pith. sign in

arxiv: 2512.08982 · v3 · submitted 2025-12-05 · 💻 cs.CV · cs.AI

Consist-Retinex: One-Step Noise-Emphasized Consistency Training Accelerates High-Quality Retinex Enhancement

Jian Xu , Wei Chen , Shigui Li , Delu Zeng , John Paisley , Qibin Zhao This is my paper

Pith reviewed 2026-05-17 00:35 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords low-light image enhancementRetinex decompositionconsistency modelsone-step inferenceimage restorationreflectance illumination separation
0
0 comments X

The pith

Consist-Retinex reaches stable one-step Retinex low-light enhancement by pairing trajectory consistency with ground-truth component alignment and noise-emphasized sampling.

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

The paper aims to solve the slow iterative sampling that limits generative Retinex methods for low-light enhancement. Direct use of consistency models fails because one-step inference occurs at the high-noise end where training provides little guidance and plain self-consistency does not pin down the correct reflectance and illumination targets. The proposed approach first runs a Retinex Transformer Decomposition Network to produce paired reflectance and illumination maps, then trains two conditional consistency models under a dual objective that adds explicit alignment to ground-truth components and uses adaptive sampling that places more points near the high-noise endpoint. Theoretical results on endpoint error bounds and anchoring propagation explain why the combination succeeds, and experiments confirm competitive or superior scores on paired and unpaired benchmarks with far fewer sampling steps and lower training cost.

Core claim

A Retinex Transformer Decomposition Network first extracts reflectance and illumination, after which two conditional consistency models are trained with a Retinex-aware dual objective that merges trajectory consistency and paired ground-truth alignment under adaptive noise-emphasized fixed-point sampling; this yields stable one-step inference whose correctness is supported by an endpoint error bound, an anchoring-propagation result, and a high-noise sample-allocation analysis.

What carries the argument

Retinex-aware dual objective that joins trajectory consistency to paired ground-truth component alignment, combined with adaptive noise-emphasized fixed-point sampling that concentrates training near the inference endpoint.

If this is right

  • Delivers the highest VE-LOL-L scores among compared methods under one-step inference.
  • Stays competitive with prior methods on the standard LOL paired benchmark.
  • Cuts the number of sampling steps and the consistency-stage training cost relative to standard consistency-model adaptations.

Where Pith is reading between the lines

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

  • The same decomposition-plus-dual-objective pattern could be tested on other factorized restoration problems such as shadow removal where multiple image components must remain consistent.
  • Endpoint-focused sampling schedules may shorten training for any consistency-model application whose inference point lies at high noise.
  • If the reported cost reduction holds on mobile hardware, the method could enable real-time low-light processing without dedicated accelerators.

Load-bearing premise

The dual objective of trajectory consistency plus ground-truth alignment, together with noise-emphasized sampling, produces stable one-step Retinex inference without the instability that normally appears when consistency models are applied directly to reflectance-illumination factorization.

What would settle it

If ablating the ground-truth alignment term from the dual objective causes one-step output quality on the VE-LOL-L benchmark to fall below the multi-step baselines reported in the paper, the claim that the full objective is required for stability would be falsified.

Figures

Figures reproduced from arXiv: 2512.08982 by Delu Zeng, Jian Xu, John Paisley, Qibin Zhao, Shigui Li, Wei Chen.

Figure 1
Figure 1. Figure 1: Illustration of consistency mapping. Green ODE trajec [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Consist-Retinex framework. Given a low-light image [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Task-driven sampling strategy comparison. Top: Task formulations differ fundamentally—unconditional genera￾tion learns the full data distribution from pure noise, while our conditional enhancement performs one-step mapping from con￾catenated inputs of pure noise (σmaxϵ) and degraded image (Il). Middle: Standard log-uniform sampling concentrates on the data manifold, optimal for unconditional synthesis wher… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison with the state-of-the-art low-light image enhancement methods on the LOL dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison with the state-of-the-art low-light image enhancement methods on the VE-LOL [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison with the state-of-the-art low-light image enhancement methods on the DICM and VV dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
read the original abstract

