Pith. sign in

REVIEW 5 major objections 4 minor 51 references

A new real-world benchmark for all-in-one image restoration ranks 20 solutions, with the top three separating from the field.

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 ↗

T0 review

2026-08-01 08:22 UTC pith:DAPQEVG3

load-bearing objection The second LoViF challenge report is a genuinely useful survey of 20 restoration recipes, but the disclosed test leaks mean the leaderboard cannot yet be read as a clean benchmark. the 5 major comments →

arxiv 2607.21118 v1 pith:DAPQEVG3 submitted 2026-07-23 cs.CV

The Second LoViF 2026 Challenge on Real-World All-in-One Image Restoration: Methods and Results

Xiang Chen , Hao Li , Jiangxin Dong , Jinshan Pan , Xin Li , Hongbo Ding , Junpeng Jiang , Xingyu Qiu
show 84 more authors
Yilian Zhong Yuxiang Chen Shibo Yin Zixuan Huang Yushun Fang Xilei Zhu Yahui Wang Chen Lu Xiaodong Zhou Qingyue Cao Changwei Gong Jingyun Liu Xingchen Yi Hansen Shi Ruiyi Liu Jirui Xie Tao Liu Wenzhuo Ma Hongzhen Li Yongyong Chen Zheng Zhou Jingyong Su Jie Liu Haijin Zeng Cheng Li Peishuai Zha Ziyi Wang Jian Tang Yan Chen Long Bao Heng Sun Jiyuan Zhang Shuai Liu Wei Ding Chengjun Guo Yibin Huang Xiaotao Wang Dongqing Zou Lei Lei Xiaofeng Wang Xiao Liu Yulin Wu Yuhan Zhao Shurui Peng Chao Ren Yu-Kai Wang Kosuke Shigematsu Asuka Shin Rong-Lin Jian Cheng-Jun Kang Jin-Hui Jiang Jialin Zhou Kuo Yuan Songyu Zhang E B Benson Ashfaq Hussain Pruthvikanth AC Qirui Chen Jinyuan Chen Jun Zhang Xu Zhang Xuhui Cao Jiaqi Ma Laibin Chang Yuchun Miao Yichu Xu Yuanzhi Yao Shi Chen Yuning Cui Huan Zhang Lefei Zhang Saeed Ahmad Ik Hyun Lee Jun Young Park Ji Hwan Yoon Shangquan Sun Behrooz Nobahar-Moghanlou Majid Edalatjou Karim Shahi-Niyar Ruibo Zhang Dexiang Hong Xinyan Liu Shengeng Tang Weidong Chen
This is my paper
classification cs.CV
keywords all-in-one image restorationreal-world benchmarkimage degradationchallenge reportcomposite scorerestoration architecture surveycross-degradation generalizationfrequency-domain processing
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.

This paper reports the second edition of a challenge on real-world all-in-one image restoration, in which a single network must remove blur, low-light, haze, rain, and snow from genuine photographs. It claims that the curated benchmark — 24,500 paired training images, 500 validation images, and 500 test images, scored by a composite of PSNR, SSIM, and LPIPS — is a valid updated reference for comparing unified restoration methods. The reported results place three solutions in a distinct leading group (composite scores 42.28, 41.81, 41.40), well ahead of fourth place (34.81) and the rest of the field (34.58–28.47). The paper further argues that the submitted methods reveal effective design strategies — wavelet frequency decomposition, state-space blocks, one-step pixel-level diffusion, and mixture-of-experts routing — for handling multiple degradations in one network.

Core claim

The paper's central claim is that this challenge establishes an updated, reproducible benchmark for real-world all-in-one restoration. The evaluation rests on a composite score, PSNR plus 10 times SSIM minus 5 times LPIPS, computed on the luminance channel of 512-by-512 images spanning five degradation categories. Under this metric, the top three submissions form a distinct leading cluster — scores of 42.28, 41.81, and 41.40 — while the fourth-place method trails at 34.81 and the remaining 16 teams fall between 34.58 and 28.47. The paper asserts that the large gap indicates the leading solutions exploit shared representations across degradations more effectively, and it surveys each architec

What carries the argument

The central mechanism is the benchmark itself: a balanced multi-task dataset of 24,500 real paired low-quality and ground-truth images (4,900 per degradation), 500 validation images, and 500 test images, judged by a composite score of PSNR + 10×SSIM − 5×LPIPS on the YCbCr luminance channel. This single scalar is the load-bearing instrument: it grades fidelity, structure, and perceptual quality at once, and the paper uses it to rank 20 reproduced submissions. The second key component is the common evaluation protocol that all teams follow, which the paper treats as guaranteeing a fair cross-method comparison.

