Pith. sign in

REVIEW 3 major objections 5 minor 70 references

From Enhancement to Understanding: Build a Generalized Bridge for Low-light Vision via Semantically Consistent Unsupervised Fine-tuning

T0 review · 3 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 A single unsupervised enhancer, trained on roughly a thousand unpaired images, lets pretrained daytime models classify, detect, and segment dark scenes directly.

desk verdict Useful evaluation protocol and careful empirical work, but the semantic-consistency mechanism is under-tested and the headline segmentation result rests on 34 images without error bars. read the letter →

arxiv 2507.08380 v1 pith:FULFSNKY submitted 2025-07-11 cs.CV

Sen Wang , Shao Zeng , Tianjun Gu , Zhizhong Zhang , Ruixin Zhang , Shouhong Ding , Jingyun Zhang , Jun Wang
show 3 more authors
Xin Tan Yuan Xie Lizhuang Ma
This is my paper · ORCID
classification cs.CV
keywords low-lightimageenhancementsemanticconsistencyunsupervisedfine-tuningdiffusionmodelszero-shotgeneralizationday-nightdomainadaptationforunderstanding
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 argues that low-light enhancement and high-level visual understanding should be joined by one generalized bridge: rather than training task-specific adapters for dark images, a single enhancer should make dark scenes directly readable by existing pretrained daytime models. It proposes SCUF, a semantically consistent unsupervised fine-tuning method that trains a diffusion-based generator on roughly 1k unpaired low- and normal-light images using a lighten-darken cycle. The central claim is that the enhanced images, tested immediately on pretrained models, outperform both dedicated low-light enhancement methods and zero-shot day-night domain adaptation for classification, face detection, and semantic segmentation. If the claim holds, one enhancement model could serve many downstream tasks without retraining, and low-light evaluation would shift from pixel-level quality to semantic usefulness.

What carries the argument

The load-bearing mechanism is a cycle-consistent lighten-darken generation trained with semantic-consistency losses on unpaired images. The V channel of the HSV representation gives an illumination-aware image prompt, and the cycle-attention adapter queries that prompt through two cross-attention layers so image-level illumination semantics are not reduced to a text description. Caption consistency uses a BLIP-generated caption to align high-level abstract semantics between the lightening and darkening stages, while reflectance consistency uses a decoder to align a RetinexNet reflectance map, preserving spatial structure. LoRA adapters fine-tune the frozen diffusion generator under cycle, identity, adversarial, caption, and reflectance losses.

What would settle it

Run the trained enhancer on held-out extreme-low-light night-driving images with glare and sensor noise, and measure both the cycle reconstruction error and the mIoU of a pretrained segmenter on the enhanced images; if the darkening stage has learned a degenerate inverse, the cycle loss will stay low while enhanced images show blur or color casts and the downstream mIoU gains disappear.

Watch

Extended reading notes

Core claim

The paper's central claim is the GEFU paradigm: enhance a low-light image once, then feed it directly to any pretrained normal-light model, without retraining the downstream task. The proposed SCUF framework realizes this with a cycle generation built on the distilled SD-Turbo diffusion model, where a lightening stage converts a low-light image to normal light and a darkening stage reconstructs the original. An illumination-aware image prompt, obtained from the inverted V channel of the HSV image, is routed through a cycle-attention adapter so that illumination and spatial details guide generation; caption consistency aligns the cycle's latent features with a BLIP caption, and reflectance consistency aligns a RetinexNet-derived reflectance map across the two stages. On CODaN dark-image classification, DARK FACE face detection, and BDD100k-night semantic segmentation, SCUF reports the best direct-test scores among all compared enhancement and day-night adaptation methods (Top-1 60.92%, mAP 16.9%, mIoU 20.1%).

Load-bearing premise

The cycle-consistency objective assumes the darkening stage is a faithful inverse of lightening, so the L1 cycle loss cannot be satisfied by blurry or color-shifted reconstructions, and it assumes RetinexNet's reflectance map is a reliable target for spatial semantic consistency.

