Pith. sign in

REVIEW 3 major objections 6 minor 94 references

QuReC claims that replacing a single image-level degradation prompt with per-query prototype matching, plus calibrating local-global attention, makes one restoration model the best across three all-in-one benchmark suites—including all 11 c

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 · deepseek-v4-flash

2026-08-02 00:08 UTC pith:XPMVJN4M

load-bearing objection A well-ablated, genuinely new query-specific prompt mechanism for all-in-one restoration; the SOTA margins are plausible but rest on single unseeded runs and undocumented baseline provenance. the 3 major comments →

arxiv 2607.15097 v2 pith:XPMVJN4M submitted 2026-07-16 cs.CV

QuReC: All-in-One Image Restoration with Query-Specific Guidance and Local-Global Response Calibration

Shen Zhou , Jinghui Zhang , Wenbo Huang , Xuwei Qian , Zhen Wu , Guangwen Peng , Zhiyuan Li , Ding Ding
show 2 more authors
Dian Shen Fang Dong
This is my paper
classification cs.CV
keywords all-in-one image restorationquery-specific degradation guidanceprototype matchinglocal-global attentionresponse calibrationcomposite degradationdenoising/deraining/dehazing/deblurring/low-lightCDD11
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 paper sets out to fix a blind spot in 'all-in-one' image-restoration models—networks that clear noise, rain, haze, blur, and low light with a single set of weights. Existing prompt-based models send the whole image one shared hint about what is wrong; QuReC instead gives every spatial location its own degradation-aware query, then calibrates how that query pulls together local details and global context. The paper reports that this query-specific guidance plus calibrated aggregation reaches the best average PSNR/SSIM on three benchmark suites (three-task, five-task, and the 11-scenario CDD11 composite set), including best marks in all 11 CDD11 settings. The two new modules—DQRM for query reconstruction via prototype matching, and LGRCM for local-global response calibration—each contribute ablation gains, and their combination plus a prototype-matching learning objective yields the reported results. If the numbers hold under a common training protocol, the paper would show that finer-grained degradation guidance, not bigger backbones, is the lever for unified restoration.

Core claim

The central claim is that replacing one image-level degradation prompt with per-query degradation semantics, and then re-weighting local and global attention outputs with learnable priors, lets a single restoration network outperform dedicated and all-in-one baselines on denoising, deraining, dehazing, deblurring, low-light enhancement, and composite mixes of them. DQRM matches each spatial query against a bank of degradation prototypes (built from text embeddings of degradation names) to reconstruct a query-specific prompt, aided by a weakly supervised loss that balances prototype usage and pulls the average matching distribution toward the degradations actually present in the image. LGRCM

What carries the argument

The load-bearing mechanism is the combination of (1) a Degradation-Guided Query Reconstruction Module (DQRM), in which each spatial query soft-matches against a frozen text-embedding prototype bank and adds the matched prototype signal plus a shared anchor to its base query; and (2) a Local-Global Response Calibration Module (LGRCM), which computes local-window and global attention in a shared softmax space and gates the outputs with prior responses retrieved from the base query. The two are inserted as Query Reconstruction and Response Calibration Blocks (QRCBs) at three decoder stages, so guidance is refined progressively from coarse to fine. A weakly supervised matching loss (load balanci

Load-bearing premise

The central claim assumes that the baseline numbers in Tables 1–3 were produced under the same all-in-one training/evaluation protocol as QuReC; the paper does not state whether baselines were retrained in-house or how many seeds were used, and some winning margins are smaller than typical run-to-run variation.

What would settle it

Run the compared baselines and QuReC under one identical training schedule, dataset splits, patch sizes, and evaluation script with multiple seeds; if margins on SOTS, BSD68-denosing, and other settings shrink to near zero or flip, the state-of-the-art claim would not survive. A targeted behavioral falsifier: synthesize images that are half haze and half rain (or haze plus local rain patches) and check whether the model applies qualitatively different restoration in the two regions—if the output looks like a single global prompt was used, the query-specific guidance claim is undercut.

