Pith. sign in

REVIEW 2 major objections 2 minor 25 references

Overlapped Wavelet Diffusion for Low-Light Image Enhancement

T0 review · 2 major / 2 minor · reviewed 2026-06-27 · grok-4.3

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

Pith's one-line read An overlapped wavelet transform and guided high-frequency block enable artifact-free low-light image enhancement in diffusion models.

desk verdict The paper adds overlapped WT and low-freq-guided HFEBlock to DiffLL but the gains over the baseline are not isolated to those changes. read the letter →

arxiv 2606.10280 v1 pith:4MIH4455 submitted 2026-06-09 eess.IV cs.CV

Fen Peng , Taizo Suzuki , Seisuke Kyochi This is my paper · ORCID
classification eess.IVcs.CV
keywords low-lightimageenhancementwavelettransformdiffusionmodelsblockingartifactshigh-frequencyrestorationprocessingLOLdataset
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 introduces OWDiff, a diffusion framework for low-light image enhancement that uses an Overlapped Wavelet Transform to eliminate blocking artifacts and a low-frequency-guided High-Frequency Enhance Block to improve texture and edge recovery. It claims these changes overcome the limitations of the Haar wavelet and previous high-frequency modules in methods like DiffLL. Experiments on LOLv1 and LOLv2-real show consistent outperformance with gains in standard metrics. This matters because low-light enhancement is widely used in photography and computer vision, where artifacts can degrade usability.

What carries the argument

The Overlapped Wavelet Transform (OWT) that prevents blocking by using neighboring correlations, paired with the low-frequency-guided High-Frequency Enhance Block (HFEBlock) for detail strengthening.

What would settle it

Running an experiment where only the wavelet transform and high-frequency module are swapped into the baseline DiffLL model and checking if the same performance gains are obtained on the same test sets.

Watch

Extended reading notes

Core claim

The framework addresses structural limitations of the Haar Wavelet Transform and the High-Frequency Restoration Module by introducing an Overlapped WT that incorporates correlations across neighboring regions and a low-frequency-guided HFEBlock, resulting in blocking artifact-free enhancement with sharper edges and reliable textures, as demonstrated by superior performance on LOLv1 and LOLv2-real datasets.

Load-bearing premise

The observed improvements are due to the Overlapped WT and HFEBlock rather than variations in training, model size, or other factors.

Editorial extensions

If this is right

  • OWDiff outperforms existing LLIE methods on LOLv1 and LOLv2-real datasets in visual quality and metrics.
  • It achieves an average PSNR gain of 0.58 dB, 1.64% SSIM improvement, and 5.9% LPIPS reduction compared to DiffLL.
  • The method maintains computational efficiency while providing superior results.
  • Structural prevention of blocking artifacts and improved detail recovery are achieved through the new components.

Reading between the lines

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

  • If the OWT generalizes beyond this diffusion setup, similar overlapping strategies could reduce artifacts in other frequency-based image tasks.
  • The reliance on low-frequency guidance for high-frequency enhancement may suggest broader uses in multi-scale image processing.
  • Further tests on additional datasets would help confirm if the gains are robust across different low-light conditions.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The paper proposes OWDiff, an overlapped wavelet diffusion framework for low-light image enhancement. It introduces an Overlapped Wavelet Transform (OWT) to structurally prevent blocking artifacts from the standard Haar WT and a low-frequency-guided High-Frequency Enhance Block (HFEBlock) to improve detail recovery over the prior HFRM in DiffLL. The central empirical claim is consistent outperformance on LOLv1 and LOLv2-real, with average gains of 0.58 dB PSNR, 1.64% relative SSIM improvement, and 5.9% relative LPIPS reduction versus DiffLL while preserving computational efficiency.

Significance. If the metric gains can be isolated to the OWT and HFEBlock, the work would offer a targeted structural fix for known artifacts in wavelet-based diffusion LLIE methods and strengthen the case for overlapped transforms in this domain.