Editorial extensions

If this is right

  • One enhanced image can be shared across classification, detection, and segmentation models, replacing per-task day-night domain adaptation with a single preprocessing step.
  • Because the training is unsupervised and only needs roughly 1k unpaired images, adding a new low-light domain reduces to collecting unpaired image pairs, not task labels.
  • Prompt-based diffusion enhancement can be steered by image-level illumination cues rather than by text alone, which is important for scenes text cannot describe.
  • The reported numbers imply that semantic consistency losses are a usable training signal when pixel-level metrics and downstream recognition performance diverge.

Reading between the lines

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

  • If GEFU holds in general, enhancement benchmarks should add semantic metrics; methods that score high on PSNR and SSIM but low on recognition tasks would be revealed as optimizing the wrong objective.
  • The caption- and reflectance-consistency losses are generic semantic-preservation regularizers that could transfer to other unpaired image translation settings, such as dehazing or deraining.
  • A testable extension is to train SCUF on one low-light dataset and evaluate on several unseen night-driving datasets; the paper's Ours-LSRW variant hints that training-data choice measurably changes the gain, so a systematic cross-dataset sweep would pin down the zero-shot boundary.
  • The framework's dependence on cycle reconstruction could be probed by adding glare and sensor-noise degradations at test time, since those break the assumption that darkening inverts lightening.
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

3 major / 5 minor

Summary. The paper proposes a new evaluation paradigm, Generalized Enhancement For Understanding (GEFU), in which low-light images are enhanced and then directly tested with pretrained normal-light models on classification, detection, and semantic segmentation. To instantiate this paradigm, the authors propose Semantically Consistent Unsupervised Fine-tuning (SCUF), an unsupervised cycle-consistency fine-tuning of SD-Turbo. SCUF introduces an illumination-aware image prompt injected through a new Cycle-Attention Adapter, together with caption consistency and reflectance consistency losses intended to preserve high-level and spatial semantics. The method is evaluated on CODaN, DARK FACE, BDD100k-night, LOL, and LSRW, with ablations and comparisons against low-light enhancement and day-night domain adaptation methods.

Significance. If the underlying mechanism holds, GEFU is a useful benchmark and SCUF would demonstrate that a single enhancer trained on about 1k unpaired images can improve multiple downstream visual tasks without task-specific training. The experimental protocol has notable strengths: LightenDiffusion is retrained on the same EnlightenGAN data, the training sets of compared methods are disclosed in the supplementary material, and the ablation study covers adapter placement and individual loss components. The code is promised. However, the load-bearing claim that cycle training enforces semantic consistency is not currently verified, and some headline numbers rely on small or insufficiently qualified comparisons. The paper is a solid candidate after major revision.