Retinex-based low-light image enhancement benefits from separating reflectance and illumination, yet recent generative approaches often rely on iterative sampling and are difficult to deploy under strict latency budgets. Consistency models offer a natural route to one-step restoration, but direct adaptation to Retinex-factorized enhancement is unstable: one-step inference is evaluated at the high-noise endpoint, whereas standard training schedules provide little supervision there, and temporal self-consistency alone does not determine the correct conditional target. We propose Consist-Retinex, which first uses a Retinex Transformer Decomposition Network (TDN) to obtain paired reflectance and illumination maps, then trains two conditional consistency models with a Retinex-aware dual objective and adaptive noise-emphasized fixed-point sampling. The dual objective combines trajectory consistency with paired ground-truth component alignment, while the sampling rule concentrates supervision near the inference endpoint without discarding full-range noise coverage. We further provide an endpoint error bound, an anchoring-propagation result, and a high-noise sample-allocation analysis that explain why endpoint supervision and temporal consistency are complementary for one-step Retinex enhancement. Experiments on paired and unpaired low-light benchmarks show that Consist-Retinex obtains the best VE-LOL-L scores among the compared methods under one-step inference and remains competitive on LOL, with substantially reduced sampling and consistency-stage training cost in the reported setup.

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 Consist-Retinex for one-step Retinex-based low-light enhancement. It first applies a Retinex Transformer Decomposition Network (TDN) to obtain reflectance and illumination maps, then trains conditional consistency models using a Retinex-aware dual objective (trajectory consistency plus paired ground-truth component alignment) and adaptive noise-emphasized fixed-point sampling. Theoretical results include an endpoint error bound, anchoring-propagation analysis, and high-noise sample allocation. Experiments claim best VE-LOL-L scores among compared methods under one-step inference, competitive LOL results, and substantially lower sampling/training cost.

Significance. If the empirical claims hold after addressing the noted gaps, the work offers a practical route to low-latency Retinex enhancement by stabilizing one-step consistency-model inference on decomposed components. The dual objective and sampling schedule, together with the provided analytic bounds, constitute a concrete technical contribution that could reduce deployment barriers for generative enhancement models. Credit is due for the explicit theoretical analysis of endpoint supervision complementarity and the reported cost reductions.

major comments (2)
  1. Experiments: the claim of obtaining the best VE-LOL-L scores under one-step inference is presented without specification of exact baselines, statistical significance, error bars, or data-split details, leaving the strength of support for the central performance claim moderate.
  2. Method (Retinex-aware dual objective and adaptive sampling): the stability of one-step inference at the high-noise endpoint is attributed to the paired ground-truth component alignment term, yet no ablation isolates its contribution versus the sampling schedule or TDN decomposition; removing this term and re-running one-step inference would directly test whether instability (mode collapse or color-shift artifacts) reappears.
minor comments (2)
  1. The precise formulation of the adaptive noise-emphasized fixed-point sampling rule and its hyper-parameters should be stated explicitly (e.g., as an algorithm box or equation) to support reproducibility.
  2. Notation for the two conditional consistency models (reflectance and illumination) and the weighting between the two terms of the dual objective could be clarified in the method section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed review and the recognition of our technical contributions in one-step Retinex enhancement. We address each major comment below and outline the revisions planned for the manuscript.

read point-by-point responses

  1. Referee: Experiments: the claim of obtaining the best VE-LOL-L scores under one-step inference is presented without specification of exact baselines, statistical significance, error bars, or data-split details, leaving the strength of support for the central performance claim moderate.

    Authors: We agree with the referee that additional details on the experimental setup and results would strengthen the presentation of our empirical findings. In the revised manuscript, we will expand the experimental section to specify all baseline methods and their implementations, include statistical significance tests or report standard deviations across multiple runs, add error bars to performance plots, and provide explicit information on the data splits for the VE-LOL-L and LOL benchmarks. This will better support the claim of achieving the best scores under one-step inference. revision: yes

  2. Referee: Method (Retinex-aware dual objective and adaptive sampling): the stability of one-step inference at the high-noise endpoint is attributed to the paired ground-truth component alignment term, yet no ablation isolates its contribution versus the sampling schedule or TDN decomposition; removing this term and re-running one-step inference would directly test whether instability (mode collapse or color-shift artifacts) reappears.

