; Jinshan Pan; Cong Wang

cs.CVOct 24, 2026

Bidirectional Cross-frequency Domain Interaction Network for UHD Low-Light Image Enhancement

Authors: Bo Fu, Hai Chi, Liyan Wang, Shilin Fu, Wei Wang, Kin-Man Lam, Jinshan Pan, Cong Wang

Organizations: School of Computer and Artificial Intelligence, Liaoning Normal University, Dalian 116081, China · School of Mathematical Sciences, Dalian University of Technology, Dalian 116024, China · School of Computer Sciences, Dalian University of Technology, Dalian 116024, China · School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, Shenzhen, China ·

Abstract

Existing Ultra-High-Definition (UHD) low-light image enhancement (LLIE) methods typically rely on either spatial- or frequency-domain modeling, or independently process different frequency components for efficient restoration. However, they often fail to jointly exploit complementary spatial--frequency information and the inherent relationship between illumination and structural information across frequencies, leading to inconsistent illumination, residual noise, and detail artifacts. In this paper, we propose a novel Bidirectional Cross-frequency Domain Interaction Network (BiFreqNet) for UHD LLIE. Specifically, BiFreqNet incorporates a Dual-domain Feature Interaction (DFI) module into the multi-scale representations of a pretrained variational autoencoder (VAE) to perform reciprocal enhancement between low- and high-frequency features. DFI consists of a Spatial and Spectral-Aware Low-Frequency Enhancement (S $^2$ LFE) block and a Cross-Frequency Enhancement (CFE) block. S $^2$ LFE exploits high-frequency structural cues to guide low-frequency restoration and jointly models spatial and spectral information to recover global exposure and spatially varying illumination while suppressing sensor noise. CFE further leverages the enhanced low-frequency contextual information to guide adaptive high-frequency refinement, enabling faithful reconstruction of structures and fine details. Extensive experiments on four general and UHD LLIE benchmarks demonstrate that BiFreqNet achieves competitive or superior performance compared with state-of-the-art methods.

Figures & tables

Panel of Figure 1 — Fig. 1: Visual comparison of UHD low-light image enhancement. Existing spatial-/frequency-domain methods (b)–(d) and frequency-decomposition-based methods (e)–(f) suffer from suboptimal restoration results due to insufficient cross-frequency interaction. In contrast, our BiFreqNet (g) produces visually pleasing results through bidirectional interaction between low- and high-frequency representations.

Fig. 2: Overview of our proposed bidirectional cross-frequency domain interaction network (BiFreqNet), with two training stages. In Stage I (a), we pre-train a VAE to establish a compact latent space and extract reliable multi-scale image features for subsequent low-light enhancement. In Stage II (b), given a low-light image $\mathbf{I}_{lq}$ , we first extract multi-scale features $\{\mathbf{F}_{i}\}_{i=1}^{4}$ using the frozen VAE encoder and enhance them with the proposed DFI modules, yielding $\{\widetilde{\mathbf{F}}_{i}\}_{i=1}^{4}$ . We then fuse each $\widetilde{\mathbf{F}}_{i}$ with its corresponding $\mathbf{F}_{i}$ via concatenation and a $1\times 1$ convolution to obtain $\{\mathbf{F}^{fus}_{i}\}_{i=1}^{4}$ , while deriving the latent representation $\mathbf{Z}$ from $\boldsymbol{\mu}$ and $\log\boldsymbol{\sigma}^{2}$ via reparameterization. Finally, $\mathbf{Z}$ and $\{\mathbf{F}^{fus}_{i}\}_{i=1}^{4}$ are progressively decoded by the frozen VAE decoder to generate the enhanced image $\mathbf{I}_{out}$ .

Fig. 3: Overview of our proposed dual-domain feature interaction (DFI) module. It consists of a spatial- and spectral-aware low-frequency enhancement (S 2 LFE) block and a cross-frequency enhancement (CFE) block for reciprocally enhancing the low- and high-frequency features.

Fig. 4: Overview of the global spectral refinement (GSR) block (a) and the Mamba for image restoration (MaIR) block (b). The GSR block aims to correct global exposure by refining the amplitude and phase spectra in the Fourier domain, while the MaIR block aims to model spatially varying illumination and suppress sensor noise through a nested S-shaped scanning strategy.

Type	Method	Venue	UHD-LL [ 24 ]		UHD-LOL4K [ 25 ]		Params (M) $\downarrow$	FLOPs (G) $\downarrow$
Type	Method	Venue	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	Params (M) $\downarrow$	FLOPs (G) $\downarrow$
non-UHD	Zero-DCE++ [ 7 ]	TPAMI’22	16.41	0.630	15.58	0.835	0.001	1.32
	RUAS [ 8 ]	CVPR’21	13.56	0.749	14.68	0.758	0.0003	0.83
	Uformer [ 46 ]	CVPR’22	19.28	0.849	29.99	0.980	20.63	43.86
	Restormer [ 47 ]	CVPR’22	22.25	0.871	36.91	0.988	26.13	141.75
	DiffLL [ 10 ]	SIGGRAPH’23	21.36	0.872	36.95	0.989	17.29	29.74

TABLE I: Quantitative comparison with state-of-the-art methods on the UHD-LL [ 24 ] and UHD-LOL4K [ 25 ] datasets. Higher PSNR and SSIM values indicate better performance, whereas lower Params and FLOPs values are preferred. FLOPs are computed on a

512\times 512

input.