major comments (3)
  1. [Sec. 3.4 / Eq. (6) / Algorithm 1] The cycle-consistency objective Eq. (6) can be satisfied by a copy shortcut because the darkening stage is conditioned on the original low-light image. In Algorithm 1, the darkening step for the low-light cycle is I'_l = Dd(U(Ed(În), (Td, Il,v))), with Il,v the V channel of the original Il; the identity pass in Eq. (7) is conditioned on the same Il,v. Thus Lcycle,Il can be driven to zero by copying the illumination structure from the conditioning prompt rather than by checking that În retained the objects, layout, and reflectance needed for classification, detection, or segmentation. The caption and reflectance losses (Eqs. (4)-(5), Algorithm 2) are anchored to the original Il (BLIP caption and RetinexNet reflectance), so they reinforce the same original-image shortcut. The paper reports no cycle-reconstruction fidelity metric on held-out or night-driving images, so this possibility is not ruled out. This point is load-bearing because the paper's central claim is that semantic consistency is learned through unsupervised cycle training. I request: (i) report cycle reconstruction error and a semantic-consistency measure (e.g., caption or detection agreement) between Il and I'_l on held-out night images; (ii) ablate the darkening-conditioning by using V(În) or an illumination estimate from În instead of Il,v in the cycle loss while retaining the identity loss; (iii) show that the downstream gains survive that change.
  2. [Sec. 4.2 / Tables 2 and 3] The BDD100k-night segmentation result rests on only 34 test images. The mIoU differences between the proposed method (20.1) and Sim-MinMax (18.6) or DAI-Net (18.2) are reported without standard deviations, confidence intervals, or significance tests. With 34 images, these differences may be within sampling noise. Because the segmentation result is one of the three headline claims for the GEFU paradigm, the paper should add per-image or bootstrap variance estimates and, ideally, a paired significance test for the comparisons in Tables 2 and 3.
  3. [Abstract / Sec. 4.3 / Table 4] The abstract's claim that the method 'outperforms current state-of-the-art methods in traditional image quality' is not supported by Table 4. On LOL, original LightenDiffusion achieves higher PSNR and SSIM and lower LPIPS than Ours (22.79/0.8540/0.1666 vs. 21.32/0.8073/0.1928). On LSRW, NeRCo has higher PSNR and SSIM (19.46/0.5506) and CIDNet has lower LPIPS (0.2962 vs. 0.2974) than Ours. The row Ours-LSRW is trained on LSRW, the same dataset used for evaluation, so it does not support the zero-shot generalization claim unless a clear train/test split is specified. Please narrow the image-quality claim to the comparisons that are statistically supported, or state explicitly that the claim is relative to baselines retrained on the same training data.
minor comments (5)
  1. [Sec. 4.4 / Table 7] The inference time for Ours in Table 7 (0.348 s) conflicts with the per-task times in Table 2 (0.18/0.91/0.94 s); please specify the resolution, device, and what exactly is being timed.
  2. [Algorithm 1] Line 2 contains a typo: 'draken' should be 'darken'.
  3. [Table 1] The header 'Image Quatily' should be 'Image Quality'.
  4. [Sec. 4.1 / Sec. 3.2] For reproducibility, please specify the SD-Turbo checkpoint, LoRA rank and application points, the reflectance decoder architecture, and the BLIP caption settings; these details are currently missing.
  5. [References] Reference [62] has a malformed author list ('Zhangyang Wang, , and et al.'); please correct it.

Circularity Check

0 steps flagged · score 1.0 of 10

No material circularity: SCUF's downstream gains are measured on external pretrained models and held-out benchmarks; the few self-citations are background references, not load-bearing.

full rationale

Walking the derivation chain, SCUF is trained with unsupervised cycle-consistency, caption-consistency, and reflectance-consistency losses (Eqs. 6-8) on unpaired EnlightenGAN images, and then its enhanced outputs are fed directly to pretrained ResNet-18, YOLOv3/DSFD, and RefineNet models evaluated on CODaN-dark, DARK FACE, and BDD100k-night (Sec. 4.2, Tables 2-3). None of these downstream numbers is produced by fitting a parameter to the benchmark being reported; the pretrained models are fixed and the test sets are external to the training data. The cycle loss (Eq. 6) and identity loss (Eq. 7) are training objectives, not predictions, so the fact that the darkening stage is conditioned on the original V channel (Algorithm 1, lines 5-6) is a potential shortcut or robustness concern about whether semantic consistency is actually enforced, not a case of the claimed result being equivalent to its input by construction. The paper's self-citations ([9], [49], [56], [57]) appear only as background on night-time scene parsing or in the BDD100k-night evaluation setup and do not carry the central claim; no uniqueness theorem or prior work by the same authors is invoked to forbid alternatives or justify the architecture. The GEFU protocol is a benchmark and evaluation naming rather than a derivation that renames a known result into a prediction. Thus the central empirical claims are self-contained against external benchmarks, and no circular step can be exhibited from the paper's equations.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on transfer assumptions from three pretrained models (SD-Turbo, BLIP, RetinexNet) and on the cycle-consistency objective. Hyperparameters are standard and hand-chosen; none are fitted to the downstream benchmarks. No new physical or conceptual entities are introduced beyond the architectural components described in the method.

