Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Unsupervised Exposure Correction

T0 review · 4 major / 5 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 Exposure correction can be learned without manual annotations by letting multi-exposure images serve as each other's ground truth.

desk verdict A genuinely tiny self-supervised exposure model with a useful dataset, but the abstract overclaims and the cross-scene transfer argument needs work. read the letter →

arxiv 2507.17252 v1 pith:VHQERSYE submitted 2025-07-23 cs.CV

Ruodai Cui , Li Niu , Guosheng Hu This is my paper
classification cs.CV
keywords unsupervisedexposurecorrectionmulti-exposuresequencesradiometryimageenhancementfeatureencoderlow-levelvisionedgedetectionself-supervisedlearning
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

This paper tries to establish that exposure correction, adjusting an image that is too dark or too bright, can be learned without any human-annotated pairs, using only multi-exposure sequences of the same scene as free supervision. The authors argue that because these sequences vary only in exposure, a network trained on them learns to modify exposure and nothing else, so the model can later correct new scenes. They report that this unsupervised model is competitive with the supervised state of the art on a standard benchmark, generalizes better to a different low-light dataset, and preserves fine details better for downstream edge detection, all with about 19,000 parameters, roughly 0.01% of a leading supervised baseline's 182 million. They also contribute a new radiometry-focused dataset built by varying exposure while holding editing style constant, to support training and evaluation without stylistic bias.

What carries the argument

The mechanism is a three-module pipeline: an Exposure Feature Encoder that pools channel statistics such as maximum, average, and standard deviation into a 96-dimensional descriptor; a Parameter Predictor that computes the scalar exposure difference $\Delta E = d(e(I_1), e(I_2))$ in that latent space; and an Exposure Corrector that applies a pixel-wise interpolation $I_{out}(x,y) = \lambda I_{in}(x,y) + (1-\lambda) h(I_{in}(x,y))$, where $h(\cdot)$ is implemented by $1\times1$ convolution layers and the operation is repeated three times. Training is driven by a restoration loss on same-sequence pairs, a monopoly loss that preserves relative brightness order when the references come from a different scene, and a total-variation semantic loss. The scalar latent difference is what lets the model transfer an exposure adjustment from one scene to another without seeing paired examples.

What would settle it

Train the model on multi-exposure sequences from one collection of scenes, then apply it to images from a very different scene distribution and compare outputs with the inputs at matched exposure. If object edges, textures, or object-level colors change when only brightness should shift, the exposure-only assumption fails; a cleaner numerical check is computing $d(e(I_a), e(I_b))$ for two different scenes at the same exposure and checking whether it is near zero.

Watch

Extended reading notes

Core claim

The paper's central claim is that exposure correction can be learned without manually annotated ground truth by exploiting multi-exposure sequences of the same scene, generated freely through an emulated ISP pipeline. Images in one sequence are made to serve as mutual ground truth: a style encoder maps each image to a latent exposure feature, a difference function collapses the gap between two images to a single scalar, and a correction network transforms one image by that scalar. A second cross-scene training signal enforces a Monopoly Principle: when two reference images of known exposure order are applied to the same input, the brighter reference must produce a brighter output, which aligns exposures across scenes without paired supervision. The paper reports that this unsupervised model matches or improves on the supervised state of the art on a standard exposure benchmark, transfers better to a low-light dataset, and preserves edges better in downstream edge detection, while using roughly 0.01% of the baseline's parameters.

Load-bearing premise

The model can correct new scenes without altering their content rests on the assumption that, because training pairs vary only in exposure, the learned latent exposure feature captures exposure alone and ignores scene content and color statistics.

Editorial extensions

If this is right

  • Exposure correction can be trained at scale from freely generated multi-exposure sequences, removing the cost and subjectivity of expert retouching.
  • A model of about 19,000 parameters can run in real time, making exposure correction practical for on-device camera pipelines.
  • Radiometry-only correction preserves low-level features, so corrected images feed edge detection and similar tasks better than transformation-heavy baselines.
  • Training on a style-consistent radiometry dataset improves cross-dataset generalization compared with training on expert-retouched ground truth.

