Multinex: Lightweight Low-light Image Enhancement via Multi-prior Retinex

Alexandru Brateanu; Tingting Mu; Codruta Ancuti; Cosmin Ancuti

cs.CVApr 11, 2026

Multinex: Lightweight Low-light Image Enhancement via Multi-prior Retinex

Authors: Alexandru Brateanu, Tingting Mu, Codruta Ancuti, Cosmin Ancuti

Organizations: Department of Computer Science, University of Manchester, Manchester, United Kingdom · ETcTI, University Politehnica Timisoara, Timisoara, Romania · West University of Timisoara, Timisoara, Romania

Abstract

Low-light image enhancement (LLIE) aims to restore natural visibility, color fidelity, and structural detail under severe illumination degradation. State-of-the-art (SOTA) LLIE techniques often rely on large models and multi-stage training, limiting practicality for edge deployment. Moreover, their dependence on a single color space introduces instability and visible exposure or color artifacts. To address these, we propose Multinex, an ultra-lightweight structured framework that integrates multiple fine-grained representations within a principled Retinex residual formulation. It decomposes an image into illumination and color prior stacks derived from distinct analytic representations, and learns to fuse these representations into luminance and reflectance adjustments required to correct exposure. By prioritizing enhancement over reconstruction and exploiting lightweight neural operations, Multinex significantly reduces computational cost, exemplified by its lightweight (45K parameters) and nano (0.7K parameters) versions. Extensive benchmarks show that all lightweight variants significantly outperform their corresponding lightweight SOTA models, and reach comparable performance to heavy models. Paper page available at https://albrateanu.github.io/multinex.

Figures & tables

Figure 2: Illustration of Multinex Architecture. Fusion modules $f_{\mathcal{L}}$ and $f_{\mathcal{R}}$ use illumination and reflectance guidance stacks $\mathcal{S}_{\mathcal{L}}$ and $\mathcal{S}_{\mathcal{R}}$ to produce color and luminance correction terms $\boldsymbol{\Delta}_{L}$ and $\boldsymbol{\Delta}_{R}$ . Retinex-like fusion then yields the Enhancement Delta $\boldsymbol{\Delta}_{I}$ .

		LOLv1 Dataset			LOLv2-real Dataset			LOLv2-syn Dataset			Model Size
	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Param. (M) $\downarrow$	GFLOPs $\downarrow$
	QuadPrior [ 38 ]	22.85	0.800	0.201	20.59	0.811	0.202	16.11	0.758	0.144	1252.75	1103.2
	GLARE [ 51 ]	23.55	0.863	0.086	22.51	0.871	0.105	18.21	0.842	0.178	59.48	508.42
	LLFlow [ 39 ]	21.15	0.854	0.116	17.43	0.831	0.176	24.81	0.919	0.067	17.42	358.4
Heavy ( $>10$ M)	GSAD [ 11 ]	22.77	0.852	0.102	20.15	0.846	0.113	24.47	0.929	0.051	17.36	442.02
	KinD [ 49 ]	17.65	0.775	0.207	17.54	0.669	0.375	18.32	0.796	0.252	8.02	34.99

Table 1: Results on LOLv1 [ 41 ] , and LOLv2 [ 45 ] datasets (real and synthetic) across four model groups of parameter scales. Best performance in each group is highlighted in bold .

Figure 3: Qualitative comparison of SOTA light-weight approaches [ 41 , 6 , 7 , 1 , 21 , 33 , 27 ] and Multinex on reference data.