free parameters (4)
  • loss weight λidt = 0.5
    Hand-chosen balance for identity regularization in Eq. (8); no sensitivity analysis reported.
  • loss weight λGAN = 1.0
    Hand-chosen balance for the adversarial loss in Eq. (8); no sensitivity analysis reported.
  • training iterations = 25,000
    Training duration on the EnlightenGAN dataset; no convergence curve or ablation on this choice is reported.
  • learning rate = 1e-5
    Standard choice for LoRA fine-tuning; not tuned or ablated.
assumptions (5)
  • domain assumption Pretrained SD-Turbo provides a strong semantic prior for low-light enhancement.
    The method relies on the diffusion model's generative prior to produce normal-light images; failure of this prior would collapse the enhancer (Section 3).
  • domain assumption Cycle-consistency with L1 loss preserves image content.
    Eq. (6) assumes that brightening then darkening returns the original image, which is only approximately true for real low-light degradations.
  • domain assumption BLIP-generated captions for dark images are semantically accurate.
    Caption consistency (Section 3.2) uses BLIP captions as the target; wrong captions would enforce incorrect high-level semantics.
  • domain assumption RetinexNet reflectance maps are reliable semantic supervision.
    Reflectance consistency (Section 3.3) uses reflectance maps from a model trained on LOL; errors on unseen domains are not quantified.
  • domain assumption Directly testing on pretrained normal-light models is a valid measure of low-light understanding.
    The GEFU protocol assumes these models are good judges of semantic quality after enhancement; suboptimal models could bias the ranking of enhancers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Enhancement to Understanding: Build a Generalized Bridge for Low-light Vision via Semantically Consistent Unsupervised Fine-tuning." pith.science (2026). https://pith.science/paper/FULFSNKY

@misc{pith2026250708380,
  author       = {Pith},
  title        = {Pith review of: From Enhancement to Understanding: Build a Generalized Bridge for Low-light Vision via Semantically Consistent Unsupervised Fine-tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FULFSNKY}},
  note         = {Machine review of arXiv:2507.08380}
}
read the original abstract

Low-level enhancement and high-level visual understanding in low-light vision have traditionally been treated separately. Low-light enhancement improves image quality for downstream tasks, but existing methods rely on physical or geometric priors, limiting generalization. Evaluation mainly focuses on visual quality rather than downstream performance. Low-light visual understanding, constrained by scarce labeled data, primarily uses task-specific domain adaptation, which lacks scalability. To address these challenges, we build a generalized bridge between low-light enhancement and low-light understanding, which we term Generalized Enhancement For Understanding (GEFU). This paradigm improves both generalization and scalability. To address the diverse causes of low-light degradation, we leverage pretrained generative diffusion models to optimize images, achieving zero-shot generalization performance. Building on this, we propose Semantically Consistent Unsupervised Fine-tuning (SCUF). Specifically, to overcome text prompt limitations, we introduce an illumination-aware image prompt to explicitly guide image generation and propose a cycle-attention adapter to maximize its semantic potential. To mitigate semantic degradation in unsupervised training, we propose caption and reflectance consistency to learn high-level semantics and image-level spatial semantics. Extensive experiments demonstrate that our proposed method outperforms current state-of-the-art methods in traditional image quality and GEFU tasks including classification, detection, and semantic segmentation.

Figures

Figures reproduced from arXiv: 2507.08380 by the authors.

Figure 1
Figure 1. We propose “Generalized Enhancement For Understand [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Compared with other unsupervised low-light enhance [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The framework of SCUF. The low-light image [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Difference between Image Prompt Adapter[ [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of the proposed method with other state-of-the-art methods on nighttime semantic segmentation. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Visual quality comparison between the proposed method and other state-of-the-art methods on the LOL[ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visual analysis of the different components of the pro [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Visual quality comparison between the proposed method and other state-of-the-art methods on the LSRW[ [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Visual quality comparison between the proposed method and other state-of-the-art methods on the LSRW[ [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Visual quality comparison between the proposed method and other state-of-the-art methods on the LOL[ [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Qualitative comparison of the proposed method with other state-of-the-art low-light enhancement methods on night image [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison of the proposed method with other state-of-the-art low-light enhancement methods on dark face detection [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]
Figure 13
Figure 13. Figure 13: Qualitative comparison of the proposed method with other state-of-the-art low-light enhancement methods on dark face detection [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Qualitative comparison of the proposed method with other state-of-the-art low-light enhancement methods on nighttime image [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]
Figure 15
Figure 15. Figure 15: Visual analysis of the different adapters. [PITH_FULL_IMAGE:figures/full_fig_p022_15.png]
Figure 16
Figure 16. Figure 16: Fail cases [PITH_FULL_IMAGE:figures/full_fig_p022_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