Reading between the lines

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

  • If the exposure-only transfer assumption holds, the same mutual-supervision scheme could extend to other radiometric transforms, such as white balance or denoising, by generating sequences that vary only in that factor.
  • The Monopoly Principle suggests a testable extension: using more than two references with known EV orderings could supply denser supervision and finer control over the output exposure level.
  • A direct stress test would be to compute $d(e(I_a), e(I_b))$ for same-exposure images of different scenes; near-zero values would confirm the encoder isolates exposure, while large values would reveal content leakage.
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 / 5 minor

Summary. The paper proposes an unsupervised exposure correction method (UEC) that trains on multi-exposure sequences generated by an emulated ISP pipeline without manually annotated pairs. A restoration loss treats images from the same sequence as mutual ground truth, while a Monopoly loss enforces relative brightness ordering across different scenes. At test time, a single fixed reference image calibrates the exposure of all inputs. The paper also introduces a Radiometry Correction Dataset and reports results on MSEC, LOL, and edge detection, with a model about 0.01% the size of the supervised ECM baseline. The central claim is that this is the first unsupervised exposure correction method and that it is competitive with or superior to supervised state-of-the-art methods.

Significance. The paper's premise is attractive: if exposure correction can be trained without manual annotation and with a tiny model, it would be valuable for scalable and on-device use. The authors also release code and a large radiometry-focused dataset, which are potentially useful community resources. The edge-detection experiments provide an interesting perspective on downstream low-level vision. However, the manuscript's significance is undercut by an unsupported transfer assumption, an unqualified performance claim contradicted by its own main benchmark table, and a test-time protocol that requires a reference image, which is not surfaced in the abstract. These issues need to be resolved before the practical value of the method can be assessed.

major comments (4)
  1. [Sec. 3.1, Eqs. (3)-(5)] The paper states that combining Eqs. (3) and (4) yields the pixel-wise inequality in Eq. (5). This is not a logical consequence: Eq. (4) merely defines the two outputs, and no property of the network or the difference function guarantees that a brighter reference leads to a brighter output. In fact, Eq. (5) is exactly the constraint that the Monopoly loss in Eq. (8) enforces during training. The manuscript should present Eq. (5) as a training objective (the Monopoly principle) rather than as a derived fact, because this distinction is critical for readers to understand that the cross-scene signal is only an ordinal constraint.
  2. [Sec. 3.1 and Sec. 3.2] The load-bearing transfer premise is asserted, not validated. The sentence 'Since our method learn from data which varies solely in exposure, the trained f(·) can only modify exposures' is a non-sequitur: training on same-scene pairs does not by itself guarantee that the learned transformation will change only exposure when applied to new scenes. The Exposure Feature Encoder computes global statistics (mean, maximum, standard deviation) that can encode scene content and color distribution, and the Monopoly loss enforces only a relative brightness order. Nothing in the losses separates exposure information from content information. The paper needs an isolation experiment, for example, applying the trained model to pairs of images with identical EV but different content and measuring whether output content changes, or comparing the predicted exposure difference against ground-truth EV on a calibration set. Without such evidence, the central claim that the method corrects only exposure is unsupported.
  3. [Abstract and Sec. 5.2, Table 1a] The abstract claims the method 'outperforms state-of-the-art supervised methods [12], while utilizing only 0.01% of their parameters,' but Table 1a shows that on the MSEC benchmark, ECM achieves PSNR 20.874 and SSIM 0.877 while the proposed method achieves PSNR 18.756 and SSIM 0.812, i.e., lower on both metrics. The paper does outperform ECM on the Radiometry Correction Dataset (Table 2a, average PSNR 20.548 vs. 20.445, SSIM 0.868 vs. 0.744) and in the edge-detection experiment, but the abstract's unqualified 'outperforms' statement is contradicted by the paper's own main benchmark. The performance claim should be restricted to the specific dataset and metric, or the abstract should be corrected.
  4. [Sec. 3.4 and Sec. 5.1] The testing protocol requires a user-supplied reference image: 'we hard code the exposure features derived from this image across all test cases' and in the experiments 'We select the second image from left to right in Fig. 9 as our reference.' This reference acts as a global calibration target, so the method is not fully automatic at inference. The abstract and introduction describe the method as 'unsupervised' and as eliminating the need for manual annotations, which is potentially misleading because a well-exposed reference image must still be chosen for each deployment scenario. The paper should clearly state this requirement in the abstract and discuss how the choice of reference affects the output; the robustness shown in Fig. 9 is only demonstrated on a few examples and does not cover failure cases such as a poorly exposed or low-quality reference.
minor comments (5)
  1. [Fig. 2 caption] The caption contains a typo: 'Schematic diagrams of traing UEC' should read 'of training UEC.'
  2. [Sec. 4] The sentence 'while while freezing other post-processing ISP procedures' has a duplicated 'while.'
  3. [Sec. 3.3, Eq. (8)] The variables Iout1 and Iout2 are defined only in the text after the equation; it would be clearer to define them immediately after Eq. (8), matching the notation used in Sec. 3.1.
  4. [Sec. 3.2] The iterative application of the exposure correction is mentioned as 'This iterative process is repeated three times' but the recurrence is not specified in the main text. Please provide the explicit iterative formulation or refer more precisely to the supplementary material.
  5. [Sec. 5.2] The term 'Monopoly Principle' is not explained beyond the sentence about brightness monopoly; the connection between the name and the mathematical constraint is unclear and should be defined more precisely.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the training losses are direct self-supervision on multi-exposure images, and reported gains are benchmarked against external baselines and datasets.