major comments (2)
  1. [Abstract / Experiments] Abstract and Experiments section: the attribution of the 0.58 dB PSNR / 1.64% SSIM / 5.9% LPIPS gains specifically to OWT and HFEBlock is load-bearing for the central claim, yet no ablation studies, controlled replacement experiments, or matched training-protocol comparisons with DiffLL are reported; without these, alternative explanations (differences in optimizer, loss weighting, capacity, or data augmentation) cannot be ruled out.
  2. [Methods] Methods section: the description of OWT and HFEBlock integration does not include quantitative verification (e.g., artifact maps or frequency-domain analysis) that the overlapped structure eliminates blocking independently of the diffusion schedule or other modules.
minor comments (2)
  1. Figure captions and tables should explicitly state whether results are averaged over both datasets or reported separately to allow direct comparison with the abstract numbers.
  2. Notation for the low-frequency guidance signal inside HFEBlock should be defined once at first use to avoid ambiguity in the block diagram.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight opportunities to strengthen the empirical support for our claims. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract / Experiments] Abstract and Experiments section: the attribution of the 0.58 dB PSNR / 1.64% SSIM / 5.9% LPIPS gains specifically to OWT and HFEBlock is load-bearing for the central claim, yet no ablation studies, controlled replacement experiments, or matched training-protocol comparisons with DiffLL are reported; without these, alternative explanations (differences in optimizer, loss weighting, capacity, or data augmentation) cannot be ruled out.

    Authors: We agree that the manuscript does not currently include ablation studies or controlled replacement experiments that isolate the contributions of OWT and HFEBlock, nor does it provide matched training-protocol comparisons against the original DiffLL implementation. The reported gains are based on end-to-end comparisons on LOLv1 and LOLv2-real. In the revised manuscript we will add ablation tables that (i) replace OWT with standard non-overlapped Haar WT while retaining all other modules and training settings, and (ii) replace HFEBlock with the original HFRM under identical optimizer, loss weights, and data-augmentation protocols. These controlled experiments will directly address the possibility of confounding factors. revision: yes

  2. Referee: [Methods] Methods section: the description of OWT and HFEBlock integration does not include quantitative verification (e.g., artifact maps or frequency-domain analysis) that the overlapped structure eliminates blocking independently of the diffusion schedule or other modules.

    Authors: The manuscript motivates OWT by noting that overlap incorporates cross-region correlations and thereby structurally avoids the blocking artifacts inherent to non-overlapped Haar WT. However, we acknowledge that the current text provides no quantitative verification (artifact maps, frequency spectra, or edge-sharpness metrics) demonstrating this effect independently of the diffusion schedule. In revision we will add (i) visual artifact maps comparing OWT versus Haar WT outputs at matched diffusion steps and (ii) frequency-domain analysis (power-spectrum ratios and high-frequency energy preservation) to quantify the blocking reduction attributable to the overlapped structure. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity; empirical method with external baselines

full rationale

The manuscript proposes OWDiff as an empirical framework combining Overlapped WT and low-frequency-guided HFEBlock, then reports PSNR/SSIM/LPIPS gains versus the external DiffLL baseline on LOLv1 and LOLv2-real. No derivation chain, first-principles prediction, or fitted parameter is presented that reduces to its own inputs by construction. All quantitative claims are externally falsifiable via replication on the cited public datasets and are not justified by self-citation load-bearing steps or ansatz smuggling. The work is therefore self-contained against external benchmarks.

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

Abstract-only review; no explicit free parameters, axioms, or invented entities are stated. The method implicitly relies on standard assumptions of diffusion models and wavelet transforms from prior literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Overlapped Wavelet Diffusion for Low-Light Image Enhancement." pith.science (2026). https://pith.science/paper/4MIH4455

@misc{pith2026260610280,
  author       = {Pith},
  title        = {Pith review of: Overlapped Wavelet Diffusion for Low-Light Image Enhancement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4MIH4455}},
  note         = {Machine review of arXiv:2606.10280}
}
read the original abstract

In this study, we propose an overlapped wavelet diffusion framework for Low-Light Image Enhancement (LLIE), which incorporates two complementary components to achieve blocking artifact-free and detail-preserving enhancement. Although recent diffusion-based LLIE methods have demonstrated remarkable performance compared with traditional approaches, DiffLL still suffers from blocking artifacts caused by the Haar Wavelet Transform (WT) and blurred edges or over-smoothed textures due to the limitations of its High-Frequency Restoration Module (HFRM). To overcome these issues, we introduce an Overlapped WT (OWT) that incorporates correlations across neighboring regions, thereby structurally preventing blocking artifacts. Furthermore, we integrate a low-frequency-guided High-Frequency Enhance Block (HFEBlock) to strengthen detail recovery, yielding sharper edges and more reliable textures. Extensive experiments on the LOLv1 and LOLv2-real datasets demonstrate that our framework, termed OWDiff, consistently outperforms existing LLIE methods both qualitatively and quantitatively, achieving superior visual quality while maintaining computational efficiency. OWDiff effectively addresses the structural limitations of the Haar WT and the HFRM, achieving an average PSNR gain of 0.58 dB, along with a 1.64% relative improvement in SSIM and a 5.9% relative reduction in LPIPS, compared to DiffLL across both the LOLv1 and LOLv2-real datasets.