70 extracted references · 55 canonical work pages

  1. [1]

    Handbook of image and video processing

    Alan C Bovik. Handbook of image and video processing . Academic press, 2010. 3

    work page 2010

  2. [2]

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

    Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Tim- ofte, and Yulun Zhang. Retinexformer: One-stage retinex- based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12504–12513, 2023. 1, 2, 3, 6, 7, 4

    work page 2023

  3. [3]

    The cityscapes dataset for semantic urban scene understanding

    Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 3213–3223, 2016. 6

    2016

  4. [4]

    Multitask aet with orthogonal tangent regularity for dark object detection

    Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogonal tangent regularity for dark object detection. In Proceedings of the IEEE/CVF international conference on computer vi- sion, pages 2553–2562, 2021. 3, 5

    work page 2021

  5. [5]

    You only need 90k parameters to adapt light: a light weight trans- former for image enhancement and exposure correction

    Ziteng Cui, Kunchang Li, Lin Gu, Shenghan Su, Peng Gao, ZhengKai Jiang, Yu Qiao, and Tatsuya Harada. You only need 90k parameters to adapt light: a light weight trans- former for image enhancement and exposure correction. In 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022. BMV A Press, 2022. 3

    work page 2022

  6. [6]

    Boosting object detection with zero-shot day-night domain adaptation

    Zhipeng Du, Miaojing Shi, and Jiankang Deng. Boosting object detection with zero-shot day-night domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12666–12676, 2024. 1, 2, 3, 5, 6

    work page 2024

  7. [7]

    Yolov3: An incremental improvement

    Ali Farhadi and Joseph Redmon. Yolov3: An incremental improvement. In Computer vision and pattern recognition , pages 1–6. Springer Berlin/Heidelberg, Germany, 2018. 5

    work page 2018

  8. [8]

    Learning a simple low-light image enhancer from paired low-light instances

    Zhenqi Fu, Yan Yang, Xiaotong Tu, Yue Huang, Xinghao Ding, and Kai-Kuang Ma. Learning a simple low-light image enhancer from paired low-light instances. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22252–22261, 2023. 2, 3, 6, 7, 4

    work page 2023

Show all 70 references
  1. [9]

    Scene point cloud understanding and recon- struction technologies in 3d space

    Jingyu Gong, Yujing Lou, Fengqi Liu, Zhiwei Zhang, Haoming Chen, Zhizhong Zhang, Xin Tan, Yuan Xie, and Lizhuang Ma. Scene point cloud understanding and recon- struction technologies in 3d space. Journal of Image and Graphics, 28(06):1741–1766, 2023. 1

    2023

  2. [10]

    Digital image process- ing

    Rafael C Gonzales and Paul Wintz. Digital image process- ing. Addison-Wesley Longman Publishing Co., Inc., 1987. 4

    1987

  3. [11]

    Zero-reference deep curve estimation for low-light image enhancement

    Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 1780–1789, 2020. 1,...

    2020

  4. [12]

    Lime: Low-light im- age enhancement via illumination map estimation

    Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light im- age enhancement via illumination map estimation. IEEE Transactions on image processing, 26(2):982–993, 2016. 3, 6, 7, 4

    2016

  5. [13]

    R2rnet: Low-light image enhancement via real-low to real-normal network

    Jiang Hai, Zhu Xuan, Ren Yang, Yutong Hao, Fengzhu Zou, Fang Lin, and Songchen Han. R2rnet: Low-light image enhancement via real-low to real-normal network. Jour- nal of Visual Communication and Image Representation, 90: 103712, 2023. 1, 7, 4, 5, 6

    2023

  6. [14]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 5

    2016

  7. [15]

    Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation

    Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9924–9935, 2022. 3

    2022

  8. [16]

    Hrda: Context-aware high-resolution domain-adaptive semantic segmentation

    Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Hrda: Context-aware high-resolution domain-adaptive semantic segmentation. In European conference on computer vision , pages 372–391. Springer, 2022. 3

