Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

SAIGFormer: A Spatially-Adaptive Illumination-Guided Network for Low-Light Image Enhancement

T0 review · 4 major / 6 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 SAIGFormer claims that feeding a transformer a per-pixel adaptive illumination estimate, computed from a dynamic integral image, yields state-of-the-art low-light enhancement, especially under non-uniform lighting.

desk verdict A real architecture paper with a genuinely new illumination estimator, but the headline gains are thin, the adaptivity evidence is mostly indirect, and Eq. (5) needs fixing before I'd trust a reimplementation. read the letter →

arxiv 2507.15520 v1 pith:RZCNWKO5 submitted 2025-07-21 cs.CV

Hanting Li , Fei Zhou , Xin Sun , Yang Hua , Jungong Han , Liang-Jie Zhang This is my paper · ORCID
classification cs.CV
keywords Low-lightimageenhancementIlluminationestimationTransformerSpatiallyadaptivefilteringIntegralNon-uniformAttentionmechanism
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 tries to establish that low-light image enhancement improves when the network estimates illumination per pixel with an adaptive filter, then lets that estimate steer the transformer's attention. It introduces SAIGFormer, whose SAI²E module computes a summed-area table and learns, for each location, a rectangular integration window and a modulation coefficient, giving an illumination map at $O(1)$ per-pixel cost. That map is injected into the query path of channel-wise self-attention (IG-MSA), so the network concentrates its capacity on lightness-relevant features. On six benchmarks the authors report consistent gains over recent transformers, including 0.33 dB on LOL-v1, 0.24 dB on LOL-v2-Syn, 0.23 dB on SMID, and 0.14 dB on LOL-Blur. If correct, the work shows that explicit per-region illumination adaptivity, not just global brightness priors, is what resolves over-exposure and shadow artifacts in low-light restoration.

What carries the argument

The central object is the Spatially-Adaptive Integral Illumination Estimator (SAI²E): a summed-area table of the input image plus two small convolutional subnets that predict, per pixel, a rectangular integration window (top/left/bottom/right offsets) and a modulation coefficient. Summing the precomputed integral over that window and dividing by the window area yields a local brightness estimate with $O(1)$ cost per pixel. The illumination map then enters the Illumination-Guided Multi-head Self-Attention (IG-MSA), where it is adaptively downsampled, aligned, and concatenated with the layer-normalized query so that the channel-affinity computation is conditioned on lighting. This combination is what lets a fixed-cost filter vary its support across the image.

What would settle it

On a synthetic test set with known per-pixel ground-truth illumination, compute the correlation between SAI²E's predicted integration window area and the true local brightness; a weak or negative correlation would mean the offsets are not tracking illumination, and a fixed-window control with matched parameter count would then determine whether the reported gains come from adaptivity or from added capacity.

Watch

Extended reading notes

Core claim

The central claim is that a spatially adaptive illumination estimate, computed in the image's early processing stage, can guide a transformer to restore non-uniform illumination more accurately than global priors or retinex decompositions. The authors demonstrate this with SAIGFormer, in which SAI²E assembles a dynamic integral image: a tiny Offset-Net predicts four edge offsets per pixel, defining a rectangular integration region whose area-normalized sum estimates local lighting; a Modulation-Net rescales that estimate. Because the integral image is precomputed, every pixel performs constant work, making the estimator lightweight. The resulting illumination map is downsampled and concatenated into the query of each transformer block's multi-head self-attention, forcing the attention to weight channels by their relevance to illumination. Across LOL-v1, LOL-v2, SID, SMID, and LOL-Blur, the paper reports the best or near-best PSNR/SSIM, with the largest margins on scenes with uneven lighting; ablation shows replacing the adaptive windows with a mean-illumination prior or fixed average pooling loses 0.6-1 dB.

Load-bearing premise

The learned per-pixel offsets and modulation coefficients are assumed to capture the scene's true illumination structure, not just training-set brightness statistics; the paper validates them only indirectly through end-task PSNR/SSIM and one visualization, never by comparing predicted window sizes to ground-truth lighting.

