Pith. sign in

REVIEW 4 major objections 3 minor 49 references

Uncertainty-Aware Spatial Color Correlation for Low-Light Image Enhancement

T0 review · 4 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that low-light image enhancement is best treated as an uncertainty-entropy minimization problem, and that entropy-guided dual-domain denoising plus asymmetric causal color modeling yields state-of-the-art results with a 0.4

desk verdict A credible lightweight low-light enhancer with honest ablations, but the entropy-minimization theory in Eqs. 5–6 is asserted rather than derived, and the missing supplement is a real problem. read the letter →

arxiv 2508.04176 v1 pith:DEBQFUC5 submitted 2025-08-06 cs.CV

Jin Kuang , Dong Liu , Yukuang Zhang , Shengsheng Wang This is my paper
classification cs.CV
keywords low-lightimageenhancementuncertaintyentropydual-domaindenoisingfrequency-domainGaussianfilteringstatespacemodelcolorconsistencycausalmodelingrestoration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that low-light enhancement fails in extremely dark regions because feature uncertainty is ignored: noisy, high-entropy features dominate gradients and break color and detail consistency. It proposes U2CLLIE, which computes an entropy map over features and uses it to guide dual-domain (spatial and frequency) denoising, plus two asymmetric causal modules that enforce neighborhood and color consistency. The claimed payoff is a lightweight 0.48M-parameter network that reports the highest PSNR and SSIM on LOLv2-Real, LOLv2-Synthetic, and LSRW-Huawei, and the lowest NIQE on five unpaired datasets. The central theoretical step is the claim that entropy gating bounds how much noise the network can fit during training.

What carries the argument

The load-bearing quantity is the uncertainty entropy $H(x)=-\frac{1}{\log c}\sum_{c=1}^{C} p_c(x)\log p_c(x)$, and the central asserted identity is Eq. (6), $E_{\text{noise}}\le\gamma\,\|\partial F^i/\partial\theta\|$ with $\gamma\in(0,1)$. In the network, G2AF forms Gaussian masks in the Fourier domain with learnable radii, NeCo adds neighborhood offsets to the state-space recurrence $h_t=\bar{A}h_{t-1}+\bar{B}(x_t+\tau_p)$, and AsC selects top-$k$ neighboring feature patches to compute color-correlation weights. These mechanisms carry the entropy signal into both gradient weighting and color reconstruction.

What would settle it

On LOLv2-Real, record per-patch fitted noise and $\|\partial F^i/\partial\theta\|$ during training; if any high-entropy dark patch violates $E_{\text{noise}}\le\gamma\|\partial F^i/\partial\theta\|$ with $\gamma<1$, the bound is refuted. Also swap UaD for a fixed Gaussian low-pass filter at equal parameter count; if the reported PSNR/SSIM gains persist, entropy gating is not the active cause.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that treating low-light enhancement as entropy minimization is the reason U2CLLIE works. The UaD module quantifies per-feature entropy, uses a Gaussian-guided frequency filter (G2AF) to separate low- and high-frequency structure, and gates the denoised features by entropy before recombining them. The NeCo module adds a local causal offset to a state-space scan so neighborhood structure survives in dark areas, while AsC calibrates colors by comparing each feature to its nearest neighbors. The paper reports top scores on three paired benchmarks and five unpaired datasets, and its ablations attribute the gains to the joint action of entropy guidance and causal

Load-bearing premise

The paper's explanation for why entropy gating helps is Eq. (6), an asserted inequality, $E_{\text{noise}}\le\gamma\,\|\partial F^i/\partial\theta\|$ with $\gamma\in(0,1)$, that is never proved; if the inequality does not hold, the claimed mechanism dissolves into an unexplained combination of modules.