	MEF		LIME		DICM		NPE		Mean		Model Size
Method	NIQE $\downarrow$	BRISQUE $\downarrow$	NIQE $\downarrow$	BRISQUE $\downarrow$	NIQE $\downarrow$	BRISQUE $\downarrow$	NIQE $\downarrow$	BRISQUE $\downarrow$	NIQE $\downarrow$	BRISQUE $\downarrow$	Params (M) $\downarrow$	GFLOPs $\downarrow$
SNR-Net [ 43 ]	4.14	31.28	5.51	39.22	4.62	37.35	4.36	26.65	4.66	33.63	26.35	4.01
LLFlow [ 39 ]	3.92	30.27	5.29	27.06	3.78	26.36	4.16	28.86	4.29	28.14	17.42	358.40
CIDNet [ 44 ]	3.56	13.77	4.13	16.25	3.79	21.47	3.74	18.92	3.81	17.60	1.88	7.57
ZeroDCE [ 7 ]	4.93	17.32	5.82	20.44	4.58	27.56	4.53	20.72	4.97	21.51	0.075	4.83
Multinex	3.46	13.82	3.76	14.60	3.64	16.39	3.71	12.52	3.64	14.33	0.0446	2.50

Table 2: No-reference results on MEF, LIME, DICM, and NPE using NIQE [ 29 ] and BRISQUE [ 28 ] . Lower is better. † indicates results obtained with newly trained models using the same architecture, training configuration, and evaluation protocol as the conference version.

Method	Bic.	Boa.	Bot.	Bus	Car	Cat	Cha.	Cup	Dog	Mot.	Ppl.	Tab.	mAP50%	P (K)
YOLOv3 [ 32 ]	79.8	75.3	78.1	92.3	83.0	68.0	69.0	79.0	78.0	77.3	81.5	55.5	76.4	–
KinD [ 49 ]	80.9	75.0	75.8	93.3	82.4	69.4	69.2	79.0	76.9	76.3	79.6	55.4	76.1	8000
ZDCE [ 7 ]	81.2	75.0	75.7	93.4	83.2	67.7	70.2	76.4	74.1	77.7	81.3	55.5	75.9	79
Pair [ 6 ]	80.8	78.3	76.8	90.5	84.5	66.8	69.1	75.6	78.9	73.7	80.3	54.5	75.8	342
DENet [ 31 ]	80.9	79.2	80.1	90.7	84.5	70.7	72.0	79.3	80.1	76.7	82.4	58.0	77.9	45
PE-Y [ 47 ]	84.7	79.2	79.3	92.5	83.9	71.5	71.7	79.7	79.7	77.3	81.8	55.3	78.0	91

Table 3: Low-light object detection on ExDark [ 22 ] : per-class AP% and overall mAP50% , with LLIE model parameter count ( P ). Best results are in bold , second best are underlined .

(a) Multinex Priors			(b) CWA/MSEF Analysis				(c) CWA and Attention Variants
Formulation	H-Dim	PSNR $\uparrow$	CWA	MSEF	H-Dim	PSNR $\uparrow$	Attention Operation	H-Dim	PSNR $\uparrow$
$\hat{\mathbf{I}}_{i}=f_{i}(\mathbf{I}_{i})$	87	14.15	✗	✗	55	19.92	CBAM ${}_{\text{S}}$	39	20.44
$\hat{\mathbf{I}}_{i}=f_{\mathcal{L}}(\mathcal{S}_{\mathcal{L}},\theta_{\mathcal{L}})+\mathbf{I}_{i}$	87	20.57	✓	✗	55	21.34	MHSA + Pooling	39	21.67
$\hat{\mathbf{I}}_{i}=f_{\mathcal{R}_{i}}(\mathcal{S}_{\mathcal{R}},\theta_{\mathcal{R}})+\mathbf{I}_{i}$	87	18.50	✗	✓	39	22.27	MDTA	39	22.39
$\hat{\mathbf{I}}_{i}=f_{\mathcal{L}}(\mathcal{S}_{\mathcal{L}},\theta_{\mathcal{L}})\odot f_{\mathcal{R}_{i}}(\mathcal{S}_{\mathcal{R}},\theta_{\mathcal{R}})+\mathbf{I}_{i}$	39	23.19	✓	✓	39	23.19	CWA	39	23.19

Table 4: Results of ablation studies (a), (b) and (c) validating design elements of Multinex.