Figures

Figures reproduced from arXiv: 2606.10280 by the authors.

Figure 1
Figure 1. Subband and reconstructed images with high [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overall pipeline of our OWDiff with the 2-level decomposition. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Architecture of WCDM (red and blue arrows indicate [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Qualitative comparison among LLIE methods ( [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 7
Figure 7. Figure 7: Qualitative comparison highlighting the contribu [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 1 canonical work pages

  1. [1]

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

    Y .P . Loh and C.S. Chan, “Getting to know low-light images with the exclusively dark dataset,” Comput. Vis. Image Underst., vol.178, pp.30–42, Jan. 2019

    2019

  2. [2]

    Recurrent exposure generation for low-light face detec- tion,

    X.J. Liang, W.J. Wang, H.Y . Quan, Y .T. Chen, Y .J. Liu, B.H. Ling, and Y . Xu, “Recurrent exposure generation for low-light face detec- tion,” IEEE Trans. Multimedia, vol.24, pp.1609–1621, Mar. 2022

    2022

  3. [3]

    A deep learn- ing based image enhancement approach for autonomous driving at night,

    F.G. Li, F.Y . Y ang, D.X. Qu, P .D. Cao, and Q.K. Li, “A deep learn- ing based image enhancement approach for autonomous driving at night,” Knowl.-Based Syst., vol.213, no.15, pp.1–14, Feb. 2021. 8 IEICE TRANS. ??, VOL.Exx–??, NO.xx XXXX 200x

    2021

  4. [4]

    Adaptive image contrast enhancement using generaliza- tions of histogram equalization,

    J.A. Stark, “Adaptive image contrast enhancement using generaliza- tions of histogram equalization,” IEEE Trans. Image Process., vol.9, no.5, pp.889–896, May 2000

    2000

  5. [5]

    A simple and effective histogram equal- ization approach to image enhancement,

    D.H. Cheng and J.X. Shi, “A simple and effective histogram equal- ization approach to image enhancement,” Digit. Signal Process., vol.14, no.2, pp.158–170, Mar. 2004

    2004

  6. [6]

    A dynamic histogram equalization for image contrast enhance- ment,

    M. Abdullah-Al-Wadud, M.H. Kabir, M.A.A. Dewan, and O. Chae, “A dynamic histogram equalization for image contrast enhance- ment,” IEEE Trans. Consum. Electron., vol.53, no.2, pp.593–600, May 2007

    2007

  7. [7]

    A total variation model for retinex,

    M.K. Ng and W. Wang, “A total variation model for retinex,” SIAM J. Imaging Sci., vol.4, no.1, pp.345–365, Jan. 2011

    2011

  8. [8]

    LIME: Low-light image enhance- ment via illumination map estimation,

    J.X. Guo, Y . Li, and B.H. Ling, “LIME: Low-light image enhance- ment via illumination map estimation,” IEEE Trans. Image Process., vol.26, no.2, pp.982–993, Feb. 2017

    2017

Show all 25 references
  1. [9]

    A real-time image proces- sor with combining dynamic contrast ratio enhancement and inverse gamma correction for PDP,

    G.Z. Wang, H.Z. Liang, and L.C. Liu, “A real-time image proces- sor with combining dynamic contrast ratio enhancement and inverse gamma correction for PDP,” Displays, vol.30, no.3, pp.133–139, Jul. 2009

    2009

  2. [10]

    Efficient contrast enhance- ment using adaptive gamma correction with weighting distribution,

    C.S. Huang, C.F. Cheng, and S.Y . Chiu, “Efficient contrast enhance- ment using adaptive gamma correction with weighting distribution,” IEEE Trans. Image Process., vol.22, no.3, pp.1032–1041, Mar. 2013

    2013

  3. [11]

    An adaptive gamma correction for image enhancement,

    S. Rahman, M.M. Rahman, M. Abdullah-Al-Wadud, G.D. Al- Quaderi, and M. Shoyaib, “An adaptive gamma correction for image enhancement,” EURASIP J. Image Video Process., vol.2016, no.1, pp.1–13, Oct. 2016

    2016

  4. [12]

    Deep retinex decom- position for low-light enhancement,

    C. Wei, J.W. Wang, H.W. Y ang, and Y .J. Liu, “Deep retinex decom- position for low-light enhancement,” arXiv preprint arXiv:1808.04560, Aug. 2018

