Pith. sign in

REVIEW 4 major objections 4 minor 80 references

UBLLIE: Unified Backlight and Low-Light Image Enhancement

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A single unsupervised network trained with CLIP text prompts matches or beats supervised baselines on both backlit and low-light enhancement.

desk verdict A straightforward CLIP-LIT extension with a residual U-Net and ASPP, but the numerical inconsistencies and the overclaimed 'consistently outperforms' abstract make the empirical claims shaky until fixed. read the letter →

arxiv 2608.04429 v1 pith:UFBCXIQ7 submitted 2026-08-05 cs.CV

Yasmin Yasin , Muhammad Usman , Ibrahim Radwan , Saeed Anwar This is my paper
classification cs.CV
keywords backlitimageenhancementlow-lightunsupervisedlearningCLIPpromptsemanticguidanceresidualU-Netatrousspatialpyramidpoolingrefinement
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 show that one unsupervised model can fix two related but distinct lighting failures—backlit scenes with strong exposure imbalance and low-light scenes that are globally dark—without any paired ground-truth images. The idea is to replace human-labeled training pairs with semantic supervision: CLIP's text encoder produces learned positive and negative prompts ('well-lit' versus poorly lit), and the enhancement network is trained to push its outputs toward the positive prompt and away from the negative one. The paper reports that this recipe, combined with a symmetric residual U-Net and an atrous spatial pyramid pooling module, yields the best fidelity and perceptual scores among all compared methods on BAID, Backlit300, LOL, and VE-LOL-L. If the claim holds, it would mean common lighting defects can be corrected with the same scalable, unpaired training pipeline.

What carries the argument

The load-bearing mechanism is a pair of learned textual prompts, $P^+$ ('well-lit') and $P^-$ ('poorly lit'), embedded with CLIP's frozen text encoder and used as semantic anchors. Training minimizes a contrastive cross-entropy loss on cosine similarities between CLIP image embeddings of the enhanced outputs and the two prompt embeddings, pushing outputs toward $P^+$ and away from $P^-$; a soft ranking loss then refines the prompts using the network's own output, and an identity loss weighted by $\lambda=0.9$ preserves structure. The image generator is a symmetric residual U-Net: three encoder blocks and three decoder blocks with skip connections and residual blocks of two $3\times3$ ReLU convolutions, plus an ASPP module after the second encoder stage with dilation rates 1, 6, 12, and 18 and a global-average-pooling path, giving the network the multi-scale context needed for spatially heterogeneous illumination.

What would settle it

Train the identical pipeline once with the positive prompt replaced by a semantically unrelated caption (for example 'an orange cat') and keep $\lambda=0.9$; if BAID PSNR and SSIM stay close to 22.0 and 0.897, then the 'well-lit' semantics are not what drives the gains, while a collapse would confirm the semantic signal is load-bearing.

Watch

Extended reading notes

Core claim

On the paper's own terms, UBLLIE establishes that a CLIP-guided prompt-learning framework can perform unified, fully unsupervised enhancement of both backlit and low-light images, outperforming state-of-the-art supervised and unsupervised baselines. Concretely, Table 1 reports the best PSNR (22.017), SSIM (0.897), LPIPS (0.153), and MUSIQ (55.691) on the BAID test set, and the best MUSIQ (53.252) on the unpaired Backlit300 set; Table 2 reports the best LOE of 107.45 on LOL and 102.56 on VE-LOL-L together with strong PSNR and SSIM values. The mechanism is two-stage: first initialize positive and negative prompts by contrastive CLIP alignment against well-lit and degraded reference images, train the enhancement network against those prompts, then refine the prompts with a soft ranking loss that orders input, enhanced output, and reference by their similarity to the positive prompt, and fine-tune the network. The paper attributes the gains to a symmetric residual U-Net augmented with an ASPP module, whose multi-scale receptive fields let the network correct uneven illumination without the halos and over-brightening typical of prior methods.

Load-bearing premise

The whole unsupervised scheme depends on CLIP's cosine similarity to a 'well-lit' text prompt being a trustworthy measure of real illumination quality; if that correlation is weak, the prompts can guide the network toward images that score high in CLIP space but are not genuinely better lit.

Editorial extensions

If this is right

  • Because training never uses the available ground-truth pairs, the same pipeline can be applied to datasets where paired references do not exist, such as the real-world Backlit300 set.
  • A single trained model covers both backlit and low-light images, so a deployed system would not need separate models for each degradation or even knowledge of which degradation is present.
  • At inference the network is feed-forward and needs neither CLIP nor prompt embeddings, so the method is lightweight at deployment time.
  • The reported numbers on BAID and Backlit300 give new baselines for backlit enhancement, a task with far fewer established benchmarks than low-light enhancement.
  • Outperforming supervised baselines without labels suggests that semantic text guidance can substitute for curated expert ground truth in illumination correction.

