Pith. sign in

REVIEW 43 references

UHD Image Dehazing via anDehazeFormer with Atmospheric-aware KV Cache

T0 review · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read anDehazeFormer combines a physics-guided KV cache with per-block adaptive normalization for ultra-high-definition dehazing, but its stated real-time speed and 5x training speedup are not supported by the reported data.

arxiv 2505.14010 v1 pith:2R423TZI submitted 2025-05-20 cs.CV

Pu Wang , Pengwen Dai , Chen Wu , Yeying Jin , Dianjie Lu , Guijuan Zhang , Youshan Zhang , Zhuoran Zheng This is my paper · ORCID
classification cs.CV
keywords dehazingimagetextbftrainingapproacharchitecturemechanismmemory
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

Removing haze from very large photos (4K and 8K) usually costs a lot of memory and time. The paper proposes a transformer model that uses two tricks. First, it normalizes each block of the network using values derived from an estimated haze map, borrowing an idea from a faster training method called nGPT. Second, it keeps a cache of previously processed image features and throws away the parts that seem least useful based on a simple haze physics model: the darker an area's dark channel, the more haze is assumed, and the more aggressively the cache is pruned.

The authors report that the model dehazes 512x512 patches at 53 frames per second, and that on three benchmarks it reaches PSNR/SSIM values at or above a recent method called DehazeXL. The memory use in an ablation drops from 18.3 GB to 15.6 GB. The interpretability add-on, a gradient-based attribution map called PAAM, is a minor contribution and is not evaluated quantitatively.

The paper has internal problems. The abstract claims real-time processing of 50 high-resolution images per second, but Table 1 shows 0.22 seconds per 4K image, about 4.5 frames per second. The claimed 5x training convergence speedup is not backed by any learning curve. The statement that the method achieves the highest scores for both PSNR and SSIM is false for one row: on O-HAZE the SSIM is 0.7333, slightly below DehazeXL's 0.7348. No error bars are reported anywhere.

Extended reading notes

Core claim

The proposed anDehazeFormer 'enables real-time processing of 50 high-resolution images per second on an RTX4090 GPU', 'improves the training convergence speed by 5x', and 'achieves state-of-the-art dehazing quality' with the highest PSNR and SSIM on 8KDehaze, 4KID, O-HAZE, and I-HAZE. If correct, it would be a practical UHD dehazing system with a large efficiency advantage over DehazeXL.

Load-bearing premise

The paper assumes that FPS and memory numbers measured on 512x512 patches, combined with sliding-window inference, transfer directly to full 4K/8K output without seam artifacts or additional memory cost, and that the KV cache pruning guided by the learned coefficient c_a preserves quality when cache sequences are interpolated across resolutions. Section 4.1 states that inference uses a sliding window, but no full-image memory footprint, stitching quality, or end-to-end timing at native resolution is measured. If this assumption fails, the real-time and memory claims collapse.

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.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The method relies on several learned parameters (c_a, r, b) and hand-tuned hyperparameters (eta, alpha, beta, tau, loss weights, t_mid) plus standard domain assumptions about the atmospheric scattering model and dark channel prior. The learned parameters are fitted to data, not derived from first principles.

free parameters (6)
  • c_a = learned from data
    Per-image learned scaling coefficient mapping dark channel D_hat to transmission t = 1 - c_a * D_hat (Eq. 5); central to the atmospheric model.
  • r, b = learned from data
    Adaptive normalization scaling and bias derived from physical parameters (Eq. 6); both are network outputs fitted to data.
  • eta (cache decay rate) = 0.5 (default)
    Algorithm 1 hyperparameter controlling cache retention gamma = 1 - mean(c_a) * eta; chosen by hand with no sensitivity study.
  • alpha, beta, tau (window size) = alpha=8, beta=4, tau=1024
    Hand-tuned parameters in Eq. 7 that determine dynamic window sizing; no ablation is reported for these values.
  • loss weights = 0.8 L1, 0.1 MSE, 0.1 SSIM
    Section 4.1 loss combination; weights are chosen without sensitivity analysis.
  • t_mid = 0.7
    Baseline haze level for the attribution path in Eq. 12; fixed arbitrarily.
assumptions (5)
  • domain assumption Atmospheric scattering model I = J*t + A*(1-t)
    Used in Eq. 5, Eq. 9, and Eq. 11; assumes a single global atmospheric light per image and no multiple scattering.
  • domain assumption Dark channel prior estimates transmission
    Section 3.1 assumes dark channel intensity is inversely related to haze density, which is known to fail in some sky and white regions.
  • domain assumption Adaptive normalization inspired by nGPT preserves training stability
    Section 3.2 claims 5x faster convergence without providing convergence curves or a training time comparison.
  • domain assumption KV cache techniques from LLM inference transfer to vision windows
    Section 2 and Algorithm 1 build on [7] but do not analyze cache dimension growth or overhead across network depth.
  • domain assumption Synthetic benchmarks 8KDehaze and 4KID generate haze using the same atmospheric model
    Section 4.1; this is favorable to methods that embed the same physical prior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UHD Image Dehazing via anDehazeFormer with Atmospheric-aware KV Cache." pith.science (2026). https://pith.science/paper/2R423TZI

