Pith. sign in

REVIEW 4 major objections 7 minor 42 references

A frequency-filtering state-space model wins both tracks of the new high-FPS motion deblurring challenge.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

The AIM 2025 challenge ranks 9 deblurring solutions on new high-FPS motion blur datasets, with VPEG placing first in both moderate and extreme tracks.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A solid challenge report with transparent rankings; the main caveat is that the 'significant progress' claim rests on synthetic blur realism that this paper never actually argues for. the 4 major comments →

arxiv 2509.06793 v1 pith:LJ5X46QB submitted 2025-09-08 cs.CV

AIM 2025 Challenge on High FPS Motion Deblurring: Methods and Results

George Ciubotariu , Florin-Alexandru Vasluianu , Zhuyun Zhou , Nancy Mehta , Radu Timofte , Ke Wu , Long Sun , Lingshun Kong
show 39 more authors
Zhongbao Yang Jinshan Pan Jiangxin Dong Jinhui Tang Hao Chen Yinghui Fang Dafeng Zhang Yongqi Song Jiangbo Guo Shuhua Jin Zeyu Xiao Rui Zhao Zhuoyuan Li Cong Zhang Yufeng Peng Xin Lu Zhijing Sun Chengjie Ge Zihao Li Zishun Liao Ziang Zhou Qiyu Kang Xueyang Fu Zheng-Jun Zha Yuqian Zhang Shuai Liu Jie Liu Zhuhao Zhang Lishen Qu Zhihao Liu Shihao Zhou Yaqi Luo Juncheng Zhou Jufeng Yang Qianfeng Yang Qiyuan Guan Xiang Chen Guiyue Jin Jiyu Jin
This is my paper
classification cs.CV
keywords motion deblurringMIOReVAR-MIOReimage restoration benchmarkstate-space modelfrequency-domain attentionLPIPSPSNR/SSIM
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 reports a two-track competition on single-image motion deblurring, using a new pair of datasets built from high-frame-rate video frames that are synthetically blurred to different severities. Track 1 uses moderate blur plus some defocus; Track 2 uses up to four times stronger blur. The challenge's stated aim is to find networks that restore sharp, visually plausible images under both regimes, and it ranks nine submitted methods by a composite of PSNR, SSIM, and LPIPS. The leading solution, a visual state-space model trained with progressively larger image patches, wins both tracks, with the best fidelity scores (34.5 dB PSNR on Track 1, 30.3 dB on Track 2), while a generative diffusion entry obtains the best perceptual-distance score in both tracks. The report thus establishes a current snapshot of which architectural families lead under a common metric and provides a reusable benchmark for the field.

Core claim

The central claim of the paper is that a shared, controlled evaluation—moderate blur in MIORe and extreme blur in VAR-MIORe—can rank single-image motion deblurring systems across difficulty levels, and that the winner of both tracks is a frequency-filtering visual state-space model (EVSSM) trained in stages with increasing patch size. The reported numbers anchor the claim: the winning method reaches 34.484 dB PSNR / 0.9026 SSIM on Track 1 and 30.287 dB / 0.8434 on Track 2, ahead of the second place by roughly 0.3–1.7 dB. The paper also documents the solution space: frequency-domain attention, prompt-guided two-stage restoration, progressive multiscale training, pretrained model fusion, and d

What carries the argument

The load-bearing objects are the MIORe and VAR-MIORe datasets and the evaluation protocol itself. MIORe supplies 7,860 full-HD synthetic blur/sharp pairs with mild to moderate blur and occasional defocus; VAR-MIORe supplies 1,299 pairs with blur severity scaled up to four times. Rankings are computed by averaging each team's rank on PSNR, SSIM, and LPIPS (with ImageNet-pretrained deep features), which makes the composite final rank the operational definition of success in the challenge. On the method side, the winning pipeline is EVSSM, a visual state-space architecture whose core blocks alternate transposition and flipping to preserve spatial structure while filtering frequencies in the Fou

Load-bearing premise

The rankings stand on the assumption that blur generated synthetically from high-FPS frames behaves like real motion blur, including the up-to-4x severity used in Track 2; if that generation misses real-world effects such as camera shake, rolling shutter, or occlusion, the benchmark results may not transfer.