Reading between the lines

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

  • Editorial inference: the same prompt-optimization recipe could be lifted to other spatially varying degradations—haze, rain, shadow, or blur—by swapping the semantic label, since the architecture itself is degradation-agnostic.
  • Editorial inference: the ranking loss uses the network's own output to refine the prompts, so the anchors and the generator can co-adapt; a testable safeguard is to freeze the refined prompts and verify the gains survive, or to check for prompt collapse by measuring the margin between $P^+$ and $P^-$.
  • Editorial inference: most of the no-reference evidence on Backlit300 comes from MUSIQ, so the generalization claim would be strengthened by pairing those scores with human preference ratings or a second no-reference metric.
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 / 4 minor

Summary. The paper proposes UBLLIE, an unsupervised framework for joint backlit and low-light image enhancement. It uses CLIP-guided prompt learning with learned positive and negative text prompts, a symmetric residual U-Net with an ASPP module, and an iterative prompt-refinement procedure based on a ranking loss. Training is unsupervised and uses BAID, DIV2K, LOL, and VE-LOL-L; evaluation is on BAID, Backlit300, LOL, and VE-LOL-L with full-reference and no-reference metrics. The central claim is that the method consistently outperforms supervised and unsupervised baselines in fidelity, perceptual quality, and generalization.

Significance. If the reported results are accurate, the paper would make a useful contribution: a single unsupervised framework that handles both backlit and low-light images, a multi-scale backbone that appears to improve on the CLIP-LIT architecture, and a simple prompt-refinement mechanism. The paper also usefully draws attention to benchmarking gaps for backlit enhancement. However, the paper does not release code or checkpoints, and the empirical support for the headline claim contains several internal inconsistencies that must be resolved before the contribution can be assessed.

major comments (4)
  1. [§5.1, Table 1] The Backlit300 MUSIQ result is reported inconsistently: Section 5.1 states that the method achieves the highest MUSIQ score of 55.05, but Table 1 lists UBLLIE's Backlit300 MUSIQ as 53.252. These values are different and cannot both describe the same experiment; the source of the discrepancy should be identified and the correct number reported.
  2. [Table 2, abstract] On LOL, Table 2 reports UBLLIE SSIM = 0.82, which is lower than TBFEN's 0.833. This directly contradicts the abstract's claim that the framework 'consistently outperforms' state-of-the-art methods in terms of fidelity, if SSIM is included as a fidelity metric, and also conflicts with the statement in Section 5.3 that the model achieves 'consistent improvements' across datasets. The claim should be qualified (e.g., 'competitive SSIM with higher PSNR and lower LOE') or the contradiction resolved.
  3. [Table 1] Several baseline entries in Table 1 are internally inconsistent: Zero-DCE appears twice with different numerical values (19.740/0.871/0.183/51.804 versus 18.553/0.863/0.194/49.436), Zero-DCE++ appears twice with different values, and the retrained SCI row (16.639, 0.768, 0.197) is numerically identical to the Input row. These entries suggest copy-and-paste errors in the table and undermine confidence in the comparison; every row should be verified against the original baselines or their released checkpoints.
  4. [§3.3, Eqs. (3)-(4)] The prompt-refinement mechanism is a potential source of circularity: the ranking loss in Eqs. (3)-(4) uses the enhancement network's own output to update the prompts, and the updated prompts are then used as supervision for the same network. The paper provides no evidence that CLIP similarity to the learned prompts tracks illumination quality rather than unrelated semantic content. A concrete test would be to measure the correlation between the CLIP prompt similarity and human or full-reference quality scores on a held-out set, or to ablate prompt refinement and show that the final metrics change in the expected direction.
minor comments (4)
  1. [§5.1] The text says the method attains the 'highest LOE scores' on LOL and VE-LOL-L, but LOE is an error metric where lower is better; the reported values (107.45 and 102.56) are in fact the lowest among the compared methods, so the wording should be corrected.
  2. [§3.4, Eq. (7)] The text states that λ = 0.9 assigns 'equal importance' to the CLIP loss and the identity loss; an identity loss weighted by 0.9 is not an equal weighting, and the sentence should be rephrased.
  3. [References] References [76] and [77] appear to be the same paper and should be consolidated or distinguished.
  4. [General] No code or checkpoints are released, which makes it impossible for readers to verify the corrected numbers or reproduce the qualitative results; the authors should state their intended release plan.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the training loop is self-referential in an algorithmic sense, but the paper's performance claims are evaluated on external metrics and are not derived from the training losses by construction.

