Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Learning to See in the Extremely Dark

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a synthesis pipeline can create calibrated RAW images down to 0.0001 lux and that a diffusion model with adaptive illumination correction restores them to state-of-the-art quality.

desk verdict A useful first 0.0001-lux benchmark and a solid diffusion baseline, but the dataset's 'well-calibrated' claim leans on histogram matching and a hand-fitted factor; the realism evidence needs to be much stronger before this becomes a standard testbed. read the letter →

arxiv 2506.21132 v2 pith:LJHLXE6H submitted 2025-06-26 cs.CV

Hai Jiang , Binhao Guan , Zhen Liu , Xiaohong Liu , Jian Yu , Zheng Liu , Songchen Han , Shuaicheng Liu This is my paper
classification cs.CV
keywords extremelylow-lightenhancementRAWimagerestorationpaireddatasetsynthesisilluminancecalibrationdiffusionmodelsadaptiveilluminationcorrectionnoisemodelingcolorconsistencyloss
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to push learning-based low-light RAW enhancement into truly extreme darkness, where illuminance falls to 0.0001 lux, a regime existing datasets cannot capture because long exposures would blur or remain noisy. To do this it builds a paired-to-paired synthesis pipeline that aligns captured short-exposure RAW images to three precisely measured illuminance ranges and adds a calibrated Gaussian-plus-Poisson-plus-dark-frame noise model, producing the See-in-the-Extremely-Dark (SIED) dataset with high-quality sRGB references. It also introduces a diffusion-based enhancement framework with an Adaptive Illumination Correction Module and a color consistency loss. The paper reports that this method outperforms prior single-stage and multi-stage approaches on both SIED and the existing SID benchmark across distortion and perceptual metrics.

What carries the argument

The load-bearing object is the paired-to-paired synthesis pipeline: illumination alignment via $I_{syn} = I_{cap}\cdot(\mathrm{Expo}(I_{st})/\mathrm{Expo}(I_{cap}) + \eta)$, where $\eta$ is hand-fitted to match illumination histograms (KL divergence below 0.06), followed by an ISO-dependent calibrated noise model that combines Gaussian, Poisson, and dark-frame components so the synthetic RAW matches real sensor statistics. The enhancement side rests on a latent-space diffusion model where an Adaptive Illumination Correction Module (AICM) estimates per-channel amplification coefficients from the RAW feature itself, and a color consistency loss aligns the color histogram of the generated sRGB feature with the reference.

What would settle it

Capture a static scene under a controlled light source metered at 0.001 lux with a Sony or Canon camera using a very long exposure to obtain a near-clean reference, then compare the noise statistics (per-channel variance versus intensity, dark-frame histograms, and illumination histograms) of a real short-exposure capture at the same lux against a SIED-synthesized image; if the variance-versus-intensity slope or the histogram KL divergence exceeds the reported tolerance of about 0.06, the calibrated synthesis premise fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that extremely dark RAW images, at illuminances as low as 0.0001 lux, can be realistically synthesized from paired real captures by scaling intensities with a measured exposure ratio plus a manually fitted correction factor, then adding sensor noise calibrated in an optical laboratory, and that a diffusion model guided by an adaptive illumination-corrected RAW feature can restore such inputs to visually pleasing sRGB images. The claimed result is state-of-the-art performance on the new SIED dataset and on the existing SID dataset, with the adaptive module removing the need to know the reference image's exposure during inference.

Load-bearing premise

The synthesized extremely dark RAW images faithfully replicate what a real camera would record at 0.0001 to 0.1 lux, which depends on the exposure-ratio scaling plus a hand-fitted factor and the calibrated Gaussian, Poisson, and dark-frame noise model matching true sensor behavior.

Editorial extensions

If this is right

  • Researchers gain a benchmark with three precisely calibrated illuminance ranges down to 0.0001 lux, enabling fair comparison of methods under extreme low-light conditions that previous datasets did not cover.
  • The adaptive illumination correction removes the dependence on ground-truth exposure values during inference, a limitation of earlier pre-amplification approaches such as SID and DNF.
  • The diffusion framework trained on SIED generalizes to real-world extremely dark scenes, as demonstrated by qualitative results on images whose illuminance is matched to laboratory standards.
  • The combination of a generative diffusion prior with explicit color-histogram alignment offers a recipe for other RAW-to-sRGB tasks where both fidelity and vivid color are required.
  • Multi-stage training, first for encoders/decoders and AICM then for the diffusion model, proves important for stable convergence and final quality.