Editorial extensions

If this is right

  • A 0.48M-parameter enhancement network at 6.18 GFLOPs (256x256 input) could plausibly run on cameras and edge devices rather than requiring server-class GPUs.
  • Stable enhancement in extremely dark regions could make downstream tasks such as nighttime detection and driving perception more reliable, since the method suppresses artifact-prone high-entropy zones.
  • Lowest NIQE on five unpaired datasets suggests the model transfers to real photographs without ground-truth pairs, not just to paired benchmarks.
  • The asymmetric design—neighborhood constraints in the encoder, color calibration in the decoder—offers a cheap template for local consistency without global self-attention.
  • If the entropy bound is genuine, the same gating principle could be reused as a regularizer in any restoration model that must suppress noise in spatially varying SNR regions.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if entropy guidance is the true operative cause, the same gating should transfer to other spatially varying degradations such as haze, rain, or defocus, where confidence also varies across the image.
  • Editorial inference: Eq. (6) could be settled by a Lipschitz-continuity analysis of the feature map; until then, the reported gains might be explained by simpler mechanisms such as denoising regularization or the extra loss terms rather than by the entropy bound itself.
  • Editorial inference: the entropy map itself could serve as an interpretable confidence output, showing users where the enhancement is unreliable—an application the paper does not explore.
  • Editorial inference: replacing AsC's top-k neighbor selection with an explicit color-space transform would isolate how much of the gain comes from causal modeling versus color-space geometry.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. The paper proposes U2CLLIE, a lightweight low-light image enhancement network combining a luminance enhancement branch (LEN), an uncertainty-aware dual-domain denoising module (UaD) with Gaussian-guided frequency filtering (G2AF), and two asymmetric causal modules (NeCo, AsC). The central claim is that reformulating enhancement as entropy-guided uncertainty minimization suppresses noise and stabilizes features, yielding state-of-the-art PSNR/SSIM on LOLv2-Real, LOLv2-Synthetic, and LSRW-Huawei, plus strong perceptual quality on unpaired datasets. The manuscript supplies a U-shaped architecture, a seven-term loss, benchmark comparisons, and component ablations.

Significance. If the empirical claims hold, U2CLLIE is a useful lightweight baseline: it reports 0.48M parameters and competitive PSNR/SSIM on three reference datasets, and the unpaired-dataset NIQE results are consistently good. The paper is also honest in reporting some limitations, such as not using GT-Mean. However, the central theoretical contribution — that uncertainty entropy provably bounds noise fitting — rests on Eqs. (5)–(6), which are not valid derivations as written, and the manuscript repeatedly defers key formal material to a supplementary file that is not provided. The empirical SOTA claim is also overstated because Table 2 shows LPIPS on LOLv2-Real is substantially worse than FourLLIE. The architecture may still be useful, but the causal mechanism and the SOTA statement need to be reworked.