full rationale

The paper's derivation chain is an iterative prompt-learning and enhancement-training procedure, not a logical derivation in which a prediction is defined as its input. Prompt initialization (Eq. 1) fits positive and negative text embeddings P+ and P− to CLIP similarity on reference images; the enhancement network is then trained (Eq. 2) to align its output with P+. Prompt refinement (Eqs. 3–4) uses the current network output I_out, the input I_in, and a reference I_ref to update the prompts, and the final network is trained with the refined prompts plus an identity loss (Eqs. 5–7). This is a co-adaptation loop between the prompt embeddings and the network, but it is not circular in the sense that the final claim reduces to the training objective: the reported quality measures (PSNR, SSIM, LPIPS, MUSIQ, LOE) are all external to the CLIP-based losses, and at inference the prompts and CLIP encoders are not used. The frozen CLIP model is an external pretrained component, and the cited CLIP-LIT prior work is not authored by the present authors, so there is no load-bearing self-citation chain. The paper's tables contain internal inconsistencies (e.g., Section 5.1 reports a Backlit300 MUSIQ of 55.05 while Table 1 lists 53.252; Table 2 shows UBLLIE SSIM 0.82 below TBFEN's 0.833 on LOL; the retrained SCI row in Table 1 nearly equals the Input row). These are empirical correctness and reporting issues, not circularity: they do not show that any output was made equal to its input by definition. Accordingly, the circularity score is 0.

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

The central claim rests on several fitted training components (learned prompts, thresholds, loss weight) and on the unverified assumption that CLIP semantic similarity is a valid supervisory signal for illumination quality. No new physical or conceptual entities are introduced.

free parameters (4)
  • Loss balancing coefficient lambda = 0.9
    Weight on the identity loss in Eq. 7, chosen by hand and following the CLIP-LIT setting, to balance semantic and structural terms.
  • Filtering threshold = 90
    Threshold controlling prompt supervision confidence during prompt learning; value chosen by hand and the mechanism is not explained.
  • Activation threshold = 60
    Threshold determining whether prompt guidance is applied; chosen by hand and the mechanism is not explained.
  • Learned positive and negative prompts P+ and P- = Optimized text-embedding vectors
    These prompts are fitted to reference well-lit and degraded images via the contrastive loss in Eq. 1, and then serve as the training target for the enhancement network. They are parameters fitted to data that define the supervision.
assumptions (4)
  • domain assumption CLIP embeddings align with human perception of lighting quality
    The whole training signal is CLIP similarity to text prompts; the paper provides no evidence that higher similarity to a 'well-lit' prompt corresponds to better PSNR/SSIM or perceptual quality.
  • domain assumption Reference image sets represent the well-lit and poorly-lit domains
    DIV2K images are used as 'well-lit' exemplars and degraded images from BAID/LOL as 'poorly-lit' exemplars. If these sets are unrepresentative, the learned prompts would be anchored incorrectly.
  • domain assumption Ranking order of embedding similarities is a valid training objective
    Equations 3-4 assume that the desired ordering sim(I_ref,P+) > sim(I_out,P+) > sim(I_in,P+) is achievable and beneficial, but the network output I_out feeds back into prompt refinement, creating a moving target.
  • domain assumption Standard full-reference metrics are valid proxies for enhancement quality
    The evaluation claims superiority based on PSNR, SSIM, LPIPS, LOE, and MUSIQ; the paper treats these as ground truth for enhancement quality without questioning their applicability to backlit or low-light images.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UBLLIE: Unified Backlight and Low-Light Image Enhancement." pith.science (2026). https://pith.science/paper/UFBCXIQ7

@misc{pith2026260804429,
  author       = {Pith},
  title        = {Pith review of: UBLLIE: Unified Backlight and Low-Light Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UFBCXIQ7}},
  note         = {Machine review of arXiv:2608.04429}
}
read the original abstract

Backlit and low-light images often suffer from severe exposure imbalance or global underexposure, presenting significant challenges for both visual perception and downstream computer vision tasks. In this paper, we propose a unified, unsupervised enhancement framework that addresses both types of degradation without relying on paired ground-truth data. Our approach builds on CLIP-guided prompt learning to semantically supervise enhancement using learned positive and negative textual prompts. To improve the quality of our improvements over prior work, we design a symmetric residual U-Net backbone augmented with an Atrous Spatial Pyramid Pooling module. This architecture captures multi-scale contextual information, enabling adaptive correction under spatially heterogeneous illumination. During training, the enhancement network is guided by CLIP-based semantic similarity losses and refined via an iterative prompt optimization mechanism. Extensive experiments on both paired and unpaired datasets, including BAID, Backlit300, LOL, and VE-LOL-L, demonstrate that our framework consistently outperforms state-of-the-art supervised and unsupervised methods in terms of fidelity, perceptual quality, and generalization. Furthermore, our work emphasizes the need for stronger benchmarking protocols for backlit enhancement, a relatively underexplored area. The proposed framework provides a robust, scalable solution for real-world illumination enhancement across diverse lighting conditions.