Editorial extensions

If this is right

  • Non-uniform illumination, including backlit and shadowed regions, can be handled by spatially varying low-pass windows rather than by global brightness adjustment.
  • An illumination estimator placed at the network's front end can guide every transformer stage, since the estimated map is injected into each block's attention query.
  • The same framework transfers across domains: training on standard low-light sets improves joint low-light enhancement and deblurring on LOL-Blur by 0.14 dB over the prior best.
  • Replacing the adaptive estimator with a mean-illumination prior or a fixed two-by-two average pool drops PSNR by 0.6-1.0 dB on LOL-v2-Real, indicating the adaptivity itself carries the gain.
  • The per-pixel cost of computing each illumination estimate is constant, so the guidance does not scale with window size, keeping the model at 12.35M parameters.

Reading between the lines

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

  • If the learned window sizes truly track scene illumination, SAI²E offers a general-purpose, parameter-efficient way to inject local smoothness priors into any restoration transformer; denoising and dehazing could adopt the same adaptive integral-image guidance.
  • A direct test of the paper's mechanism would be to regress the predicted window sizes against ground-truth illumination maps, for example from synthetic relighting, and check for monotonic correlation; the paper does not report such a check.
  • The uniform residual brightening shown in the variant without IG-MSA suggests illumination guidance acts as a spatial mask on the learned residual, so one could measure how much of the gain survives when the illumination map is replaced by a fixed smooth prior at the same resolution.
  • Because the offset and modulation subnets are only two convolutions each, one could freeze them and use analytically chosen windows, such as scale-space maxima, to separate the contribution of dynamic prediction from the integral-image representation itself.
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 / 6 minor

Summary. The paper proposes SAIGFormer, a Transformer-based low-light image enhancement framework whose central novelty is a Spatially-Adaptive Integral Illumination Estimator (SAI²E) built on a learned dynamic integral image, together with an Illumination-Guided Multi-head Self-Attention (IG-MSA) mechanism. The authors report state-of-the-art PSNR/SSIM on several benchmarks, with claimed gains of 0.14–0.33 dB over the next-best method on four datasets, and provide ablations on LOL-v2-real, visualizations, and a code link. The central claim is that spatially adaptive illumination estimation, rather than added network capacity, drives the improvements, and that the method generalizes to cross-domain tasks such as LOL-Blur.

Significance. If the central claim holds, the paper makes a useful contribution: SAI²E is a lightweight and conceptually novel mechanism for spatially varying illumination estimation, and the idea of injecting illumination into the query path of channel-wise attention is simple and potentially reusable. The paper is also strong in evaluation breadth: six datasets, including a cross-domain benchmark, and a public code release. The dynamic integral image representation is a genuine architectural idea, and the O(1) per-pixel cost after the integral image is attractive. However, the current evidence does not establish that spatial adaptivity is the cause of the reported gains: the ablations change capacity and branch structure simultaneously, the formulas in Eqs. (3)–(5) appear dimensionally inconsistent, one headline claim in Table I is contradicted by the table, and the reported margins are small and unreplicated. These issues are load-bearing for the paper's novelty and SOTA claims.