major comments (4)
  1. [Methodology, Eqs. (5)-(6)] The derivation of the central noise-fitting bound is not valid as written. Eq. (5) jumps to a proportionality for ∂L_recon/∂θ without specifying which term of the seven-term loss is differentiated or how the entropy map F^{i+1} enters the chain rule. Eq. (6) then asserts E_noise ≤ γ ||∂F^i/∂θ|| with γ∈(0,1), but E_noise is never defined, the left side is a loss/noise quantity while the right side is a feature-gradient norm, and the factor (Î_recon − I_high) is signed, so the claim that it is >0 is false in general. The text defers to a supplementary derivation that is not included. This is the sole formal support for the paper's main theoretical contribution; the authors must either provide a rigorous derivation or explicitly restate the entropy guidance as a heuristic without the claimed bound.
  2. [Experiment, Table 2] The abstract and Section 4 claim state-of-the-art performance across multiple benchmarks, but Table 2 contradicts this on LPIPS: on LOLv2-Real, U2CLLIE obtains LPIPS 0.1005 versus 0.0573 for FourLLIE. The PSNR/SSIM advantages over FourLLIE are also small (22.63 vs 22.34 dB; 0.851 vs 0.847). A blanket SOTA claim is therefore not supported across all reported metrics and datasets. Please restrict the claim to PSNR/SSIM (or a subset of datasets) and discuss the LPIPS trade-off in the text.
  3. [Experiment, Tables 4-5] The claim that entropy minimization drives the improvement is only supported by ablations on the same LSRW-Huawei dataset used to select the architecture and loss weights. The decisive comparison in Table 5, 'G2AF → DMFourLLIE', differs from the full model by 0.05 dB PSNR and 0.001 SSIM, with no standard deviations or multiple seeds. This does not establish that the entropy map F^{i+1} is the active mechanism. A direct controlled test is needed — e.g., replacing the learned entropy map with a constant, random, or unlearned map and evaluating on held-out benchmarks.
  4. [Methodology and Experiment] Several load-bearing items are deferred to a missing supplemental file: the formal analysis of Problem 2 ('See the supplement for formal analysis'), the derivation of Eq. (6), the CscO details in Eq. (9), and the derivations of L_global, L_color, and L_grad. Because these items are explicitly cited as support for the paper's central claims, the supplement is not optional. The manuscript must either include the supplementary material or remove the citations to it.
minor comments (3)
  1. [Eq. (1)] The notation is confusing: the text says LEN predicts Î_diff = (Y_high−Y_low)/Y_high, but the loss compares Î_diff with F_ill while Eq. (1) defines F_ill as the network output. Clarify which quantity is the target and which is the prediction.
  2. [Figure 2] The figure is very dense. Symbols such as ReM, CBA, L2G, and 'Conv_i applied i times' are not fully defined before use. A legend or expanded caption would help reproducibility.
  3. [Experiment, Tables 1-2] FLOPs are reported as '6.18 G' but the input resolution is only mentioned informally as 256×256 in the text. Please state the resolution in every complexity table. Also explain the 'GT-Mean' footnote, since it affects comparability of LPIPS/PSNR.

Circularity Check

1 steps flagged · score 5.0 of 10

Eq. 6's noise-fitting upper bound restates the entropy-gate design rather than deriving it; external benchmarks keep the central empirical claim independent, so circularity is partial.

  1. self definitional [Methodology, 'How UaD Addresses Problem 1', Eq. (6)]
    "⇒ F i+1 | {z } ∈(0,1) · ( ˆIrecon − Ihigh)| {z } >0 · ∂ϕ(H(F i))/∂H · ∂H(F i)/∂F i · ∂F i/∂θ| {z } ̸=0 ⇒ Enoise ≤ γ · ||∂F i/∂θ|| Upper Bound, with γ ∈ (0, 1). In Eq. 6, the left side reflects the network fitting noise, while the right side demonstrates that uncertainty entropy guides the noise toward a likelihood upper bound."

    The inequality is not derived from the chain rule; the only factor shown to lie in (0,1) is the entropy gate F^{i+1}, a normalization property of the softmax/log-probability construction. Every other factor is silently absorbed into γ, which is asserted to be in (0,1) without proof. If E_noise is understood as the left-hand product, the 'upper bound' is exactly the statement that the entropy gate is a bounded fraction of the gradient — i.e., a restatement of the architecture's design, not an independent theoretical result. The paper's claimed causal mechanism (entropy guidance constrains noise fitting) therefore reduces, by construction, to the decision to multiply the gradient by a normalized entropy map.

full rationale

The paper's empirical contribution is largely self-contained: it compares against many published methods on external benchmarks (LOLv2, LSRW, and unpaired sets), reports full ablations, and the SOTA claim does not depend on the proof of Eq. 6. The only self-citation of note is Yang et al. 2021 (SGM) involving co-author S. Wang; it is used as a baseline, not as load-bearing support, so it does not raise the score. The central theoretical claim, however, rests on Eq. 6, which is asserted after a non-sequitur chain and explicitly deferred to a supplemental file that is not provided ('See supplementary for derivations and symbol definitions'). Because the conclusion of Eq. 6 is equivalent to the design choice of a normalized entropy gate in the attention path, it is partially circular: the 'prediction' that entropy bounds noise fitting is a restatement of the architecture, not an independent first-principles result. This is a correctness/verifiability concern as much as a circularity concern, and it is confined to the theoretical framing; the benchmark tables would still support a conditional empirical claim. Score 5 reflects partial circularity in the mechanism explanation with independent external evaluation.