Reading between the lines

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

  • The synthesis pipeline may transfer to other sensors beyond Sony and Canon, provided the noise model can be recalibrated for each sensor's photon transfer curve and dark-frame characteristics; the paper does not report cross-camera noise calibration.
  • Because the validation relies on synthetic images as ground truth, a stronger test would be capturing truly long-exposure references at the same extreme illuminance levels, something the paper argues is impractical; an indirect test would compare synthesized noise statistics against real dark-frame captures at the same ISO.
  • The hand-fitted factor $\eta$ used in illumination alignment is sensor- and scene-dependent; a fully automatic estimation, perhaps from the histogram matching itself, could make the pipeline more general and remove a potential source of bias.
  • The color consistency loss operates on global histograms and may not preserve spatial color fidelity; a spatial color-aware variant could further improve results in scenes with localized colored lighting.
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 / 6 minor

Summary. The paper introduces SIED, a paired dataset for extremely low-light RAW image enhancement, containing synthesized RAW images at three nominal illuminance ranges (0.01–0.1 lux, 0.001–0.01 lux, and 0.0001–0.001 lux) together with high-quality sRGB references. The synthesis pipeline first captures low-light and normal-light image pairs in real scenes, then scales the low-light RAW intensity by an exposure ratio plus a hand-fitted factor η (Eq. 1), and finally adds an ISO-dependent calibrated Gaussian-plus-Poisson-plus-dark-frame noise model. The paper also proposes a diffusion-based restoration method with an Adaptive Illumination Correction Module (AICM) and a color consistency loss. The method is evaluated on SIED and on the SID benchmark, where the authors report state-of-the-art results and show qualitative comparisons on real extremely dark scenes.

Significance. If the SIED calibration is trustworthy, the dataset would fill a real gap: no existing benchmark covers illuminance down to 0.0001 lux with paired references. The paper is commendable for retraining all comparison methods on SIED using official code, for evaluating on the independent SID benchmark, and for releasing code and dataset. The proposed diffusion method is technically sound and shows consistent gains on SID, which is independent evidence of its effectiveness. However, the central value of SIED rests on the realism of its synthetic low-light RAW images, and that premise is currently under-validated: the only quantitative calibration evidence is a Y-channel histogram KL divergence (Table 1), and the noise model is described qualitatively without reported parameters or direct validation against real captures at the target ISOs and illuminances. The real-world generalization test in Sec. 5.2 is also self-referential because it uses the same histogram-matching procedure to assign illuminance.

major comments (3)
  1. [§3, Illumination Alignment (Eq. 1)]
  2. [§3, Noise Addition]
  3. [§5.2, Real-world Generalization]
minor comments (6)
  1. [Table 3]
  2. [Eq. (3)]
  3. [§3, Illumination Alignment]
  4. [Fig. 2]
  5. [General]
  6. [Dataset documentation]

Circularity Check

2 steps flagged · score 6.0 of 10

SIED's 'well-calibrated' and real-world generalization claims are validated by the same histogram-matching procedure used to fit the synthesis; the enhancement method retains independent SID support.

  1. fitted input called prediction [Section 3, Illumination Alignment and Noise Addition (Eq. 1, Table 1)]
    "To refine the illumination of generated images, we adopt a fixed ISP pipeline to transform the generated RAW image and standard images into YUV space, and manually finetune η to match the illumination histograms of the two images in the Y channel, which represents the illumination information of images [8]. As shown in Table 1, the mean KL divergence of the histogram distributions between our generated images and standard laboratory images is less than 0.06 across three illuminance levels."