major comments (4)
  1. [Section III-C, Eq. (5)] The area normalization in Eq. (5) is dimensionally inconsistent with the scaled coordinates in Eq. (3). With N_w = w/W and N_h = h/H, the rectangle spanned by the predicted offsets has width (r+l)·w/W and height (t+b)·h/H, so the true area in image coordinates is (r+l)(t+b)·h·w/(H·W), not (t+b)(l+r)·h·w/4. Additionally, the integral image in Eq. (1) uses the inclusive definition Iii(x,y)=sum_{x'≤x,y'≤y} I(x',y'), for which the rectangle sum in Eq. (4) should involve Iii(tl-(1,1)) rather than Iii(tl). As written, SAI²E does not compute the mean intensity of the predicted integration region, which is the mechanism the paper claims. Please correct these formulas and confirm that the released implementation matches the corrected definitions.
  2. [Section IV-C, Table VI] The ablation study does not isolate spatial adaptivity from added capacity. The avgpool-2×2 variant removes both Offset-Net and Modulation-Net, and the 'w/o modulation map' variant removes one branch while keeping the other, so the number of parameters and the network topology change simultaneously with the adaptive mechanism. The 0.89 dB gap between the w/o-modulation variant and full SAI²E, and the drop from baseline to avgpool, could therefore be a capacity effect rather than evidence for illumination-guided adaptivity. Please report parameter counts for every variant and add a matched-capacity control, e.g. fixed learned offsets or an Offset-Net whose outputs are frozen to constants, and provide a quantitative check that the predicted integration areas correlate with local illumination on held-out images. The two min-max-normalized heatmaps in Fig. 8 and the training statistics in Fig. 9 do not demonstrate that the final offset mapping is illumination-dependent or that it transfers across datasets.
  3. [Section IV-B, LOL-v1 paragraph] The text claims 'we attain first place in PSNR and SSIM' on LOL-v1, but Table I shows CIDNet with SSIM 0.870 versus SAIGFormer's 0.863, so SAIGFormer is first in PSNR and second in SSIM. This is a factual error in a headline result. Please correct this statement and adjust the abstract and introduction if they rely on the same claim.
  4. [Section IV-B, Tables I–IV] The reported gains over the next-best method are 0.33, 0.24, 0.23, and 0.14 dB on LOL-v1, LOL-v2-Syn, SMID, and LOL-Blur, respectively. No error bars, repeated runs, or significance tests are reported; for LOL-v1 the test set contains only 15 images. These margins are close to typical run-to-run variation for stochastic training, so the 'significantly outperforms' wording is not yet supported. Please provide variance estimates or the number of runs and observed spread, or temper the significance claims.
minor comments (6)
  1. [Table II] The header entry 'CIDNetSAIGFormer' is missing a separator and the bracket for [14]; it should read 'CIDNet [14] SAIGFormer (Ours)'.
  2. [Section III-D, Eq. (11)] The dual-gated feed-forward network uses W_p1 and W_p2, but these projection weights are not defined in the text; a short sentence introducing them would improve readability.
  3. [Section III-C] The complexity claim 'O(1) computational complexity' should specify that this is per-pixel after the O(HW) integral image and offset computations; otherwise the statement is misleading.
  4. [Section III-C / Contribution list] The claim of being 'the first work to propose dynamic integral image representation in deep learning' is stronger than the evidence presented; please scope it to low-light enhancement or remove the global first claim.
  5. [Fig. 8 caption] The caption reports PSNR/SSIM pairs for the residual maps but does not specify which variant corresponds to which number in enough detail; please clarify the setup.
  6. [Fig. 2 caption] 'efficient' appears to be a typo for 'effect'; please proofread the captions.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: SAI²E and IG-MSA are trained end-to-end on external benchmarks, with only non-load-bearing self-citations.

full rationale

The paper's central claims are supported by end-to-end training of SAIGFormer against paired low/normal-light images on standard external benchmarks, so the reported PSNR/SSIM improvements are not constructed from the method's own inputs or definitions. SAI²E predicts offset and modulation maps from the input image using two small conv heads (Eq. 2) and combines them with an integral image to produce an illumination estimate (Eqs. 3-5); no ground-truth illumination map is used as a fitting target, so the illumination estimate is not a fitted quantity renamed as a prediction. IG-MSA incorporates this learned illumination into the attention query (Eqs. 8-10), which is a network design choice rather than a circular derivation. The only author self-citations are Refs. [32]-[33], mentioned in the related-work section as exhibiting 'conceptual relevance' to spatially adaptive processing; they are not load-bearing for the method's correctness and are not used to forbid alternatives. The ablations in Tables V and VI do compare the full SAI²E against a fixed 2x2 avgpool and a no-modulation variant, but changing parameter counts and branches simultaneously is an experimental confound, not a definitional circularity; the absence of a quantitative illumination-correlation metric is a missing-evidence concern, not a circularity. Since no passage in the manuscript asserts a limitation that reveals a self-referential derivation, and no load-bearing step reduces to the paper's own assumptions by construction, the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on learned sub-networks and hand-chosen hyperparameters rather than on a first-principles derivation. The method introduces no new physical entities or forces; the dynamic integral image is a computational representation.