Load-bearing premise

The whole ranking is valid only if the test images were never seen by any submitted model — no near-duplicate test/training pairs, no routing by test-file order, no fine-tuning on test inputs — but the paper's own sections 4.9, 4.10, and 4.12 report all three practices, so the leaderboard may partly measure leakage rather than cross-degradation generalization.

What would settle it

Take a freshly collected, never-released set of 500 real degraded images from the same five categories, run the top three submissions exactly as described, and compare the composite-score gap to fourth place; if the 42.28/41.81/41.40 versus 34.81 gap collapses or the ordering changes, the original leaderboard measured test-set leakage, not cross-degradation skill.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the benchmark is valid, new all-in-one restoration methods can be compared against a fixed real-world reference rather than separate synthetic task benchmarks.
  • The 6.59-point gap between third and fourth place indicates the leading methods have crossed a quality threshold that most architectures have not yet reached.
  • The composite metric, by weighting SSIM tenfold and penalizing LPIPS, rewards methods that balance structural fidelity and perceptual quality rather than maximizing PSNR alone.
  • The leading solutions' use of wavelet decomposition and frequency-aware fusion suggests that explicit frequency separation is an effective inductive bias for handling heterogeneous degradations in a single network.

Where Pith is reading between the lines

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

  • The paper's own method sections report three practices that compromise the 'unseen test set' assumption: one submission routes every test image by the officially published test-index-to-degradation mapping, another fine-tunes on the official test inputs with pseudo-labels, and a third outputs training ground truth for 13 test images identified as near-duplicates of training pairs.
  • The leaderboard therefore probably measures test-set leakage as well as restoration skill; the top-three gap is partly driven by aggressive test-time self-ensembling and checkpoint fusion, so the architectural advantage may be smaller than the score gap suggests.
  • A cleaner benchmark would blind the test set — no public degradation order, no duplicate-pair allowance, no test-input fine-tuning — and would ask for a single forward pass per image; re-running the top methods under that protocol would show how much of the ranking is genuine.

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

5 major / 4 minor

Summary. The paper reports the Second LoViF 2026 Challenge on real-world all-in-one image restoration. It describes the FoundIR-LoViF benchmark (24,500 paired training images, 500 validation and 500 test images across blur, low-light, haze, rain, and snow), the evaluation protocol based on the composite score Score = PSNR + 10×SSIM − 5×LPIPS (Eq. 1), and the final ranking of 20 teams. The top three teams (Re:Pixel, REDnoteMediaLab, LucidWorld) achieve final scores 42.28, 41.81, and 41.40, with a 6.59-point gap to the 4th-ranked team, which the paper interprets as a distinct leading group. The paper also provides a method summary and training/testing details for each team.

Significance. If the benchmark is valid, the paper provides a useful updated comparison of all-in-one restoration methods on real-world paired data and documents current design strategies (state-space models, diffusion transformers, mixture-of-experts, degradation routing). Strengths include the use of a public dataset based on FoundIR and WeatherBench, a composite metric adopted from an established protocol, and the paper's unusual transparency in disclosing some team practices that affect evaluation fairness. However, these same disclosures show that the benchmark does not currently measure cross-degradation generalization as claimed: test-set leaks, test-time adaptation, and index-based routing are present in several submissions. The significance of the reported leaderboard therefore depends on whether the organizers re-analyze the results under a clean protocol.