    The only quantitative evidence for the dataset's central 'well-calibrated ... 0.0001 lux' claim is Table 1, a KL divergence between synthesized and laboratory images. But η in Eq. (1) is the free parameter of the synthesis, and the text states it was manually finetuned to minimize exactly this Y-channel histogram mismatch. The validation metric is therefore the fitting objective, so low KL values are enforced by construction rather than discovered. This is a fitted parameter being offered as independent calibration evidence: the 'precise illuminance ranges' are not verified by direct measurement of the synthesized images, only by agreement with the images used to tune the synthesizer.

  2. self definitional [Section 5.2, Real-world Generalization]
    "Real-world Generalization. To validate the generalization ability of our method, we collect several images in realistic extremely dark scenes using the Sony camera and determine their illuminance level according to the matching of the illumination histograms in the Y-channel to standard laboratory images."

    The illuminance labels for the real-world test scenes are assigned by the same Y-channel histogram-matching procedure used in Sec. 3 to define and synthesize the SIED illuminance ranges (finetuning η to match the same standard lab histograms). Therefore the real-world test is not an external check: it re-applies the calibration criterion to select and label inputs. The claim that the dataset supports real-world generalization is thus supported only by consistency with the same matching rule, not by independent ground-truth illuminance or noise measurements.

full rationale

The paper contains two distinct claims: (1) the SIED dataset provides well-calibrated extremely low-light RAW images, and (2) the diffusion method achieves state-of-the-art restoration. Claim (2) is supported by supervised experiments on SIED and on the external SID benchmark, where comparison methods are retrained with official code, so the enhancement method itself is not circular. Claim (1), however, contains two circular validation steps. The illuminance alignment uses Eq. (1) with a hand-fitted η, and η is manually finetuned to minimize the Y-channel histogram KL divergence between synthesized and standard lab images; the same KL divergence is then reported in Table 1 as evidence of calibration. This is a fitted parameter being offered as validation, so the low KL values are enforced by construction. Similarly, the real-world generalization test labels real scenes by the same Y-channel histogram-matching criterion used in synthesis, making the test self-referential rather than an independent test of physical realism. The calibrated noise model and the optical-laboratory standard captures provide some independent grounding, and the SID results are external, so the circularity is partial rather than total. Score 6 reflects that the central dataset-calibration claim reduces by construction even though the enhancement method itself is not circular.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the calibrated noise model, the hand-fitted illumination factor eta, and the histogram-matching criterion. No new physical entities are introduced. The learned network weights in the encoders, AICM, and diffusion model are supervised and fitted to data, which is standard practice and not listed as free parameters.

free parameters (2)
  • eta (illumination alignment factor) = not explicitly reported; manually defined and finetuned per illuminance range
    In Eq. (1), eta is introduced to align the illumination of captured images to the laboratory standard. It is manually finetuned by matching YUV Y-channel histograms (Sec. 3, Illumination Alignment), so the 'well-calibrated' claim depends on this hand-fitted factor.
  • Sensor noise model parameters = not reported in the paper
    Gaussian, Poisson, and dark-frame noise parameters are estimated per camera and ISO in the optical laboratory (Sec. 3, Noise Addition). These numbers are fitted to measured noise and are central to the realism of the synthetic extreme-low-light inputs.
assumptions (5)
  • standard math Diffusion forward/reverse process with a fixed variance schedule and DDIM sampling
    The diffusion formulation in Eqs. (3) and (4) follows standard DDPM/DDIM theory from Ho et al. and Song et al., cited in Sec. 4.3.
  • domain assumption An image can be decomposed into illumination and reflectance maps
    The illumination correction loss in Eq. (2) relies on Retinex decomposition from LIME (Guo et al. 2016). This is a modeling assumption about the physical image formation.
  • domain assumption Extreme-low-light RAW noise is additive Gaussian plus Poisson plus dark-frame distribution
    Sec. 3, Noise Addition, justifies the noise model by citing Wei et al. and Zhang et al., but the specific applicability to 0.0001 lux is assumed.
  • domain assumption Scaling RAW intensities by an exposure ratio approximates changing scene illuminance
    Eq. (1) assumes that the mean Bayer channel value is a surrogate for illuminance and that linear scaling of RAW values reproduces the illumination levels of a different scene. This is standard photometric modeling but untested for extreme darkness.
  • domain assumption Matching YUV Y-channel histograms is a sufficient proxy for illumination realism
    Used in calibration (Table 1) and in assigning illuminance levels to real-world test images (Sec. 5.2). The histogram distance is a weak proxy that ignores chroma and noise texture.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning to See in the Extremely Dark." pith.science (2026). https://pith.science/paper/LJHLXE6H