What would settle it

Measure optical-flow magnitude and blur-kernel statistics of real blurred videos captured with the same high-speed camera in natural scenes and compare them to MIORe and VAR-MIORe. If the synthetic data's flow distribution or blur length differs markedly from the real captures, or if method rankings on a small real-blur test set diverge from the challenge rankings, the central claim is falsified.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • The MIORe and VAR-MIORe splits give future work a fixed ground for comparing deblurring methods under moderate and 4x-severe blur, with public training data and a common scoring rule.
  • The roughly 4.2 dB PSNR drop between the winning Track 1 and Track 2 scores quantifies how much harder the extreme regime is, so new methods can be judged by how much of this gap they close.
  • Progressive training with growing patch size is shared by the two top-scoring entries, suggesting it is a transferable recipe for high-resolution deblurring.
  • Frequency-domain attention and state-space models currently lead the fidelity metrics, while a diffusion-based generative approach leads LPIPS, so a combined or application-specific objective would choose different winners.
  • Participants who transferred weights from existing real-world deblurring datasets and then fine-tuned on the challenge data reached the high ranks, supporting the transfer-then-finetune recipe on synthetic benchmarks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the final rank averages PSNR, SSIM, and LPIPS with no runtime or parameter-count term, a deployment-oriented ranking could reorder the field: the winning 17M-parameter model takes about 52 seconds per image on one GPU, while several lower-ranked entries run in under a second.
  • The diffusion entry's best-in-class LPIPS with low PSNR suggests the composite metric is mission-dependent; a challenge aimed at visually compelling output might weight perceptual distance more heavily, which would change the winner.
  • The synthetic blur-generation pipeline, if it tracks real high-FPS capture, could be extended to generate arbitrarily long blur trajectories on demand, turning the fixed 4x ceiling into a continuous stress test of deblurring robustness.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 7 minor

Summary. The paper reports the AIM 2025 High FPS Non-Uniform Motion Deblurring Challenge. Two tracks are organized around the new synthetic datasets MIORe (moderate blur, 7,860 training pairs) and VAR-MIORe (severe blur, 1,299 training pairs). Nine teams submitted valid entries; results are evaluated on held-out test splits using PSNR, SSIM, and LPIPS. The organizers report VPEG as the top-ranked method in both tracks (Track 1 PSNR 34.484, SSIM 0.9026; Track 2 PSNR 30.287, SSIM 0.8434). Section 7 summarizes each team's architecture and training procedure, and the conclusion states that the challenge showcases significant progress in high-FPS single-image motion deblurring.

Significance. If the results are taken at face value, the paper provides a useful archival record of the challenge and identifies a strong practical solution (EVSSM-based VPEG) on the MIORe/VAR-MIORe benchmarks. The public benchmark, the Codabench evaluation server, and the per-team method descriptions are concrete strengths. However, the scientific significance is limited by the absence of external baselines (so the 'state-of-the-art' claim is not demonstrated against prior methods), the lack of detail on the synthetic blur generation process, an ambiguity about test-set fine-tuning by one team, and an underspecified composite ranking formula. These issues are fixable, but they currently weaken the paper's central claims of 'significant progress' and generalizable deblurring performance.