free parameters (4)
  • Offset-Net weights = learned in training
    Predict the dynamic integration region sizes (t, l, b, r) in Eqs. (2)-(3); these offsets are not derived but fitted to the training data.
  • Modulation-Net weights = learned in training
    Predict the modulation map M in Eqs. (2) and (5); this controls the intensity of the estimated illumination.
  • Attention scaling parameter alpha = learned
    Learnable scaling in Eq. (10) for the attention logits; fitted to data.
  • Architecture hyperparameters (block counts, embedding dim) = [4, 6, 6, 8, 6, 6, 4, 4], C=32
    Chosen by hand without ablation or justification; they affect capacity and performance.
assumptions (3)
  • domain assumption Illumination is primarily encoded in low-frequency components of the image.
    Stated in Section III-A to justify the use of local averaging for illumination estimation.
  • domain assumption Retinex-style decomposition (image approximates illumination times reflectance) is valid for low-light images.
    Implicitly assumed when the network output is added as a residual to the input; not explicitly stated but foundational to the illumination guidance design.
  • domain assumption The evaluation datasets are representative of real-world low-light conditions.
    Generalization claims in Section IV-B rely on this assumption without a dedicated cross-dataset transfer test.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAIGFormer: A Spatially-Adaptive Illumination-Guided Network for Low-Light Image Enhancement." pith.science (2026). https://pith.science/paper/RZCNWKO5

@misc{pith2026250715520,
  author       = {Pith},
  title        = {Pith review of: SAIGFormer: A Spatially-Adaptive Illumination-Guided Network for Low-Light Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RZCNWKO5}},
  note         = {Machine review of arXiv:2507.15520}
}
abstract

Recent Transformer-based low-light enhancement methods have made promising progress in recovering global illumination. However, they still struggle with non-uniform lighting scenarios, such as backlit and shadow, appearing as over-exposure or inadequate brightness restoration. To address this challenge, we present a Spatially-Adaptive Illumination-Guided Transformer (SAIGFormer) framework that enables accurate illumination restoration. Specifically, we propose a dynamic integral image representation to model the spatially-varying illumination, and further construct a novel Spatially-Adaptive Integral Illumination Estimator ($\text{SAI}^2\text{E}$). Moreover, we introduce an Illumination-Guided Multi-head Self-Attention (IG-MSA) mechanism, which leverages the illumination to calibrate the lightness-relevant features toward visual-pleased illumination enhancement. Extensive experiments on five standard low-light datasets and a cross-domain benchmark (LOL-Blur) demonstrate that our SAIGFormer significantly outperforms state-of-the-art methods in both quantitative and qualitative metrics. In particular, our method achieves superior performance in non-uniform illumination enhancement while exhibiting strong generalization capabilities across multiple datasets. Code is available at https://github.com/LHTcode/SAIGFormer.git.

Figures

Figures reproduced from arXiv: 2507.15520 by the authors.

Figure 1
Figure 1. Illustration of the efficient of two representative SOTA methods and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the mechanism of the proposed SAI²E, i.e., spatial [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the SAIGFormer framework. (a) SAI²E adaptively extracts spatially-varying illumination information from the input image. (b) The [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visual comparison to SOTA models on LOL-v1 [9] dataset(top) and LOL-v2-real [36] dataset(bottom). [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Visual comparison to SOTA models on SID dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison to SOTA models on SMID dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Visual comparison to SOTA models on LOL-Blur dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Visual evidence of the effectiveness of the SAI²E module. (b) and (c) illustrate the relationship between the illumination prior of the low-light image [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Training dynamics of the SAI²E module. The figure illustrates the [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Overcoming Attention Drift: Homogeneity-Heterogeneity Guided Feature Aggregation for Low-Light Remote Sensing Image Enhancement

    cs.CV 2026-08 conditional novelty 6.0 of 10

    A dual-prior attention module adds semantic homogeneity and geometric boundary constraints to low-light remote sensing enhancement and outperforms prior methods on most tested benchmarks.

Reference graph

Works this paper leans on

50 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Low-light image and video enhancement using deep learning: A survey,

    C. Li, C. Guo, L. Han, J. Jiang, M.-M. Cheng, J. Gu, and C. C. Loy, “Low-light image and video enhancement using deep learning: A survey,”IEEE transactions on pattern analysis and machine intelligence, vol. 44, no. 12, pp. 9396–9416, 2021

    work page 2021

  2. [2]

    Single-stage face detection under extremely low-light conditions,

    J. Yu, X. Hao, and P. He, “Single-stage face detection under extremely low-light conditions,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 3523–3532

    work page 2021

  3. [3]

    Multi-object tracking in the dark,

    X. Wang, K. Ma, Q. Liu, Y . Zou, and Y . Fu, “Multi-object tracking in the dark,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 382–392

    work page 2024

  4. [4]

