pith. sign in

arxiv: 2604.10359 · v2 · submitted 2026-04-11 · 💻 cs.CV · cs.AI

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

Alexandru Brateanu , Tingting Mu , Codruta Ancuti , Cosmin Ancuti This is my paper

Pith reviewed 2026-05-10 15:29 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords low-light image enhancementRetinex modellightweight neural networkmulti-prior fusionillumination priorreflectance adjustmentexposure correctionedge deployment
0
0 comments X

The pith

Multinex fuses multiple analytic priors inside a Retinex residual model to perform high-quality low-light enhancement with as few as 0.7K parameters.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents Multinex as an ultra-lightweight network that decomposes a low-light image into separate stacks of illumination and color priors drawn from several distinct analytic representations. These priors are then fused inside a residual Retinex formulation to produce the luminance and reflectance corrections needed to restore natural exposure and color. By focusing on enhancement rather than full reconstruction and using only lightweight neural operations, the method achieves strong results with far fewer parameters than typical SOTA models. A sympathetic reader would care because the approach removes the need for heavy compute or multi-stage training while still matching or exceeding the quality of larger networks on standard benchmarks.

Core claim

By integrating multiple fine-grained priors derived from distinct analytic representations into a principled Retinex residual formulation, Multinex learns to fuse illumination and color information into the exact luminance and reflectance adjustments required to correct exposure, yielding artifact-reduced results in networks ranging from 45K down to 0.7K parameters that outperform other lightweight models and approach the performance of much heavier ones.

What carries the argument

The multi-prior Retinex residual formulation that decomposes the input into illumination and color prior stacks and learns their fusion for luminance and reflectance adjustments.

If this is right

  • Low-light enhancement becomes practical for edge devices because parameter counts drop to thousands rather than millions.
  • Single-color-space instability is avoided by fusing priors from multiple analytic representations before correction.
  • Benchmark results show all lightweight variants exceed corresponding lightweight SOTA models while nearing heavy-model quality.
  • Prioritizing enhancement over reconstruction and using lightweight operations cuts computational cost without sacrificing output fidelity.
  • The same prior-fusion structure can be instantiated at different scales from nano to lightweight while retaining the performance gains.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method may transfer to real-time video if temporal coherence between consecutive frames is added to the prior stacks.
  • Similar multi-prior fusion could reduce model size in other restoration tasks such as denoising or dehazing where analytic priors already exist.
  • Hardware deployment on mobile cameras becomes feasible, allowing on-device correction without cloud offload.
  • A natural next measurement is whether the same framework maintains quality when input noise levels vary widely.

Load-bearing premise

The fusion of multiple fine-grained analytic priors inside the Retinex residual model will generalize to produce natural, artifact-free enhancements across varied real-world low-light scenes without further tuning.

What would settle it

A controlled test on a diverse set of unseen low-light scenes that reveals either visible color or exposure artifacts in the output or a clear drop below the performance of heavy competing models on the same images.

Figures

Figures reproduced from arXiv: 2604.10359 by Alexandru Brateanu, Codruta Ancuti, Cosmin Ancuti, Tingting Mu.