    2022

  9. [17]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 5

    2021 arXiv

  10. [18]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3

    2022

  11. [19]

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

    Hai Jiang, Ao Luo, Xiaohong Liu, Songchen Han, and Shuaicheng Liu. Lightendiffusion: Unsupervised low-light image enhancement with latent-retinex diffusion models. arXiv preprint arXiv:2407.08939, 2024. 2, 6, 7, 4

    2024 arXiv

  12. [20]

    Enlightengan: Deep light enhancement without paired supervision

    Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE transactions on image processing, 30:2340–2349, 2021. 2, 3, 4, 5, 6, 7

    2021

  13. [21]

    Ensembling off-the-shelf models for gan training

    Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 10651–10662, 2022. 5

    2022

  14. [22]

    The retinex theory of color vision

    Edwin H Land. The retinex theory of color vision. Scientific american, 237(6):108–129, 1977. 3, 5

    1977

  15. [23]

    Zero-shot day-night domain adaptation with a physics prior

    Attila Lengyel, Sourav Garg, Michael Milford, and Jan C van Gemert. Zero-shot day-night domain adaptation with a physics prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4399–4409, 2021. 1, 2, 3, 5, 6, 4, 8

    2021

  16. [24]

    Learning to enhance low-light image via zero-reference deep curve esti- mation

    Chongyi Li, Chunle Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve esti- mation. IEEE transactions on pattern analysis and machine intelligence, 44(8):4225–4238, 2021. 2, 3, 6, 7, 4

    2021

  17. [25]

    Dsfd: dual shot face detector

    Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jian- jun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060–5069, 2019. 6

    2019

  18. [26]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna- tional conference on machine learning, pages 12888–12900. PMLR, 2022. 5

    2022

  19. [27]

    Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving

    Jinlong Li, Baolu Li, Zhengzhong Tu, Xinyu Liu, Qing Guo, Felix Juefei-Xu, Runsheng Xu, and Hongkai Yu. Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving. In Proceed- ings of the IEEE/CVF Conference on Computer Visi...

    2024

  20. [28]

    Iterative prompt learning for unsupervised backlit image enhancement

    Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 8094–8103, 2023. 2, 6, 7, 4

    2023

  21. [29]

    Refinenet: Multi-path refinement networks for high- resolution semantic segmentation

    Guosheng Lin, Anton Milan, Chunhua Shen, and Ian Reid. Refinenet: Multi-path refinement networks for high- resolution semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recogni- tion, pages 1925–1934, 2017. 6

    1925

  22. [30]

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

    Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongx- uan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10561–10570, 202...

    2021

  23. [31]

    Getting to know low- light images with the exclusively dark dataset

    Yuen Peng Loh and Chee Seng Chan. Getting to know low- light images with the exclusively dark dataset. Computer Vision and Image Understanding, 178:30–42, 2019. 1

    2019

  24. [32]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. In International Conference on Learning Representations, 2019. 5

    2019

  25. [33]

    Similarity min-max: Zero-shot day-night domain adap- tation

    Rundong Luo, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Similarity min-max: Zero-shot day-night domain adap- tation. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 8104–8114, 2023. 1, 2, 3, 6

    2023

  26. [34]

    Taming diffusion model for exemplar-based image translation

    Hao Ma, Jingyuan Yang, and Hui Huang. Taming diffusion model for exemplar-based image translation. Computational Visual Media, 10(6):1031–1043, 2024. 3

    2024

  27. [35]

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

    Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongx- uan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 5637– 5646, 2022. 1, 2, 3, 6, 7, 4

    2022

  28. [36]

    Histogram-based transformation function estimation for low-light image enhancement

    Jaemin Park, An Gia Vien, Jin-Hwan Kim, and Chul Lee. Histogram-based transformation function estimation for low-light image enhancement. In 2022 IEEE International Conference on Image Processing (ICIP) , pages 1–5. IEEE,

    2022

  29. [37]

    One-step image translation with text-to-image models

    Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. arXiv preprint arXiv:2403.12036, 2024. 5, 7

    2024 arXiv

  30. [38]

    Portraitbooth: A versatile portrait model for fast identity-preserved personalization

    Xu Peng, Junwei Zhu, Boyuan Jiang, Ying Tai, Donghao Luo, Jiangning Zhang, Wei Lin, Taisong Jin, Chengjie Wang, and Rongrong Ji. Portraitbooth: A versatile portrait model for fast identity-preserved personalization. InProceedings of the IEEE/CVF Conference on Computer Vision a...