major comments (4)
  1. [Sec. 2 and Sec. 8] The datasets are introduced as 'synthetically generated from sharp ground-truth frames recorded with a high-frame-rate CHRONOS 2.1-HD camera', with details deferred to reference [5]. The paper never describes the synthesis model: is blur produced by temporal averaging of the high-FPS burst, by warping/averaging with optical flow, or by learned/new synthetic kernels? Nor is 'up to four times stronger' blur in VAR-MIORe defined. Since the abstract and conclusion claim that the challenge identifies networks for 'diverse and challenging conditions' and shows 'significant progress' in high-FPS motion deblurring, the realism of the synthetic distribution is load-bearing. Please specify the generation procedure or explicitly restrict the claims to the synthetic benchmark.
  2. [Sec. 6, Tables 1 and 2] The final ranking compares only the nine submitted methods. There is no evaluation of standard non-participant baselines (e.g., Restormer, NAFNet, EVSSM, or a simple prior method) under the same test protocol. Consequently, the abstract's 'state-of-the-art advances' and the conclusion's 'significant progress' are not supported: the tables only rank the challenge participants. Adding a few reference baselines, or tempering the claims, is necessary for the paper's central message.
  3. [Sec. 7.9 vs. Sec. 5] MagicBlur's method description says the team fine-tunes the model on test-set inputs using pseudo ground-truth images produced by the same model pre-trained on the training set. Section 5, however, states only that 'the fine-tuning of the test set was restricted by limiting user submissions', which is vague and appears inconsistent with allowing test-input fine-tuning. Please clarify the official rule: was test-input adaptation explicitly permitted for all teams? If so, state it in the evaluation protocol. If not, MagicBlur's result should be flagged as non-comparable or excluded. Even though MagicBlur ranks last, the ambiguity affects the integrity of the comparison protocol.
  4. [Sec. 4, Tables 1 and 2] The 'Final Rank' is described only as 'a composite metric, derived from the average ranking'. The reader can infer that the three per-metric rank positions are averaged, but the exact formula, tie-breaking rule, and whether the average is computed before or after standard competition ranking are not specified. Since the final ranking is the paper's primary output, this should be defined precisely in Section 4 (e.g., final score = (r_PSNR + r_SSIM + r_LPIPS)/3, with ties assigned by competition ranking).
minor comments (7)
  1. [Sec. 4] Typo: 'Structured Similarity Index' should be 'Structural Similarity Index Measure'.
  2. [Sec. 2 and throughout] The dataset name is inconsistently typeset as 'V AR-MIORe' with a space; use a consistent form, e.g., 'VAR-MIORe'.
  3. [Sec. 7.1] Typo: 'the the employed EVSSM' should read 'the employed EVSSM'.
  4. [Tables 1 and 2 captions] Define 'rk.' and 'Final Rank' in the captions, since the table is not self-contained without Section 4's formula.
  5. [Sec. 7.5] The heading 'BlurKing Sharper' is confusing because Table 1 uses 'BlurKing' and Table 2 uses 'Sharper'. Please clarify that these are the same team's two track submissions.
  6. [Sec. 5] The sentence 'The fine-tuning of the test set was restricted by limiting user submissions' is unclear. It should state whether fine-tuning on test inputs was allowed, and what 'limited submissions' means operationally.
  7. [Tables 1 and 2] The test sets are small (75 and 63 images), and the PSNR gap between the top two teams (Track 1: 0.329 dB) is small. Reporting confidence intervals or a bootstrap significance test would help the reader judge whether the ranking differences are meaningful.

Circularity Check

0 steps flagged

No significant circularity; empirical challenge rankings are measured, not derived from fitted inputs.

full rationale

The paper is a challenge report rather than a derivation. The central claim — VPEG ranked first on the MIORe and VAR-MIORe test splits — is an empirical measurement obtained by evaluating submitted images against private test ground truths using fixed metrics (Section 4, Tables 1–2). The organizers did not fit the test labels or the ranking criteria to any participant's method, and participants were not given test ground truths during development or validation. The only self-referential element is that the benchmark datasets MIORe and VAR-MIORe are introduced by the same research group and cited as [5] in Section 2. This is a normal dataset citation and is not load-bearing for the ranking outcome: the dataset provides the testbed, but the scores are measured, not derived from the dataset definition or from any fitted parameter. The abstract's claim of 'significant progress' depends on the synthetic blur being representative of real high-FPS motion blur, but representativeness is a generalization/validity concern, not a circularity of the reported measurements. No equation in the paper reduces a prediction to its input by construction, no fitted parameter is renamed as a prediction, and no self-citation is invoked to force the choice of the winning method. Therefore no circular step is identified; the score reflects only the minor self-referential dataset citation.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The paper makes no mathematical derivation; the central outputs are empirical benchmark numbers. The only hand-chosen element is the composite ranking aggregation method, which affects the official standings. The background assumptions are domain assumptions about dataset representativeness, metric validity, and evaluation protocol integrity.