Figure 4: Qualitative ablation on Multinex Priors study (a).

Figure 5: The example image used by Sec. B .

Figure 6: LRA visualization of chrominance candidates $\{[\mathbf{C}_{b},\mathbf{C}_{r}],[\mathbf{U},\mathbf{V}],[\mathbf{O}_{1},\mathbf{O}_{2}],[\mathbf{r},\mathbf{g}],\mathbf{S}\}$ .

Figure 7: Similarity heatmaps between all candidate grouped reflectance-guidance components.

Figure 8: LRA visualization of combinations of the chosen feature maps for the reflectance guidance stack $\mathcal{S}_{\mathcal{R}}$ .

Figure 9: Top Row: Candidate illumination feature descriptors $\{\mathbf{Y}_{\text{Rec.709}},\mathbf{Y}_{\text{mean}},\mathbf{Y}_{\text{YCgCo}},\mathbf{Y}_{\text{vmax}},\mathbf{Y}_{\text{lightness}},\mathbf{Y}_{\text{L}_{2}}\}$ . Middle Row: $\boldsymbol{\Delta}_{E}(c)$ maps of the same candidate descriptors, where $c\in\{1,2,\ldots,6\}$ . Bottom row: $\boldsymbol{\Delta}_{G}(c)$ maps of the same candidate descriptors.

Figure 10: Similarity heatmap between all candidate individual illumination-guidance components.

Prior	$\boldsymbol{\Delta}_{E}(c)\uparrow$ / Rank	$\boldsymbol{\Delta}_{G}(c)\uparrow$ / Rank	Avg. Rank
$\mathbf{Y}_{\text{Rec.709}}$	0.0132 / 1	0.0007 / 3	2.0
$\mathbf{Y}_{\text{vmax}}$	0.0107 / 2	0.0019 / 2	2.0
$\mathbf{Y}_{\text{L}_{2}}$	0.0012 / 6	0.0194 / 1	3.5
$\mathbf{Y}_{\text{lightness}}$	0.0029 / 4	0.0002 / 4	4.0
$\mathbf{Y}_{\text{YCgCo}}$	0.0038 / 3	0.0000 / 6	4.5
$\mathbf{Y}_{\text{mean}}$	0.0017 / 5	0.0000 / 5	5.0

Table 5: Importance ranking of candidate luminance priors. Higher values indicate greater unique contribution to the stack.

$\mathcal{L}_{\text{MSE}}$	$\mathcal{L}_{\text{MS-SSIM}}$	$\mathcal{L}_{\text{Perc}}$	PSNR $\uparrow$	SSIM $\uparrow$
✓	✗	✗	22.31	0.815
✗	✓	✗	19.04	0.820
✗	✗	✓	18.80	0.793
✓	✓	✗	22.43	0.830
✓	✗	✓	22.80	0.821
✗	✓	✓	19.74	0.838

Table 6: Ablation of loss functions.

$\mathcal{S}_{\mathcal{L}}$ Components				#Ch	PSNR $\uparrow$	SSIM $\uparrow$
$\mathbf{Y}_{\text{Rec.709}}$	$\mathbf{Y}_{\text{vmax}}$	$\mathbf{Y}_{\text{lightness}}$	$\mathbf{Y}_{\text{L}_{2}}$
✓	✗	✗	✗	1	18.80	0.753
✗	✓	✗	✗	1	19.10	0.744
✗	✗	✓	✗	1	19.23	0748
✗	✗	✗	✓	1	19.05	0.757
✓	✓	✗	✗	2	19.83	0.762

Table 7: Ablation studies on feature maps of luminance guidance stack, supported by a complete Multinex architecture.

$\mathcal{S}_{\mathcal{R}}$ Components			#Ch	PSNR $\uparrow$	SSIM $\uparrow$
$[\mathbf{C_{b},C_{r}}]$	$[\mathbf{r,g}]$	$\mathbf{S}$
✓	✗	✗	2	21.55	0.792
✗	✓	✗	2	21.90	0.805
✗	✗	✓	1	20.62	0.779
✓	✓	✗	4	22.98	0.835
✓	✗	✓	3	22.12	0.819

Table 8: Ablation studies on feature maps of reflectance guidance stack, supported by a complete Multinex architecture.