Figure 1
Figure 1. Figure 1: Qualitative and quantitative comparison between [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of Multinex Architecture. Fusion modules [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of SOTA light-weight approaches [ [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative ablation on Multinex Priors study (a). [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The example image used by Sec. B. where W ∈ R D×C is computed by W = (Z ⊤Z + λID) −1Z ⊤Tc, (21) with ID being the identity matrix of size D and λ > 0 a small ridge regularization parameter. Using reshape(·) to restore a flattened signal to spatial dimensions H ×W ×C, the final reconstructed target is given by Tˆ (X) = reshape(ZW + µT). (22) Feature Visualization. By setting T as the low-light im￾age input,… view at source ↗
Figure 6
Figure 6. Figure 6: LRA visualization of chrominance candidates [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Similarity heatmaps between all candidate grouped [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: LRA visualization of combinations of the chosen feature maps for the reflectance guidance stack [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Top Row: Candidate illumination feature descriptors {YRec.709, Ymean, YYCgCo, Yvmax, Ylightness, YL2 }. Middle Row: ∆E(c) maps of the same candidate descriptors, where c ∈ {1, 2, . . . , 6}. Bottom row: ∆G(c) maps of the same candidate descriptors [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Similarity heatmap between all candidate individual [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: MSEF module architecture. puted by a two-layer excitation bottleneck, i.e., w = σtanhW2σReLU(W1GAP ◦ LN(X)) , (36) where the linear projection matrices W1 ∈ R d×C and W2 ∈ R C×d (d < C) form a feature compression￾expansion pair. Each adaptive weight wi is then used to recalibrate the normalized features from the corresponding channel, denoted as LNi(X), by multiplication, i.e., Zi = wiLNi(X), for i = 1, … view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparison on reference dataset LOL-v1 [ [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Qualitative comparison on reference dataset LOL-v1 [ [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Qualitative comparison on reference dataset LOL-v1 [ [PITH_FULL_IMAGE:figures/full_fig_p021_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Qualitative comparison on reference dataset LOL-v2-real [ [PITH_FULL_IMAGE:figures/full_fig_p022_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Qualitative comparison on no-reference datasets MEF [ [PITH_FULL_IMAGE:figures/full_fig_p023_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Additional results on no-reference dataset MEF [ [PITH_FULL_IMAGE:figures/full_fig_p024_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Additional results on no-reference dataset LIME [ [PITH_FULL_IMAGE:figures/full_fig_p025_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Additional results on no-reference dataset DICM [ [PITH_FULL_IMAGE:figures/full_fig_p026_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Additional results on no-reference dataset NPE [ [PITH_FULL_IMAGE:figures/full_fig_p027_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: A few challenging cases from DICM and MEF datasets. [PITH_FULL_IMAGE:figures/full_fig_p028_21.png] view at source ↗
read the original 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes Multinex, an ultra-lightweight Retinex-based framework for low-light image enhancement. It decomposes input images into illumination and color prior stacks derived from multiple distinct analytic representations, then employs lightweight neural networks to fuse these priors into luminance and reflectance adjustments within a residual formulation. Two variants are presented (45K parameters and 0.7K parameters), with the claim that extensive benchmarks demonstrate significant outperformance over lightweight SOTA models and performance comparable to much heavier models.

Significance. If the reported benchmark results hold under rigorous scrutiny, the work offers a meaningful advance for practical LLIE on edge devices by combining analytic priors with minimal neural fusion, addressing both computational cost and artifact issues from single-color-space methods. The parameter counts are consistent with the lightweight claim and the structured multi-prior approach could generalize if the fusion mechanism proves robust.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (Experiments): the central claim that 'all lightweight variants significantly outperform their corresponding lightweight SOTA models, and reach comparable performance to heavy models' is unsupported by any quantitative metrics, dataset names, ablation tables, or error analysis in the provided text; without these the performance assertion cannot be evaluated and is load-bearing for the paper's contribution.
  2. [§3] §3 (Method): the fusion of multiple fine-grained analytic priors into luminance/reflectance adjustments is described at a high level but lacks explicit equations showing how the residual corrections are computed from the prior stacks; this makes it impossible to verify whether the formulation is parameter-free or reduces to fitted quantities as asserted.
minor comments (2)
  1. [§2] §2 (Related Work): add explicit citations and parameter counts for the 'lightweight SOTA' baselines being compared against to allow direct replication of the claimed improvements.
  2. [Figure 1 and §3.2] Figure 1 and §3.2: the diagram of the prior-stack decomposition would benefit from clearer labeling of the analytic operations (e.g., which color spaces or filters produce each stack) to improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address the major comments point by point below and will update the manuscript to strengthen clarity and evidentiary support for the claims.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Experiments): the central claim that 'all lightweight variants significantly outperform their corresponding lightweight SOTA models, and reach comparable performance to heavy models' is unsupported by any quantitative metrics, dataset names, ablation tables, or error analysis in the provided text; without these the performance assertion cannot be evaluated and is load-bearing for the paper's contribution.

    Authors: We agree that the abstract states the performance claim at a summary level. Section 4 of the submitted manuscript already contains the supporting quantitative results: PSNR/SSIM/LPIPS tables on LOLv1, LOLv2, and MIT-Adobe FiveK, direct comparisons against lightweight baselines (e.g., Zero-DCE, EnlightenGAN variants) and heavy models (e.g., RetinexNet, KinD), plus ablation tables on prior-stack combinations. To make the claim immediately verifiable without requiring the reader to reach §4, we will revise the abstract to report the key numeric improvements (e.g., average PSNR gains) and will add a concise summary table or error-analysis paragraph at the start of §4. These changes will be present in the next revision. revision: yes

  2. Referee: [§3] §3 (Method): the fusion of multiple fine-grained analytic priors into luminance/reflectance adjustments is described at a high level but lacks explicit equations showing how the residual corrections are computed from the prior stacks; this makes it impossible to verify whether the formulation is parameter-free or reduces to fitted quantities as asserted.