Assumptions & free parameters 3 free parameters · 3 assumptions · 4 invented entities

The central claim relies on hand-tuned loss weights and hyperparameters (k, lambda), on the unverified domain assumption that softmax-based entropy is a proxy for noise, and on the paper's own asserted gradient bound in Eq. 6. No external benchmark validates these assumptions beyond the reported tables.

free parameters (3)
  • Loss weights lambda1:7 = [0.95, 0.01, 0.01, 0.1, 0.5, 0.1, 0.1]
    Hand-tuned on the benchmark validation sets; the final metric trade-off depends directly on these values, and no sensitivity analysis is provided.
  • G2AF balance lambda = 0.5
    Manual balance between low- and high-frequency branches (Eq. 4); no ablation on this value is reported.
  • AsC neighborhood size top-k = 8
    Chosen ad hoc for the color correlation module; no ablation on k is provided.
assumptions (3)
  • standard math FFT/iFFT linearity, Gaussian filtering, softmax, cross-attention, and SSM recurrence are used as standard operations.
    Invoked throughout the architecture (Eqs. 2-4, 7-8); these are accepted tools and are not questioned here.
  • domain assumption Feature entropy computed from softmax channel probabilities is a reliable proxy for noise and uncertainty in low-light features.
    The motivation and UaD module (Eq. 2) depend on this; no external validation or comparison with other uncertainty measures is provided.
  • ad hoc to paper The gradient modulation mechanism in Eqs. 5-6, where the entropy factor bounds the noise fitting term, is asserted to hold.
    The inequality E_noise <= gamma * ||dF/dtheta||_upper is not derived; it is the paper's own theoretical claim and is central to contribution 1.
invented entities (4)
  • UaD (Uncertainty-Aware Dual-domain Denoise) module
    purpose: Dual-domain denoising under entropy guidance
    Effectiveness is shown only through in-paper ablations (Table 4); no external prediction or cross-task validation.
  • G2AF (Gaussian-Guided Adaptive Frequency-Domain Feature Enhancement)
    purpose: Adaptively filter frequency-domain noise
    Ablations replace it with DMFourLLIE and WaveMamba modules; no external handle is provided.
  • NeCo (Neighborhood Correlation State Space) module
    purpose: Model neighborhood causal dependencies in the encoder
    Only in-paper ablations support its contribution; the causal interpretation is not externally tested.
  • AsC (Adaptive Spatial-Color Calibration) module
    purpose: Enforce color consistency in the decoder
    Only in-paper ablations support its contribution; the causal interpretation is not externally tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-Aware Spatial Color Correlation for Low-Light Image Enhancement." pith.science (2026). https://pith.science/paper/DEBQFUC5

@misc{pith2026250804176,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Aware Spatial Color Correlation for Low-Light Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DEBQFUC5}},
  note         = {Machine review of arXiv:2508.04176}
}
read the original abstract

Most existing low-light image enhancement approaches primarily focus on architectural innovations, while often overlooking the intrinsic uncertainty within feature representations particularly under extremely dark conditions where degraded gradient and noise dominance severely impair model reliability and causal reasoning. To address these issues, we propose U2CLLIE, a novel framework that integrates uncertainty-aware enhancement and spatial-color causal correlation modeling. From the perspective of entropy-based uncertainty, our framework introduces two key components: (1) An Uncertainty-Aware Dual-domain Denoise (UaD) Module, which leverages Gaussian-Guided Adaptive Frequency Domain Feature Enhancement (G2AF) to suppress frequency-domain noise and optimize entropy-driven representations. This module enhances spatial texture extraction and frequency-domain noise suppression/structure refinement, effectively mitigating gradient vanishing and noise dominance. (2) A hierarchical causality-aware framework, where a Luminance Enhancement Network (LEN) first performs coarse brightness enhancement on dark regions. Then, during the encoder-decoder phase, two asymmetric causal correlation modeling modules Neighborhood Correlation State Space (NeCo) and Adaptive Spatial-Color Calibration (AsC) collaboratively construct hierarchical causal constraints. These modules reconstruct and reinforce neighborhood structure and color consistency in the feature space. Extensive experiments demonstrate that U2CLLIE achieves state-of-the-art performance across multiple benchmark datasets, exhibiting robust performance and strong generalization across various scenes.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 38 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bai, J.; Yin, Y.; He, Q.; Li, Y.; and Zhang, X. 2024. Retinexmamba: Retinex-based mamba for low-light image enhancement. In International Conference on Neural Information Processing, 427--442. Springer