    2024

  31. [39]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of th...

    2021

  32. [40]

    Hierarchical text-conditional image gener- ation with clip latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image gener- ation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3

    2022 arXiv

  33. [41]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 4

    2022

  34. [42]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 3

    2022

  35. [43]

    Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2250...

    2023

  36. [44]

    Photorealistic text-to-image diffusion models with deep language understanding

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information...

    2022

  37. [45]

    Guided curriculum model adaptation and uncertainty-aware evalua- tion for semantic nighttime image segmentation

    Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Guided curriculum model adaptation and uncertainty-aware evalua- tion for semantic nighttime image segmentation. InProceed- ings of the IEEE/CVF international conference on computer vision, pages 7374–7383, 2019. 3

    2019

  38. [46]

    Adversarial diffusion distillation, 2023

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation, 2023. 3

    2023

  39. [47]

    Adversarial diffusion distillation

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision , pages 87–103. Springer,

  40. [48]

    Zero-ig: Zero-shot illumination-guided joint denoising and adaptive enhancement for low-light images

    Yiqi Shi, Duo Liu, Liguo Zhang, Ye Tian, Xuezhi Xia, and Xiaojing Fu. Zero-ig: Zero-shot illumination-guided joint denoising and adaptive enhancement for low-light images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3015–3024, 20...

    2024

  41. [49]

    Xin Tan, Ke Xu, Ying Cao, Yiheng Zhang, Lizhuang Ma, and Rynson W. H. Lau. Night-time scene parsing with a large real dataset. IEEE Transactions on Image Processing, 30:9085–9098, 2021. 6

    2021

  42. [50]

    Hla-face: Joint high-low adaptation for low light face detection

    Wenjing Wang, Wenhan Yang, and Jiaying Liu. Hla-face: Joint high-low adaptation for low light face detection. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 16195–16204, 2021. 3

    2021

  43. [51]

    Noise4denoise: Leveraging noise for unsupervised point cloud denoising

    Weijia Wang, Xiao Liu, Hailing Zhou, Lei Wei, Zhigang Deng, Manzur Murshed, and Xuequan Lu. Noise4denoise: Leveraging noise for unsupervised point cloud denoising. Computational Visual Media, 10(4):659–669, 2024. 3

    2024

  44. [52]

    Unsupervised illumination adaptation for low-light vi- sion

    Wenjing Wang, Rundong Luo, Wenhan Yang, and Jiaying Liu. Unsupervised illumination adaptation for low-light vi- sion. IEEE Transactions on Pattern Analysis & Machine In- telligence, (01):1–15, 2024. 3

    2024

  45. [53]

    Zero-reference low-light enhancement via physical quadru- ple priors

    Wenjing Wang, Huan Yang, Jianlong Fu, and Jiaying Liu. Zero-reference low-light enhancement via physical quadru- ple priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26057– 26066, 2024. 2, 3, 6, 7, 8, 4

    2024

  46. [54]

    Deep retinex decomposition for low-light enhancement,

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement,

  47. [55]

    Learning semantic-aware knowledge guidance for low-light image enhancement

    Yuhui Wu, Chen Pan, Guoqing Wang, Yang Yang, Jiwei Wei, Chongyi Li, and Heng Tao Shen. Learning semantic-aware knowledge guidance for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 1662–1671, 2023. 3

    2023

  48. [56]

    Boosting night-time scene pars- ing with learnable frequency

    Zhifeng Xie, Sen Wang, Ke Xu, Zhizhong Zhang, Xin Tan, Yuan Xie, and Lizhuang Ma. Boosting night-time scene pars- ing with learnable frequency. IEEE Transactions on Image Processing, 32:2386–2398, 2023. 1, 6