    Authors: We thank the referee for this insightful suggestion. The manuscript currently presents ablations on the overall dual objective and sampling strategy, but we concur that a more isolated ablation—specifically removing the paired ground-truth component alignment while retaining the trajectory consistency, adaptive noise-emphasized sampling, and TDN decomposition—would provide clearer evidence of its role in preventing instability during one-step inference. We will conduct this experiment in the revision, re-evaluate one-step performance without the alignment term, and document any observed mode collapse or color-shift artifacts to validate our analysis. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation introduces independent components and analytic bounds

full rationale

The paper defines a new Retinex Transformer Decomposition Network (TDN), a Retinex-aware dual objective (trajectory consistency plus paired ground-truth component alignment), and adaptive noise-emphasized fixed-point sampling. It then supplies fresh endpoint error bounds, anchoring-propagation results, and high-noise allocation analysis to justify why these elements are complementary for one-step inference. None of these steps reduce by construction to fitted parameters, self-referential definitions, or prior self-citations; the analytic results are derived from the proposed loss and sampling rule rather than presupposing the final performance. Experimental validation on VE-LOL-L and LOL benchmarks supplies independent empirical content. The chain therefore remains self-contained with no load-bearing reduction to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that Retinex decomposition is accurate and that the new dual objective plus endpoint emphasis supplies the missing supervision for one-step consistency training; no explicit free parameters or invented entities are named in the abstract.

axioms (1)
  • domain assumption Retinex decomposition into reflectance and illumination maps is a valid and useful factorization for low-light enhancement
    Invoked by the Retinex Transformer Decomposition Network (TDN) and the subsequent conditional consistency training.

pith-pipeline@v0.9.0 · 5555 in / 1285 out tokens · 31182 ms · 2026-05-17T00:35:17.064840+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

51 extracted references · 51 canonical work pages · 2 internal anchors

  1. [1]

    The perception- distortion tradeoff

    Yochai Blau and Tomer Michaeli. The perception- distortion tradeoff. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 6228–6237, 2018. 6

    work page 2018

  2. [2]

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

    Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retinex-based transformer for low-light im- age enhancement. InProceedings of the IEEE/CVF international conference on computer vision, pages 12504–12513, 2023. 2, 5, 6

    work page 2023

  3. [3]

    Anlightendiff: Anchoring diffusion probabilistic model on low light image enhancement

    Cheuk-Yiu Chan, Wan-Chi Siu, Yuk-Hee Chan, and H Anthony Chan. Anlightendiff: Anchoring diffusion probabilistic model on low light image enhancement. IEEE Transactions on Image Processing, 2024. 5

    work page 2024

  4. [4]

    Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780– 8794, 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780– 8794, 2021. 4

    work page 2021

  5. [5]

    Generative diffusion prior for unified image restoration and enhancement

    Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9935–9946, 2023. 5, 6, 7

    work page 2023

  6. [6]

    A weighted variational model for simultaneous reflectance and illumination estimation

    Xueyang Fu, Delu Zeng, Yue Huang, Xiao-Ping Zhang, and Xinghao Ding. A weighted variational model for simultaneous reflectance and illumination estimation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2782– 2790, 2016. 2

    work page 2016

  7. [7]

    Zero-reference deep curve estimation for low-light im- age 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 im- age enhancement. InProceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 1780–1789, 2020. 5, 6

    work page 2020

  8. [8]

    Lime: Low-light image enhancement via illumination map estimation

    Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 26(2):982– 993, 2016. 2, 5, 6

    work page 2016

  9. [9]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in Neural Infor- mation Processing Systems, 30, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Un- terthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in Neural Infor- mation Processing Systems, 30, 2017. 6

    work page 2017

  10. [10]

    Denoising diffusion probabilistic models.Advances in Neural In- formation Processing Systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in Neural In- formation Processing Systems, 33:6840–6851, 2020. 2

    work page 2020

  11. [11]

    Global structure-aware dif- fusion process for low-light image enhancement.Ad- vances in Neural Information Processing Systems, 36: 79734–79747, 2023

    Jinhui Hou, Zhiyu Zhu, Junhui Hou, Hui Liu, Huan- qiang Zeng, and Hui Yuan. Global structure-aware dif- fusion process for low-light image enhancement.Ad- vances in Neural Information Processing Systems, 36: 79734–79747, 2023. 2, 5, 6

    work page 2023

  12. [12]

    Low-light image enhancement with wavelet-based diffusion models.ACM Transactions on Graphics (TOG), 42(6):1–14, 2023

    Hai Jiang, Ao Luo, Haoqiang Fan, Songchen Han, and Shuaicheng Liu. Low-light image enhancement with wavelet-based diffusion models.ACM Transactions on Graphics (TOG), 42(6):1–14, 2023. 6, 7

    work page 2023

  13. [13]

    Lightendiffusion: Unsupervised low- light image enhancement with latent-retinex diffusion models

    Hai Jiang, Ao Luo, Xiaohong Liu, Songchen Han, and Shuaicheng Liu. Lightendiffusion: Unsupervised low- light image enhancement with latent-retinex diffusion models. InEuropean Conference on Computer Vision, pages 161–179. Springer, 2024. 6, 7

    work page 2024

  14. [14]

