Lightweight Low-Light Image Enhancement via Distribution-Normalizing Preprocessing and Depthwise U-Net

Shimon Murai; Teppei Kurita; Ryuta Satoh; Yusuke Moriuchi

cs.CVApr 13, 2026

Lightweight Low-Light Image Enhancement via Distribution-Normalizing Preprocessing and Depthwise U-Net

Authors: Shimon Murai, Teppei Kurita, Ryuta Satoh, Yusuke Moriuchi

Organizations: Sony Semiconductor Solutions Corporation

Abstract

We present a lightweight two-stage framework for low-light image enhancement (LLIE) that achieves competitive perceptual quality with significantly fewer parameters than existing methods. Our approach combines frozen algorithm-based preprocessing with a compact U-Net built entirely from depthwise-separable convolutions. The preprocessing normalizes the input distribution by providing complementary brightness-corrected views, enabling the trainable network to focus on residual color correction. Our method achieved 3rd place in the CVPR 2026 NTIRE Efficient Low-Light Image Enhancement Challenge. We further provide extended benchmarks and ablations to demonstrate the general effectiveness of our methods.

Figures & tables

Figure 1: Parameter count vs. LPIPS on LOLv1 test (lower-left is better). Our Mid model (859K) achieves lower LPIPS than RetinexFormer (1.6M) and matches HVI-CIDNet w/o perc (2M). All “Ours” parameter counts include the frozen CPGA-Net preprocessor.

Figure 2: Qualitative comparison on LOLv1 test images. Our Mid model (859K params) produces visually competitive results with less color distortion than RetinexFormer and comparable detail to HVI-CIDNet.

Dataset	Variant	$\bar{\mu}$	$\sigma_{\text{inter}}$	$\bar{\sigma}$	$\sigma_{\text{intra}}$
LOLv1	Low	15.5	10.7	10.4	6.9
	CPGA	117.1	16.3	38.6	10.2
	CLAHE	29.2	13.3	16.9	8.9
	Gamma	61.7	19.7	22.9	6.9
	GT	116.9	25.9	46.0	13.0
NTIRE LLIE 2026	Low	20.4	20.4	28.9	20.6

Table 1: Distribution statistics of low-light images and their preprocessed versions.

\sigma_{\text{inter}}

: std of per-image mean brightness (inter-image variability).

\sigma_{\text{intra}}

: std of per-image contrast (intra-image variability). Lower values indicate a more uniform, normalized distribution.

Figure 3: Distribution of per-image mean brightness before and after preprocessing. On LOLv1 (top), CPGA shifts the distribution close to GT, while CLAHE provides intermediate normalization. On NTIRE LLIE 2026 (bottom), the input distribution is much wider, and preprocessing provides partial but complementary normalization.

Figure 4: Overview of our two-stage pipeline. Stage 1 applies frozen preprocessing (CPGA-Net and CLAHE) to the low-light input, producing a 9-channel tensor. Stage 2 is a lightweight 3-level U-Net with depthwise-separable convolution (DWConv) blocks that learns residual color correction. C denotes channel-wise concatenation and + denotes element-wise addition (global residual).

Large models
Method	Params	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$	LIQE $\uparrow$	MUSIQ $\uparrow$
HVI-CIDNet † (w/ perc) [ 18 ]	2.0M	23.81	0.857	0.086	0.078	4.233	71.91
HVI-CIDNet † (w/o perc) [ 18 ]	2.0M	23.50	0.870	0.105	0.107	3.765	70.84
BREAD † [ 7 ]	2.1M	22.95	0.837	0.160	0.130	3.631	70.57
Ours-Mid	859K	22.84	0.816	0.099	0.108	3.733	68.69
Ours-Large	2.3M	22.82	0.828	0.094	0.101	3.807	69.82

Table 2: Quantitative comparison on LOLv1 test set, grouped by model size. Full-reference metrics are reported without GT-mean normalization. Methods are split by the NTIRE 2026 ELLIE constraint of

<

1 MB in fp16 storage. Best in bold , second best underlined within each group.

\dagger

: official pretrained weights.

*

: unsupervised (not trained on LOLv1 pairs).

\ddagger

: trained by us with the same protocol as Ours.

Loss	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
L1	22.53	0.787	0.224	0.163
L1 + 0.5 LPIPS	22.42	0.806	0.113	0.115
L1 + 1.0 LPIPS	22.84	0.816	0.099	0.108

Table 3: Ablation on loss function (Mid model, 859K params, 9ch input).

	Method	Params	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Real	HVI-CIDNet †	2.0M	23.35	0.845	0.125
	RetinexFormer †	1.6M	23.35	0.838	0.170
	Ours-Mid	859K	23.02	0.832	0.132
	Ours-Tiny	338K	21.21	0.797	0.145
Synth.	HVI-CIDNet †	2.0M	25.70	0.942	0.047
	RetinexFormer †	1.6M	21.46	0.887	0.098

Table 4: Results on LOLv2-Real and LOLv2-Synthetic.

\dagger

: official pretrained weights. Our models use

\mathcal{L}_{\ell_{1}}{+}\mathcal{L}_{\text{LPIPS}}

Method	Params	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
Ours-Tiny	338K	22.30	0.789	0.124	0.124
HVI-CIDNet-Tiny	328K	22.09	0.801	0.127	0.130
RetinexFormer-Tiny	331K	20.97	0.797	0.129	0.130

Table 5: Parameter-matched comparison (

{\sim}

328–338K params) on LOLv1 test. All models trained from scratch with

\mathcal{L}_{\ell_{1}}+\mathcal{L}_{\text{LPIPS}}

Input 1	Input 2	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	DISTS $\downarrow$
None	None	22.12	0.809	0.128	0.136
CPGA	HE	22.46	0.790	0.120	0.120
Gamma	CLAHE	22.32	0.781	0.115	0.126
CPGA	CLAHE	22.30	0.789	0.124	0.124
Zero-DCE	CLAHE	22.07	0.782	0.119	0.126
CPGA	Gamma	21.68	0.790	0.117	0.121

Table 6: Preprocessing ablation (Tiny, 338K params, LOLv1). The 9ch input is [Input 1, low-light RGB, Input 2]. “None” denotes no preprocessing (raw 3ch RGB input without global residual).

Explore similar work

Multinex: Lightweight Low-light Image Enhancement via Multi-prior Retinex
Apr 11, 2026Alexandru Brateanu, Tingting Mu, Codruta Ancuti +1Low-Light Image Enhancement Image Enhancement
Department of Computer Science, University of Manchester, Manchester, United Kingdom · ETcTI, University Politehnica Timisoara, Timisoara, Romania · West University of Timisoara, Timisoara, Romania
WREN: Low Light Image Enhancement Using Retinex theory-based Double U-Net-like Structures
Jul 17, 2026Reina Kaneko, Junya Hara, Hiroshi Higashi +1Image Enhancement U-Net
The University of Osaka, Osaka, Japan · Kansai University, Osaka, Japan
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

Apr 11, 2026J/K move · Enter open · S save

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

Alexandru Brateanu, Tingting Mu, Codruta Ancuti, Cosmin Ancuti

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

Low-Light Image Enhancement Image Enhancement

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

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

WREN: Low Light Image Enhancement Using Retinex theory-based Double U-Net-like Structures

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