    2023

  49. [57]

    Pig: Prompt images guidance for night-time scene parsing

    Zhifeng Xie, Rui Qiu, Sen Wang, Xin Tan, Yuan Xie, and Lizhuang Ma. Pig: Prompt images guidance for night-time scene parsing. IEEE Transactions on Image Processing, 33: 3921–3934, 2024. 3

    2024

  50. [58]

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

    Qingsen Yan, Yixu Feng, Cheng Zhang, Guansong Pang, Kangbiao Shi, Peng Wu, Wei Dong, Jinqiu Sun, and Yan- ning Zhang. Hvi: A new color space for low-light image enhancement. arXiv preprint arXiv:2502.20272, 2025. 2, 6, 7, 4

    2025 arXiv

  51. [59]

    Learnable differencing center for night- time depth perception

    Zhiqiang Yan, Yupeng Zheng, Deng-Ping Fan, Xiang Li, Jun Li, and Jian Yang. Learnable differencing center for night- time depth perception. Visual Intelligence, 2(1):15, 2024. 3

    2024

  52. [60]

    Wider face: A face detection benchmark

    Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaoou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE conference on computer vision and pattern recog- nition, pages 5525–5533, 2016. 5

    2016

  53. [61]

    Implicit neural representation for coopera- tive low-light image enhancement

    Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for coopera- tive low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision , pages 12918–12927, 2023. 2, 3, 6, 7, 4

    2023

  54. [62]

    Scheirer, Zhangyang Wang, , and et al

    Wenhan Yang, Ye Yuan, Wenqi Ren, Jiaying Liu, Walter J. Scheirer, Zhangyang Wang, , and et al. Advancing image understanding in poor visibility environments: A collective benchmark study. IEEE Transactions on Image Processing, 29:5737–5752, 2020. 1, 2, 5, 6, 4, 9, 10

    2020

  55. [63]

    Sparse gradient regularized deep retinex network for robust low-light image enhancement

    Wenhan Yang, Wenjing Wang, Haofeng Huang, Shiqi Wang, and Jiaying Liu. Sparse gradient regularized deep retinex network for robust low-light image enhancement. IEEE Transactions on Image Processing, 30:2072–2086, 2021. 7, 4

    2021

  56. [64]

    Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip- adapter: Text compatible image prompt adapter for text-to- image diffusion models. arXiv preprint arXiv:2308.06721,

    arXiv

  57. [65]

    Bdd100k: A diverse driving dataset for heterogeneous multitask learning

    Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Dar- rell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition ...

    2020

  58. [66]

    Dual illu- mination estimation for robust exposure correction

    Qing Zhang, Yongwei Nie, and Wei-Shi Zheng. Dual illu- mination estimation for robust exposure correction. In Com- puter graphics forum, pages 243–252. Wiley Online Library,

  59. [67]

    Kindling the darkness: A practical low-light image enhancer

    Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Pro- ceedings of the 27th ACM international conference on mul- timedia, pages 1632–1640, 2019. 3

    2019

  60. [68]

    Dsd-matchingnet: Deformable sparse-to-dense feature matching for learning accurate correspondences

    Yicheng Zhao, Han Zhang, Ping Lu, Ping Li, Enhua Wu, and Bin Sheng. Dsd-matchingnet: Deformable sparse-to-dense feature matching for learning accurate correspondences. Vir- tual Reality & Intelligent Hardware, 4(5):432–443, 2022. 3

    2022

  61. [69]

    Learning semantic degradation-aware guidance for recognition-driven unsupervised low-light im- age enhancement

    Naishan Zheng, Jie Huang, Man Zhou, Zizheng Yang, Qi Zhu, and Feng Zhao. Learning semantic degradation-aware guidance for recognition-driven unsupervised low-light im- age enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3678–3686, 2023. 2, 6, 7, 4

    2023

  62. [70]

    Unpaired image-to-image translation using cycle- consistent adversarial networks

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle- consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision , pages 2223– 2232, 2017. 5 A. Implementation details A.1. Hyp...

    2017

Pith tools

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