full rationale

The core derivation is not circular. The restoration loss (Eq. 7) directly supervises the network output against a different-exposure image from the same multi-exposure sequence, which is a standard self-supervised regression target rather than a hidden reuse of the evaluation metric. The Monopoly loss (Eq. 8) enforces a relative brightness ordering between two outputs generated from the same input, which is again a direct training signal. No parameter is fitted to the reported PSNR or SSIM values, and the single test-time reference image is an explicit design choice described in Sec. 3.4 rather than a test-set-derived constant. Comparisons against ECM [12] and Afifi et al. [2] use external implementations and external datasets (MSEC, LOL), so the central claims are not reduced to the paper's own inputs. The Sec. 3.1 transfer premise - that f trained on same-scene exposure pairs modifies only exposure on new scenes - is an unproven empirical assumption, and Eq. (5) may not follow deductively from Eqs. (3) and (4); however, that is a correctness or robustness gap, not circularity. There is no self-citation chain or uniqueness theorem doing load-bearing work, and no known empirical result is merely renamed as a new contribution.

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

The central claim relies on synthetic multi-exposure data generation, a pixel-wise monotonicity assumption, a latent-space exposure scalar, and a single reference image as the definition of correct exposure. None of these is independently validated; together they carry much of the method's weight.

free parameters (3)
  • loss weights alpha1, alpha2, alpha3 = alpha1=1, alpha2=1, alpha3=0.1
    Hand-set in Sec. 5.1 Implementation Details without sensitivity analysis; the balance between restoration, monopoly, and TV losses directly shapes output quality.
  • test reference image = second image from left to right in Fig. 9
    At test time one reference image defines the target exposure; the choice is manual and although Fig. 9 suggests robustness, no objective selection criterion or variance is reported.
  • iterative correction count = 3
    The correction block is repeated three times (Sec. 3.2); no ablation of the iteration count is provided.
assumptions (4)
  • domain assumption Pixel-wise brightness is monotonic in exposure value within a multi-exposure sequence: for EV(J1)>EV(J2), every pixel satisfies J1(x,y) >= J2(x,y).
    Eq. (3) in Sec. 3.1. This underpins the Monopoly loss; clipping, tone mapping, or noise in the emulated ISP can violate it.
  • domain assumption An emulated ISP pipeline can generate realistic multi-exposure sRGB pairs that behave like real exposure changes.
    Sec. 1 and Sec. 4 rely on this generation for free supervision; no validation of realism is provided.
  • ad hoc to paper The latent exposure difference computed in a 96D feature space transfers across different scenes and reduces to a single scalar that drives correction.
    Sec. 3.1 claims 'this approach enables the difference function to adapt images from different scenes to a certain degree'; no proof or diagnostic is given.
  • domain assumption Images edited by ExpertC in MIT-Adobe FiveK serve as a suitable universal well-exposed reference.
    Sec. 4 selects ExpertC as reference for the dataset and for test-time calibration; this imports one editor's style into the definition of correct exposure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unsupervised Exposure Correction." pith.science (2026). https://pith.science/paper/VHQERSYE