free parameters (1)
  • Composite final rank aggregation = unweighted average of per-metric ranks
    Tables 1 and 2 report a Final Rank derived from average ranking, but the exact aggregation (e.g., whether LPIPS is weighted equally) is not specified; this choice affects the standings, e.g., Mier has best LPIPS but ranks 6th overall.
axioms (4)
  • domain assumption Synthetic blur in MIORe and VAR-MIORe is representative of real high-FPS motion blur.
    Sec. 2: data are synthetically generated from sharp high-frame-rate frames; no validation against real blur shown in this report.
  • domain assumption PSNR, SSIM, and LPIPS are adequate and sufficient metrics to rank deblurring quality.
    Sec. 4 defines the evaluation solely by these metrics; no human study or other fidelity measure is used.
  • domain assumption The Codabench evaluation protocol correctly compares submitted images to private ground truths without data leaks or exploits.
    Sec. 5 describes phases; the protocol is trusted as-is, but the MagicBlur method (Sec. 7.9) uses test inputs for fine-tuning, suggesting the protocol permits transductive use.
  • domain assumption The composite Final Rank, derived from an unweighted average of per-metric ranks, is a meaningful way to determine winners.
    Sec. 6: 'Final Rank represents a composite metric, derived from the average ranking'; the aggregation choice is not justified or compared to alternatives.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of AIM 2025 Challenge on High FPS Motion Deblurring: Methods and Results." pith.science (2026). https://pith.science/paper/LJ5X46QB

@misc{pith2026250906793,
  author       = {Pith},
  title        = {Pith review of: AIM 2025 Challenge on High FPS Motion Deblurring: Methods and Results},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LJ5X46QB}},
  note         = {Machine review of arXiv:2509.06793}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper presents a comprehensive review of the AIM 2025 High FPS Non-Uniform Motion Deblurring Challenge, highlighting the proposed solutions and final results. The objective of this challenge is to identify effective networks capable of producing clearer and visually compelling images in diverse and challenging conditions, by learning representative visual cues for complex aggregations of motion types. A total of 68 participants registered for the competition, and 9 teams ultimately submitted valid entries. This paper thoroughly evaluates the state-of-the-art advances in high-FPS single image motion deblurring, showcasing the significant progress in the field, while leveraging samples of the novel dataset, MIORe, that introduces challenging examples of movement patterns.

Figures

Figures reproduced from arXiv: 2509.06793 by Chengjie Ge, Cong Zhang, Dafeng Zhang, Florin-Alexandru Vasluianu, George Ciubotariu, Guiyue Jin, Hao Chen, Jiangbo Guo, Jiangxin Dong, Jie Liu, Jinhui Tang, Jinshan Pan, Jiyu Jin, Jufeng Yang, Juncheng Zhou, Ke Wu, Lingshun Kong, Lishen Qu, Long Sun, Nancy Mehta, Qianfeng Yang, Qiyuan Guan, Qiyu Kang, Radu Timofte, Rui Zhao, Shihao Zhou, Shuai Liu, Shuhua Jin, Xiang Chen, Xin Lu, Xueyang Fu, Yaqi Luo, Yinghui Fang, Yongqi Song, Yufeng Peng, Yuqian Zhang, Zeyu Xiao, Zheng-Jun Zha, Zhihao Liu, Zhijing Sun, Zhongbao Yang, Zhuhao Zhang, Zhuoyuan Li, Zhuyun Zhou, Ziang Zhou, Zihao Li, Zishun Liao.

Figure 1
Figure 1. Figure 1: Visual comparison of the solutions proposed by top teams on samples from the AIM 2025 High FPS Motion Deblurring Challenge [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visual comparison of the solutions proposed by top teams on samples from the AIM 2025 High FPS Motion Deblurring Challenge [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: EVSSM architecture and inferences piplines [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FFTformer+ architectures. (a) The proposed asymmetric encoder-decoder network that only contains EDFFN in the encoder [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the technique proposed by team SRC-B for deblurring. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of the technique proposed by team X-L for deblurring. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Overview of the technique proposed by team BlurKing / [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Details of the Nankai-CVLab team’s architectural ap [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 33 canonical work pages · 1 internal anchor

  1. [1]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Shai Avi- dan, Gabriel J. Brostow, Moustapha Ciss ´e, Giovanni Maria Farinella, and Tal Hassner, editors,Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, Octo- ber 23-27, 2022, Proceedings, Part VII, volume 13667 of Lecture Notes in ...

    work page 2022

  2. [2]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. InEuropean confer- ence on computer vision, pages 17–33. Springer, 2022. 6

    2022

  3. [3]