    Enlightengan: Deep light enhance- ment without paired supervision.IEEE transactions on image processing, 30:2340–2349, 2021

    Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhance- ment without paired supervision.IEEE transactions on image processing, 30:2340–2349, 2021. 5, 6

    work page 2021

  15. [15]

    A multiscale retinex for bridging the gap between color images and the human observation of scenes

    Daniel J Jobson, Zia-ur Rahman, and Glenn A Wood- ell. 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. 2

    work page 1997

  16. [16]

    Properties and performance of a center/surround retinex.IEEE transactions on image processing, 6(3): 451–462, 1997

    Daniel J Jobson, Zia-ur Rahman, and Glenn A Wood- ell. Properties and performance of a center/surround retinex.IEEE transactions on image processing, 6(3): 451–462, 1997. 2

    work page 1997

  17. [17]

    Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35:26565–26577, 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35:26565–26577, 2022. 2

    work page 2022

  18. [18]

    The retinex theory of color vision

    Edwin H Land. The retinex theory of color vision. Scientific american, 237(6):108–129, 1977. 1, 2

    work page 1977

  19. [19]

    EVOD- iff: Entropy-aware variance optimized diffusion infer- ence

    Shigui Li, Wei Chen, and Delu Zeng. EVOD- iff: Entropy-aware variance optimized diffusion infer- ence. InThe Thirty-ninth Annual Conference on Neu- ral Information Processing Systems, 2025. 1

    work page 2025

  20. [20]

    Benchmarking low-light image enhancement and beyond.International Journal of Computer Vision, 129:1153–1184, 2021

    Jiaying Liu, Dejia Xu, Wenhan Yang, Minhao Fan, and Haofeng Huang. Benchmarking low-light image enhancement and beyond.International Journal of Computer Vision, 129:1153–1184, 2021. 5, 8

    work page 2021

  21. [21]

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

    Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with co- operative prior architecture search for low-light image enhancement. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 10561–10570, 2021. 5

    work page 2021

  22. [22]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decou- pled weight decay regularization.arXiv preprint arXiv:1711.05101, 2017. 5

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  23. [23]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in neural information pro- cessing systems, 35:5775–5787, 2022

    Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps.Advances in neural information pro- cessing systems, 35:5775–5787, 2022. 1

    work page 2022

  24. [24]

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

    Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 5637–5646, 2022. 5

    work page 2022

  25. [25]

    completely blind

    Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209–212, 2012. 6

    work page 2012

  26. [26]

    Pytorch: An imperative style, high- performance deep learning library.Advances in Neu- ral Information Processing Systems, 32, 2019

    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.Advances in Neu- ral Information Processing Systems, 32, 2019. 5

    work page 2019

  27. [27]

    Joint enhancement and denoising method via sequential decomposition

    Xutong Ren, Mading Li, Wen-Huang Cheng, and Jiay- ing Liu. Joint enhancement and denoising method via sequential decomposition. In2018 IEEE international symposium on circuits and systems (ISCAS), pages 1–

    work page

  28. [28]

    Progressive distilla- tion for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distilla- tion for fast sampling of diffusion models. InInterna- tional Conference on Learning Representations, 2022. 1

    work page 2022

  29. [29]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. InInternational Conference on Learning Representations, 2021. 1

    work page 2021

  30. [30]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. InInternational Con- ference on Learning Representations, 2021. 2

    work page 2021

  31. [31]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. InInternational Con- ference on Machine Learning, pages 32211–32252. PMLR, 2023. 1, 2, 3, 8

    work page 2023

  32. [32]

    Lightening network for low-light im- age enhancement.IEEE Transactions on Image Pro- cessing, 29:7984–7996, 2020

    Li-Wen Wang, Zhi-Song Liu, Wan-Chi Siu, and Daniel PK Lun. Lightening network for low-light im- age enhancement.IEEE Transactions on Image Pro- cessing, 29:7984–7996, 2020. 5

    work page 2020

  33. [33]

    Naturalness preserved enhancement algorithm for non-uniform illumination images.IEEE transactions on image processing, 22(9):3538–3548, 2013

    Shuhang Wang, Jin Zheng, Hai-Miao Hu, and Bo Li. Naturalness preserved enhancement algorithm for non-uniform illumination images.IEEE transactions on image processing, 22(9):3538–3548, 2013. 5

    work page 2013

  34. [34]