@misc{pith2026250514010,
  author       = {Pith},
  title        = {Pith review of: UHD Image Dehazing via anDehazeFormer with Atmospheric-aware KV Cache},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2R423TZI}},
  note         = {Machine review of arXiv:2505.14010}
}
abstract

In this paper, we propose an efficient visual transformer framework for ultra-high-definition (UHD) image dehazing that addresses the key challenges of slow training speed and high memory consumption for existing methods. Our approach introduces two key innovations: 1) an \textbf{a}daptive \textbf{n}ormalization mechanism inspired by the nGPT architecture that enables ultra-fast and stable training with a network with a restricted range of parameter expressions; and 2) we devise an atmospheric scattering-aware KV caching mechanism that dynamically optimizes feature preservation based on the physical haze formation model. The proposed architecture improves the training convergence speed by \textbf{5 $\times$} while reducing memory overhead, enabling real-time processing of 50 high-resolution images per second on an RTX4090 GPU. Experimental results show that our approach maintains state-of-the-art dehazing quality while significantly improving computational efficiency for 4K/8K image restoration tasks. Furthermore, we provide a new dehazing image interpretable method with the help of an integrated gradient attribution map. Our code can be found here: https://anonymous.4open.science/r/anDehazeFormer-632E/README.md.

Figures

Figures reproduced from arXiv: 2505.14010 by the authors.

Figure 1
Figure 1. Improvement of our model over the SOTA approaches in 8KDehaze [ [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The framework integrates four key components. (a) Atmospheric Estimator generating [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the baseline image and the path function. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Dehazed results on the O-HAZE [1] and I-HAZE [2] dataset. The proposed anDehaze￾Former demonstrates higher color fidelity and restores more details compared with other SOTA methods [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Dehazed results on the 4KID [42] dataset. The proposed anDehazeFormer effectively avoids ghosting and blurring, achieving full-resolution inference with remarkable clarity and precision [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Dehazed results on the 8KDehaze [6] dataset. The proposed anDehazeFormer effectively eliminates segmentation artifacts and achieves superior visual quality. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Dehazed results in the real world. Quantitative Evaluation [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages

  1. [1]

    O-haze: a dehazing benchmark with real hazy and haze-free outdoor images

    Codruta O Ancuti, Cosmin Ancuti, Radu Timofte, and Christophe De Vleeschouwer. O-haze: a dehazing benchmark with real hazy and haze-free outdoor images. InProceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 754–762, 2018

    2018

  2. [2]

    I-haze: A dehazing benchmark with real hazy and haze-free indoor images

    Cosmin Ancuti, Codruta O Ancuti, Radu Timofte, and Christophe De Vleeschouwer. I-haze: A dehazing benchmark with real hazy and haze-free indoor images. InAdvanced Concepts for Intelligent Vision Systems: 19th International Conference, ACIVS 2018, Poitiers, France, September 24–27, 2018, Proceedings 19, pages 620–631. Springer, 2018

    work page 2018

  3. [3]

    Precomputed atmospheric scattering

    Eric Bruneton and Fabrice Neyret. Precomputed atmospheric scattering. InComputer graphics forum, volume 27, pages 1079–1086. Wiley Online Library, 2008

    work page 2008

  4. [4]