Figures

Figures reproduced from arXiv: 2608.04429 by the authors.

Figure 1
Figure 1. Visual comparison on backlit images sampled from the BAID test dataset. ensuring clearer images that enable vision-based applications [52] to perform accurately across fields such as autonomous driving [31], medical imaging, and security surveillance [53]. Poorly lit images degrade the performance of these algorithms, leading to inefficiencies and errors [54]. Traditional methods, including Histogram Equalization (H… view at source ↗
Figure 2
Figure 2. Overview of our proposed unified enhancement framework. (a) During the first stage, prompts for both well-lit and poorly lit conditions are initialized using CLIP similarity scores, and a preliminary enhancement network is trained to align its outputs with the positive prompt. (b) In the second stage, the prompts are refined using a ranking loss that compares the input, enhanced, and reference image embeddings. The … view at source ↗
Figure 3
Figure 3. Visual comparison on backlit images sampled from the Backlit300 [34] test dataset. 5.2 Qualitative Comparisons Visual comparisons further confirm the superiority of our method. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual comparison of enhancement results on LOL [1] and VE-LOL-L [36] datasets. The top two rows correspond to LOL, and bottom two correspond to VE-LOL￾L. Our method (highlighted in red) achieves better visibility and color restoration while maintaining structural inte…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

80 extracted references · 76 canonical work pages

  1. [1]

    Deep Retinex decomposition for low-light enhancement (2018) 10, 12, 14

    work page 2018

  2. [2]

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

    Afifi, M., Derpanis, K.G., Ommer, B., Brown, M.S.: Learning multi-scale photo exposure correction. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9157–9167 (2021) 2, 11, 13

    work page 2021

  3. [3]

    In: CVPRW

    Agustsson, E., Timofte, R.: Ntire 2017 challenge on single image super-resolution: Dataset and study. In: CVPRW. pp. 6, 8 (July 2017) 10

    work page 2017

  4. [4]

    IEEE Access 11, 121231–121245 (2023) 2

    Akai, M., Ueda, Y., Koga, T., Suetake, N.: Low-artifact and fast backlit image enhancement method based on suppression of lightness order error. IEEE Access 11, 121231–121245 (2023) 2

    work page 2023

  5. [5]

    Signal Processing: Image Communication 100, 116527 (2022) 5

    Al Sobbahi, R., Tekli, J.: Low-light homomorphic filtering network for integrating image enhancement and classification. Signal Processing: Image Communication 100, 116527 (2022) 5

    work page 2022

  6. [6]

    IEEE Transactions on Instrumentation and Measurement (2023) 4

    Bose, S., Nawale, S., Khut, D., Kolekar, M.H.: Luminet: Multi-spatial attention generative adversarial network for backlit image enhancement. IEEE Transactions on Instrumentation and Measurement (2023) 4

    work page 2023

  7. [7]

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

    Cai, Y., Bian, H., Lin, J., Wang, H., Timofte, R., Zhang, Y.: Retinexformer: One- stage retinex-based transformer for low-light image enhancement. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12504–12513 (2023) 3

    work page 2023

  8. [8]

    arXiv preprint arXiv:1706.05587 (2017),https: //arxiv.org/abs/1706.055875, 9

    Chen, L.C., Papandreou, G., Schroff, F., Adam, H.: Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587 (2017),https: //arxiv.org/abs/1706.055875, 9

    arXiv 2017

Show all 80 references
  1. [9]

    illuminate: Zero-shot low- light image enhancement with diffusion prior

    Cho, J., Aghajanzadeh, S., Zhu, Z., Forsyth, D.: Mr. illuminate: Zero-shot low- light image enhancement with diffusion prior. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 8460–8470 (June 2026) 3

    2026

  2. [10]

    IEEE signal processing mag- azine35(1), 53–65 (2018) 3

    Creswell, A., White, T., Dumoulin, V., Arulkumaran, K., Sengupta, B., Bharath, A.A.: Generative adversarial networks: An overview. IEEE signal processing mag- azine35(1), 53–65 (2018) 3

    2018

  3. [11]

    Engineering Applications of Ar- tificial Intelligence116, 105411 (2022) 3

    Cui, H., Li, J., Hua, Z., Fan, L.: Tpet: two-stage perceptual enhancement trans- former network for low-light image enhancement. Engineering Applications of Ar- tificial Intelligence116, 105411 (2022) 3

    2022

  4. [12]

    Engineering Applications of Artificial Intelligence117, 105632 (2023) 2, 5

    Fan,S., Liang,W.,Ding,D.,Yu,H.: Lacn:Alightweightattention-guidedconvnext network for low-light image enhancement. Engineering Applications of Artificial Intelligence117, 105632 (2023) 2, 5