@misc{pith2026250717252,
  author       = {Pith},
  title        = {Pith review of: Unsupervised Exposure Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VHQERSYE}},
  note         = {Machine review of arXiv:2507.17252}
}
read the original abstract

Current exposure correction methods have three challenges, labor-intensive paired data annotation, limited generalizability, and performance degradation in low-level computer vision tasks. In this work, we introduce an innovative Unsupervised Exposure Correction (UEC) method that eliminates the need for manual annotations, offers improved generalizability, and enhances performance in low-level downstream tasks. Our model is trained using freely available paired data from an emulated Image Signal Processing (ISP) pipeline. This approach does not need expensive manual annotations, thereby minimizing individual style biases from the annotation and consequently improving its generalizability. Furthermore, we present a large-scale Radiometry Correction Dataset, specifically designed to emphasize exposure variations, to facilitate unsupervised learning. In addition, we develop a transformation function that preserves image details and outperforms state-of-the-art supervised methods [12], while utilizing only 0.01% of their parameters. Our work further investigates the broader impact of exposure correction on downstream tasks, including edge detection, demonstrating its effectiveness in mitigating the adverse effects of poor exposure on low-level features. The source code and dataset are publicly available at https://github.com/BeyondHeaven/uec_code.

Figures

Figures reproduced from arXiv: 2507.17252 by the authors.

Figure 1
Figure 1. Visual Comparison: ECM [12] vs. Our UEC method. of sRGB images remains crucial. The advent of deep learning has spurred ex￾tensive research in this field, with numerous studies proposing models for expo￾sure correction [1, 4, 14, 16, 19, 24, 25, 27, 32, 34, 42, 44], demonstrating significant achievements. Nevertheless, these methods still face three challenges. (1) A primary challenge arises from the dependence on t… view at source ↗
Figure 2
Figure 2. Schematic diagrams of traing UEC. Summarization. We design a transformation to adjust exposures, which can be optimized by two principles. Firstly, we select images from a single multi￾exposure sequence, establishing a baseline for exposure manipulation. Secondly, we aim to adapt to varied scenes. This involves selecting two reference images from different exposures, but differing from the sequence where we sample t… view at source ↗
Figure 3
Figure 3. The architecture of our UEC method. plemented by 1 × 1 convolution layers. This iterative process is repeated three times to enhance the correction effectiveness. For more details, please refer to supplementary material. 3.3 Loss Function We jointly train the Exposure Feature Encoder, Parameter Predictor, and Ex￾posure Corrector in an end-to-end fashion. We employ three types of losses and compute their weighted sum… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visualization of Datasets: (a) Afifi et al’s [2] vs. (b) Ours. to select the most suitable image from this sequence. To enhance efficiency, our method bypasses the need to generate and then choose from multiple exposures, instead directly yielding the image with the be…
Figure 5
Figure 5. Figure 5: Visual Comparison: ECM [12] vs. Our UEC method. Implementation Details. The weights in Eq. (10) are empirically set α1 = α2 = 1 and α3 = 0.1. We use only one well-exposed reference image during testing. We select the second image from left to right in [PITH_FULL_IMAGE…
Figure 6
Figure 6. Figure 6: Results on MSEC Dataset [2]. We take images from [12] and compare with our model. LOL dataset comprises 485 pairs of training images, with each pair featuring one low-light image and one corresponding normal-light image. For our assessment, we exclusively employed the …
Figure 7
Figure 7. Figure 7: Generalizability performance by ECM [12] and our UEC. We pretrain models on MSEC Dataset [2] and assess their performance on LOL Dataset [40]. ECM [12] Ours EV PSNR SSIM PSNR SSIM -2 20.122 0.718 19.475 0.812 -1 20.301 0.735 20.144 0.847 0 20.546 0.751 20.968 0.884 +1 …
Figure 8
Figure 8. Figure 8: Edge detection performance by ECM [12] and our UEC. Results on Edge Detection. Exposure correction not only enhances image aesthetics but also supports downstream tasks in computer vision. Since many algorithms rely on images with standard exposure, variations in real-…
Figure 9
Figure 9. Figure 9: Impact of Reference Images on UEC Results. Method Parameters Model Size Time(GPU) Time(CPU) ECM. [12] 182,443,267 695 MB 7.08ms 149.39ms ours 19,388 0.079 MB 1.46ms 6.38ms [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 41 canonical work pages