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

    Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra-high- definition low-light image enhancement: A bench- mark and transformer-based method. InProceed- ings of the AAAI Conference on Artificial Intelligence, pages 2654–2662, 2023. 5, 6

    work page 2023

  35. [35]

    Zero-reference low-light enhancement via physi- cal quadruple priors

    Wenjing Wang, Huan Yang, Jianlong Fu, and Jiaying Liu. Zero-reference low-light enhancement via physi- cal quadruple priors. InProceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 26057–26066, 2024. 5

    work page 2024

  36. [36]

    Image quality assessment: from error visibility to structural similarity.IEEE transac- tions on image processing, 13(4):600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE transac- tions on image processing, 13(4):600–612, 2004. 6

    work page 2004

  37. [37]

    Deep Retinex Decomposition for Low-Light Enhancement

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiay- ing Liu. Deep retinex decomposition for low-light en- hancement.arXiv preprint arXiv:1808.04560, 2018. 2, 5, 6

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  38. [38]

    Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement

    Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages 5901–5910, 2022. 2, 5, 6

    work page 2022

  39. [39]

    Cretinex: A progressive color-shift aware retinex model for low-light image enhancement.International Journal of Computer Vision, 132(9):3610–3632, 2024

    Han Xu, Hao Zhang, Xunpeng Yi, and Jiayi Ma. Cretinex: A progressive color-shift aware retinex model for low-light image enhancement.International Journal of Computer Vision, 132(9):3610–3632, 2024. 5, 6

    work page 2024

  40. [40]

    Implicit neural representation for cooperative low-light image enhancement

    Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. InPro- ceedings of the IEEE/CVF international conference on computer vision, pages 12918–12927, 2023. 5

    work page 2023

  41. [41]

    From fidelity to perceptual quality: A semi-supervised approach for low-light image en- hancement

    Wenhan Yang, Shiqi Wang, Yuming Fang, Yue Wang, and Jiaying Liu. From fidelity to perceptual quality: A semi-supervised approach for low-light image en- hancement. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 3063–3072, 2020. 5

    work page 2020

  42. [42]

    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. In Proceedings of the IEEE/CVF international confer- ence on computer vision, pages 12302–12311, 2023. 1, 2, 3, 4, 5, 6, 7, 8

    work page 2023

  43. [43]

    Diff-retinex++: Retinex-driven rein- forced diffusion model for low-light image enhance- ment.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 2025

    Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex++: Retinex-driven rein- forced diffusion model for low-light image enhance- ment.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 2025. 2, 3, 5, 6, 7, 8

    work page 2025

  44. [44]

    Cle diffusion: Con- trollable light enhancement diffusion model

    Yuyang Yin, Dejia Xu, Chuangchuang Tan, Ping Liu, Yao Zhao, and Yunchao Wei. Cle diffusion: Con- trollable light enhancement diffusion model. InPro- ceedings of the 31st ACM International Conference on Multimedia, pages 8145–8156, 2023. 2, 5, 6, 7

    work page 2023

  45. [45]

    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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728–5739, 2022. 4, 5, 6

    work page 2022

  46. [46]

    The unreasonable ef- fectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable ef- fectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vi- sion and Pattern Recognition, pages 586–595, 2018. 6

    work page 2018

  47. [47]

    Kin- dling the darkness: A practical low-light image en- hancer

    Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kin- dling the darkness: A practical low-light image en- hancer. InProceedings of the ACM International Con- ference on Multimedia, pages 1632–1640, 2019. 2, 5, 6

    work page 2019

  48. [48]

    Beyond brightening low-light images

    Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision, 129:1013– 1037, 2021. 2, 5, 6

    work page 2021

  49. [49]

    Empow- ering low-light image enhancer through customized learnable priors

    Naishan Zheng, Man Zhou, Yanmeng Dong, Xiangyu Rui, Jie Huang, Chongyi Li, and Feng Zhao. Empow- ering low-light image enhancer through customized learnable priors. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12559–12569, 2023. 5

    work page 2023

  50. [50]

    Semantic-guided zero-shot learning for low-light image/video enhance- ment

    Shen Zheng and Gaurav Gupta. Semantic-guided zero-shot learning for low-light image/video enhance- ment. InProceedings of the IEEE/CVF Winter confer- ence on applications of computer vision, pages 581– 590, 2022. 5

    work page 2022

  51. [51]

    Pyramid diffusion models for low-light image enhancement

    D Zhou, Z Yang, and Y Yang. Pyramid diffusion models for low-light image enhancement. InInterna- tional Joint Conference on Artificial Intelligence. IJ- CAI, 2023. 5, 6, 7

    work page 2023