major comments (5)
  1. [Section 4.9, Table 1] The final submission of team jason0411202 uses the official training ground-truth images as outputs for 13 test images identified as near-duplicates of training images, with only a JPEG-block dither. This is a direct test-set leak. The paper reports the team's rank (9th) and score (32.65) in Table 1 without excluding or re-ranking, and without quantifying the effect of these 13 images on the score. The claim that the benchmark measures generalization to unseen degradations is therefore not valid for this submission. The authors should re-evaluate all submissions under a protocol that forbids using training ground truth for test images, or at least report the leaderboard with and without this correction.
  2. [Section 4.12] Team MoEFlow performs transductive pseudo-label fine-tuning on the official test inputs and blends the pseudo-finetuned model into the final output (15% weight for rain/snow, 20% for other categories). The paper concedes that 'pseudo labels may introduce test-set bias' but still reports the team's rank (12th) and score (32.17) without quantifying the bias. This practice violates the standard benchmark assumption that test inputs remain unseen during training. The authors need to either disqualify such adaptation or provide a sensitivity analysis showing its effect on the leaderboard.
  3. [Sections 4.9, 4.10, 4.12 (also 4.6)] Several teams use the officially published test-image index ordering (blur=0001-0100, low-light=0101-0200, etc.) to determine the degradation category at inference instead of learning a classifier. Since the composite score in Eq. (1) rewards category-specific restoration, this protocol leak gives these teams an unfair advantage and undermines the 'all-in-one blind restoration' claim. The paper should list every team that used index routing and assess whether the leaderboard would change if this information were withheld.
  4. [Section 4.6] Team MiVideo states that 'the final checkpoint is selected according to its score on the online leaderboard.' The paper does not define whether the online leaderboard reflects validation or test scores. If it reflects test-phase feedback, this is another form of test-set-based model selection. Please clarify and, if it is the test set, treat it as a protocol violation or justify its permissibility.
  5. [Section 3] The paper states that the 20 included teams 'were successfully reproduced and verified' but provides no details of the verification procedure (e.g., whether code was re-run, whether outputs matched, what tolerance was used). Without this information, the reader cannot assess the reliability of the reported scores. Please describe the verification protocol or soften the claim.
minor comments (4)
  1. [Table 1 vs Section 4.4] The team name appears as 'Pheonix' in Table 1 and Section 4.4 but 'Phoenix' in the team list and affiliations. Please unify the spelling.
  2. [Section 4.6 vs Table 1] Section 4.6 reports 947.22 GFLOPs, while Table 1 lists 974.22 G. Please reconcile the discrepancy.
  3. [Section 4.9, Figure 8] The figure mentions 'DC dither' for the duplicate-pair patch, but the term is not defined in the text. Please explain the perturbation applied to the 1/16 JPEG blocks.
  4. [Section 4.16 and References] Section 4.16 reports '10,661 million activations'; please use '10.66 billion' for clarity. Also, Reference [7] is cited as a CVPR 2026 paper, but the challenge is held at ECCV 2026; please correct the venue or clarify.

Circularity Check

0 steps flagged

No circularity: the leaderboard is an external measurement; the composite score is an adopted evaluation rule, not a fitted or derived quantity.

full rationale

This is a challenge report rather than a derivation-driven paper. The only closed-form quantity is the composite score in Eq. (1), Score = PSNR(Y) + 10*SSIM(Y) - 5*LPIPS, which is explicitly adopted from the previous edition's protocol rather than fitted to the current results. The final rankings are direct measurements of submitted outputs against held-out ground-truth images; no parameter of Eq. (1) is inferred from Table 1, and no method's score is used to define another method's score. The claims about a leading group and the gap to fourth place are arithmetic descriptions of the measured table, not predictions derived from the inputs. The disclosed practices in Sections 4.9, 4.10, and 4.12—duplicate-pair ground-truth output for 13 blur images, transductive pseudo-label fine-tuning on test inputs, and index-based degradation routing—are real threats to the validity of the benchmark as a measure of cross-degradation generalization, but they are protocol-integrity concerns rather than instances of a result being equivalent to its inputs by construction. Citations to the organizers' prior FoundIR work and to the previous LoViF edition provide data and evaluation-protocol provenance, but the measured scores are external observations and do not reduce to those citations. No self-definitional step, fitted-input-as-prediction, or author-imported uniqueness argument appears in the paper.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The paper introduces no new physical or mathematical entities. Its free parameters are the composite-score weights and a set of benchmark-integrity assumptions about alignment, metric validity, and test-set separation that are load-bearing for the leaderboard claim.

free parameters (1)
  • Composite score weights (w_PSNR, w_SSIM, w_LPIPS) = 1, 10, -5
    Equation (1). These hand-chosen/adopted weights determine the final ranking; alternative weights could reorder teams outside the top cluster. They are protocol constants, not fitted to data, but they are still arbitrary choices the central claim depends on.
axioms (4)
  • domain assumption Ground-truth clean images in FoundIR/WeatherBench are correctly aligned and represent the true scene content.
    Section 2 states the benchmark is curated from FoundIR and WeatherBench; if the paired alignment or reference quality is wrong, every leaderboard score is suspect.
  • domain assumption The composite Score = PSNR + 10*SSIM - 5*LPIPS is a valid ordinal measure of restoration quality.
    Equation (1). The weights are adopted from the previous edition; the paper provides no validation that this weighted sum matches human preference or downstream utility.
  • domain assumption The test set is free of near-duplicates of training images and teams do not use test-set information beyond the degraded pixels.
    Section 2 describes a standard train/validation/test split; Section 4.9 contradicts this by outputting training ground truths for 13 near-identical test images, and Section 4.12 contradicts it by pseudo-label fine-tuning on test inputs.
  • domain assumption The fixed test-image order does not reveal degradation categories in a way that changes the all-in-one evaluation.
    Sections 4.9, 4.10, and 4.12 use the competition-defined image order (e.g., 0001-0100 blur, 0101-0200 low-light) to route inputs, which means those entries are not blind unified models.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of The Second LoViF 2026 Challenge on Real-World All-in-One Image Restoration: Methods and Results." pith.science (2026). https://pith.science/paper/DAPQEVG3