    Hierarchical integration diffusion model for realistic image deblurring.NeurIPS, 2024

    Zheng Chen, Yulun Zhang, Ding Liu, Jinjin Gu, Linghe Kong, Xin Yuan, et al. Hierarchical integration diffusion model for realistic image deblurring.NeurIPS, 2024. 6

    work page 2024

  4. [4]

    AIM 2025 high FPS non-uniform motion deblurring challenge report

    George Ciubotariu, Florin-Alexandru Vasluianu, Zhuyun Zhou, Nancy Mehta, Radu Timofte, et al. AIM 2025 high FPS non-uniform motion deblurring challenge report. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025. 1

    work page 2025

  5. [5]

    MIORe & V AR-MIORe: Benchmarks to push the boundaries of restoration

    George Ciubotariu, Zhuyun Zhou, Zongwei Wu, and Radu Timofte. MIORe & V AR-MIORe: Benchmarks to push the boundaries of restoration. InProceedings of the IEEE In- ternational Conference on Computer Vision (ICCV). IEEE Computer Society, 2025. 1, 3, 6, 8

    work page 2025

  6. [6]

    AIM 2025 rip current segmentation (RipSeg) challenge report

    Andrei Dumitriu, Florin Miron, Florin Tatui, Radu Tudor Ionescu, Radu Timofte, Aakash Ralhan, Florin-Alexandru Vasluianu, et al. AIM 2025 rip current segmentation (RipSeg) challenge report. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Work- shops, 2025. 1

    work page 2025

  7. [7]

    Efficient real-world deblurring using single images: AIM 2025 chal- lenge report

    Daniel Feijoo, Paula Garrido, Marcos Conde, Jaesung Rim, Alvaro Garcia, Sunghyun Cho, Radu Timofte, et al. Efficient real-world deblurring using single images: AIM 2025 chal- lenge report. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025. 1

    work page 2025

  8. [8]

    4K image super-resolution on mobile NPUs: Mobile AI & AIM 2025 challenge report

    Andrey Ignatov, Georgy Perevozchikov, Radu Timofte, et al. 4K image super-resolution on mobile NPUs: Mobile AI & AIM 2025 challenge report. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Work- shops, 2025. 1

    work page 2025

  9. [9]

    Adapting stable diffusion for on-device inference: Mobile AI & AIM 2025 challenge report

    Andrey Ignatov, Georgy Perevozchikov, Radu Timofte, et al. Adapting stable diffusion for on-device inference: Mobile AI & AIM 2025 challenge report. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025. 1

    work page 2025

  10. [10]

    Efficient image denoising on smartphone GPUs: Mobile AI & AIM 2025 challenge report

    Andrey Ignatov, Georgy Perevozchikov, Radu Timofte, et al. Efficient image denoising on smartphone GPUs: Mobile AI & AIM 2025 challenge report. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025. 1

    work page 2025

  11. [11]

    Efficient learned smartphone ISP on mobile GPUs: Mo- bile AI & AIM 2025 challenge report

    Andrey Ignatov, Georgy Perevozchikov, Radu Timofte, et al. Efficient learned smartphone ISP on mobile GPUs: Mo- bile AI & AIM 2025 challenge report. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025. 1

    work page 2025

  12. [12]

    AIM 2025 challenge on robust offline video super-resolution: Dataset, methods and results

    Nikolai Karetin, Ivan Molodetskikh, Dmitry Vatolin, Radu Timofte, et al. AIM 2025 challenge on robust offline video super-resolution: Dataset, methods and results. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision (ICCV) Workshops, 2025. 1

    work page 2025

  13. [13]

    Efficient frequency domain-based trans- formers for high-quality image deblurring

    Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based trans- formers for high-quality image deblurring. InCVPR, 2023. 3

    work page 2023

  14. [14]

    Efficient visual state space model for image deblurring

    Lingshun Kong, Jiangxin Dong, Jinhui Tang, Ming-Hsuan Yang, and Jinshan Pan. Efficient visual state space model for image deblurring. InCVPR, 2025. 2, 3, 6

    work page 2025

  15. [15]

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural net- works.Commun. ACM, 60(6):84–90, may 2017. 2

    work page 2017

  16. [16]

    Flux.1 kontext: Flow matching for in-context image generation and editing in latent space,

    Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dock- horn, Jack English, Zion English, Patrick Esser, Sumith Ku- lal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas M¨uller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. Flux.1 kontext: Flow matching for in-context i...