  1. [12]

    IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops2022-June, 675–685 (2022)

    Eyiokur, F.I., Yaman, D., Ekenel, H.K., Waibel, A.: Exposure Correction Model to Enhance Image Quality. IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops2022-June, 675–685 (2022)

    work page 2022

  2. [1]

    Learning Multi-Scale Photo Exposure Correction

    Afifi, M., Derpanis, K.G., Ommer, B., Brown, M.S.: Learning to correct overex- posed and underexposed photos. arXiv preprint arXiv:2003.1159613 (2020)

    work page Pith review arXiv 2020

  3. [2]

    Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition pp

    Afifi, M., Derpanis, K.G., Ommer, B., Brown, M.S.: Learning Multi-Scale Photo Exposure Correction. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition pp. 9153–9163 (2021)

    work page 2021

  4. [3]

    IEEE Transactions on Image Processing14(10), 1647– 1659 (2005)

    Aly, H.A., Dubois, E.: Image up-sampling using total-variation regularization with a new observation model. IEEE Transactions on Image Processing14(10), 1647– 1659 (2005)

    work page 2005

  5. [4]

    IEEE Access10, 57942–57961 (2022)

    Bhattacharya, J., Modi, S., Gregorat, L., Ramponi, G.: D2bgan: A dark to bright image conversion model for quality enhancement and analysis tasks without paired supervision. IEEE Access10, 57942–57961 (2022)

    work page 2022

  6. [5]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Bychkovsky, V., Paris, S., Chan, E., Durand, F.: Learning photographic global tonal adjustment with a database of input/output image pairs. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 97–104. IEEE (2011)

    work page 2011

  7. [6]

    IEEE Transactions on Image Processing27(4), 2049–2062 (2018)

    Cai, J., Gu, S., Zhang, L.: Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing27(4), 2049–2062 (2018)

    work page 2018

  8. [7]

    In: The IEEE Winter Conference on Applications of Computer Vision (WACV)

    Chai, Y., Giryes, R., Wolf, L.: Supervised and unsupervised learning of parame- terized color enhancement. In: The IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 992–1000 (2020)

    work page 2020

Show all 44 references
  1. [8]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Chen,Y.S.,Wang,Y.C.,Kao,M.H.,Chuang,Y.Y.:Deepphotoenhancer:Unpaired learning for image enhancement from photographs with gans. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 6306–6314 (2018)

    2018

  2. [9]

    In: Pro- ceedings of the IEEE conference on computer vision and pattern recognition

    Choi, Y., Choi, M., Kim, M., Ha, J.W., Kim, S., Choo, J.: Stargan: Unified gener- ative adversarial networks for multi-domain image-to-image translation. In: Pro- ceedings of the IEEE conference on computer vision and pattern recognition. pp. 8789–8797 (2018)

    2018

  3. [10]

    John Wiley & Sons (2010)

    Dayley, L.D., Dayley, B.: Photoshop CS5 Bible. John Wiley & Sons (2010)

    2010

  4. [11]

    In: 2018 ACM Multimedia Conference on Multimedia Conference (MM)

    Deng, Y., Loy, C., Tang, X.: Aesthetic-driven image enhancement by adversarial learning. In: 2018 ACM Multimedia Conference on Multimedia Conference (MM). pp. 870–878. ACM (2018)

    2018

  5. [13]

    ACM Transactions on Graphics (TOG) 36(4), 118 (2017)

    Gharbi, M., Chen, J., Barron, J.T., Hasinoff, S.W., Durand, F.: Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (TOG) 36(4), 118 (2017)

    2017

  6. [14]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Guo, C., Li, C., Guo, J., Loy, C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1780–1789 (2020)

    2020

  7. [15]

    In: European Conference on Computer Vision (ECCV)