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

If this is right

  • A single model can now handle spatially heterogeneous and mixed degradations without needing a separate branch per degradation type.
  • Because queries are reconstructed per location, the same prototype bank can in principle scale to more degradation categories by adding prototypes, assuming the matching loss is updated accordingly.
  • Joint normalization of local and global attention forces the two branches to compete in one probability space, which the ablations show improves coordination over independent branches.
  • The weak semantic supervision makes routing interpretable: t-SNE shows separation of degradation clusters, enabling region-level diagnostics of what the model thinks is wrong.
  • The gains on the composite CDD11 set (e.g., about 3.7 dB over the prior best model on the haze setting) suggest composite degradations benefit most from per-query guidance.

Where Pith is reading between the lines

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

  • The prototype bank is fixed to the degradations seen in training; extending toward open-set or never-seen degradations via dynamic prototype creation is a natural next step that the paper's own OOD experiment shows is currently limited.
  • The reported margins on some tasks (e.g., 0.08 dB on SOTS in the five-task table) are smaller than typical run-to-run noise; a controlled re-benchmark with identical training protocol and multiple seeds would make the superiority claim robust.
  • Because the matching loss uses only image-level degradation labels, the model learns spatial attribution without dense annotation; one could test whether the learned attribution matches human-annotated degradation maps, which the qualitative visualizations suggest but do not quantify.
  • The extra overhead of DQRM over LGRCM is small (0.89M params, 3.27G FLOPs, about 0.001s), so the per-query mechanism may transfer to real-time or mobile restoration pipelines if the base attention cost is acceptable.

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

3 major / 6 minor

Summary. The paper proposes QuReC, a unified all-in-one image restoration framework. Two modules are introduced: (i) DQRM, which builds a degradation prototype bank from CLIP text embeddings and reconstructs query-specific degradation-aware queries by softly matching each spatial token against the prototypes, regularized by a load-balancing loss and an image-level weak matching loss; and (ii) LGRCM, which performs joint local-global attention with learnable prior-gated calibration. The model is evaluated on three benchmark suites: three-task all-in-one (denoising, deraining, dehazing), five-task all-in-one (additionally deblurring and low-light enhancement), and the CDD11 composite-degradation benchmark. The paper reports state-of-the-art average PSNR/SSIM on all three suites and claims the best PSNR/SSIM in nearly all individual settings. Ablations show that both modules and the matching objective contribute to the reported gains.

Significance. If the reported results are reproducible, the architecture is a meaningful contribution: it moves beyond image-level prompts to spatially adaptive degradation guidance and introduces a local-global attention calibration mechanism that is broadly applicable to all-in-one restoration. The paper has several strengths: a clean module design, consistent ablation results, supplementary sensitivity analyses, and a promise of code release. However, the central claim of state-of-the-art performance rests on experimental comparisons whose rigor is not yet established: no seed or error-bar information is provided, baseline protocols are not documented, and several reported margins are extremely small. These issues are fixable but currently prevent the paper's headline claim from being fully verified.

major comments (3)
  1. [Section 4.2, Tables 1-3] The central SOTA claim is based on single-run numbers without error bars, number of seeds, or significance tests. Key margins are very small: Table 2 shows a 0.08 dB gain on SOTS (31.72 vs. 31.64) and Table 1 shows a 0.09 dB gain on BSD68 sigma=50 (28.34 vs. 28.25). In the all-in-one restoration literature, such differences are commonly within run-to-run variation under a fixed schedule. Please report mean and standard deviation over at least three seeds for QuReC and, ideally, for the main baselines; otherwise the claim of 'best in all five evaluation settings' is not established.
  2. [Section 4.1 / Table 4 protocol] The manuscript does not state whether baseline numbers in Tables 1-3 were produced by retraining under the same merged-data, 150-epoch protocol or taken from original papers. The baseline Restormer value in Table 4 matches Table 1, but this does not document the protocol for the other methods. Differences in training schedule, crop size, loss, or data mixing can easily account for margins of 0.1-0.3 dB. Please clarify for each baseline whether it was retrained in-house, and if numbers are cited, provide the source and state the exact training setting used for the comparison.
  3. [Supplementary Table 16] The Urban100 columns appear misaligned: the IRCNN row lists PSNR 27.59 for sigma=15, 31.20 for sigma=25, and 27.70 for sigma=50, which is an implausible ordering for a denoising method. This suggests a formatting or transcription error in the table. Since the supplementary tables are used to support the model's task-specific competitiveness, please correct the alignment and verify every number against the original sources.
minor comments (6)
  1. [Section 3.2, Eq. (5)] Equation (5) is typeset ambiguously: the expression 'K \sum (u)^2 - 1' appears to be a garbled form of the squared coefficient of variation used in Algorithm 1 (line 11). Please rewrite the equation in standard notation so that the load-balancing objective is unambiguous.
  2. [Section 3.2 / Supplementary Algorithm 1] The weak matching loss in Eq. (7) uses image-level degradation labels y_b, which in the all-in-one setting are exactly the degradation labels present in each training image. Calling this 'weakly supervised' is misleading; it is image-level supervision, not token-level. Suggest renaming it 'image-level matching supervision' to avoid confusion.
  3. [Section 2 (Related Work)] Typo: 'arichitectural' should be 'architectural'.
  4. [Figure 2] The figure contains garbled labels in the QRCB diagram, e.g., 'Spilit' and malformed resolution markers. Please regenerate the figure with clean text.
  5. [Table 1] The Restormer row reports SSIM 0.865 for both BSD68 sigma=15 and sigma=25, which is inconsistent with published Restormer results (sigma=15 should be substantially higher). If this is not a typo, please explain the discrepancy; if it is, correct it.
  6. [Section 3.2] The CLIP text encoder is said to be frozen, but the specific CLIP variant (e.g., ViT-B/32) is not stated. Please provide the exact model and text prompt templates used for the prototype bank.