    Dehazenet: An end-to-end system for single image haze removal.IEEE transactions on image processing, 25(11):5187– 5198, 2016

    Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal.IEEE transactions on image processing, 25(11):5187– 5198, 2016

    work page 2016

  5. [5]

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

    Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinex- former: One-stage retinex-based transformer for low-light image enhancement. InProceedings of the IEEE/CVF international conference on computer vision, pages 12504–12513, 2023

    work page 2023

  6. [6]

    Tokenize Image Patches: Global Context Fusion for Effective Haze Removal in Large Images

    Jiuchen Chen, Xinyu Yan, Qizhi Xu, and Kaiqi Li. Tokenize image patches: Global context fusion for effective haze removal in large images.arXiv preprint arXiv:2504.09621, 2025

    work page Pith review arXiv 2025

  7. [7]

    Ultra- high-definition dynamic multi-exposure image fusion via infinite pixel learning

    Xingchi Chen, Zhuoran Zheng, Xuerui Li, Yuying Chen, Shu Wang, and Wenqi Ren. Ultra- high-definition dynamic multi-exposure image fusion via infinite pixel learning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 2248–2255, 2025

    work page 2025

  8. [8]

    Dea-net: Single image dehazing based on detail- enhanced convolution and content-guided attention.IEEE Transactions on Image Processing, 33:1002–1015, 2024

    Zixuan Chen, Zewei He, and Zhe-Ming Lu. Dea-net: Single image dehazing based on detail- enhanced convolution and content-guided attention.IEEE Transactions on Image Processing, 33:1002–1015, 2024

    work page 2024

Show all 43 references
  1. [9]

    Ffa-net: fast feature aggregation network for 3d point cloud segmentation.Machine Vision and Applications, 34(5):80, 2023

    Ruting Cheng, Hui Zeng, Baoqing Zhang, Xuan Wang, and Tianmeng Zhao. Ffa-net: fast feature aggregation network for 3d point cloud segmentation.Machine Vision and Applications, 34(5):80, 2023

    2023

  2. [10]

    Image dehazing using deep learning techniques.Procedia Computer Science, 167:1110–1119, 2020

    Ravi Raj Choudhary, KK Jisnu, and Gaurav Meena. Image dehazing using deep learning techniques.Procedia Computer Science, 167:1110–1119, 2020

    2020

  3. [11]

    Revitalizing convolutional network for image restoration.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Revitalizing convolutional network for image restoration.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  4. [12]

    Image melding: Combining inconsistent images using patch-based synthesis.ACM Transactions on graphics (TOG), 31(4):1–10, 2012

    Soheil Darabi, Eli Shechtman, Connelly Barnes, Dan B Goldman, and Pradeep Sen. Image melding: Combining inconsistent images using patch-based synthesis.ACM Transactions on graphics (TOG), 31(4):1–10, 2012. 10

    2012

  5. [13]

    Single image dehazing using extended local dark channel prior.Image and vision computing, 136:104747, 2023

    Pulkit Dwivedi and Soumendu Chakraborty. Single image dehazing using extended local dark channel prior.Image and vision computing, 136:104747, 2023

    2023

  6. [14]

    Image dehazing transformer with transmission-aware 3d position embedding

    Chun-Le Guo, Qixin Yan, Saeed Anwar, Runmin Cong, Wenqi Ren, and Chongyi Li. Image dehazing transformer with transmission-aware 3d position embedding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5812–5820, June 2022

    2022

  7. [15]

    The University of North Carolina at Chapel Hill, 2003

    Mark Jason Harris.Real-time cloud simulation and rendering. The University of North Carolina at Chapel Hill, 2003

    2003

  8. [16]

    Single image haze removal using dark channel prior

    Kaiming He, Jian Sun, and Xiaoou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341–2353, 2010

    2010

  9. [17]

    Multi-matrix factorization attention.arXiv preprint arXiv:2412.19255, 2024

    Jingcheng Hu, Houyi Li, Yinmin Zhang, Zili Wang, Shuigeng Zhou, Xiangyu Zhang, Heung-Yeung Shum, and Daxin Jiang. Multi-matrix factorization attention.arXiv preprint arXiv:2412.19255, 2024