    Authors: We appreciate the request for mathematical precision. The current §3 describes the multi-prior decomposition and residual fusion conceptually to highlight the analytic-plus-learned structure. In the revised manuscript we will insert explicit equations: let P_illum = [P1, P2, …, Pk] be the illumination prior stack from distinct analytic operators; the luminance residual is then ΔL = f_θ(P_illum) where f_θ is the lightweight CNN (45 k or 0.7 k parameters). An analogous equation holds for the reflectance residual ΔR = g_φ(P_color). The priors themselves remain closed-form and parameter-free; only the fusion networks are learned. These equations will be added to §3 together with a short derivation showing the overall parameter count. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper describes an empirical architecture: multi-prior Retinex residual decomposition into analytic illumination/color stacks, fused via lightweight neural operations (45K/0.7K params). Performance claims rest on external benchmarks against SOTA models rather than any internal derivation that reduces to fitted parameters or self-citations by construction. No load-bearing step equates a prediction to its own inputs; the framework is presented as a practical fusion of known Retinex priors with neural enhancement, self-contained against external validation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the Retinex decomposition being a useful prior for low-light images and on the learnable fusion of multiple analytic representations being sufficient to correct exposure without introducing new artifacts.

axioms (1)
  • domain assumption Retinex decomposition into illumination and reflectance (or color) components is a valid and useful model for low-light image correction.
    The entire framework is built on this classic image formation assumption.

pith-pipeline@v0.9.0 · 5493 in / 1121 out tokens · 38142 ms · 2026-05-10T15:29:04.786072+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

55 extracted references · 55 canonical work pages

  1. [1]

    Tech- nical report, International Telecommunication Union, 2019

    Itu-r bt.709: Parameter values for the hdtv standards. Tech- nical report, International Telecommunication Union, 2019

    work page 2019

  2. [2]

    Lyt-net: Lightweight yuv transformer-based network for low-light image enhance- ment.IEEE Signal Processing Letters, pages 1–5, 2025

    Alexandru Brateanu, Raul Balmez, Adrian Avram, Ciprian Orhei, and Cosmin Ancuti. Lyt-net: Lightweight yuv transformer-based network for low-light image enhance- ment.IEEE Signal Processing Letters, pages 1–5, 2025

    work page 2025

  3. [3]

    Enhancing low-light images with kolmogorov–arnold networks in transformer attention

    Alexandru Brateanu, Raul Balmez, Ciprian Orhei, Cosmin Ancuti, and Codruta Ancuti. Enhancing low-light images with kolmogorov–arnold networks in transformer attention. Sensors, 25(2):327, 2025

    work page 2025

  4. [4]

    Retinexformer: One-stage retinex- based transformer for low-light image enhancement

    Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Tim- ofte, and Yulun Zhang. Retinexformer: One-stage retinex- based transformer for low-light image enhancement. In IEEE/CVF International Conference on Computer Vision, pages 12504–12513, 2023

    work page 2023

  5. [5]

    You only need 90k parameters to adapt light: A lightweight trans- former for image enhancement and exposure correction

    Ziteng Cui, Kunchang Li, Lin Gu, Shenghan Su, Peng Gao, ZhengKai Jiang, Yu Qiao, and Tatsuya Harada. You only need 90k parameters to adapt light: A lightweight trans- former for image enhancement and exposure correction. In British Machine Vision Conference, 2022

    work page 2022

  6. [6]

    Emv toward highly efficient semantic-guided machine vi- sion for low-light object detection

    Xin Feng, Junxian Zeng, Siping Wang, and Zhenwei He. Emv toward highly efficient semantic-guided machine vi- sion for low-light object detection. InBritish Machine Vision Conference, 2024

    work page 2024

  7. [7]