Circularity Check

0 steps flagged

No significant circularity: QuReC is an empirical architecture paper whose reported gains come from supervised training against clean targets; the few self-citations are not load-bearing.

full rationale

QuReC does not derive a prediction from a fitted input. Its modules are defined by explicit equations: DQRM reconstructs queries as q̂_n = q_n + p_g + p_n (Eq. 2) via a soft mixture of CLIP prototypes (Eq. 1), and LGRCM computes calibrated attention outputs (Eqs. 9–13). The final training objective L = L_res + L_pm (Eq. 16) is dominated by L_res = L1 + 0.1 L_fft against clean images, so the reported PSNR/SSIM values are not forced by construction. The auxiliary prototype-matching loss L_match (Eq. 7) uses image-level degradation labels y_b (Eq. 6); this is ordinary weak supervision of an internal routing/assignment mechanism, not the restoration target being encoded into the input. The ablation tables (Tables 4–6, and supplementary Tables 7–10) show incremental gains from each component under a fixed evaluation protocol, which is consistent with a genuine architectural contribution rather than a renamed known result. References [70] and [84] share an author with this paper, but they appear only as related-work examples of prompt-based and dehazing methods and carry no argumentative weight; no uniqueness theorem or ansatz is imported from them, and no design choice is justified solely by those citations. The main weaknesses of the paper—small margins (e.g., 0.08 dB on SOTS in Table 2), unreported seeds, and unspecified baseline training protocols—are experimental-validity concerns about benchmark comparability, not circularity: they do not show that any stated quantity is equal to its input by definition. Under the given rubric, which requires exhibiting a specific reduction (Eq. X = Eq. Y by construction, or a fitted parameter renamed as a prediction), no circular step is present.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 2 invented entities

The central empirical claim rests on standard deep-learning hyperparameters, a CLIP-text semantic assumption, and the comparability of baseline numbers. There are no physically invented entities, so the invented-entity list is limited to architectural components with no independent evidence.