    2023

  5. [13]

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

    Fei, B., Lyu, Z., Pan, L., Zhang, J., Yang, W., Luo, T., Zhang, B., Dai, B.: Genera- tive diffusion prior for unified image restoration and enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9935–9946 (2023) 4

    2023

  6. [14]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Feng, Y., Hou, S., Lin, H., Zhu, Y., Wu, P., Dong, W., Sun, J., Yan, Q., Zhang, Y.: Difflight: Integrating content and detail for low-light image enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. pp. 6143–6152 (2024) 4

    2024

  7. [15]

    In: European conference on computer vi- sion

    Gaintseva, T., Benning, M., Slabaugh, G.: Rave: Residual vector embedding for clip-guided backlit image enhancement. In: European conference on computer vi- sion. pp. 412–428. Springer (2024) 3

    2024

  8. [16]

    In: Proceedings of the UBLLIE: Unified Backlight and Low-Light Image Enhancement 17 IEEE/CVF conference on computer vision and pattern recognition

    Guo, C., Li, C., Guo, J., Loy, C.C., Hou, J., Kwong, S., Cong, R.: Zero-reference deep curve estimation for low-light image enhancement. In: Proceedings of the UBLLIE: Unified Backlight and Low-Light Image Enhancement 17 IEEE/CVF conference on computer vision and pattern recog...

    2020

  9. [17]

    International Journal of Computer Vision131(1), 48–66 (2023) 5

    Guo, X., Hu, Q.: Low-light image enhancement via breaking down the darkness. International Journal of Computer Vision131(1), 48–66 (2023) 5

    2023

  10. [18]

    IEEE Transactions on Image Processing26(2), 982–993 (2016) 1

    Guo, X., Li, Y., Ling, H.: Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on Image Processing26(2), 982–993 (2016) 1

    2016

  11. [19]

    Journal of Visual Communication and Image Representation90, 103712 (2023) 3

    Hai, J., Xuan, Z., Yang, R., Hao, Y., Zou, F., Lin, F., Han, S.: R2rnet: Low- light image enhancement via real-low to real-normal network. Journal of Visual Communication and Image Representation90, 103712 (2023) 3

    2023

  12. [20]

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

    He, C., Zhang, R., Xiao, F., Fang, C., Tang, L., Zhang, R., Farsiu, S.: Unfoldir: Rethinking deep unfolding network in illumination degradation image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5003–5013 (2026) 2

    2026

  13. [21]

    He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 770–778 (2016) 9

    2016

  14. [22]

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

    Hua,Z.,Li,H.,Wang,Y.,Wang,S.,Zhu,C.,Zheng,C.:Bi-bridge:Bidirectionaldif- fusion bridges for low-light image enhancement. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 37455– 37464 (June 2026) 3

    2026

  15. [23]

    ACM Transactions on Graphics (TOG)42(6), 1– 14 (2023) 3, 4

    Jiang, H., Luo, A., Fan, H., Han, S., Liu, S.: Low-light image enhancement with wavelet-based diffusion models. ACM Transactions on Graphics (TOG)42(6), 1– 14 (2023) 3, 4

    2023

  16. [24]

    In: Proceedings of the AAAI conference on artificial intelligence

    Jiang, K., Wang, Z., Wang, Z., Chen, C., Yi, P., Lu, T., Lin, C.W.: Degrade is upgrade: Learning degradation for low-light image enhancement. In: Proceedings of the AAAI conference on artificial intelligence. pp. 1078–1086 (2022) 4

    2022

  17. [25]

    IEEE Transac- tions on Intelligent Transportation Systems23(10), 19440–19455 (2022) 5

    Jiang, Q., Mao, Y., Cong, R., Ren, W., Huang, C., Shao, F.: Unsupervised decom- position and correction network for low-light image enhancement. IEEE Transac- tions on Intelligent Transportation Systems23(10), 19440–19455 (2022) 5

    2022

  18. [26]

    IEEE Transactions on Image Processing30, 2340–2349 (2021) 2, 11, 12, 13

    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) 2, 11, 12, 13

    2021

  19. [27]

    IEEE Transactions on Image processing6(7), 965–976 (1997) 2

    Jobson, D.J., Rahman, Z.u., Woodell, G.A.: A multiscale retinex for bridging the gap between color images and the human observation of scenes. IEEE Transactions on Image processing6(7), 965–976 (1997) 2

    1997

  20. [28]

    IEEE transactions on image processing6(3), 451–462 (1997) 2

    Jobson, D.J., Rahman, Z.u., Woodell, G.A.: Properties and performance of a cen- ter/surround retinex. IEEE transactions on image processing6(3), 451–462 (1997) 2