@misc{pith2026250621132,
  author       = {Pith},
  title        = {Pith review of: Learning to See in the Extremely Dark},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LJHLXE6H}},
  note         = {Machine review of arXiv:2506.21132}
}
read the original abstract

Learning-based methods have made promising advances in low-light RAW image enhancement, while their capability to extremely dark scenes where the environmental illuminance drops as low as 0.0001 lux remains to be explored due to the lack of corresponding datasets. To this end, we propose a paired-to-paired data synthesis pipeline capable of generating well-calibrated extremely low-light RAW images at three precise illuminance ranges of 0.01-0.1 lux, 0.001-0.01 lux, and 0.0001-0.001 lux, together with high-quality sRGB references to comprise a large-scale paired dataset named See-in-the-Extremely-Dark (SIED) to benchmark low-light RAW image enhancement approaches. Furthermore, we propose a diffusion-based framework that leverages the generative ability and intrinsic denoising property of diffusion models to restore visually pleasing results from extremely low-SNR RAW inputs, in which an Adaptive Illumination Correction Module (AICM) and a color consistency loss are introduced to ensure accurate exposure correction and color restoration. Extensive experiments on the proposed SIED and publicly available benchmarks demonstrate the effectiveness of our method. The code and dataset are available at https://github.com/JianghaiSCU/SIED.

Figures

Figures reproduced from arXiv: 2506.21132 by the authors.

Figure 1
Figure 1. An illustration of our synthesized low-light RAW im [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) illustrates the optical laboratory, collected standard RAW images, and camera settings adopted for image capturing. (b) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The overall pipeline of our proposed framework. We first employ a RAW encoder [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Examples from the Canon and Sony subsets of our SIED [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: The detailed architecture of our proposed AICM. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of our method and competitive methods [ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison of our method and competitive methods [ [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Qualitative comparison of our method and competitive [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. AU-IQA: A Benchmark Dataset for Perceptual Quality Assessment of AI-Enhanced User-Generated Content

    cs.CV 2025-08 conditional novelty 6.0 of 10

    AU-IQA is a new 4,800-image benchmark showing existing quality models, mainly those trained on ordinary user content, only partially predict human ratings of AI-enhanced photos.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages · cited by 1 Pith paper

  1. [1]

    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 ICCV, pages 12504–12513, 2023. 1

    work page 2023

  2. [2]

    Learning to see in the dark

    Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In CVPR, pages 3291–3300,

    work page

  3. [3]

    Seeing motion in the dark

    Chen Chen, Qifeng Chen, Minh N Do, and Vladlen Koltun. Seeing motion in the dark. InICCV, pages 3185–3194, 2019. 2, 3

    work page 2019

  4. [4]

    Retinex-RAWMamba: Bridging Demosaicing and Denoising for Low-Light RAW Image Enhancement

    Xianmin Chen, Peiliang Huang, Xiaoxu Feng, Dingwen Zhang, Longfei Han, and Junwei Han. Retinex-rawmamba: Bridging demosaicing and denoising for low-light raw image enhancement. arXiv preprint arXiv:2409.07040, 2024. 1, 2, 3, 5, 6, 7, 8

    work page Pith review arXiv 2024

  5. [5]

    Come-closer-diffuse-faster: Accelerating conditional diffu- sion models for inverse problems through stochastic contrac- tion

    Hyungjin Chung, Byeongsu Sim, and Jong Chul Ye. Come-closer-diffuse-faster: Accelerating conditional diffu- sion models for inverse problems through stochastic contrac- tion. In CVPR, pages 12413–12422, 2022. 5

    work page 2022

  6. [6]