free parameters (6)
  • lambda_bal = 0.01
    Weight for the mixed load-balancing loss; hand-set, with sensitivity tested in Table 9.
  • lambda_match = 0.05
    Weight for the weak matching supervision loss; hand-set, with sensitivity tested in Table 9.
  • alpha_s = 0.7
    Weight for the soft prototype-usage term in L_bal; default chosen in Section A, sensitivity in Table 10.
  • alpha_h = 0.3
    Weight for the hard prototype-assignment term in L_bal; default chosen in Section A, sensitivity in Table 10.
  • local window size k = 3
    Local attention window size in LGRCM; default chosen, sensitivity in Table 8.
  • L_fft weight = 0.1
    Weight multiplying the frequency-domain loss in Eq. (14); hand-set, not ablated.
axioms (4)
  • domain assumption CLIP text embeddings of degradation names provide semantically meaningful anchors for pixel-level restoration guidance.
    DQRM matching in Section 3.2 relies on frozen CLIP text prototypes being informative for degradation semantics; the sensitivity study in Table 7 suggests wording robustness, but does not validate the deeper assumption.
  • domain assumption Image-level degradation labels are available for every training image and the prototype categories match the evaluation distributions.
    The weak matching loss in Eq. (6) uses the set of degradation categories in each image; this is training supervision, and it assumes the benchmark protocol provides such labels.
  • domain assumption Baseline numbers from Table 1-3 are comparable to QuReC under the same training/evaluation protocol.
    The paper does not document how baseline PSNR/SSIM values were produced; the SOTA claim depends on this comparability.
  • domain assumption PSNR and SSIM on the chosen benchmarks are accepted measures of restoration quality.
    The paper evaluates only distortion metrics plus LPIPS/DISTS in the supplement; this is standard in the field, but the conclusion inherits the limits of those benchmarks.
invented entities (2)
  • Degradation prototype bank T no independent evidence
    purpose: Provides per-query soft-matching anchors that reconstruct degradation-aware queries in DQRM.
    An internal model component; its utility is shown only through ablations and t-SNE visualization, with no external falsifiable prediction.
  • Learnable local/global prior banks B_l and B_g no independent evidence
    purpose: Retrieved by the base query to gate attention-guided aggregation in LGRCM.
    Learned internal tables; no evidence outside the paper ties them to measurable phenomena.

pith-pipeline@v1.3.0-alltime-deepseek · 26948 in / 10049 out tokens · 108106 ms · 2026-08-02T00:08:44.237056+00:00 · methodology

0 comments
read the original abstract

All-in-one image restoration aims to recover clean images degraded by multiple corruption types using a single unified model. Existing methods typically rely on image-level prompts or shared guidance to handle diverse degradations. However, such a paradigm becomes inadequate when degradations are spatially heterogeneous or even coexist in mixed forms within a single image. Yet spatially adaptive guidance alone is not sufficient, since accurate restoration also requires each spatial query to reliably aggregate complementary information from local neighborhoods and global contexts. To this end, we propose QuReC, a unified framework for all-in-one image restoration. QuReC consists of a Degradation-Guided Query Reconstruction Module (DQRM) and a Local-Global Response Calibration Module (LGRCM). Specifically, DQRM matches each spatial query against a degradation prototype space to reconstruct a query-specific degradation-aware representation, thereby providing fine-grained spatially adaptive restoration guidance. To further stabilize this query-wise matching process, we introduce a weakly supervised prototype matching learning strategy to improve optimization stability and degradation semantic consistency. Meanwhile, LGRCM performs local-global dual-branch aggregation and calibrates the aggregated responses with learnable priors, improving the reliability of feature aggregation and the coordination between local detail modeling and global context modeling. Extensive experiments demonstrate that QuReC achieves superior performance on multiple all-in-one image restoration benchmarks. The code is released at https://github.com/zhoushen1/QuReC.

Figures

Figures reproduced from arXiv: 2607.15097 by Dian Shen, Ding Ding, Fang Dong, Guangwen Peng, Jinghui Zhang, Shen Zhou, Wenbo Huang, Xuwei Qian, Zhen Wu, Zhiyuan Li.

Figure 1
Figure 1. Figure 1: Comparison of existing prompt-based methods and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed QuReC framework. The upper part presents the overall encoder-decoder architecture, in [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the Degradation-Guided Query Re [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Given a degraded input image I𝑖𝑛, our goal is to recover the corresponding clean image I𝑜𝑢𝑡 using the unified restoration frame￾work QuReC. The model adopts an asymmetric four-stage encoder￾decoder architecture. A 3 × 3 convolution first extracts shallow fea￾tures from I𝑖𝑛. The encoder is composed of hierarchically stacked Transformer blocks, which progressively transform high-resolution features into comp… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparisons of different methods under the all-in-one setting with three degradations. QuReC produces [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The t-SNE visualization of routing logits shows that [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Spatial visualization of query-wise prototype re [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional qualitative comparisons of different methods on the denoising task under the all-in-one setting with three [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional qualitative comparisons of different methods on the deraining task under the all-in-one setting with three [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional qualitative comparisons of different methods on the dehazing task under the all-in-one setting with three [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparisons of different methods across all 11 degradation cases in the CDD11 dataset. QuReC consistently [PITH_FULL_IMAGE:figures/full_fig_p021_10.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

94 extracted references · 5 linked inside Pith

  1. [1]