    Naturalness preserved enhancement algorithm for non-uniform illumination images,

    S. Wang, J. Zheng, H.-M. Hu, and B. Li, “Naturalness preserved enhancement algorithm for non-uniform illumination images,”IEEE transactions on image processing, vol. 22, no. 9, pp. 3538–3548, 2013

    work page 2013

  5. [5]

    Low-light image enhancement with semi-decoupled decomposition,

    S. Hao, X. Han, Y . Guo, X. Xu, and M. Wang, “Low-light image enhancement with semi-decoupled decomposition,”IEEE transactions on multimedia, vol. 22, no. 12, pp. 3025–3038, 2020

    work page 2020

  6. [6]

    Learning to see moving objects in the dark,

    H. Jiang and Y . Zheng, “Learning to see moving objects in the dark,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 7324–7333. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2025 11

    work page 2019

  7. [7]

    Zero- shot restoration of back-lit images using deep internal learning,

    L. Zhang, L. Zhang, X. Liu, Y . Shen, S. Zhang, and S. Zhao, “Zero- shot restoration of back-lit images using deep internal learning,” in Proceedings of the 27th ACM international conference on multimedia, 2019, pp. 1623–1631

    work page 2019

  8. [8]

    Dslr: Deep stacked laplacian restorer for low- light image enhancement,

    S. Lim and W. Kim, “Dslr: Deep stacked laplacian restorer for low- light image enhancement,”IEEE Transactions on Multimedia, vol. 23, pp. 4272–4284, 2020

    2020

Show all 50 references
  1. [9]

    Deep retinex decomposition for low-light enhancement,

    C. Wei, W. Wang, W. Yang, and J. Liu, “Deep retinex decomposition for low-light enhancement,” inBritish Machine Vision Conference, 2018

    2018

  2. [10]

    Kindling the darkness: A practical low-light image enhancer,

    Y . Zhang, J. Zhang, and X. Guo, “Kindling the darkness: A practical low-light image enhancer,” inProceedings of the 27th ACM Interna- tional Conference on Multimedia, 2019, p. 1632–1640

    2019

  3. [11]

    Beyond brightening low-light images,

    Y . Zhang, X. Guo, J. Ma, W. Liu, and J. Zhang, “Beyond brightening low-light images,”International Journal of Computer Vision, vol. 129, pp. 1013–1037, 2021

    2021

  4. [12]

    Snr-aware low-light image en- hancement,

    X. Xu, R. Wang, C.-W. Fu, and J. Jia, “Snr-aware low-light image en- hancement,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 17 714–17 724

    2022

  5. [13]

    Retinex- former: One-stage retinex-based transformer for low-light image en- hancement,

    Y . Cai, H. Bian, J. Lin, H. Wang, R. Timofte, and Y . Zhang, “Retinex- former: One-stage retinex-based transformer for low-light image en- hancement,” inICCV, 2023

    2023

  6. [14]

    Hvi: A new color space for low-light image enhancement,

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

    2025

  7. [15]

    Decoupled low-light image enhancement,

    S. Hao, X. Han, Y . Guo, and M. Wang, “Decoupled low-light image enhancement,”ACM Transactions on Multimedia Computing, Commu- nications, and Applications (TOMM), vol. 18, no. 4, pp. 1–19, 2022

    2022

  8. [16]

    Uretinex- net: Retinex-based deep unfolding network for low-light image enhance- ment,

    W. Wu, J. Weng, P. Zhang, X. Wang, W. Yang, and J. Jiang, “Uretinex- net: Retinex-based deep unfolding network for low-light image enhance- ment,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 5901–5910

    2022

  9. [17]

    Lime: Low-light image enhancement via illumination map estimation,

    X. Guo, Y . Li, and H. Ling, “Lime: Low-light image enhancement via illumination map estimation,”IEEE Transactions on Image Processing, vol. 26, no. 2, pp. 982–993, 2017

    2017

  10. [18]

    Contrast enhancement based on layered difference representation of 2d histograms,

    C. Lee, C. Lee, and C.-S. Kim, “Contrast enhancement based on layered difference representation of 2d histograms,”IEEE Transactions on Image Processing, vol. 22, no. 12, pp. 5372–5384, 2013

    2013

  11. [19]