@misc{pith2026260721118,
  author       = {Pith},
  title        = {Pith review of: The Second LoViF 2026 Challenge on Real-World All-in-One Image Restoration: Methods and Results},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DAPQEVG3}},
  note         = {Machine review of arXiv:2607.21118}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

This paper presents a review of the second LoViF Challenge on Real-World All-in-One Image Restoration. The challenge aims to advance unified image restoration under diverse real-world degradation conditions, including blur, low-light, haze, rain, and snow. It provides a common benchmark for evaluating the restoration accuracy, robustness, and generalization capability of models across multiple degradation categories within a unified framework. The competition attracted 158 registered participants, and 20 teams were included in the final ranking after their submitted results were successfully reproduced and verified. This report provides a comprehensive analysis of the submitted solutions and corresponding results, highlighting recent advances in real-world all-in-one image restoration. The summarized methods and empirical findings reveal effective design strategies and establish an updated benchmark for future research in real-world low-level vision.

Figures

Figures reproduced from arXiv: 2607.21118 by Ashfaq Hussain, Asuka Shin, Behrooz Nobahar-Moghanlou, Changwei Gong, Chao Ren, Chengjun Guo, Cheng-Jun Kang, Cheng Li, Chen Lu, Dexiang Hong, Dongqing Zou, E B Benson, Haijin Zeng, Hansen Shi, Hao Li, Heng Sun, Hongbo Ding, Hongzhen Li, Huan Zhang, Ik Hyun Lee, Jialin Zhou, Jiangxin Dong, Jian Tang, Jiaqi Ma, Jie Liu, Ji Hwan Yoon, Jingyong Su, Jingyun Liu, Jin-Hui Jiang, Jinshan Pan, Jinyuan Chen, Jirui Xie, Jiyuan Zhang, Junpeng Jiang, Jun Young Park, Jun Zhang, Karim Shahi-Niyar, Kosuke Shigematsu, Kuo Yuan, Laibin Chang, Lefei Zhang, Lei Lei, Long Bao, Majid Edalatjou, Peishuai Zha, Pruthvikanth AC, Qingyue Cao, Qirui Chen, Rong-Lin Jian, Ruibo Zhang, Ruiyi Liu, Saeed Ahmad, Shangquan Sun, Shengeng Tang, Shibo Yin, Shi Chen, Shuai Liu, Shurui Peng, Songyu Zhang, Tao Liu, Wei Ding, Weidong Chen, Wenzhuo Ma, Xiang Chen, Xiaodong Zhou, Xiaofeng Wang, Xiao Liu, Xiaotao Wang, Xilei Zhu, Xingchen Yi, Xingyu Qiu, Xin Li, Xinyan Liu, Xuhui Cao, Xu Zhang, Yahui Wang, Yan Chen, Yibin Huang, Yichu Xu, Yilian Zhong, Yongyong Chen, Yuanzhi Yao, Yuchun Miao, Yuhan Zhao, Yu-Kai Wang, Yulin Wu, Yuning Cui, Yushun Fang, Yuxiang Chen, Zheng Zhou, Zixuan Huang, Ziyi Wang.