    work page 2024

  4. [4]

    Cai, Y.; Bian, H.; Lin, J.; Wang, H.; Timofte, R.; and Zhang, Y. 2023. Retinexformer: One-stage retinex-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision, 12504--12513

    work page 2023

  5. [5]

    Chen, H.; Wang, Y.; Guo, T.; Xu, C.; Deng, Y.; Liu, Z.; Ma, S.; Xu, C.; Xu, C.; and Gao, W. 2021. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12299--12310

    work page 2021

  6. [6]

    Gu, A.; and Dao, T. 2023. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752

    arXiv 2023

  7. [7]

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

    work page 2023

  8. [8]

    Guo, X.; Li, Y.; and Ling, H. 2016. LIME: Low-light image enhancement via illumination map estimation. IEEE Transactions on image processing, 26(2): 982--993

    work page 2016

Show all 49 references
  1. [9]

    Hai, J.; Xuan, Z.; Yang, R.; Hao, Y.; Zou, F.; Lin, F.; and Han, S. 2023. R2rnet: Low-light image enhancement via real-low to real-normal network. Journal of Visual Communication and Image Representation, 90: 103712

    2023

  2. [10]

    Hou, J.; Zhu, Z.; Hou, J.; Liu, H.; Zeng, H.; and Yuan, H. 2023. Global structure-aware diffusion process for low-light image enhancement. Advances in Neural Information Processing Systems, 36: 79734--79747

    2023

  3. [11]

    Hu, Z.; Hu, Y.; Li, X.; Tang, S.; and Duan, L.-Y. 2025. Beyond Entropy: Region Confidence Proxy for Wild Test-Time Adaptation. arXiv preprint arXiv:2505.20704

    2025 arXiv

  4. [12]

    Huang, J.; Liu, Y.; Zhao, F.; Yan, K.; Zhang, J.; Huang, Y.; Zhou, M.; and Xiong, Z. 2022. Deep fourier-based exposure correction network with spatial-frequency interaction. In European Conference on Computer Vision, 163--180. Springer

    2022

  5. [13]

    Jain, H.; Zepeda, J.; P \'e rez, P.; and Gribonval, R. 2017. Subic: A supervised, structured binary code for image search. In Proceedings of the IEEE international conference on computer vision, 833--842

    2017

  6. [14]

    Jiang, Y.; Gong, X.; Liu, D.; Cheng, Y.; Fang, C.; Shen, X.; Yang, J.; Zhou, P.; and Wang, Z. 2021. Enlightengan: Deep light enhancement without paired supervision. IEEE transactions on image processing, 30: 2340--2349

    2021

  7. [15]

    Lee, C.; Lee, C.; and Kim, C.-S. 2013. Contrast enhancement based on layered difference representation of 2D histograms. IEEE transactions on image processing, 22(12): 5372--5384

    2013

  8. [16]

    Lee, J.; Jung, D.; Lee, S.; Park, J.; Shin, J.; Hwang, U.; and Yoon, S. 2024. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. arXiv preprint arXiv:2403.07366

    2024 arXiv

  9. [17]

    Li, C.; Guo, C.-L.; Zhou, M.; Liang, Z.; Zhou, S.; Feng, R.; and Loy, C. C. 2023. Embedding fourier for ultra-high-definition low-light image enhancement. arXiv preprint arXiv:2302.11831