    Abandoning the bayer-filter to see in the dark

    Xingbo Dong, Wanyan Xu, Zhihui Miao, Lan Ma, Chao Zhang, Jiewen Yang, Zhe Jin, Andrew Beng Jin Teoh, and Jiajun Shen. Abandoning the bayer-filter to see in the dark. In CVPR, pages 17431–17440, 2022. 1, 2, 6, 7

    work page 2022

  7. [7]

    Self-guided network for fast image denoising

    Shuhang Gu, Yawei Li, Luc Van Gool, and Radu Timofte. Self-guided network for fast image denoising. In ICCV, pages 2511–2520, 2019. 1, 2, 6, 7

    work page 2019

  8. [8]

    Low-light image enhancement via breaking down the darkness

    Xiaojie Guo and Qiming Hu. Low-light image enhancement via breaking down the darkness. IJCV, 131(1):48–66, 2023. 1, 4

    work page 2023

Show all 43 references
  1. [9]

    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 TIP, 26(2):982–993, 2016. 5

    2016

  2. [10]

    Fuzzy color histogram and its use in color image retrieval.IEEE TIP, 11(8):944–952, 2002

    Ju Han and Kai-Kuang Ma. Fuzzy color histogram and its use in color image retrieval.IEEE TIP, 11(8):944–952, 2002. 5

    2002

  3. [11]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. In NeurIPS, pages 6840–6851,

  4. [12]

    Global structure-aware diffusion pro- cess for low-light image enhancement

    Jinhui Hou, Zhiyu Zhu, Junhui Hou, Hui Liu, Huanqiang Zeng, and Hui Yuan. Global structure-aware diffusion pro- cess for low-light image enhancement. In NeurIPS, pages 79734–79747, 2023. 1

    2023

  5. [13]

    Towards low light enhancement with raw images

    Haofeng Huang, Wenhan Yang, Yueyu Hu, Jiaying Liu, and Ling-Yu Duan. Towards low light enhancement with raw images. IEEE TIP, 31:1391–1405, 2022. 2

    2022

  6. [14]

    Low-light image enhancement with wavelet-based diffusion models

    Hai Jiang, Ao Luo, Songchen Han, Haoqiang Fan, and Shuaicheng Liu. Low-light image enhancement with wavelet-based diffusion models. ACM TOG , 42(6):1–14,

  7. [15]

    Dnf: Decouple and feedback network for seeing in the dark

    Xin Jin, Ling-Hao Han, Zhen Li, Chun-Le Guo, Zhi Chai, and Chongyi Li. Dnf: Decouple and feedback network for seeing in the dark. In CVPR, pages 18135–18144, 2023. 1, 2, 3, 5, 6, 7, 8

    2023

  8. [16]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 ,

    arXiv

  9. [17]

    Restoring extremely dark images in real time

    Mohit Lamba and Kaushik Mitra. Restoring extremely dark images in real time. In CVPR, pages 3487–3497, 2021. 1, 2, 6, 7

    2021

  10. [18]

    Towards fast and light-weight restoration of dark images

    Mohit Lamba, Atul Balaji, and Kaushik Mitra. Towards fast and light-weight restoration of dark images. arXiv preprint arXiv:2011.14133, 2020. 1, 2, 6, 7

    2011 arXiv

  11. [19]

    The retinex theory of color vision

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

    1977

  12. [20]

    Embedding fourier for ultra-high-definition low-light image enhancement

    Chongyi Li, Chun-Le Guo, Man Zhou, Zhexin Liang, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Embedding fourier for ultra-high-definition low-light image enhancement. In ICLR, 2023. 1

    2023

  13. [21]

    Dmhomo: Learning ho- mography with diffusion models

    Haipeng Li, Hai Jiang, Ao Luo, Ping Tan, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. Dmhomo: Learning ho- mography with diffusion models. ACM TOG, 43(3):1–16,

  14. [22]

    Flowdiffuser: Advancing optical flow es- timation with diffusion models

    Ao Luo, Xin Li, Fan Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Flowdiffuser: Advancing optical flow es- timation with diffusion models. In CVPR, pages 19167– 19176, 2024. 5

    2024

  15. [23]