Figure 1
Figure 1. Figure 1: The overall architecture of ReMamba proposed by Team Re:Pixel. are decomposed into LL, LH, HL, and HH subbands. ReMamba consists of a High-Frequency State-Aware Block, a Low-Frequency Information Enhancement Block, and a Wavelet Fusion Block. The High-Frequency State-Aware Block com￾bines patch attention with a Vision State Space Module to enhance spatially non￾uniform high-frequency degradations, while a … view at source ↗
Figure 2
Figure 2. Figure 2: The overall architecture of DreamIR proposed by Team REDnoteMediaLab. DreamIR jointly processes text, degraded-image features, timestep information, and image patch tokens to produce a restored image in a single forward pass. second stage, the model is fine-tuned for 100 epochs on the official competition training set using only flow-matching supervision. In the third stage, it is fur￾ther trained for 50 e… view at source ↗
Figure 3
Figure 3. Figure 3: The overall framework of HyRoute proposed by Team Phoenix. dicted category is used for hard routing to the corresponding restoration ex￾pert. This design allows individual experts to specialize in different degradation patterns and reduces interference among heterogeneous restoration tasks. The second stage introduces a task-shared mixture-of-experts module for content￾adaptive refinement. Multiple shared … view at source ↗
Figure 4
Figure 4. Figure 4: The overall framework of RIRNet proposed by Team SeeIR. A ConvNeXt-based DINOv3 backbone extracts high-level semantic representations, while the multi-scale encoder–decoder branch progressively restores image structures and details. 4.5 SeeIR This team proposes RIRNet, a unified representation-based image restoration network for real-world all-in-one image restoration. RIRNet combines a pre￾trained ConvNeX… view at source ↗
Figure 5
Figure 5. Figure 5: The overall architecture of AdaIGFormer proposed by Team MiVideo. categories using supervised contrastive learning together with an auxiliary clas￾sification loss. The encoder produces a 128-dimensional condition embedding and is trained for 2,500 iterations using AdamW with a learning rate of 0.00005. In the second stage, the restoration network is trained on approximately 72,357 image pairs sampled from … view at source ↗
Figure 6
Figure 6. Figure 6: The real-world all-in-one restoration framework proposed by Team Mi￾Algo_LM. image is produced by a single model without degradation-specific checkpoint selection. The factsheet does not report the use of test-time augmentation, model ensembling, or output-level fusion. The final checkpoint is selected according to its score on the online leaderboard. The model requires approximately 433 milliseconds to pr… view at source ↗
Figure 7
Figure 7. Figure 7: The overall framework proposed by Team Pamy. batch size of 5. Adam is adopted with an initial learning rate of 0.0003, beta val￾ues of 0.9 and 0.99, and cosine annealing to 0.000001. The objective combines a PSNR-oriented reconstruction loss with multi-scale uncertainty supervision. The auxiliary reconstruction term is assigned a weight of 0.5, while the uncertainty variance regularization term is assigned… view at source ↗
Figure 8
Figure 8. Figure 8: UniNAF-MoRE inference pipeline. One architecture, one parameter file; the routing signal is the officially published index-to-degradation mapping. Training details. All experts are initialized using the publicly available NAFNet￾REDS-width64 pretrained weights. The blur, low-light, and haze experts are independently trained using the corresponding 4,900 paired images from the of￾ficial LoViF training set. … view at source ↗
Figure 9
Figure 9. Figure 9: The overall framework of Team Zzz. The degradation classifier routes high￾confidence blur inputs to the NAFNet expert, while the remaining inputs are restored by the DATPRL-IR-based all-in-one model. Mild CLAHE preprocessing is applied to high-confidence haze inputs. DATPRL-IR-based all-in-one model. For high-confidence haze inputs, the team additionally applies mild contrast-limited adaptive histogram equ… view at source ↗
Figure 10
Figure 10. Figure 10: The overall framework of Team BaseLess. FFC-Restore combines an encoder￾decoder architecture, Fourier Feature Convolution blocks, pixel attention, and a mixture-of-experts bottleneck for unified image restoration. 96. The method does not use model ensembling, test-time augmentation, multi￾model fusion, or degradation-specific output selection. The two internal experts are combined through the learned soft… view at source ↗
Figure 11
Figure 11. Figure 11: The PromptCIR-style backbone adopted by Team IK Lab. The original at￾tention blocks are replaced with Mamba-based state-space blocks, while degradation routing, routed prompts, wavelet refinement, and a gated deblurring branch are added to the restoration network. assigned to blur, low-light, and haze because these categories are more difficult on the internal validation split. Progressive patch-size trai… view at source ↗
Figure 12
Figure 12. Figure 12: The final inference pipeline of Team IK Lab. The backbone predicts both the restored image and degradation probabilities, after which a content-aware gate selects the backbone result, the EVSSM blur expert, or the fused low-light expert. approximately 3.21 seconds per image on a single NVIDIA RTX 4090 GPU, with a peak memory consumption of approximately 8.8 GB. Their final submission achieves an overall t… view at source ↗
Figure 13
Figure 13. Figure 13: The overall architecture of SupMoE proposed by Team Ipara. of the main-stream output with 5 percent of the auxiliary NAFNet output. The restored images are saved as RGB JPEG files with a quality value of 96, disabled chroma subsampling, and optimized encoding. The reported runtime is approx￾imately 0.20 seconds per image on a GPU. Their final submission achieves an overall testing score of 29.86. 4.19 Ipa… 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