    Pablo Arbeláez, Michael Maire, Charless Fowlkes, and Jitendra Malik. 2011. Con- tour Detection and Hierarchical Image Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence 33, 5 (2011), 898–916

    2011

  2. [2]

    Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. 2016. De- hazeNet: An End-to-End System for Single Image Haze Removal. IEEE Transac- tions on Image Processing 25, 11 (2016), 5187–5198

    2016

  3. [3]

    Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang

  4. [4]

    Anthony Chan

    Cheuk-Yiu Chan, Wan-Chi Siu, Yuk-Hee Chan, and H. Anthony Chan. 2024. AnlightenDiff: Anchoring Diffusion Probabilistic Model on Low Light Image Enhancement. IEEE Transactions on Image Processing 33 (2024), 6324–6339

    2024

  5. [5]

    Chenghao Chen and Hao Li. 2021. Robust representation learning with feedback for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 7742–7751

    2021

  6. [6]

    Jun Chen, Liling Yang, Wei Yu, Wenping Gong, Zhanchuan Cai, and Jiayi Ma

  7. [7]

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

    2022

  8. [8]

    Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian- Jiun Ding, Sy-Yen Kuo, et al. 2021. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and con- tradict channel loss. In Proceedings of the IEEE/CVF international conference on computer vision. 4196–4205

    2021

  9. [9]

    Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. 2022. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 17653–17662

    2022

  10. [10]

    Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. 2021. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision . 4641–4650

    2021

  11. [11]

    Marcos V Conde, Gregor Geigle, and Radu Timofte. 2024. Instructir: High-quality image restoration following human instructions. In European Conference on Computer Vision. 1–21

    2024

  12. [12]

    Yuning Cui, Yi Tao, Wenqi Ren, and Alois Knoll. 2023. Dual-domain attention for image deblurring. In Proceedings of the AAAI conference on artificial intelligence . 479–487

    2023

  13. [13]

    Yuning Cui, Syed Waqas Zamir, Salman Khan, Alois Knoll, Mubarak Shah, and Fahad Shahbaz Khan. 2025. Adair: Adaptive all-in-one image restoration via frequency mining and modulation. In 13th international conference on learning representations, ICLR 2025. 57335–57356

    2025

  14. [14]

    Shifei Ding, Qidong Wang, Lili Guo, Xuan Li, Ling Ding, and Xindong Wu

  15. [15]

    Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming- Hsuan Yang. 2020. Multi-scale boosted dehazing network with dense feature fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2157–2167

    2020

  16. [16]

    Yu Dong, Yihao Liu, He Zhang, Shifeng Chen, and Yu Qiao. 2020. FD-GAN: Generative adversarial networks with fusion-discriminator for single image dehazing. In Proceedings of the AAAI conference on artificial intelligence . 10729– 10736

    2020

  17. [17]

    Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. 2019. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 3848–3856

    2019

  18. [18]

    Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. 2020. Zero-reference deep curve estimation for low- light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 1780–1789

    2020

  19. [19]

    Yu Guo, Yuan Gao, Yuxu Lu, Huilin Zhu, Ryan Wen Liu, and Shengfeng He. 2024. Onerestore: A universal restoration framework for composite degradation. In European conference on computer vision . 255–272

    2024

  20. [20]

    Chunming He, Chengyu Fang, Yulun Zhang, Longxiang Tang, Jinfa Huang, Kai Li, zhenhua guo, Xiu Li, and Sina Farsiu. 2025. Reti-Diff: Illumination Degradation Image Restoration with Retinex-based Latent Diffusion Model. In International Conference on Learning Representations . 43332–43352

    2025

  21. [21]

    Shengkai Hu, Jiaqi Ma, Jun Wan, Wenwen Min, Yongcheng Jing, Lefei Zhang, and Dacheng Tao. 2025. ClusIR: Towards Cluster-Guided All-in-One Image Restoration. arXiv preprint arXiv:2512.10948 (2025)

    Pith/arXiv arXiv 2025

  22. [22]

    Tao Huang, Songjiang Li, Xu Jia, Huchuan Lu, and Jianzhuang Liu. 2021. Neigh- bor2neighbor: Self-supervised denoising from single noisy images. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 14781– 14790

    2021

  23. [23]