    Learning a simple low-light im- age enhancer from paired low-light instances

    Zhenqi Fu, Yan Yang, Xiaotong Tu, Yue Huang, Xinghao Ding, and Kai-Kuang Ma. Learning a simple low-light im- age enhancer from paired low-light instances. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22252–22261, 2023

    work page 2023

  8. [8]

    Zero-reference deep curve estimation for low-light image enhancement

    Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1780–1789, 2020

    work page 2020

  9. [9]

    Low-light image enhancement via breaking down the darkness.International Journal of Computer Vision, 131(1):48–66, 2023

    Xiaojie Guo and Qiming Hu. Low-light image enhancement via breaking down the darkness.International Journal of Computer Vision, 131(1):48–66, 2023

    work page 2023

  10. [10]

    Lime: Low-light im- age enhancement via illumination map estimation.IEEE Transactions on Image Processing, 26(2):982–993, 2016

    Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light im- age enhancement via illumination map estimation.IEEE Transactions on Image Processing, 26(2):982–993, 2016

    work page 2016

  11. [11]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. InAdvances in Neural Informa- tion Processing Systems, pages 6840–6851, 2020

    work page 2020

  12. [12]

    Global structure-aware diffusion pro- cess for low-light image enhancement

    Jinhui Hou, Zhiyu Zhu, Junhui Hou, Hui Liu, Huanqiang Zeng, and Hui Yuan. Global structure-aware diffusion pro- cess for low-light image enhancement. InAdvances in Neu- ral Information Processing Systems, 2024

    work page 2024

  13. [13]

    Squeeze-and-excitation net- works

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation net- works. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7132–7141, 2018

    work page 2018

  14. [14]

    En- lightengan: Deep light enhancement without paired super- vision.IEEE Transactions on Image Processing, 30:2340– 2349, 2021

    Yifan Jiang, Wenhan Yang, Jiayi Fu, Haowen Ren, Zihan Wang, Yingyong Si, Jiaying Liu, and Zhengguo Zhang. En- lightengan: Deep light enhancement without paired super- vision.IEEE Transactions on Image Processing, 30:2340– 2349, 2021

    work page 2021

  15. [15]

    Joblove and Robert Greenberg

    David A. Joblove and Robert Greenberg. Color spaces for computer graphics. InACM SIGGRAPH, pages 20–25, 1978

    work page 1978

  16. [16]

    Jobson, Zia-ur Rahman, and Glenn A

    Daniel J. Jobson, Zia-ur Rahman, and Glenn A. Woodell. A multiscale retinex for bridging the gap between color images and the human observation of scenes.IEEE Transactions on Image Processing, 6(7):965–976, 1997

    work page 1997

  17. [17]

    Jobson, Zia-ur Rahman, and Glenn A

    Daniel J. Jobson, Zia-ur Rahman, and Glenn A. Woodell. Properties and performance of a center/surround retinex. IEEE Transactions on Image Processing, 6(3):451–462, 1997

    work page 1997

  18. [18]

    Perceptual losses for real-time style transfer and super-resolution

    Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, 2016

    work page 2016

  19. [19]

    Diederik P. Kingma. Adam: A method for stochastic opti- mization.arXiv preprint, 2014

    work page 2014

  20. [20]

    Contrast en- hancement based on layered difference representation of 2d histograms.IEEE Transactions on Image Processing, 22 (12):5372–5384, 2013

    Chulwoo Lee, Chul Lee, and Chang-Su Kim. Contrast en- hancement based on layered difference representation of 2d histograms.IEEE Transactions on Image Processing, 22 (12):5372–5384, 2013

    work page 2013

  21. [21]

    Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement

    Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongx- uan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021

    work page 2021

  22. [22]

    Getting to know low- light images with the exclusively dark dataset.Computer Vision and Image Understanding, 178:30–42, 2019

    Yuen Peng Loh and Chee Seng Chan. Getting to know low- light images with the exclusively dark dataset.Computer Vision and Image Understanding, 178:30–42, 2019

    work page 2019

  23. [23]

    Ll- net: A deep autoencoder approach to natural low-light image enhancement.Pattern Recognition, 61:650–662, 2017

    Kin Gwn Lore, Adedotun Akintayo, and Soumik Sarkar. Ll- net: A deep autoencoder approach to natural low-light image enhancement.Pattern Recognition, 61:650–662, 2017

    work page 2017

  24. [24]

    Sgdr: Stochastic gradi- ent descent with warm restarts.International Conference on Learning Representations, 2017

    Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradi- ent descent with warm restarts.International Conference on Learning Representations, 2017

    work page 2017

  25. [25]

    Mbllen: Low-light image/video enhancement using cnns

    Feifan Lv, Feng Lu, Jianhua Wu, and Chongsoon Lim. Mbllen: Low-light image/video enhancement using cnns. In British Machine Vision Conference, 2018

    work page 2018

  26. [26]

    Perceptual quality assessment for multi-exposure image fusion.IEEE Transac- tions on Image Processing, 24(11):3345–3356, 2015

    Kede Ma, Kai Zeng, and Zhou Wang. Perceptual quality assessment for multi-exposure image fusion.IEEE Transac- tions on Image Processing, 24(11):3345–3356, 2015

    work page 2015

  27. [27]

    Toward fast, flexible, and robust low-light image enhancement

    Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongx- uan Luo. Toward fast, flexible, and robust low-light image enhancement. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    work page 2022

  28. [28]

    No-reference image quality assessment in the spa- tial domain.IEEE Transactions on Image Processing, 21 (12):4695–4708, 2012

    Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spa- tial domain.IEEE Transactions on Image Processing, 21 (12):4695–4708, 2012

    work page 2012

  29. [29]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Mak- ing a “completely blind” image quality analyzer.IEEE Sig- nal Processing Letters, 20(3):209–212, 2012

    work page 2012

  30. [30]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019

    work page 2019

  31. [31]

    Denet: Detection-driven enhancement network for ob- ject detection under adverse weather conditions

    Qingpao Qin, Kan Chang, Mengyuan Huang, and Guiqing Li. Denet: Detection-driven enhancement network for ob- ject detection under adverse weather conditions. InAsian Conference on Computer Vision, pages 2813–2829, 2022

    work page 2022

  32. [32]

    Yolov3: An incremental improvement.arXiv preprint, 2018

    Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement.arXiv preprint, 2018

    work page 2018

  33. [33]

    Saurabh Saini and P. J. Narayanan. Specularity factoriza- tion for low-light enhancement. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    work page 2024

  34. [34]

    Zero-ig: Zero-shot illumination-guided joint denoising and adaptive enhancement for low-light images

    Yiqi Shi, Duo Liu, Liguo Zhang, Ye Tian, Xuezhi Xia, and Xiaojing Fu. Zero-ig: Zero-shot illumination-guided joint denoising and adaptive enhancement for low-light images. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3015–3024, 2024

    work page 2024

  35. [35]

    Underexposed photo enhance- ment using deep illumination estimation

    Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhance- ment using deep illumination estimation. InIEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2019

    work page 2019

  36. [36]

    Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment

    Ruixing Wang, Xiaoyu Xu, Chi-Wing Fu, Jie Lu, Baochang Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In IEEE/CVF International Conference on Computer Vision, 2021

    work page 2021

  37. [37]

    Nat- uralness preserved enhancement algorithm for non-uniform illumination images.IEEE Transactions on Image Process- ing, 22(9):3538–3548, 2013

    Shuhang Wang, Jin Zheng, Hai-Miao Hu, and Bo Li. Nat- uralness preserved enhancement algorithm for non-uniform illumination images.IEEE Transactions on Image Process- ing, 22(9):3538–3548, 2013

    work page 2013

  38. [38]

    Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method

    Tingyang Wang, Kai Zhang, Tao Shen, Wenzi Luo, Bj”orn Stenger, and Tao Lu. Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. InAAAI Conference on Artificial Intelligence, 2023

    work page 2023

  39. [39]

    Zero-reference low-light enhancement via physical quadru- ple priors

    Wenjing Wang, Huan Yang, Jianlong Fu, and Jiaying Liu. Zero-reference low-light enhancement via physical quadru- ple priors. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    work page 2024

  40. [40]

    Yufei Wang, Renjie Wan, Wenhan Yang, Haoliang Li, Chau Lap-pui, and Alex C. Kot. Low-light image enhancement with normalizing flow. InAAAI Conference on Artificial In- telligence, 2022

    work page 2022

  41. [41]

    Simoncelli, and Alan C

    Zhou Wang, Eero P. Simoncelli, and Alan C. Bovik. Multi- scale structural similarity for image quality assessment. In IEEE Asilomar Conference on Signals, Systems, and Com- puters, 2003

    work page 2003

  42. [42]