51 extracted references · 7 linked inside Pith

  1. [1]

    In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

    Ainslie, J., Lee-Thorp, J., De Jong, M., Zemlyanskiy, Y., Lebrón, F., Sanghai, S.: Gqa: Training generalized multi-query transformer models from multi-head check- points. In: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. pp. 4895–4901 (2023)

    2023

  2. [2]

    IEEE Transactions on Image Processing27(4), 2049–2062 (2018)

    Cai, J., Gu, S., Zhang, L.: Learning a deep single image contrast enhancer from multi-exposure images. IEEE Transactions on Image Processing27(4), 2049–2062 (2018)

    2049

  3. [3]

    arXiv preprint arXiv:2605.11061 (2026)

    Cai, Q., Chen, J., Gao, C., Gong, Z., Li, Y., Pan, Y., Peng, Y., Qiu, Z., Yu, K., Zhang, Y., et al.: Hidream-o1-image: A natively unified image generative founda- tion model with pixel-level unified transformer. arXiv preprint arXiv:2605.11061 (2026)

    Pith/arXiv arXiv 2026

  4. [4]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Cai, Y., Bian, H., Lin, J., Wang, H., Timofte, R., Zhang, Y.: Retinexformer: One- stage retinex-based transformer for low-light image enhancement. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12504–12513 (2023)

    2023

  5. [5]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 9650–9660 (2021)

    2021

  6. [6]

    In: Proceedings of the European Conference on Computer Vision (2022)

    Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: Proceedings of the European Conference on Computer Vision (2022)

    2022

  7. [7]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, X., Li, H., Dong, J., Pan, J., Li, X., He, X., Chen, N., Li, S., Liu, F., Lv, H., et al.: Lovif 2026 challenge on real-world all-in-one image restoration: Methods and results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 4927–4935 (2026)

    2026

  8. [8]

    In: CVPR (2023)

    Chen, X., Li, H., Li, M., Pan, J.: Learning a sparse transformer network for effective image deraining. In: CVPR (2023)

    2023

  9. [9]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Chen, X., Pan, J., Dong, J., Yang, J., Tang, J.: Foundir-v2: Optimizing pre- training data mixtures for image restoration foundation model. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8471–8480 (2026)

    2026

  10. [10]

    arXiv preprint arXiv:2310.11881 (2023) 34 X

    Chen, X., Li, Z., Pu, Y., Liu, Y., Zhou, J., Qiao, Y., Dong, C.: A comparative study of image restoration networks for general backbone network design. arXiv preprint arXiv:2310.11881 (2023) 34 X. Chen et al

    Pith/arXiv arXiv 2023

  11. [11]

    Advances in Neural Information Processing Systems38, 80452–80481 (2026)

    Cui, Y., Ren, W., Knoll, A.: Bio-inspired image restoration. Advances in Neural Information Processing Systems38, 80452–80481 (2026)

    2026

  12. [12]

    In: International Conference on Learning Representations

    Cui, Y., Zamir, S.W., Khan, S., Knoll, A., Shah, M., Khan, F.: Adair: Adaptive all- in-one image restoration via frequency mining and modulation. In: International Conference on Learning Representations. vol. 2025, pp. 101306–101327 (2025)

    2025

  13. [13]

    arXiv preprint arXiv:2603.01725 (2026)

    Dong, G., Li, C., Ren, C., Hu, J., Shi, Y., Zhu, X.X., Mou, L.: Learning domain- aware task prompt representations for multi-domain all-in-one image restoration. arXiv preprint arXiv:2603.01725 (2026)

    arXiv 2026

  14. [14]

    In: CVPR

    Feijoo, D., Benito, J.C., Garcia, A., Conde, M.V.: DarkIR: Robust low-light image restoration. In: CVPR. pp. 10879–10889 (2025)

    2025

  15. [15]

    arXiv:2312.00752 (2023)

    Gu, A., Dao, T.: Mamba: Linear-time sequence modeling with selective state spaces. arXiv:2312.00752 (2023)

    Pith/arXiv arXiv 2023

  16. [16]

    Advances in Neural Information Processing Systems38, 3189–3225 (2026)

    Guan, Q., Chen, X., Jin, G., Jin, J., Fan, S., Song, T., Pan, J.: Rethinking night- time image deraining via learnable color space transformation. Advances in Neural Information Processing Systems38, 3189–3225 (2026)