    Kui Jiang, Zhongyuan Wang, Peng Yi, Chen Chen, Baojin Huang, Yimin Luo, Jiayi Ma, and Junjun Jiang. 2020. Multi-scale progressive fusion network for single image deraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8346–8355

    2020

  24. [24]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  25. [25]

    Orest Kupyn, Tetiana Martyniuk, Junru Wu, and Zhangyang Wang. 2019. Deblurgan-v2: Deblurring (orders-of-magnitude) faster and better. In Proceedings of the IEEE/CVF international conference on computer vision . 8878–8887

    2019

  26. [26]

    Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. 2022. All- in-one image restoration for unknown corruption. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 17452–17462

    2022

  27. [27]

    Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. 2019. Benchmarking Single-Image Dehazing and Beyond. IEEE Transactions on Image Processing 28, 1 (2019), 492–505

    2019

  28. [28]

    Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. 2020. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 3175–3185

    2020

  29. [29]

    Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. 2023. Prompt-In-Prompt Learning for Universal Image Restoration. arXiv preprint arXiv:2312.05038 (2023)

    Pith/arXiv arXiv 2023

  30. [30]

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. 2021. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision . 1833–1844

    2021

  31. [31]

    Xin Lin, Chao Ren, Xiao Liu, Jie Huang, and Yinjie Lei. 2023. Unsupervised image denoising in real-world scenarios via self-collaboration parallel generative adversarial branches. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 12642–12652

    2023

  32. [32]

    Yun Liu, Tao Li, Chunping Tan, Wenqi Ren, Cosmin Ancuti, and Weisi Lin. 2026. IHDCP: Single Image Dehazing Using Inverted Haze Density Correction Prior. IEEE Transactions on Image Processing 35 (2026), 1448–1461

    2026

  33. [33]

    Zhaoxin Liu, Jinjian Wu, Guangming Shi, Wen Yang, Weisheng Dong, and Qing- hang Zhao. 2024. Motion-Oriented Hybrid Spiking Neural Networks for Event- Based Motion Deblurring. IEEE Transactions on Circuits and Systems for Video Technology 34, 5 (2024), 3742–3754

    2024

  34. [34]

    Gustafsson, Zheng Zhao, Jens Sjölund, and Thomas Schön

    Ziwei Luo, Fredrik K. Gustafsson, Zheng Zhao, Jens Sjölund, and Thomas Schön

  35. [35]

    Jiaqi Ma, Shengkai Hu, Xu Zhang, Jun Wan, Jiaxing Huang, Lefei Zhang, and Salman Khan. 2025. EvoIR: Towards All-in-One Image Restoration via Evolution- ary Frequency Modulation. arXiv preprint arXiv:2512.05104 (2025)

    arXiv 2025

  36. [36]

    Kede Ma, Zhengfang Duanmu, Qingbo Wu, Zhou Wang, Hongwei Yong, Hongliang Li, and Lei Zhang. 2017. Waterloo Exploration Database: New Chal- lenges for Image Quality Assessment Models. IEEE Transactions on Image Pro- cessing 26, 2 (2017), 1004–1016

    2017

  37. [37]

    Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. 2023. Intriguing findings of frequency selection for image deblurring. In Proceedings of the AAAI conference on artificial intelligence . 1905–1913

    2023

  38. [38]

    In International Conference on Learning Representations

    Controlling Vision-Language Models for Multi-Task Image Restoration. In International Conference on Learning Representations . 16226–16246

  39. [39]

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. 2017. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition . 3883–3891. MM ’26, November 10–14, 2026, Rio de Janeiro, Brazi Shen Zhou, et al

    2017

  40. [40]

    Dongwon Park, Dong Un Kang, Jisoo Kim, and Se Young Chun. 2020. Multi- temporal recurrent neural networks for progressive non-uniform single image deblurring with incremental temporal training. In European conference on com- puter vision. 327–343

    2020

  41. [41]

    Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan

  42. [42]

    Martin, C

    D. Martin, C. Fowlkes, D. Tal, and J. Malik. 2001. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Confer- ence on Computer Vision. ICCV 2001 , Vol. 2. 416–423 vol.2

    2001

  43. [43]

    Yanyun Qu, Yizi Chen, Jingying Huang, and Yuan Xie. 2019. Enhanced pix2pix dehazing network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 8160–8168

    2019

  44. [44]

    Chao Ren, Xiaohai He, Chuncheng Wang, and Zhibo Zhao. 2021. Adaptive consistency prior based deep network for image denoising. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8596–8606