    1997

  21. [29]

    In: ICCV

    Ke, J., Wang, Q., Wang, Y., Milanfar, P., Yang, F.: Musiq: Multi-scale image quality transformer. In: ICCV. p. 6 (2021) 10

    2021

  22. [30]

    Pattern Recognition p

    Khan, R., Mehmood, A., Shahid, F., Zheng, Z., Ibrahim, M.M.: Lit me up: A reference free adaptive low light image enhancement for in-the-wild conditions. Pattern Recognition p. 110490 (2024) 5

    2024

  23. [31]

    IEEE transactions on pattern analysis and machine intelligence44(12), 9396–9416 (2021) 2, 3

    Li, C., Guo, C., Han, L., Jiang, J., Cheng, M.M., Gu, J., Loy, C.C.: Low-light image and video enhancement using deep learning: A survey. IEEE transactions on pattern analysis and machine intelligence44(12), 9396–9416 (2021) 2, 3

    2021

  24. [32]

    IEEE transactions on pattern analysis and machine intelli- gence44(8), 4225–4238 (2021) 11, 12, 13 18 Yasin et al

    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 intelli- gence44(8), 4225–4238 (2021) 11, 12, 13 18 Yasin et al

    2021

  25. [33]

    Pattern recognition letters104, 15–22 (2018) 12

    Li, C., Guo, J., Porikli, F., Pang, Y.: Lightennet: A convolutional neural network for weakly illuminated image enhancement. Pattern recognition letters104, 15–22 (2018) 12

    2018

  26. [34]

    In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV)

    Liang, Z., Li, C., Zhou, S., Feng, R., Loy, C.C.: Iterative prompt learning for un- supervised backlit image enhancement. In: Proceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV). pp. 8094–8103 (October 2023) 1, 2, 3, 4, 5, 7, 9, 10, 11, 12, 13

    2023

  27. [35]

    IEEE Transactions on Multimedia23, 4272–4284 (2020) 12

    Lim, S., Kim, W.: Dslr: Deep stacked laplacian restorer for low-light image en- hancement. IEEE Transactions on Multimedia23, 4272–4284 (2020) 12

    2020

  28. [36]

    International Journal of Computer Vision129(4), 1153– 1184 (2021) 10, 12, 14

    Liu, J., Xu, D., Yang, W., Fan, M., Huang, H.: Benchmarking low-light image enhancement and beyond. International Journal of Computer Vision129(4), 1153– 1184 (2021) 10, 12, 14

    2021

  29. [37]

    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) 11

    2021

  30. [38]

    Pattern Recognition133, 109039 (2023) 4

    Liu, X., Ma, W., Ma, X., Wang, J.: Lae-net: A locally-adaptive embedding network for low-light image enhancement. Pattern Recognition133, 109039 (2023) 4

    2023

  31. [39]

    Pattern Recognition61, 650–662 (2017) 12

    Lore,K.G.,Akintayo,A.,Sarkar,S.:Llnet:Adeepautoencoderapproachtonatural low-light image enhancement. Pattern Recognition61, 650–662 (2017) 12

    2017

  32. [40]

    IEEE Transactions on Multimedia23, 4093–4105 (2020) 12

    Lu,K.,Zhang,L.:Tbefn:Atwo-branchexposure-fusionnetworkforlow-lightimage enhancement. IEEE Transactions on Multimedia23, 4093–4105 (2020) 12

    2020

  33. [41]

    In: BMVC

    Lv, F., Lu, F., Wu, J., Lim, C.S.: Mbllen: Low-light image/video enhancement using cnns. In: BMVC. p. 220 (2018) 12

    2018

  34. [42]

    Computer Vision and Image Understand- ing218, 103403 (2022) 2, 4, 10, 11, 12

    Lv, X., Zhang, S., Liu, Q., Xie, H., Zhong, B., Zhou, H.: Backlitnet: A dataset and network for backlit image enhancement. Computer Vision and Image Understand- ing218, 103403 (2022) 2, 4, 10, 11, 12

    2022

  35. [43]

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

    Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5637–5646 (2022) 11

    2022

  36. [44]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Morawski, I., He, K., Dangi, S., Hsu, W.H.: Unsupervised image prior via prompt learning and clip semantic guidance for low-light image enhancement. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5971–5981 (2024) 3, 4

    2024

  37. [45]

    IEEE Access (2024) 3

    Park, H., Lim, H., Jang, D.: Dedu: Dual-enhancing dense-unet for lowlight image enhancement and denoise. IEEE Access (2024) 3

    2024

  38. [46]

    In: Proceedings of the International Conference on Machine Learning (ICML) (2021) 4, 6

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: Proceedings of the International Conference on Machine Learning (ICML) (2021) 4, 6

    2021

  39. [47]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR)

    Ren, W., Jinjin, H., Yong, X., Weiming, Z., Tongliang, L., Changwen, L.: Lr3m: Robust reference-free image quality assessment via reinforcement learning. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR). pp. 10599–10608 (2020) 10

    2020

  40. [48]

    In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, Oc- tober 5-9, 2015, Proceedings, Part III

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, Oc- tober 5-9, 2015, Proceedings, Part III. vo...