  17. [17]

    Real-world raw de- noising using diverse cameras: AIM 2025 challenge report

    Feiran Li, Jiacheng Li, Marcos Conde, Beril Besbinar, Vlad Hosu, Daisuke Iso, Radu Timofte, et al. Real-world raw de- noising using diverse cameras: AIM 2025 challenge report. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025. 1

    work page 2025

  18. [18]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6

    2021

  19. [19]

    AIM 2025 perceptual image super-resolution chal- lenge

    Bruno Longarela, Marcos Conde, ´Alvaro Garc´ıa, Radu Tim- ofte, et al. AIM 2025 perceptual image super-resolution chal- lenge. InProceedings of the IEEE/CVF International Con- ference on Computer Vision (ICCV) Workshops, 2025. 1

    work page 2025

  20. [20]

    SGDR: stochastic gradient descent with warm restarts

    Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. InICLR, 2017. 3

    2017

  21. [21]

    Evenformer: Dynamic even transformer for real-world image restoration

    Xin Lu, Yuanfei Bao, Jiarong Yang, Anya Hu, Jie Xiao, Kunyu Wang, Dong Li, Senyan Xu, Kean Liu, Xueyang Fu, and Zheng-Jun Zha. Evenformer: Dynamic even transformer for real-world image restoration. InProceedings of the Com- puter Vision and Pattern Recognition Conference (CVPR) Workshops, pages 1081–1091, June 2025. 6

    work page 2025

  22. [22]

    Elucidating and Endowing the Diffusion Training Paradigm for General Image Restoration

    Xin Lu, Xueyang Fu, Jie Xiao, Zihao Fan, Yurui Zhu, and Zheng-Jun Zha. Elucidating and endowing the diffu- sion training paradigm for general image restoration.arXiv preprint arXiv:2506.21722, 2025. 6

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [23]

    Boosting inverse tone mapping via diffusion regularization

    Xin Lu, Yufeng Peng, Chengjie Ge, Zhijing Sun, Ziang Zhou, Zihao Li, Zishun Liao, Dong Li, Qiyu Kang, Xueyang Fu, and Zheng-Jun Zha. Boosting inverse tone mapping via diffusion regularization. In2025 IEEE/CVF International Conference on Computer Vision (ICCV) Workshops. IEEE Computer Society, 2025. 6

    work page 2025

  24. [24]

    Efficient high fps non-uniform mo- tion deblurring via progressive learning

    Xin Lu, Zhijing Sun, Chengjie Ge, Yufeng Peng, Ziang Zhou, Zihao Li, Zishun Liao, Dong Li, Qiyu Kang, Xueyang Fu, and Zheng-Jun Zha. Efficient high fps non-uniform mo- tion deblurring via progressive learning. In2025 IEEE/CVF International Conference on Computer Vision (ICCV) Work- shops. IEEE Computer Society, 2025. 6

    work page 2025

  25. [25]

    Continuous adverse weather removal via degradation-aware distillation

    Xin Lu, Jie Xiao, Yurui Zhu, and Xueyang Fu. Continuous adverse weather removal via degradation-aware distillation. InProceedings of the Computer Vision and Pattern Recog- nition Conference (CVPR), pages 28113–28123, June 2025. 6

    work page 2025

  26. [26]

    Advancing ambient lighting nor- malization via diffusion shadow generation

    Xin Lu, Jiarong Yang, Yuanfei Bao, Zihao Fan, Anya Hu, Kunyu Wang, Jie Xiao, Xi Wang, Hongjian Liu, Xueyang Fu, and Zheng-Jun Zha. Advancing ambient lighting nor- malization via diffusion shadow generation. InProceedings of the Computer Vision and Pattern Recognition Conference (CVPR) Workshops, pages 1070–1080, June 2025. 6

    work page 2025

  27. [27]

    Hirformer: Dynamic high resolution transformer for large-scale image shadow removal

    Xin Lu, Yurui Zhu, Xi Wang, Dong Li, Jie Xiao, Yun- peng Zhang, Xueyang Fu, and Zheng-Jun Zha. Hirformer: Dynamic high resolution transformer for large-scale image shadow removal. InProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 6513–6523, June 2024. 6

    work page 2024

  28. [28]