    2021

  45. [45]

    Dongwei Ren, Wangmeng Zuo, Qinghua Hu, Pengfei Zhu, and Deyu Meng

  46. [46]

    Advances in neural information processing systems 36 (2023), 71275–71293

    Promptir: Prompting for all-in-one image restoration. Advances in neural information processing systems 36 (2023), 71275–71293

    2023

  47. [47]

    Xu Qin, Zhilin Wang, Yuanchao Bai, Xiaodong Xie, and Huizhu Jia. 2020. FFA-Net: Feature fusion attention network for single image dehazing. In Proceedings of the AAAI conference on artificial intelligence . 11908–11915

    2020

  48. [48]

    Xiongfei Su, Siyuan Li, Yuning Cui, Miao Cao, Yulun Zhang, Zheng Chen, Zongliang Wu, Zedong Wang, Yuanlong Zhang, and Xin Yuan. 2025. Prior-guided hierarchical harmonization network for efficient image dehazing. In Proceedings of the AAAI conference on artificial intelligence . 7042–7050

    2025

  49. [49]

    Chunwei Tian, Yong Xu, and Wangmeng Zuo. 2020. Image denoising using deep CNN with batch renormalization. Neural Networks 121 (2020), 461–473

    2020

  50. [50]

    Xiangpeng Tian, Xiangyu Liao, Xiao Liu, Meng Li, and Chao Ren. 2025. Degradation-aware feature perturbation for all-in-one image restoration. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 28165–28175

    2025

  51. [51]

    Yuchuan Tian, Jianhong Han, Hanting Chen, Yuanyuan Xi, Ning Ding, Jie Hu, Chao Xu, and Yunhe Wang. 2024. Instruct-ipt: All-in-one image processing transformer via weight modulation. arXiv preprint arXiv:2407.00676 (2024)

    Pith/arXiv arXiv 2024

  52. [52]

    Wenqi Ren, Si Liu, Hua Zhang, Jinshan Pan, Xiaochun Cao, and Ming-Hsuan Yang. 2016. Single image dehazing via multi-scale convolutional neural networks. In European conference on computer vision . 154–169

    2016

  53. [53]

    Lingyan Ruan, Bin Chen, Jizhou Li, and Miuling Lam. 2022. Learning to deblur using light field generated and real defocus images. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 16304–16313

    2022

  54. [54]

    Cong Wang, Jinshan Pan, Wanyu Lin, Jiangxin Dong, Wei Wang, and Xiao-Ming Wu. 2024. Selfpromer: Self-prompt dehazing transformers with depth-consistency. In Proceedings of the AAAI conference on artificial intelligence , Vol. 38. 5327–5335

    2024

  55. [55]

    Yongzhen Wang, Yongjun Li, Zhuoran Zheng, Xiao-Ping Zhang, and Mingqiang Wei. 2025. M2Restore: Mixture-of-Experts-Based Mamba-CNN Fusion Framework for All-in-One Image Restoration.IEEE Transactions on Image Processing 34 (2025), 8086–8100

    2025

  56. [56]

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. 2022. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 17683–17693

    2022

  57. [57]

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. 2018. Deep retinex decomposition for low-light enhancement.arXiv preprint arXiv:1808.04560 (2018)

    Pith/arXiv arXiv 2018

  58. [58]

    Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. 2022. Tran- sweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2353–2363

    2022

  59. [59]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems 30 (2017)

    2017

  60. [60]

    Gang Wu, Junjun Jiang, Yijun Wang, Kui Jiang, and Xianming Liu. 2025. Debiased all-in-one image restoration with task uncertainty regularization. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 39. 8386–8394

    2025

  61. [61]

    Haiyan Wu, Yanyun Qu, Shaohui Lin, Jian Zhou, Ruizhi Qiao, Zhizhong Zhang, Yuan Xie, and Lizhuang Ma. 2021. Contrastive learning for compact single image dehazing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10551–10560

    2021

  62. [62]

    Jiawei Wu, Zhifei Yang, Zhe Wang, and Zhi Jin. 2025. Gradient as conditions: Rethinking HOG for all-in-one image restoration. arXiv preprint arXiv:2504.09377 (2025)

    arXiv 2025

  63. [63]