    He, J., Liu, Y., Qiao, Y., Dong, C.: Conditional sequential modulation for efficient global image retouching. In: European Conference on Computer Vision (ECCV). pp. 679–695. Springer (2020)

    2020

  8. [16]

    Electronics11(1), 32 (2021) 16 Ruodai et al

    Hu, S., Yan, J., Deng, D.: Contextual information aided generative adversarial network for low-light image enhancement. Electronics11(1), 32 (2021) 16 Ruodai et al

    2021

  9. [17]

    IEEE Transactions on Consumer Electronics 53(4), 1752–1758 (2007)

    Ibrahim, H., Kong, N.S.P.: Brightness preserving dynamic histogram equaliza- tion for image contrast enhancement. IEEE Transactions on Consumer Electronics 53(4), 1752–1758 (2007)

    2007

  10. [18]

    In: Proceedings of the IEEE international conference on computer vision

    Ignatov, A., Kobyshev, N., Timofte, R., Vanhoey, K., Van Gool, L.: Dslr-quality photos on mobile devices with deep convolutional networks. In: Proceedings of the IEEE international conference on computer vision. pp. 3277–3285 (2017)

    2017

  11. [19]

    IEEE transactions on image processing30, 2340–2349 (2021)

    Jiang, Y., Gong, X., Liu, D., Cheng, Y., Fang, C., Shen, X., Yang, J., Zhou, P., Wang, Z.: Enlightengan: Deep light enhancement without paired supervision. IEEE transactions on image processing30, 2340–2349 (2021)

    2021

  12. [20]

    In: European Conference on Computer Vision (ECCV)

    Kim, H., Koh, Y., Kim, C.: Global and local enhancement networks for paired and unpaired image enhancement. In: European Conference on Computer Vision (ECCV). pp. 339–354. Springer (2020)

    2020

  13. [21]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Kotovenko, D., Sanakoyeu, A., Lang, S., Ommer, B.: Content and style disentan- glement for artistic style transfer. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4422–4431 (2019)

    2019

  14. [22]

    Josa61(1), 1–11 (1971)

    Land, E.H., McCann, J.J.: Lightness and retinex theory. Josa61(1), 1–11 (1971)

    1971

  15. [23]

    arXiv preprint arXiv:2010.13412 (2020)

    Li, C., Guo, C., Ai, Q., Zhou, S., Loy, C.C.: Flexible piecewise curves estimation for photo enhancement. arXiv preprint arXiv:2010.13412 (2020)

    2020 arXiv

  16. [24]

    arXiv preprint arXiv:2207.14273 (2022)

    Li, C., Guo, C., Feng, R., Zhou, S., Loy, C.C.: Cudi: Curve distillation for efficient and controllable exposure adjustment. arXiv preprint arXiv:2207.14273 (2022)

    2022 arXiv

  17. [25]

    IEEE Transactions on Pattern Analysis and Machine Intel- ligence 44(8), 4225–4238 (2021)

    Li, C., Guo, C., Loy, C.C.: Learning to enhance low-light image via zero-reference deep curve estimation. IEEE Transactions on Pattern Analysis and Machine Intel- ligence 44(8), 4225–4238 (2021)

    2021

  18. [26]

    In: Proceedings of the Asian Conference on Computer Vision (2020)

    Liu, E., Li, S., Liu, S.: Color enhancement using global parameters and local fea- tures learning. In: Proceedings of the Asian Conference on Computer Vision (2020)

    2020

  19. [27]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Liu, R., Ma, L., Zhang, J., Fan, X., Luo, Z.: Retinex-inspired unrolling with coop- erative prior architecture search for low-light image enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10561–10570 (2021)

    2021

  20. [28]

    CoRR abs/2104.06279 (2021)

    Liu, Y., He, J., Chen, X., Zhang, Z., Zhao, H., Dong, C., Qiao, Y.: Very lightweight photo retouching network with conditional sequential modulation. CoRR abs/2104.06279 (2021)

    2021 arXiv

  21. [29]

    Computer Vision and Image Understanding178, 30–42 (2019)

    Loh, Y.P., Chan, C.S.: Getting to know low-light images with the exclusively dark dataset. Computer Vision and Image Understanding178, 30–42 (2019)