    Loformer: Local frequency transformer for im- age deblurring

    Xintian Mao, Jiansheng Wang, Xingran Xie, Qingli Li, and Yan Wang. Loformer: Local frequency transformer for im- age deblurring. In Jianfei Cai, Mohan S. Kankanhalli, Bal- akrishnan Prabhakaran, Susanne Boll, Ramanathan Subra- manian, Liang Zheng, Vivek K. Singh, Pablo C ´esar, Lexing Xie, and Dong Xu, editors,Proceedings of the 32nd ACM International Con...

    work page 2024

  29. [29]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3883–3891,

  30. [30]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 257–265. IEEE Computer Society,

    work page 2017

  31. [31]

    Real-world blur dataset for learning and benchmarking deblurring algorithms

    Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors,Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV, volume 12370 ofLecture Notes in Co...

    work page 2020

  32. [32]

    AIM 2025 challenge on screen-content video quality assessment: Methods and results

    Nickolay Safonov, Mikhail Rakhmanov, Dmitriy Vatolin, Radu Timofte, et al. AIM 2025 challenge on screen-content video quality assessment: Methods and results. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision (ICCV) Workshops, 2025. 1

    work page 2025

  33. [33]

    Chronos 2.1-hd high-speed camera

    Kron Technologies. Chronos 2.1-hd high-speed camera. https://www.krontech.ca/product/chronos- 2- 1- hd- high- speed- camera/, 2024. Accessed: 2024-11-11. 2

    work page 2024

  34. [34]

    AIM 2025 challenge on inverse tone mapping report: Methods and results

    Chao Wang, Francesco Banterle, Bin Ren, Radu Timofte, et al. AIM 2025 challenge on inverse tone mapping report: Methods and results. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV) Workshops,

    work page 2025

  35. [35]

    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 Si- moncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4):600–612, 2004. 2

    2004

  36. [36]

    Codabench: Flexible, easy-to-use, and reproducible meta-benchmark platform.Patterns, 3(7):100543, 2022

    Zhen Xu, Sergio Escalera, Adrien Pav ˜ao, Magali Richard, Wei-Wei Tu, Quanming Yao, Huan Zhao, and Isabelle Guyon. Codabench: Flexible, easy-to-use, and reproducible meta-benchmark platform.Patterns, 3(7):100543, 2022. 1, 2

    work page 2022

  37. [37]

    AIM 2025 low-light raw video denoising challenge: Dataset, methods and results

    Alexander Yakovenko, George Chakvetadze, Ilya Khrapov, Maksim Zhelezov, Dmitry Vatolin, Radu Timofte, et al. AIM 2025 low-light raw video denoising challenge: Dataset, methods and results. InProceedings of the IEEE/CVF Inter- national Conference on Computer Vision (ICCV) Workshops,

    work page 2025

  38. [38]

    Towards efficient and scale-robust ultra- high-definition image demoir´eing

    Xin Yu, Peng Dai, Wenbo Li, Lan Ma, Jiajun Shen, Jia Li, and Xiaojuan Qi. Towards efficient and scale-robust ultra- high-definition image demoir´eing. In Shai Avidan, Gabriel J. Brostow, Moustapha Ciss ´e, Giovanni Maria Farinella, and Tal Hassner, editors,Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proce...

    work page 2022

  39. [39]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InCVPR, 2022. 3

    2022

  40. [40]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728–5739, 2022. 8

    work page 2022

  41. [41]

    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. InCVPR, 2018. 2

    2018

  42. [42]

    Easycontrol: Adding efficient and flexible control for diffusion transformer.arXiv preprint arXiv:2503.07027, 2025

    Yuxuan Zhang, Yirui Yuan, Yiren Song, Haofan Wang, and Jiaming Liu. Easycontrol: Adding efficient and flexible control for diffusion transformer.arXiv preprint arXiv:2503.07027, 2025. 6 A. Teams and Affiliations AIM 2025 Team Title:AIM 2025 High FPS Motion Deblurring Challenge Members: George Ciubotariu 1 (george.ciubotariu@uni- wuerzburg.de), Florin Vasl...

    Pith/arXiv arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.