    Han Xu, Xunpeng Yi, Chen Lu, Guangcan Liu, and Jiayi Ma. 2025. URFusion: Un- supervised Unified Degradation-Robust Image Fusion Network.IEEE Transactions on Image Processing 34 (2025), 5803–5818

    2025

  64. [64]

    Wei Wei, Deyu Meng, Qian Zhao, Zongben Xu, and Ying Wu. 2019. Semi- supervised transfer learning for image rain removal. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 3877–3886

    2019

  65. [65]

    Gang Wu, Junjun Jiang, Kui Jiang, Xianming Liu, and Liqiang Nie. 2026. Beyond Degradation Redundancy: Contrastive Prompt Learning for All-in-One Image Restoration. IEEE Transactions on Pattern Analysis and Machine Intelligence 48, 4 (2026), 4005–4022

    2026

  66. [66]

    Wenhan Yang, Robby T Tan, Shiqi Wang, Yuming Fang, and Jiaying Liu. 2020. Single image deraining: From model-based to data-driven and beyond. IEEE Transactions on pattern analysis and machine intelligence 43, 11 (2020), 4059– 4077

    2020

  67. [67]

    Mingde Yao, Ruikang Xu, Yuanshen Guan, Jie Huang, and Zhiwei Xiong. 2024. Neural Degradation Representation Learning for All-in-One Image Restoration. IEEE Transactions on Image Processing 33 (2024), 5408–5423

    2024

  68. [68]

    Siyuan Yao, Rui Zhu, Ziqi Wang, Wenqi Ren, Yanyang Yan, and Xiaochun Cao

  69. [69]

    Rajeev Yasarla and Vishal M Patel. 2019. Uncertainty guided multi-scale residual learning-using a cycle spinning cnn for single image de-raining. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition . 8405–8414

    2019

  70. [70]

    Qingsen Yan, Yixu Feng, Cheng Zhang, Guansong Pang, Kangbiao Shi, Peng Wu, Wei Dong, Jinqiu Sun, and Yanning Zhang. 2025. Hvi: A new color space for low-light image enhancement. In Proceedings of the computer vision and pattern recognition conference. 5678–5687

    2025

  71. [71]

    Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu

    Wenhan Yang, Robby T. Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. 2020. Joint Rain Detection and Removal from a Single Image with Contextualized Deep Networks. IEEE Transactions on Pattern Analysis and Ma- chine Intelligence 42, 6 (2020), 1377–1393

    2020

  72. [72]

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shah- baz Khan, and Ming-Hsuan Yang. 2022. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition . 5728–5739

    2022

  73. [73]

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. 2021. Multi-stage progressive image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 14821–14831

    2021

  74. [74]

    Haijin Zeng, Xiangming Wang, Yongyong Chen, Jingyong Su, and Jie Liu. 2025. Vision-language gradient descent-driven all-in-one deep unfolding networks. In Proceedings of the Computer Vision and Pattern Recognition Conference. 7524–7533

    2025

  75. [75]

    In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    UMDATrack: Unified Multi-Domain Adaptive Tracking Under Adverse Weather Conditions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 6466–6475

  76. [76]

    Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. 2023. Ingredient-oriented multi-degradation learning for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5825–5835

    2023

  77. [77]

    Xiaoyan Yu, Shen Zhou, Huafeng Li, and Liehuang Zhu. 2025. Multi-Expert Adaptive Selection: Task-Balancing for All-in-One Image Restoration. IEEE Transactions on Circuits and Systems for Video Technology 35, 5 (2025), 4619–4634

    2025

  78. [78]

    Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yuedong Tan, Danda Pani Paudel, Yulun Zhang, and Radu Timofte. 2025. Complexity experts are task-discriminative learners for any image restoration. In Proceedings of the Computer Vision and Pattern Recognition Conference. 12753–12763

    2025

  79. [79]

    Kai Zhang, Wangmeng Zuo, and Lei Zhang. 2018. FFDNet: Toward a Fast and Flexible Solution for CNN-Based Image Denoising. IEEE Transactions on Image Processing 27, 9 (2018), 4608–4622

    2018

  80. [80]

    Xiaoqin Zhang, Runhua Jiang, Tao Wang, and Wenhan Luo. 2021. Single Im- age Dehazing via Dual-Path Recurrent Network. IEEE Transactions on Image Processing 30 (2021), 5211–5222

    2021

Showing first 80 references.