    2023 arXiv

  10. [18]

    Liang, D.; Li, L.; Wei, M.; Yang, S.; Zhang, L.; Yang, W.; Du, Y.; and Zhou, H. 2022 a . Semantically contrastive learning for low-light image enhancement. In Proceedings of the AAAI conference on artificial intelligence, volume 36, 1555--1563

    2022

  11. [19]

    Liang, Y.; Wang, B.; Ren, W.; Liu, J.; Wang, W.; and Zuo, W. 2022 b . Learning hierarchical dynamics with spatial adjacency for image enhancement. In Proceedings of the 30th ACM International Conference on Multimedia, 2767--2776

    2022

  12. [20]

    Liu, Y.; Tian, Y.; Zhao, Y.; Yu, H.; Xie, L.; Wang, Y.; Ye, Q.; Jiao, J.; and Liu, Y. 2024. Vmamba: Visual state space model. Advances in neural information processing systems, 37: 103031--103063

    2024

  13. [21]

    Ma, K.; Zeng, K.; and Wang, Z. 2015. Perceptual quality assessment for multi-exposure image fusion. IEEE Transactions on Image Processing, 24(11): 3345--3356

    2015

  14. [22]

    completely blind

    Mittal, A.; Soundararajan, R.; and Bovik, A. C. 2012. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3): 209--212

    2012

  15. [23]

    D.; Zong, S.; Hemminger, B

    Pisano, E. D.; Zong, S.; Hemminger, B. M.; DeLuca, M.; Johnston, R. E.; Muller, K.; Braeuning, M. P.; and Pizer, S. M. 1998. Contrast limited adaptive histogram equalization image processing to improve the detection of simulated spiculations in dense mammograms. Journal of Dig...

    1998

  16. [24]

    Springenberg, J. T. 2015. Unsupervised and semi-supervised learning with categorical generative adversarial networks. arXiv preprint arXiv:1511.06390

    2015 arXiv

  17. [25]

    Wang, C.; Pan, J.; Wang, W.; Fu, G.; Liang, S.; Wang, M.; Wu, X.-M.; and Liu, J. 2024 a . Correlation matching transformation transformers for uhd image restoration. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 5336--5344

    2024

  18. [26]

    Wang, C.; Wu, H.; and Jin, Z. 2023. Fourllie: Boosting low-light image enhancement by fourier frequency information. In Proceedings of the 31st ACM International Conference on Multimedia, 7459--7469

    2023

  19. [27]

    Wang, S.; Zheng, J.; Hu, H.-M.; and Li, B. 2013. Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE transactions on image processing, 22(9): 3538--3548

    2013

  20. [28]

    Wang, W.; Yang, H.; Fu, J.; and Liu, J. 2024 b . Zero-reference low-light enhancement via physical quadruple priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 26057--26066

    2024

  21. [29]

    C.; Sheikh, H

    Wang, Z.; Bovik, A. C.; Sheikh, H. R.; and Simoncelli, E. P. 2004. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4): 600--612

    2004

  22. [30]

    Wang, Z.; Cun, X.; Bao, J.; Zhou, W.; Liu, J.; and Li, H. 2022. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 17683--17693

    2022

  23. [31]

    Wang, Z.-G.; Liang, Z.-H.; and Liu, C.-L. 2009. A real-time image processor with combining dynamic contrast ratio enhancement and inverse gamma correction for PDP. Displays, 30(3): 133--139

    2009

  24. [32]

    Weng, J.; Yan, Z.; Tai, Y.; Qian, J.; Yang, J.; and Li, J. 2024. Mamballie: Implicit retinex-aware low light enhancement with global-then-local state space. Advances in Neural Information Processing Systems, 37: 27440--27462

    2024

  25. [33]

    Wu, Y.; Pan, C.; Wang, G.; Yang, Y.; Wei, J.; Li, C.; and Shen, H. T. 2023. Learning semantic-aware knowledge guidance for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 1662--1671