Placement	Formulation $f(\mathcal{S})=\mathrm{Conv}_{1\times 1}\circ\dots$	PSNR $\uparrow$	SSIM $\uparrow$
Before	$\mathrm{FB}^{2T}\Big(\mathrm{Conv}_{1\times 1}(\mathcal{S})\odot\mathrm{CWA}(\mathcal{S})\Big)$	22.78	0.831
Between	$\mathrm{FB}^{T}\Big(\mathrm{CWA}(\mathcal{S})\odot\mathrm{FB}^{T}\circ\mathrm{Conv}_{1\times 1}(\mathcal{S})\Big)$	23.19	0.843
After	$\mathrm{CWA}(\mathcal{S})\odot\mathrm{FB}^{2T}\circ\mathrm{Conv}_{1\times 1}(\mathcal{S})$	22.41	0.823

Table 9: Ablation on where to place CWA in fusion networks.

Panel of Figure 12 — Figure 12: Qualitative comparison on reference dataset LOL-v1 [ 41 ] between Multinex and state-of-the-art lightweight and micro scale models RetinexNet [ 41 ] , PairLIE [ 6 ] , ZeroDCE [ 7 ] , LYT-Net [ 1 ] , RUAS [ 21 ] , RSF [ 33 ] , SCI [ 27 ] .

Panel of Figure 13 — Figure 13: Qualitative comparison on reference dataset LOL-v1 [ 41 ] between Multinex and state-of-the-art lightweight and micro scale models RetinexNet [ 41 ] , PairLIE [ 6 ] , ZeroDCE [ 7 ] , LYT-Net [ 1 ] , RUAS [ 21 ] , RSF [ 33 ] , SCI [ 27 ] .

Panel of Figure 14 — Figure 14: Qualitative comparison on reference dataset LOL-v1 [ 41 ] between Multinex-Nano and state-of-the-art micro scale models RUAS [ 21 ] , RSF [ 33 ] , SCI [ 27 ] .

Panel of Figure 15 — Figure 15: Qualitative comparison on reference dataset LOL-v2-real [ 45 ] between Multinex and state-of-the-art mid-sized (1-10 M param.) models RetinexFormer [ 3 ] and CIDNet [ 44 ] , and heavy ( $>$ 10 M param.) model GLARE [ 51 ] .

Panel of Figure 16 — Figure 16: Qualitative comparison on no-reference datasets MEF [ 26 ] , LIME [ 9 ] , DICM [ 20 ] , NPE [ 36 ] between Multinex and state-of-the-art mid-sized (1-10M param.) models RetinexFormer [ 3 ] and CIDNet [ 44 ] and heavy-weight ( $>$ 10M param.) GLARE [ 51 ] .

Figure 17: Additional results on no-reference dataset MEF [ 26 ] . For corresponding images, top is input, and bottom is Multinex output.

Figure 18: Additional results on no-reference dataset LIME [ 9 ] . For corresponding images, top is input, and bottom is Multinex output.

Figure 19: Additional results on no-reference dataset DICM [ 20 ] . For corresponding images, top is input, and bottom is Multinex output.

Figure 20: Additional results on no-reference dataset NPE [ 36 ] . For corresponding images, top is input, and bottom is Multinex output.

Figure 21: A few challenging cases from DICM and MEF datasets.