    Deep retinex decomposition for low-light enhancement

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement. In British Machine Vision Conference, 2018

    work page 2018

  43. [43]

    Cbam: Convolutional block attention module

    Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In European Conference on Computer Vision, 2018

    work page 2018

  44. [44]

    Snr-aware low-light image enhancement

    Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. InIEEE/CVF Con- ference on Computer Vision and Pattern Recognition, 2022

    work page 2022

  45. [45]

    Hvi: A new color space for low-light image en- hancement

    Qingsen Yan, Yixu Feng, Cheng Zhang, Guansong Pang, Kangbiao Shi, Peng Wu, Wei Dong, Jinqiu Sun, and Yan- ning Zhang. Hvi: A new color space for low-light image en- hancement. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025

    work page 2025

  46. [46]

    Sparse gradient regularized deep retinex network for robust low-light image enhancement.IEEE Transactions on Image Processing, 30:2072–2086, 2021

    Wenhan Yang, Shiqi Wang, Yuming Fang, Yue Wang, and Ji- aying Liu. Sparse gradient regularized deep retinex network for robust low-light image enhancement.IEEE Transactions on Image Processing, 30:2072–2086, 2021

    work page 2072

  47. [47]

    Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model

    Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. InIEEE/CVF Interna- tional Conference on Computer Vision, pages 12302–12311, 2023

    work page 2023

  48. [48]

    Pe-yolo: Pyramid enhancement network for dark object detection

    Xiangchen Yin, Zhenda Yu, Zetao Fei, Wenjun Lv, and Xin Gao. Pe-yolo: Pyramid enhancement network for dark object detection. InInternational Conference on Artificial Neural Networks, pages 163–174, 2023

    work page 2023

  49. [49]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    work page 2022

  50. [50]

    Kindling the darkness: A practical low-light image enhancer

    Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. InACM International Conference on Multimedia, 2019

    work page 2019

  51. [51]

    Better than reference in low-light image enhance- ment: Conditional re-enhancement network.IEEE Transac- tions on Image Processing, 31:759–772, 2021

    Yu Zhang, Xiaoguang Di, Bin Zhang, Ruihang Ji, and Chun- hui Wang. Better than reference in low-light image enhance- ment: Conditional re-enhancement network.IEEE Transac- tions on Image Processing, 31:759–772, 2021

    work page 2021

  52. [52]

    Glare: Low light image en- hancement via generative latent feature based codebook re- trieval

    Huan Zhou, Wei Dong, Xintao Liu, Shiyu Liu, Xu Min, Guocheng Zhai, and Jing Chen. Glare: Low light image en- hancement via generative latent feature based codebook re- trieval. InEuropean Conference on Computer Vision, 2024

    work page 2024

  53. [53]

    A low-light image enhancement method based on hsv space.Imaging Science Journal, pages 1–14, 2023

    Libing Zhou, Xiaojing Chen, Baisong Ye, Xueli Jiang, Sheng Zou, Liang Ji, Zhengqian Yu, Jianjian Wei, Yexin Zhao, and Tianyu Wang. A low-light image enhancement method based on hsv space.Imaging Science Journal, pages 1–14, 2023. Multinex: Lightweight Low-light Image Enhancement via Multi-prior Retinex Supplementary Material Content Summary:Section A expl...

    work page 2023

  54. [54]

    6 visualizes how well the chrominance candidates [Cb,C r],[U,V],[O 1,O 2],[r,g], andScan reconstruct the RGB content of the low-light imageI

    Fig. 6 visualizes how well the chrominance candidates [Cb,C r],[U,V],[O 1,O 2],[r,g], andScan reconstruct the RGB content of the low-light imageI. To quantify the redundancy among the candidate chroma priors, we com- pute their pairwise Pearson correlation. Because the de- scriptors vary in channel depth, we first compute the pixel- wiseL 2 norm across th...

    work page

  55. [55]

    between” setting corresponds to the proposed design in Eq. (18). In the other two settings of “before

    that decays from2×10 −4 to1×10 −6. Multinex is trained from scratch for 150K iterations with a batch size of 8 and patch size of256×256, using the designated training splits of each dataset. C.4. Discussion on GT-Mean GT-Mean is a post-processing step used by some LLIE works [12, 40, 45, 52] when evaluating their approaches on small paired datasets, e.g.,...

    work page