Panel of Figure 6 — Fig. 5: Visual comparison of representative UHD low-light image enhancement methods on the UHD-LL dataset [ 24 ] . BiFreqNet produces more visually pleasing results with more faithful illumination and clearer details.

Panel of Figure 7 — Fig. 6: Visual comparison results on the UHD-LOL4K dataset [ 25 ] . BiFreqNet produces more visually pleasing results with more faithful illumination and clearer details.

Type	Method	Venue	LOLv1 [ 48 ]		LOLv2-Real [ 49 ]		Params (M) $\downarrow$	FLOPs (G) $\downarrow$
Type	Method	Venue	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	Params (M) $\downarrow$	FLOPs (G) $\downarrow$
non-UHD	Uformer [ 46 ]	CVPR’22	18.55	0.721	18.44	0.759	20.63	43.86
	Restormer [ 47 ]	CVPR’22	22.37	0.816	24.91	0.851	26.13	141.75
	Retinexformer [ 11 ]	ICCV’23	25.16	0.845	22.80	0.840	1.61	15.57
	MambaIR [ 41 ]	ECCV’24	22.31	0.826	21.25	0.831	4.30	60.66
	RetinexMamba [ 14 ]	ICONIP’24	24.03	0.827	22.45	0.844	4.59	42.82

TABLE II: Quantitative comparison with state-of-the-art methods on the LOLv1 [ 48 ] and LOLv2-Real [ 49 ] datasets. Higher PSNR and SSIM values indicate better performance, whereas lower Params and FLOPs values are preferred. FLOPs are computed on a

512\times 512

input.

Panel of Figure 9 — Fig. 7: Visual comparison results on the LOLv1 dataset [ 48 ] . BiFreqNet produces more visually pleasing results with more faithful illumination, clearer structures, and finer details.

Panel of Figure 10 — Fig. 8: Visual comparison results on the LOLv2-Real dataset [ 49 ] . BiFreqNet produces more visually pleasing results with more faithful illumination, clearer structures, and finer details.

ID	Options	PSNR $\uparrow$	SSIM $\uparrow$	Params $\downarrow$	FLOPs $\downarrow$
(a)	w/o DFI	22.81	0.784	1.18M	21.53G
(b)	w/o S 2 LFE	25.04	0.871	1.87M	24.30G
(c)	w/o CFE	27.03	0.923	8.57M	52.63G
(d)	w/ DFI (Ours)	27.39	0.929	9.26M	55.33G

TABLE III: Effect of the DFI module on UHD-LL [ 24 ] . The best performance is marked in bold .

Panel of Figure 12 — Fig. 9: Visual ablation of DFI, illustrating the complementary roles of S 2 LFE and CFE.

ID	Options	PSNR $\uparrow$	SSIM $\uparrow$	Params $\downarrow$	FLOPs $\downarrow$
(a)	w/o GSR; w/o MaIR	24.96	0.899	3.54M	29.12G
(b)	w/o GSR	27.21	0.916	4.32M	39.53G
(c)	w/o MaIR	26.71	0.923	8.48M	44.92G
(d)	MaIR $\rightarrow$ LFSS [ 29 ]	26.35	0.929	10.47M	59.01G
(e)	MaIR $\rightarrow$ Transformer Block [ 25 ]	26.70	0.911	12.48M	57.06G
(f)	w/ S 2 LFE (Ours)	27.39	0.929	9.26M	55.33G

TABLE IV: Internal analysis of the S 2 LFE block on UHD-LL [ 24 ] . FLOPs are measured with a

512\times 512

input. The best performance is marked in bold .

Panel of Figure 14 — Fig. 10: Visual ablation of the S 2 LFE block, highlighting its contribution to spatial–frequency feature modeling and detail restoration.

Mamba Blocks	PSNR $\uparrow$	SSIM $\uparrow$	Params $\downarrow$	FLOPs $\downarrow$
$\{2,2,4,8\}$	26.67	0.920	8.45M	47.84G
$\{2,4,4,8\}$	26.49	0.919	8.62M	52.08G
$\{2,4,6,8\}$	27.39	0.929	9.26M	55.33G
$\{2,6,6,8\}$	27.04	0.926	9.43M	59.56G

TABLE V: Ablation study on the number of low-frequency Mamba blocks on UHD-LL [ 24 ] . FLOPs are measured with a

512\times 512

input. The best performance is marked in bold .

Panel of Figure 16 — Fig. 11: Visual ablation of the number of MaIR blocks, highlighting their contribution to spatial–frequency feature modeling and detail restoration.