    The retinex theory of color vision,

    E. H. Land, “The retinex theory of color vision,”Sci. Amer., vol. 237, no. 6, pp. 108–28, 1977

    1977

  12. [20]

    Properties and performance of a center/surround retinex,

    D. Jobson, Z. Rahman, and G. Woodell, “Properties and performance of a center/surround retinex,”IEEE Transactions on Image Processing, vol. 6, no. 3, pp. 451–462, 1997

    1997

  13. [21]

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

    ——, “A multiscale retinex for bridging the gap between color images and the human observation of scenes,”IEEE Transactions on Image Processing, vol. 6, no. 7, pp. 965–976, 1997

    1997

  14. [22]

    Structure-revealing low- light image enhancement via robust retinex model,

    M. Li, J. Liu, W. Yang, X. Sun, and Z. Guo, “Structure-revealing low- light image enhancement via robust retinex model,”IEEE Transactions on Image Processing, vol. 27, no. 6, pp. 2828–2841, 2018

    2018

  15. [23]

    Learning to see in the dark,

    C. Chen, Q. Chen, J. Xu, and V . Koltun, “Learning to see in the dark,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 3291–3300

    2018

  16. [24]

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

    W. Yang, S. Wang, Y . Fang, Y . Wang, and J. Liu, “From fidelity to perceptual quality: A semi-supervised approach for low-light image en- hancement,” inIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    2020

  17. [25]

    Zero-reference deep curve estimation for low-light image enhance- ment,

    C. G. Guo, C. Li, J. Guo, C. C. Loy, J. Hou, S. Kwong, and R. Cong, “Zero-reference deep curve estimation for low-light image enhance- ment,” inProceedings of the IEEE conference on computer vision and pattern recognition (CVPR), June 2020, pp. 1780–1789

    2020

  18. [26]

    Surroundnet: Towards effective low-light image enhancement,

    F. Zhou, X. Sun, J. Dong, and X. X. Zhu, “Surroundnet: Towards effective low-light image enhancement,”Pattern Recognition, vol. 141, p. 109602, 2023

    2023

  19. [27]

    Enlightengan: Deep light enhancement without paired supervision,

    Y . Jiang, X. Gong, D. Liu, Y . Cheng, C. Fang, X. Shen, J. Yang, P. Zhou, and Z. Wang, “Enlightengan: Deep light enhancement without paired supervision,”IEEE Transactions on Image Processing, vol. 30, pp. 2340–2349, 2021

    2021

  20. [28]

    Restormer: Efficient transformer for high-resolution image restoration,

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

    2022

  21. [29]

    Glow in the dark: Low-light image enhancement with external memory,

    D. Ye, Z. Ni, W. Yang, H. Wang, S. Wang, and S. Kwong, “Glow in the dark: Low-light image enhancement with external memory,”IEEE Transactions on Multimedia, vol. 26, pp. 2148–2163, 2023

    2023

  22. [30]

    Fourllie: Boosting low-light image enhancement by fourier frequency information,

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

    2023

  23. [31]

    Wave-mamba: Wavelet state space model for ultra-high-definition low-light image enhancement,

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

    2024

  24. [32]

    Adaptive morphology filter: A lightweight module for deep hyperspectral image classification,

    F. Zhou, X. Sun, C. Sun, J. Dong, and X. X. Zhu, “Adaptive morphology filter: A lightweight module for deep hyperspectral image classification,” IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1– 16, 2023

    2023

  25. [33]

    Gaussian dynamic convolution for efficient single-image segmentation,

    X. Sun, C. Chen, X. Wang, J. Dong, H. Zhou, and S. Chen, “Gaussian dynamic convolution for efficient single-image segmentation,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 32, no. 5, pp. 2937–2948, 2021

    2021

  26. [34]

    Rapid object detection using a boosted cascade of simple features,

    P. Viola and M. Jones, “Rapid object detection using a boosted cascade of simple features,” inProceedings of the 2001 IEEE computer society conference on computer vision and pattern recognition. CVPR 2001, vol. 1. Ieee, 2001, pp. I–I

    2001

  27. [35]

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

    T. Wang, K. Zhang, T. Shen, W. Luo, B. Stenger, and T. Lu, “Ultra-high- definition low-light image enhancement: A benchmark and transformer- based method,” inProceedings of the AAAI conference on artificial intelligence, vol. 37, no. 3, 2023, pp. 2654–2662