    2024 arXiv

  10. [18]

    Ide: Image dehazing and exposure using an enhanced atmospheric scattering model.IEEE Transactions on Image Processing, 30:2180–2192, 2021

    Mingye Ju, Can Ding, Wenqi Ren, Yi Yang, Dengyin Zhang, and Y Jay Guo. Ide: Image dehazing and exposure using an enhanced atmospheric scattering model.IEEE Transactions on Image Processing, 30:2180–2192, 2021

    2021

  11. [19]

    Single image dehazing via an improved atmospheric scattering model.The Visual Computer, 33:1613–1625, 2017

    Mingye Ju, Dengyin Zhang, and Xuemei Wang. Single image dehazing via an improved atmospheric scattering model.The Visual Computer, 33:1613–1625, 2017

    2017

  12. [20]

    Deep scattering: Rendering atmospheric clouds with radiance-predicting neural networks.ACM Transactions on Graphics (TOG), 36(6):1–11, 2017

    Simon Kallweit, Thomas Müller, Brian Mcwilliams, Markus Gross, and Jan Novák. Deep scattering: Rendering atmospheric clouds with radiance-predicting neural networks.ACM Transactions on Graphics (TOG), 36(6):1–11, 2017

    2017

  13. [21]

    Yongwoo Kim, Jae-Seok Choi, and Munchurl Kim. A real-time convolutional neural network for super-resolution on fpga with applications to 4k uhd 60 fps video services.IEEE Transactions on Circuits and Systems for Video Technology, 29(8):2521–2534, 2018

    2018

  14. [22]

    Aod-net: All-in-one dehazing network

    Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng. Aod-net: All-in-one dehazing network. InProceedings of the IEEE international conference on computer vision, pages 4770–4778, 2017

    2017

  15. [23]

    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. arXiv preprint arXiv:2302.11831, 2023

    2023 arXiv

  16. [24]

    Griddehazenet: Attention-based multi- scale network for image dehazing

    Xiaohong Liu, Yongrui Ma, Zhihao Shi, and Jun Chen. Griddehazenet: Attention-based multi- scale network for image dehazing. InProceedings of the IEEE/CVF international conference on computer vision, pages 7314–7323, 2019

    2019

  17. [25]

    Patch based synthesis for single depth image super-resolution

    Oisin Mac Aodha, Neill DF Campbell, Arun Nair, and Gabriel J Brostow. Patch based synthesis for single depth image super-resolution. InComputer Vision–ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part III 12, pages 7...

    2012

  18. [26]

    A novel approach for image enhancement by using contrast limited adaptive histogram equalization method

    S Muniyappan, A Allirani, and S Saraswathi. A novel approach for image enhancement by using contrast limited adaptive histogram equalization method. In2013 fourth international conference on computing, communications and networking technologies (ICCCNT), pages 1–6. IEEE, 2013

    2013

  19. [27]

    Mb- taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing

    Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. Mb- taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12802–12813, ...

    2023

  20. [28]

    Trends and prospects of techniques for haze removal from degraded images: A survey.IEEE Transactions on Emerging Topics in Computational Intelligence, 6(4):762–782, 2022

    Geet Sahu, Ayan Seal, Debotosh Bhattacharjee, Mita Nasipuri, Peter Brida, and Ondrej Krejcar. Trends and prospects of techniques for haze removal from degraded images: A survey.IEEE Transactions on Emerging Topics in Computational Intelligence, 6(4):762–782, 2022. 11

    2022

  21. [29]

    Domain adaptation for image dehazing

    Yuanjie Shao, Lerenhan Li, Wenqi Ren, Changxin Gao, and Nong Sang. Domain adaptation for image dehazing. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2808–2817, 2020

    2020

  22. [30]

    Single image dehazing based on one dimensional linear filtering and adoptive histogram equalization method

    Ashok Shrivastava and Sanjay Jain. Single image dehazing based on one dimensional linear filtering and adoptive histogram equalization method. In2016 International Conference on Electrical, Electronics, and Optimization Techniques (ICEEOT), pages 4074–4078. IEEE, 2016

    2016

  23. [31]