ID	Options	PSNR $\uparrow$	SSIM $\uparrow$	Params $\downarrow$	FLOPs $\downarrow$
(a)	CFE $\rightarrow$ HFE [ 29 ]	27.13	0.925	9.16M	56.27G
(b)	CFE $\rightarrow$ HFRM [ 10 ]	26.96	0.922	9.25M	55.20G
(c)	w/ CFE (Ours)	27.39	0.929	9.26M	55.33G

TABLE VI: Internal analysis of the CFE block on UHD-LL [ 24 ] . The best performance is marked in bold .

Panel of Figure 18 — Fig. 12: Visual ablation of CFE, highlighting its contribution to low-frequency-guided high-frequency detail refinement.

ID	Options	PSNR $\uparrow$	SSIM $\uparrow$	Params $\downarrow$	FLOPs $\downarrow$
(a)	w/o High-Freq Guidance	27.08	0.924	8.50M	53.18G
(b)	w/o Low-Freq Guidance	27.17	0.925	9.26M	55.27G
(c)	Bidirectional Guidance	27.39	0.929	9.26M	55.33G

TABLE VII: Effect of the bidirectional guidance mechanism on UHD-LL [ 24 ] dataset. The best performance is marked in bold .

Panel of Figure 20 — Fig. 13: Visual comparison of the bidirectional guidance mechanism, showing the benefit of the ordered high-to-low-to-high interaction.

Panel of Figure 21 — Fig. 14: BiFreqNet may fail to restore the white balance and color temperature in extremely low-light scenes due to insufficient color information in low-light RGB input.

Low-Light Image Enhancement

Low-light image enhancement restores visibility, color, and detail in underexposed images while limiting noise, blur, and color distortion.

See all 95 papers

Low-light Image Enhancement via Multi-scale Attention combined with Fourier Transform
Jul 27, 2026Wenbin Du, Jian Long, Zhu CaoFrequency-Domain Feature Learning Low-Light Image Enhancement
aKey Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Shanghai, 200237, China · bCollege of Electronics and Information Engineering, Tongji University, Shanghai 201804, China · cShanghai Research Institute for Intelligent Autonomous Systems, Tongji University, Shanghai 201210, China
Unifying Amplitude-Phase Coupling and Reliable Endogenous Priors for Low-Light Image Enhancement
Oct 17, 2026Tongshun Zhang, Zixuan Zhong, Pingping Liu +4Low-Light Image Enhancement Image Restoration
College of Computer Science and Technology, Jilin University, Changchun, Jilin, 130012, China, and the Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education, Jilin University, Changchun, Jilin 130012, China. · College of Software, Jilin University, Changchun, Jilin, 130012, China. · School of Computer and Artificial Intelligence, XiangnanUniversity, Chenzhou, 423300, China. +1
IDM-Net: A Lightweight Illumination-Decoupled Modulation Network for Low-Light Image Enhancement
Sep 25, 2026Cheng-Yen Hsiao, Jing-Ming GuoLow-Light Image Enhancement
Department of Electrical Engineering National Taiwan University of Science and Technology Taipei, Taiwan

Jul 27, 2026J/K move · Enter open · S save

Low-light Image Enhancement via Multi-scale Attention combined with Fourier Transform

Wenbin Du, Jian Long, Zhu Cao

aKey Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and Technology, Shanghai, 200237, China · bCollege of Electronics and Information Engineering, Tongji University, Shanghai 201804, China · cShanghai Research Institute for Intelligent Autonomous Systems, Tongji University, Shanghai 201210, China

Frequency-Domain Feature Learning Low-Light Image Enhancement Multi-Scale Attention

Open paper AskPDF arXiv

Low-light image enhancement (LLIE) aims to improve image quality and clarity in diverse and demanding low-illumination environments. However, existing deep learning-based LLIE methods struggle to accurately capture real-world illumination and restore texture details, largely because their algorithmic strengths remain underutilized. To address these issues, we present a supervised frequency domain deep learning network for LLIE, named multi-scale attention combined with the Fourier transform (MSFT) which adopts a U-shaped, one-stage architecture that infuses guidance from low-light images into the network by channeling it through multi-scale attention. We further fuse the amplitude information from priori channels with that of the low-light image in MSFT's self-created module, and carry out multi-scale guidance along with the network. Subsequently, to better enhance the faint feature, such as fine content and textures, and to better fuse global context confidence in the decoding stage, we separately introduce a multi-shape synergistic attention and a lightweight network that effectively integrate information in high-dimensional space to embed into the superlative feature space channel containing rich texture information. Extensive experiments conducted on LOL, SID, SMID, and SDSD datasets demonstrate that MSFT significantly outperforms state-of-the-art competitors. For example, compared with Retinexformer, our method achieves a peak signal-to-noise ratio of up to 41.76 decibels on the SDSD-outdoor dataset with an increase of 11.92 decibels and a structural similarity index of 0.988 with a 13.80% improvement.

Abstract

Low-Light Image Enhancement

Explore similar work

Low-light Image Enhancement via Multi-scale Attention combined with Fourier Transform

Unifying Amplitude-Phase Coupling and Reliable Endogenous Priors for Low-Light Image Enhancement

IDM-Net: A Lightweight Illumination-Decoupled Modulation Network for Low-Light Image Enhancement