Explore similar work

May 19, 2026

InterLight: Leveraging Intrinsic Illumination Priors for Low-Light Image Enhancement

Low-Light Image Enhancement (LLIE) has long been a challenging problem in low-level vision, as insufficient illumination often leads to low contrast, detail loss, and noise. Recent studies show that deep learning-based Retinex theory can effectively decouple illumination and reflectance. However, existing methods frequently suffer from over-enhancement or color distortion, and often assume uniform noise or ideal lighting. To address these limitations, we propose InterLight, a novel framework that systematically excavates and operationalizes intrinsic illumination priors for LLIE.Our core insight is that robust enhancement requires not just estimating illumination, but constructing an illumination-aware pipeline. We first inject sensor-level illumination-response priors via physics-guided augmentation, then represent the degradation through adaptive prompts conditioned on the scene's latent illumination state. This explicit representation directly guides a luminance-gated intrinsic memory mechanism to selectively compensate for information loss, prioritizing reconstruction in dark regions while preserving fidelity in bright ones. Finally, the entire process is regularized by a self-supervised consistency objective that distills illumination-invariant features. By deeply exploiting intrinsic illumination priors, our method achieves clearer textures and more visually coherent enhancement results. Extensive experiments across multiple benchmarks demonstrate the effectiveness of our approach. Code is available at: https://github.com/House-yuyu/InterLight.

Ziqi Wang, Xu Zhang, Laibin Chang +3

National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University · Department of Computer Science, University of Macau · Department of Computer Vision, Mohamed bin Zayed University of Artificial Intelligence +1

Image Enhancement Low-Light Image Enhancement

Ask PDF

Sep 25, 2026

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

Low-light image enhancement (LLIE) remains challenging for lightweight models because illumination restoration and color fidelity are difficult to optimize simultaneously in the RGB color space. Although recent color-decoupled methods separate luminance and chrominance representations, they primarily optimize luminance as an enhancement target, leaving its potential as an explicit guidance prior largely unexplored during feature reconstruction. To address this limitation, we propose IDM-Net, a lightweight Illumination-Decoupled Modulation Network for low-light image enhancement. IDM-Net adopts a dual-encoder architecture consisting of a structure encoder that extracts multi-scale appearance features from the RGB image and a lightweight illumination encoder that learns illumination priors from the decoupled luminance (Y) channel. To effectively exploit these priors, we introduce an Illumination-Guided Modulation (IGM) module that injects multi-scale illumination cues into the decoder through spatially adaptive affine modulation, enabling accurate brightness restoration while preserving natural color consistency. Furthermore, we design a lightweight Feature Refinement Block (FRB) to progressively suppress degradation artifacts and recover fine-grained image details during reconstruction. Extensive experiments on multiple standard low-light image enhancement benchmarks demonstrate that IDM-Net achieves competitive performance among lightweight LLIE methods while maintaining an excellent balance between restoration quality and computational efficiency.

Cheng-Yen Hsiao, Jing-Ming Guo

Department of Electrical Engineering National Taiwan University of Science and Technology Taipei, Taiwan

Low-Light Image Enhancement

Ask PDF

May 11, 2026

M2Retinexformer: Multi-Modal Retinexformer for Low-Light Image Enhancement

Low-light image enhancement is challenging due to complex degradations, including amplified noise, artifacts, and color distortion. While Retinex-based deep learning methods have achieved promising results, they primarily rely on single-modality RGB information. We propose M2Retinexformer (Multi-Modal Retinexformer), a novel framework that extends Retinexformer by incorporating depth cues, luminance priors, and semantic features within a progressive refinement pipeline. Depth provides geometric context that is invariant to lighting variations, while luminance and semantic features offer explicit guidance on brightness distribution and scene understanding. Modalities are extracted at multiple scales and fused through cross-attention, with adaptive gating dynamically balancing illumination-guided self-attention and cross-attention based on the reliability of auxiliary cues. Evaluations on the LOL, SID, SMID, and SDSD benchmarks demonstrate overall improvements over Retinexformer and recent state-of-the-art methods. Code and pretrained weights are available at https://github.com/YoussefAboelwafa/M2Retinexformer

Youssef Aboelwafa, Hicham G. Elmongui, Marwan Torki

Alexandria University, Egypt

Image Enhancement Low-Light Image Enhancement

Ask PDF