    Vision transformers for single image dehazing

    Yuda Song, Zhuqing He, Hui Qian, and Xin Du. Vision transformers for single image dehazing. IEEE Transactions on Image Processing, 32:1927–1941, 2023

    1927

  24. [32]

    Visualizing the impact of feature attribution baselines.Distill, 5(1):e22, 2020

    Pascal Sturmfels, Scott Lundberg, and Su-In Lee. Visualizing the impact of feature attribution baselines.Distill, 5(1):e22, 2020

    2020

  25. [33]

    Correlation matching transformation transformers for uhd image restoration

    Cong Wang, Jinshan Pan, Wei Wang, Gang Fu, Siyuan Liang, Mengzhu Wang, Xiao-Ming Wu, and Jun Liu. Correlation matching transformation transformers for uhd image restoration. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5336–5344, 2024

    2024

  26. [34]

    Dap-led: Learning degradation-aware priors with clip for joint low-light enhancement and deblurring.arXiv preprint arXiv:2409.13496, 2024

    Ling Wang, Chen Wu, and Lin Wang. Dap-led: Learning degradation-aware priors with clip for joint low-light enhancement and deblurring.arXiv preprint arXiv:2409.13496, 2024

    2024 arXiv

  27. [35]

    Tsid-net: a two-stage single image de- hazing framework with style transfer and contrastive knowledge transfer.The Visual Computer, 41(3):1921–1938, 2025

    Shilong Wang, Qianwen Hou, Jiaang Li, and Jianlei Liu. Tsid-net: a two-stage single image de- hazing framework with style transfer and contrastive knowledge transfer.The Visual Computer, 41(3):1921–1938, 2025

    1921

  28. [36]

    Ultra- high-definition low-light image enhancement: A benchmark and transformer-based method

    Tao Wang, Kaihao Zhang, Tianrun Shen, Wenhan Luo, Bjorn Stenger, and Tong Lu. Ultra- high-definition low-light image enhancement: A benchmark and transformer-based method. In Proceedings of the AAAI conference on artificial intelligence, volume 37, pages 2654–2662, 2023

    2023

  29. [37]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600– 612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600– 612, 2004

    2004

  30. [38]

    Ultra- high definition (8k uhd) endoscope: our first clinical success.SpringerPlus, 5:1–5, 2016

    Hiromasa Yamashita, Hisae Aoki, Kenkichi Tanioka, Toshiyuki Mori, and Toshio Chiba. Ultra- high definition (8k uhd) endoscope: our first clinical success.SpringerPlus, 5:1–5, 2016

    2016

  31. [39]

    Cyclegan-based deep learning technique for artifact reduction in fundus photography.Graefe’s Archive for Clinical and Experimental Ophthalmology, 258:1631–1637, 2020

    Tae Keun Yoo, Joon Yul Choi, and Hong Kyu Kim. Cyclegan-based deep learning technique for artifact reduction in fundus photography.Graefe’s Archive for Clinical and Experimental Ophthalmology, 258:1631–1637, 2020

    2020

  32. [40]

    H2o: Heavy-hitter oracle for efficient generative inference of large language models.Advances in Neural Information Processing Systems, 36:34661–34710, 2023

    Zhenyu Zhang, Ying Sheng, Tianyi Zhou, Tianlong Chen, Lianmin Zheng, Ruisi Cai, Zhao Song, Yuandong Tian, Christopher Ré, Clark Barrett, et al. H2o: Heavy-hitter oracle for efficient generative inference of large language models.Advances in Neural Information Processing System...

    2023

  33. [41]

    4k-haze: A dehazing benchmark with 4k resolution hazy and haze-free images.arXiv preprint arXiv:2303.15848, 2023

    Zhuoran Zheng and Xiuyi Jia. 4k-haze: A dehazing benchmark with 4k resolution hazy and haze-free images.arXiv preprint arXiv:2303.15848, 2023

    2023 arXiv

  34. [42]

    Limitations

    Zhuoran Zheng, Wenqi Ren, Xiaochun Cao, Xiaobin Hu, Tao Wang, Fenglong Song, and Xiuyi Jia. Ultra-high-definition image dehazing via multi-guided bilateral learning. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16180–16189. IEEE, 2021. 12...

    2021

  35. [43]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

    2025

Pith tools

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