    2023

  28. [36]

    Sparse gradient reg- ularized deep retinex network for robust low-light image enhancement,

    W. Yang, W. Wang, H. Huang, S. Wang, and J. Liu, “Sparse gradient reg- ularized deep retinex network for robust low-light image enhancement,” IEEE Transactions on Image Processing, vol. 30, pp. 2072–2086, 2021

    2021

  29. [37]

    Uformer: A general u-shaped transformer for image restoration,

    Z. Wang, X. Cun, J. Bao, W. Zhou, J. Liu, and H. Li, “Uformer: A general u-shaped transformer for image restoration,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 17 683–17 693

    2022

  30. [38]

    Learning enriched features for fast image restoration and enhancement,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, M.-H. Yang, and L. Shao, “Learning enriched features for fast image restoration and enhancement,”IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 2, pp. 1934–1948, 2022

    1934

  31. [39]

    Low-light image enhancement with normalizing flow,

    Y . Wang, R. Wan, W. Yang, H. Li, L.-P. Chau, and A. Kot, “Low-light image enhancement with normalizing flow,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 3, pp. 2604–2612, 2022

    2022

  32. [40]

    Global structure- aware diffusion process for low-light image enhancement,

    J. Hou, Z. Zhu, J. Hou, H. Liu, H. Zeng, and H. Yuan, “Global structure- aware diffusion process for low-light image enhancement,”Advances in Neural Information Processing Systems, vol. 36, pp. 79 734–79 747, 2023

    2023

  33. [41]

    Retinexmamba: Retinex- based mamba for low-light image enhancement,

    J. Bai, Y . Yin, Q. He, Y . Li, and X. Zhang, “Retinexmamba: Retinex- based mamba for low-light image enhancement,” inInternational Con- ference on Neural Information Processing. Springer, 2025, pp. 427– 442

    2025

  34. [42]

    Urwkv: Unified rwkv model with multi-state perspective for low-light image restora- tion,

    R. Xu, Y . Niu, Y . Li, H. Xu, W. Liu, and Y . Chen, “Urwkv: Unified rwkv model with multi-state perspective for low-light image restora- tion,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 21 267–21 276

    2025

  35. [43]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”IEEE transactions on image processing, vol. 13, no. 4, pp. 600–612, 2004

    2004

  36. [44]

    Seeing motion in the dark,

    C. Chen, Q. Chen, M. N. Do, and V . Koltun, “Seeing motion in the dark,” inProceedings of the IEEE/CVF International conference on computer vision, 2019, pp. 3185–3194

    2019

  37. [45]

    Lednet: Joint low-light enhance- ment and deblurring in the dark,

    S. Zhou, C. Li, and C. Change Loy, “Lednet: Joint low-light enhance- ment and deblurring in the dark,” inEuropean conference on computer vision. Springer, 2022, pp. 573–589

    2022

  38. [46]

    Mambair: A simple baseline for image restoration with state-space model,

    H. Guo, J. Li, T. Dai, Z. Ouyang, X. Ren, and S.-T. Xia, “Mambair: A simple baseline for image restoration with state-space model,” in European conference on computer vision. Springer, 2024, pp. 222– 241

    2024

  39. [47]

    SGDR: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “SGDR: Stochastic gradient descent with warm restarts,” inInternational Conference on Learning Representa- tions, 2017

    2017

  40. [48]

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

    R. Liu, L. Ma, J. Zhang, X. Fan, and Z. Luo, “Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 10 561–10 570

    2021

  41. [49]

    Glare: Low light image enhancement via generative latent feature based codebook retrieval,

    H. Zhou, W. Dong, X. Liu, S. Liu, X. Min, G. Zhai, and J. Chen, “Glare: Low light image enhancement via generative latent feature based codebook retrieval,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 36–54

    2024

  42. [50]

    Perceptual decoupling with heterogeneous auxiliary tasks for joint low-light image enhancement and deblurring,

    Y . Li, R. Xu, Y . Niu, W. Guo, and T. Zhao, “Perceptual decoupling with heterogeneous auxiliary tasks for joint low-light image enhancement and deblurring,”IEEE Transactions on Multimedia, vol. 26, pp. 6663–6675, 2024

    2024

Pith tools

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