    2026

  17. [17]

    IEEE Transactions on Multimedia (2026)

    Guan,Q.,Fan,S.,Jin,J.,Jin,G.,Song,T.,Li,P.,Chen,X.,Jiang,K.:Harmonizing light and darkness: Nighttime image deraining using color space transformation. IEEE Transactions on Multimedia (2026)

    2026

  18. [18]

    In: Pro- ceedingsofthe33rdACMinternationalconferenceonmultimedia.pp.12607–12613 (2025)

    Guan, Q., Yang, Q., Chen, X., Song, T., Jin, G., Jin, J.: Weatherbench: A real- world benchmark dataset for all-in-one adverse weather image restoration. In: Pro- ceedingsofthe33rdACMinternationalconferenceonmultimedia.pp.12607–12613 (2025)

    2025

  19. [19]

    In: ECCV (2024)

    Guo, Y., Chen, J., Wang, Z., Jin, X., Feng, J.: Mambairv2: Attentive state space models for efficient image restoration. In: ECCV (2024)

    2024

  20. [20]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Howard, A., Sandler, M., Chu, G., Chen, L.C., Chen, B., Tan, M., Wang, W., Zhu, Y., Pang, R., Vasudevan, V., Le, Q.V., Adam, H.: Searching for mobilenetv3. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1314–1324 (2019)

    2019

  21. [21]

    IEEE TPAMI (2025)

    Jiang, J., Zuo, Z., Wu, G., Jiang, K., Liu, X.: A survey on all-in-one image restora- tion: Taxonomy, evaluation and future trends. IEEE TPAMI (2025)

    2025

  22. [22]

    In: Proceedings of the 33rd ACM International Conference on Multimedia

    Jin, W., Yang, Q., Wu, X., Chen, H., Li, P., Chen, X.: Smokebench: A real-world dataset for surveillance image desmoking in early-stage fire scenes. In: Proceedings of the 33rd ACM International Conference on Multimedia. pp. 12722–12728 (2025)

    2025

  23. [23]

    In: CVPR (2023)

    Kong, L., Dong, J., Ge, J., Li, M., Pan, J.: Efficient frequency domain-based trans- formers for high-quality image deblurring. In: CVPR (2023)

    2023

  24. [24]

    In: CVPR (2025)

    Kong, L., Dong, J., Tang, J., Yang, M.H., Pan, J.: Efficient visual state space model for image deblurring. In: CVPR (2025)

    2025

  25. [25]

    In: CVPRW (2024)

    Li, B., Li, X., Lu, Y., Feng, R., Guo, M., Zhao, S., Zhang, L., Chen, Z.: PromptCIR: Blind compressed image restoration with prompt learning. In: CVPRW (2024)

    2024

  26. [26]

    In: CVPR (2022)

    Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR (2022)

    2022

  27. [27]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Li, H., Chen, X., Dong, J., Tang, J., Pan, J.: Foundir: Unleashing million-scale training data to advance foundation models for image restoration. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 12626–12636 (2025)

    2025

  28. [28]

    In: CVPR Workshop (2025)

    Li, X., Jin, Y., Jin, X., Wu, Z., Li, B., Wang, Y., Yang, W., Li, Y., Chen, Z., Wen, B., et al.: Ntire 2025 challenge on day and night raindrop removal for dual-focused images: Methods and results. In: CVPR Workshop (2025)

    2025

  29. [29]

    CoRRabs/2504.02016(2025) Title Suppressed Due to Excessive Length 35

    Liu, Z., Zhang, F., Song, W., Li, X., Wei, W.: Fast fourier correlation is a highly efficient and accurate feature attribution algorithm from the perspective of control theory and game theory. CoRRabs/2504.02016(2025) Title Suppressed Due to Excessive Length 35

    Pith/arXiv arXiv 2025

  30. [30]

    arXiv:2504.17732 (2025)

    Liu, Z., et al.: DPMambaIR: All-in-one image restoration via degradation-aware prompt state space model. arXiv:2504.17732 (2025)

    arXiv 2025

  31. [31]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11976–11986 (2022)

    2022

  32. [32]

    In: ICLR

    Luo, Z., Gustafsson, F.K., Zhao, Z., Sjölund, J., Schön, T.: Controlling vision- language models for multi-task image restoration. In: ICLR. vol. 2024, pp. 16226– 16246 (2024)