    2019

  22. [30]

    In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Moran, S., Marza, P., McDonagh, S., Parisot, S., Slabaugh, G.: Deeplpf: Deep local parametric filters for image enhancement. In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 12826–12835 (2020)

    2020

  23. [31]

    In: 2020 25th International Conference on Pattern Recognition (ICPR)

    Moran, S., McDonagh, S., Slabaugh, G.: Curl: Neural curve layers for global im- age enhancement. In: 2020 25th International Conference on Pattern Recognition (ICPR). pp. 9796–9803. IEEE (2021)

    2021

  24. [32]

    In: Proc

    Nsampi, N.E., Hu, Z., Wang, Q.: Learning exposure correction via consistency modeling. In: Proc. Brit. Mach. Vision Conf. (2021)

    2021

  25. [33]

    Journal of Electronic imaging13(1), 100–110 (2004)

    Rahman, Z.u., Jobson, D.J., Woodell, G.A.: Retinex processing for automatic im- age enhancement. Journal of Electronic imaging13(1), 100–110 (2004)

    2004

  26. [34]

    IEEE Transactions on Image Processing 28(9), 4364–4375 (2019)

    Ren, W., Liu, S., Ma, L., Xu, Q., Xu, X., Cao, X., Du, J., Yang, M.H.: Low- light image enhancement via a deep hybrid network. IEEE Transactions on Image Processing 28(9), 4364–4375 (2019)

    2019

  27. [35]

    Journal of Computer and Communications 7(3), 8–18 (2019) Unsupervised Exposure Correction 17

    Sara, U., Akter, M., Uddin, M.S.: Image quality assessment through fsim, ssim, mse and psnr—a comparative study. Journal of Computer and Communications 7(3), 8–18 (2019) Unsupervised Exposure Correction 17

    2019

  28. [36]

    IEEE Access10, 68281–68290 (2022)

    Soria, X., Pomboza-Junez, G., Sappa, A.D.: Ldc: lightweight dense cnn for edge detection. IEEE Access10, 68281–68290 (2022)

    2022

  29. [37]

    ACM Transactions on Graphics (TOG)30(4), 1–12 (2011)

    Wang, B., Yu, Y., Xu, Y.: Example-based image color and tone style enhancement. ACM Transactions on Graphics (TOG)30(4), 1–12 (2011)

    2011

  30. [38]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Wang, T., Li, Y., Peng, J., Ma, Y., Wang, X., Song, F., Yan, Y.: Real-time im- age enhancer via learnable spatial-aware 3d lookup tables. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 2471–2480 (October 2021)

    2021

  31. [39]

    In: European Conference on Computer Vision

    Wang, Y., Li, X., Xu, K., He, D., Zhang, Q., Li, F., Ding, E.: Neural color operators for sequential image retouching. In: European Conference on Computer Vision. pp. 38–55. Springer (2022)

    2022

  32. [40]

    arXiv preprint arXiv:1808.04560 (2018)

    Wei, C., Wang, W., Yang, W., Liu, J.: Deep retinex decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560 (2018)

    2018 arXiv

  33. [41]

    ACM Transactions on Graphics (TOG)35(2), 11 (2016)

    Yan, Z., Zhang, H., Wang, B., Paris, S., Yu, Y.: Automatic photo adjustment using deep neural networks. ACM Transactions on Graphics (TOG)35(2), 11 (2016)

    2016

  34. [42]

    International Journal of Computer Vision131(4), 1022–1041 (2023)

    Yang, K.F., Cheng, C., Zhao, S.X., Yan, H.M., Zhang, X.S., Li, Y.J.: Learning to adapt to light. International Journal of Computer Vision131(4), 1022–1041 (2023)

    2023

  35. [43]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)

    Zeng, H., Cai, J., Li, L., Cao, Z., Zhang, L.: Learning image-adaptive 3d lookup tables for high-performance photo enhancement in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI) (2020)

    2020

  36. [44]

    Zhang, Y., Zhang, J., Guo, X.: Kindling the darkness: A practical low-light image enhancer.In:Proceedingsofthe27thACMinternationalconferenceonmultimedia. pp. 1632–1640 (2019)

    2019

Pith tools

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