    2018 arXiv

  5. [13]

    Underexposed photo enhancement using deep illumination es- timation,

    X.R. Wang, Q. Zhang, W.C. Fu, Y .X. Shen, S.W. Zheng, and Y .J. Jia, “Underexposed photo enhancement using deep illumination es- timation,” Proc. IEEE/CVF CVPR, Long Beach, CA, Jun. 2019

    2019

  6. [14]

    Zero-reference deep curve estimation for low-light im- age enhancement,

    L.C. Guo, Y .C. Li, C.J. Guo, C.C. Loy, H.J. Hou, S. Kwong, and M.R. Cong, “Zero-reference deep curve estimation for low-light im- age enhancement,” Proc. IEEE/CVF CVPR, Seattle, W A, pp.1780– 1789, Jun. 2020

    2020

  7. [15]

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

    H.W . Y ang, J.W. Wang, F.H. Huang, Q.S. Wang, and Y .J. Liu, “Sparse gradient regularized deep retinex network for robust low- light image enhancement,” IEEE Trans. Image Process., vol.30, pp.2072–2086, Jan. 2021

  8. [16]

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

    S.R. Liu, L. Ma, A.J. Zhang, X. Fan, and X.Z. Luo, “Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement,” Proc. IEEE/CVF CVPR, Online, pp.10561– 10570, Jun. 2021

    2021

  9. [17]

    EnlightenGAN: Deep light en- hancement without paired supervision,

    F.Y . Jiang, Y .X. Gong, D. Liu, Y . Cheng, C. Fang, H.X. Shen, C.J. Y ang, P . Zhou, and Y .Z. Wang, “EnlightenGAN: Deep light en- hancement without paired supervision,” IEEE Trans. Image Process., vol.30, pp.2340–2349, Jan. 2021

    2021

  10. [18]

    Beyond brightening low-light images,

    H.Y . Zhang, J.X. Guo, Y .J. Ma, W. Liu, and W.J. Zhang, “Beyond brightening low-light images,” Int. J. Comput. Vis., vol.129, no.4, pp.1013–1037, Jan. 2021

    2021

  11. [19]

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

    L. Ma, Y .T. Ma, S.M. Liu, X. Fan, and X.Z. Luo, “Toward fast, flexible, and robust low-light image enhancement,” Proc. IEEE/CVF CVPR, New Orleans, LA, pp.5637–5646, Jun. 2022

    2022

  12. [20]

    SNR-aware low-light image enhancement,

    G.X. Xu, X.R. Wang, W.C. Fu, and Y .J. Jia, “SNR-aware low-light image enhancement,” Proc. IEEE/CVF CVPR, New Orleans, LA, pp.17714–17724, Jun. 2022

    2022

  13. [21]

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

    H.Y . Cai, H. Bian, J. Lin, Q.H. Wang, R. Timofte, and L.Y . Zhang, “Retinexformer: One-stage retinex-based transformer for low-light image enhancement,” Proc. IEEE/CVF ICCV , Paris, France, pp.12504–12513, Oct. 2023

    2023

  14. [22]

    Global structure-aware diffusion process for low-light image en- hancement,

    H.J. Hou, Y .Z. Zhu, H.J. Hou, H. Liu, Q.H. Zeng, and H. Yuan, “Global structure-aware diffusion process for low-light image en- hancement,” Proc. NeurIPS, New Orleans, LA, pp.79734–79747, Sep. 2023

    2023

  15. [23]

    Low-light im- age enhancement with wavelet-based diffusion models,

    H. Jiang, A. Luo, Q.H. Fan, C.S. Han, and C.S. Liu, “Low-light im- age enhancement with wavelet-based diffusion models,” ACM Trans. Graph., vol.42, no.6, pp.1–14, Dec. 2023

    2023

  16. [24]

    Selective hourglass mapping for universal image restoration based on diffusion model,

    D. Zheng, M.X. Wu, Z.S. Y ang, J. Zhang, F.J. Hu, and S.W. Zheng, “Selective hourglass mapping for universal image restoration based on diffusion model,” Proc. IEEE/CVF CVPR, Seattle, W A, pp.25445–25455, Jun. 2024

    2024

  17. [25]

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

    W.B. Zou, H.X. Gao, W.P . Y ang, and T.T. Liu, “Wave-Mamba: Wavelet state space model for ultra-high-definition low-light image enhancement,” Proc. ACM MM, Melbourne, Australia, pp.4515– 4523, Jul. 2024. Fen Peng received the B.E. degree in Radio and TV Engineering from Commun...

    2024

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.