    2023

  26. [34]

    Xu, X.; Wang, R.; Fu, C.-W.; and Jia, J. 2022. Snr-aware low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 17714--17724

    2022

  27. [35]

    Xu, X.; Wang, R.; and Lu, J. 2023. Low-light image enhancement via structure modeling and guidance. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9893--9903

    2023

  28. [36]

    Yan, Q.; Feng, Y.; Zhang, C.; Pang, G.; Shi, K.; Wu, P.; Dong, W.; Sun, J.; and Zhang, Y. 2025. Hvi: A new color space for low-light image enhancement. In Proceedings of the Computer Vision and Pattern Recognition Conference, 5678--5687

    2025

  29. [37]

    Yang, S.; Ding, M.; Wu, Y.; Li, Z.; and Zhang, J. 2023. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision, 12918--12927

    2023

  30. [38]

    Yang, W.; Wang, W.; Huang, H.; Wang, S.; and Liu, J. 2021. Sparse gradient regularized deep retinex network for robust low-light image enhancement. IEEE Transactions on Image Processing, 30: 2072--2086

    2021

  31. [39]

    Yi, X.; Xu, H.; Zhang, H.; Tang, L.; and Ma, J. 2025. Diff-Retinex++: Retinex-Driven Reinforced Diffusion Model for Low-Light Image Enhancement. IEEE Transactions on Pattern Analysis and Machine Intelligence

    2025

  32. [40]

    Yin, Y.; Xu, D.; Tan, C.; Liu, P.; Zhao, Y.; and Wei, Y. 2023. Cle diffusion: Controllable light enhancement diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, 8145--8156

    2023

  33. [41]

    W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F

    Zamir, S. W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F. S.; and Yang, M.-H. 2022. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 5728--5739

    2022

  34. [42]

    W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F

    Zamir, S. W.; Arora, A.; Khan, S.; Hayat, M.; Khan, F. S.; Yang, M.-H.; and Shao, L. 2020. Learning enriched features for real image restoration and enhancement. In European conference on computer vision, 492--511. Springer

    2020

  35. [43]

    Zhang, T.; Liu, P.; Cai, M.; Zhang, Z.; Lu, Y.; and Zhou, Q. 2025 a . BSMamba: Brightness and Semantic Modeling for Long-Range Interaction in Low-Light Image Enhancement. arXiv preprint arXiv:2506.18346

    2025 arXiv

  36. [44]

    Zhang, T.; Liu, P.; Lu, Y.; Cai, M.; Zhang, Z.; Zhang, Z.; and Zhou, Q. 2025 b . CWNet: Causal Wavelet Network for Low-Light Image Enhancement. arXiv preprint arXiv:2507.10689

    2025 arXiv

  37. [45]

    Zhang, T.; Liu, P.; Zhao, M.; and Lv, H. 2024. DMFourLLIE: dual-stage and multi-branch fourier network for low-light image enhancement. In Proceedings of the 32nd ACM International Conference on Multimedia, 7434--7443

    2024

  38. [46]

    Zhang, Y.; Guo, X.; Ma, J.; Liu, W.; and Zhang, J. 2021. Beyond brightening low-light images. International Journal of Computer Vision, 129(4): 1013--1037

    2021

  39. [47]

    Zheng, D.; Wu, X.-M.; Yang, S.; Zhang, J.; Hu, J.-F.; and Zheng, W.-S. 2024. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 25445--25455

    2024

  40. [48]

    Zhou, D.; Bousquet, O.; Lal, T.; Weston, J.; and Sch \"o lkopf, B. 2003. Learning with local and global consistency. Advances in neural information processing systems, 16

    2003

  41. [49]

    Zou, W.; Gao, H.; Yang, W.; and Liu, T. 2024. Wave-mamba: Wavelet state space model for ultra-high-definition low-light image enhancement. In Proceedings of the 32nd ACM International Conference on Multimedia, 1534--1543

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.