    2015

  41. [49]

    IEEE Transactions on Computational Imaging (2024) 3

    Shi, P., Xu, X., Fan, X., Yang, X., Xin, Y.: Ll-unet++: Unet++ based nested skip connections network for low-light image enhancement. IEEE Transactions on Computational Imaging (2024) 3

    2024

  42. [50]

    IEEE Sensors Journal (2023) 3

    Tang, G., Ni, J., Chen, Y., Cao, W., Yang, S.X.: An improved cyclegan based model for low-light image enhancement. IEEE Sensors Journal (2023) 3

    2023

  43. [51]

    Sensors23(16), 7306 (2023) 3

    Tian, J., Zhang, J.: A zero-shot low light image enhancement method integrating gating mechanism. Sensors23(16), 7306 (2023) 3

    2023

  44. [52]

    Sensors23(18), 7763 (2023) 2

    Tian, Z., Qu, P., Li, J., Sun, Y., Li, G., Liang, Z., Zhang, W.: A survey of deep learning-based low-light image enhancement. Sensors23(18), 7763 (2023) 2

    2023

  45. [53]

    IEEE Access8, 71940–71950 (2020) 2

    Trongtirakul, T., Chiracharit, W., Agaian, S.S.: Single backlit image enhancement. IEEE Access8, 71940–71950 (2020) 2

    2020

  46. [54]

    In: 2020 IEEE International Conference on Image Processing (ICIP)

    Ueda, Y., Moriyama, D., Koga, T., Suetake, N.: Histogram specification-based image enhancement for backlit image. In: 2020 IEEE International Conference on Image Processing (ICIP). pp. 958–962 (2020) 2

    2020

  47. [55]

    arXiv preprint arXiv:2409.13496 (2024) 3

    Wang, L., Wu, C., Wang, L.: Dap-led: Learning degradation-aware priors with clip for joint low-light enhancement and deblurring. arXiv preprint arXiv:2409.13496 (2024) 3

    2024 arXiv

  48. [56]

    Big Data Mining and Analytics 5(2), 110–119 (2022) 3

    Wang, R., Jiang, B., Yang, C., Li, Q., Zhang, B.: Magan: Unsupervised low-light image enhancement guided by mixed-attention. Big Data Mining and Analytics 5(2), 110–119 (2022) 3

    2022

  49. [57]

    IEEE Transactions on Image Process- ing22(9), 3538–3548 (2013) 10

    Wang, S., Zheng, J., Hu, H.M., Li, B.: Naturalness preserved enhancement algo- rithm for non-uniform illumination images. IEEE Transactions on Image Process- ing22(9), 3538–3548 (2013) 10

    2013

  50. [58]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Wang, T., Zhang, K., Shen, T., Luo, W., Stenger, B., Lu, T.: Ultra-high-definition low-light image enhancement: A benchmark and transformer-based method. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 2654–2662 (2023) 5

    2023

  51. [59]

    In: Proceedings of the AAAI conference on ar- tificial intelligence

    Wang, Y., Wan, R., Yang, W., Li, H., Chau, L.P., Kot, A.: Low-light image en- hancement with normalizing flow. In: Proceedings of the AAAI conference on ar- tificial intelligence. pp. 2604–2612 (2022) 5

    2022

  52. [60]

    IEEE transactions on image processing 13(4), 600–612 (2004) 10

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600–612 (2004) 10

    2004

  53. [61]

    IEEE Transactions on Circuits and Systems for Video Technology (2026) 3

    Wu,H.,Tang,Y.,Li,C.,Jin,Z.:Control-lit:Illuminationcontrollablebacklitimage enhancement. IEEE Transactions on Circuits and Systems for Video Technology (2026) 3

    2026

  54. [62]

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

    Wu,W.,Weng,J.,Zhang,P.,Wang,X.,Yang,W.,Jiang,J.:Uretinex-net:Retinex- based deep unfolding network for low-light image enhancement. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 5901– 5910 (2022) 5, 11

    2022

  55. [63]

    arXiv preprint arXiv:2312.12826 (2023) 3

    Wu, Y., Wang, G., Wang, Z., Yang, Y., Li, T., Wang, P., Li, C., Shen, H.T.: Reco- diff: Explore retinex-based condition strategy in diffusion model for low-light image enhancement. arXiv preprint arXiv:2312.12826 (2023) 3

    2023 arXiv

  56. [64]

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

    Xu, S., Sun, Z., Liu, K., Lu, X., Jiang, R., Fu, X., Zha, Z.J.: Event-illumination col- laborative low-light image enhancement with a high-resolution real-world dataset. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22270–222...