    2024

  33. [33]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    Nah, S., Kim, T.H., Lee, K.M.: Deep multi-scale convolutional neural network for dynamic scene deblurring. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3883–3891 (2017)

    2017

  34. [34]

    arXiv preprint arXiv:2605.13258 (2026)

    Pan, Y., Cao, L., Zhu, Y., Zhu, F.: X-restormer++: 1st place solution for the ug2+ cvpr 2026 all-weather restoration challenge. arXiv preprint arXiv:2605.13258 (2026)

    Pith/arXiv arXiv 2026

  35. [35]

    In: NeurIPS (2023)

    Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: PromptIR: Prompting for all- in-one blind image restoration. In: NeurIPS (2023)

    2023

  36. [36]

    In: International Conference on Learning Representations (2026)

    Ren, B., Li, Y., Zheng, X., Fu, Y., Paudel, D.P., Liu, H., Yang, M.H., Gool, L.V., Sebe, N.: Efficient degradation-agnostic image restoration via channel-wise func- tional decomposition and manifold regularization. In: International Conference on Learning Representations (2026)

    2026

  37. [37]

    arXiv preprint arXiv:2002.05202 (2020)

    Shazeer, N.: Glu variants improve transformer. arXiv preprint arXiv:2002.05202 (2020)

    Pith/arXiv arXiv 2002

  38. [38]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Shen, Z., Wang, W., Lu, X., Shen, J., Ling, H., Xu, T., Shao, L.: Human-aware motion deblurring. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5572–5581 (2019)

    2019

  39. [39]

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., Tolan, J., Brandt, J., Couprie, C., Mairal, J., Jégou, H., Labatut, P., Bojanowski, P.: DINOv3 (2025),https://ar...

    Pith/arXiv arXiv 2025

  40. [40]

    Neurocomputing568, 127063 (2024)

    Su, J., Ahmed, M., Lu, Y., Pan, S., Bo, W., Liu, Y.: Roformer: Enhanced trans- former with rotary position embedding. Neurocomputing568, 127063 (2024)

    2024

  41. [41]

    VCLab, T.H.K.P.U.: Ggt-100k dataset.https://huggingface.co/datasets/ VCLab-PolyU/GGT-100K(2025), hugging Face Dataset

    2025

  42. [42]

    Information Processing & Management63(2, Part B), 104380 (2026)

    Wang, X., Liu, X., Yang, Y., Wang, Z., He, X., Chen, H., Li, Y., Wang, P.: Dsrir: Dynamic spatial refinement learning for progressive all-in-one image restoration. Information Processing & Management63(2, Part B), 104380 (2026)

    2026

  43. [43]

    In: CVPR (2025)

    Yan, Q., et al.: HVI: A new color space for low-light image enhancement. In: CVPR (2025)

    2025

  44. [44]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Yang, Q., Chen, X., Li, P., Guan, Q., Jin, G., Jin, J.: Rethinking rainy 3d scene reconstruction via perspective transforming and brightness tuning. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 40, pp. 11658–11666 (2026)

    2026

  45. [45]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yang, Q., Guan, Q., Chen, X., Jin, J., Jin, G., Dong, J.: Unirain: Unified image deraining with rag-based dataset distillation and multi-objective reweighted opti- mization. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 12428–12437 (2026)

    2026

  46. [46]

    The Visual Computer42(4), 193 (2026) 36 X

    Yang, Q., Li, P., Jin, J., Jin, G., Song, T., Fan, S., Hou, H.: Textual–visual inter- action for enhanced single image deraining using adapter-tuned vlms. The Visual Computer42(4), 193 (2026) 36 X. Chen et al

    2026

  47. [47]

    In: ECCV (2020)

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for real image restoration and enhancement. In: ECCV (2020)

    2020

  48. [48]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.: Restormer: Efficient transformer for high-resolution image restoration. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5718– 5729 (2022)

    2022

  49. [49]

    In: Advances in neural information processing systems

    Zhang, B., Sennrich, R.: Root mean square layer normalization. In: Advances in neural information processing systems. vol. 32 (2019)

    2019

  50. [50]

    In: CVPR

    Zhang, R., Isola, P., Efros, A.A., Shechtman, E., Wang, O.: The unreasonable effectiveness of deep features as a perceptual metric. In: CVPR. pp. 586–595 (2018)

    2018

  51. [51]

    In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion

    Zheng, D., Wu, X.M., Yang, S., Zhang, J., Hu, J.F., Zheng, W.S.: Selective hour- glass mapping for universal image restoration based on diffusion model. In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion. pp. 25445–25455 (2024)

    2024

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.