    Improving extreme low-light image denoising via residual learning

    Paras Maharjan, Li Li, Zhu Li, Ning Xu, Chongyang Ma, and Yue Li. Improving extreme low-light image denoising via residual learning. In ICME, pages 916–921, 2019. 1, 2, 5, 6, 7

    2019

  16. [24]

    Elucidating the exposure bias in diffusion models

    Mang Ning, Mingxiao Li, Jianlin Su, Albert Ali Salah, and Itir Onal Ertugrul. Elucidating the exposure bias in diffusion models. In ICLR, 2024. 5

    2024

  17. [25]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, pages 234–241, 2015. 6

    2015

  18. [26]

    Human motion diffusion as a generative prior

    Yoni Shafir, Guy Tevet, Roy Kapon, and Amit Haim Bermano. Human motion diffusion as a generative prior. In ICLR, 2022. 5

    2022

  19. [27]

    Denois- ing diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denois- ing diffusion implicit models. In ICLR, 2021. 5, 6

    2021

  20. [28]

    Seeing dynamic scene in the dark: A high- quality video dataset with mechatronic alignment

    Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high- quality video dataset with mechatronic alignment. In ICCV, pages 9700–9709, 2021. 2, 3

    2021

  21. [29]

    Practical deep raw image denoising on mobile devices

    Yuzhi Wang, Haibin Huang, Qin Xu, Jiaming Liu, Yiqun Liu, and Jue Wang. Practical deep raw image denoising on mobile devices. In ECCV, pages 1–16, 2020. 4

    2020

  22. [30]

    Low-light image enhancement with illumination-aware gamma correction and complete image modelling network

    Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In ICCV, pages 13128–13137, 2023. 1

    2023

  23. [31]

    Exposurediffusion: Learning to expose for low-light image enhancement

    Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. Exposurediffusion: Learning to expose for low-light image enhancement. In ICCV, pages 12438–12448, 2023. 1

    2023

  24. [32]

    Bovik, H.R

    Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600–612, 2004. 7

    2004

  25. [33]

    Deep retinex decomposition for low-light enhancement

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement. In BMVC, 2018. 1

    2018

  26. [34]

    Physics-based noise modeling for extreme low-light photog- raphy

    Kaixuan Wei, Ying Fu, Yinqiang Zheng, and Jiaolong Yang. Physics-based noise modeling for extreme low-light photog- raphy. IEEE TPAMI, 44(11):8520–8537, 2021. 3

    2021

  27. [35]

    Ldm-isp: Enhancing neural isp for low light with latent dif- fusion models

    Qiang Wen, Yazhou Xing, Zhefan Rao, and Qifeng Chen. Ldm-isp: Enhancing neural isp for low light with latent dif- fusion models. arXiv preprint arXiv:2312.01027, 2023. 1

    2023 arXiv

  28. [36]

    Learning to restore low-light images via decomposition-and- enhancement

    Ke Xu, Xin Yang, Baocai Yin, and Rynson WH Lau. Learning to restore low-light images via decomposition-and- enhancement. In CVPR, pages 2281–2290, 2020. 2, 6, 7

    2020

  29. [37]

    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 TIP, 30:2072–2086, 2021. 1

    2021

  30. [38]

    Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model

    Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In ICCV, pages 12302– 12311, 2023. 1

    2023

  31. [39]

    The unreasonable effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586–595,

  32. [40]

    Beyond brightening low-light images

    Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. IJCV, 129: 1013–1037, 2021. 1

    2021

  33. [41]

    Rethinking noise synthesis and modeling in raw denois- ing

    Yi Zhang, Hongwei Qin, Xiaogang Wang, and Hongsheng Li. Rethinking noise synthesis and modeling in raw denois- ing. In ICCV, pages 4593–4601, 2021. 3, 4

    2021

  34. [42]

    Pyramid diffusion models for low-light image enhancement

    Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. In IJCAI, 2023. 1

    2023

  35. [43]

    Eemefn: Low-light image enhancement via edge-enhanced multi- exposure fusion network

    Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Eemefn: Low-light image enhancement via edge-enhanced multi- exposure fusion network. In AAAI, pages 13106–13113,

Pith tools

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