    2026

  57. [65]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition

    Xu, X., Wang, R., Fu, C.W., Jia, J.: Snr-aware low-light image enhancement. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition. pp. 17714–17724 (2022) 11 20 Yasin et al

    2022

  58. [66]

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

    Xu, X., Wang, R., Lu, J.: Low-light image enhancement via structure modeling and guidance. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9893–9903 (2023) 3

    2023

  59. [67]

    arXiv preprint arXiv:2401.03788 (2024) 2, 4

    Xue,M.,He,J.,He,Y.,Liu,Z.,Wang,W.,Zhou,M.:Low-lightimageenhancement via clip-fourier guided wavelet diffusion. arXiv preprint arXiv:2401.03788 (2024) 2, 4

    2024 arXiv

  60. [68]

    IEEE Transactions on Computational Imaging 9, 29–42 (2023) 3

    Yang, S., Zhou, D., Cao, J., Guo, Y.: Lightingnet: An integrated learning method for low-light image enhancement. IEEE Transactions on Computational Imaging 9, 29–42 (2023) 3

    2023

  61. [69]

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

    Yang, S., Ding, M., Wu, Y., Li, Z., Zhang, J.: Implicit neural representation for cooperative low-light image enhancement. In: Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision (ICCV). pp. 12918–12927 (October

  62. [70]

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

    Yang, W., Wang, S., Fang, Y., Wang, Y., Liu, J.: From fidelity to perceptual qual- ity: A semi-supervised approach for low-light image enhancement. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 3063–3072 (2020) 12

    2020

  63. [71]

    In: Proceedings of the 27th ACM international conference on multimedia

    Zhang, L., Zhang, L., Liu, X., Shen, Y., Zhang, S., Zhao, S.: Zero-shot restoration of back-lit images using deep internal learning. In: Proceedings of the 27th ACM international conference on multimedia. pp. 1623–1631 (2019) 2, 11, 12, 13

    2019

  64. [72]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018) 10

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2018) 10

    2018

  65. [73]

    International Journal of Computer Vision129, 1013–1037 (2021) 12

    Zhang, Y., Guo, X., Ma, J., Liu, W., Zhang, J.: Beyond brightening low-light images. International Journal of Computer Vision129, 1013–1037 (2021) 12

    2021

  66. [74]

    Zhang, Y., Zhang, J., Guo, X.: Kindling the darkness: A practical low-light image enhancer.In:Proceedingsofthe27thACMinternationalconferenceonmultimedia. pp. 1632–1640 (2019) 12

    2019

  67. [75]

    In: Proceed- ings of the IEEE/CVF international conference on computer vision

    Zhao, L., Lu, S.P., Chen, T., Yang, Z., Shamir, A.: Deep symmetric network for underexposed image enhancement with recurrent attentional learning. In: Proceed- ings of the IEEE/CVF international conference on computer vision. pp. 12075– 12084 (2021) 11

    2021

  68. [76]

    In: Proceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision (WACV) Workshops

    Zheng, S., Gupta, G.: Semantic-guided zero-shot learning for low-light image/video enhancement. In: Proceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision (WACV) Workshops. pp. 581–590 (January 2022) 3

    2022

  69. [77]

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

    Zheng, S., Gupta, G.: Semantic-guided zero-shot learning for low-light image/video enhancement. In: Proceedings of the IEEE/CVF Winter conference on applications of computer vision. pp. 581–590 (2022) 4

    2022

  70. [78]

    arXiv preprint arXiv:2305.10028 (2023) 3, 4

    Zhou, D., Yang, Z., Yang, Y.: Pyramid diffusion models for low-light image en- hancement. arXiv preprint arXiv:2305.10028 (2023) 3, 4

    2023 arXiv

  71. [79]

    In: 2020 IEEE Inter- national Conference on Multimedia and Expo (ICME)

    Zhu, A., Zhang, L., Shen, Y., Ma, Y., Zhao, S., Zhou, Y.: Zero-shot restoration of underexposed images via robust retinex decomposition. In: 2020 IEEE Inter- national Conference on Multimedia and Expo (ICME). pp. 1–6. IEEE (2020) 12, 13

    2020

  72. [80]

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

    Zhu, Y., Zhang, K., Liang, J., Cao, J., Wen, B., Timofte, R., Van Gool, L.: De- noising diffusion models for plug-and-play image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 1219– 1229 (2023) 3

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.