pith. machine review for the scientific record. sign in

arxiv: 2605.06127 · v1 · submitted 2026-05-07 · 💻 cs.CV · cs.AI

Recognition: unknown

Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration

Haisen He, Heng Li, SongLin Dong, Xiangyu Zou, Yihong Gong, Zhiheng Ma

Authors on Pith no claims yet

Pith reviewed 2026-05-08 13:53 UTC · model grok-4.3

classification 💻 cs.CV cs.AI
keywords all-in-one image restorationlow-rank residualsinstance conditioningcross-attention adapterdynamic parameterizationtoken-wise routingimage degradation
0
0 comments X

The pith

Continuous Expert Assembly adapts a shared restoration model to unknown, spatially varying degradations by synthesizing and densely combining instance-specific low-rank residuals at each image token.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Continuous Expert Assembly to let one set of network weights handle real-world images with mixed, localized corruptions such as noise in one region and blur in another. It replaces global prompts or fixed expert pools with a lightweight adapter that reads intermediate features and creates low-rank routing bases plus residual directions on the fly. Each spatial token then forms its own update through a dense signed dot-product over those components. This matters because existing methods either lose fine local evidence through global conditioning or suffer from inefficient or unstable routing when degradations are compositional. If the approach works, restoration quality rises on benchmarks with diverse degradations while parameter count, compute, and speed remain competitive.

Core claim

CEA employs a Cross-Attention Hyper-Adapter to probe intermediate spatial features and generate instance-conditioned low-rank routing bases together with residual directions; each token then assembles its residual update via dense signed dot-product affinities over the rank-wise components, without external prompts, static expert banks, or discrete Top-k selection, and the assembly admits a linear-attention interpretation.

What carries the argument

The Cross-Attention Hyper-Adapter that generates instance-conditioned low-rank routing bases and residual directions from intermediate features, followed by dense signed dot-product token-wise assembly.

If this is right

  • Higher average restoration metrics on AIO-3, AIO-5, and CDD-11, especially where degradations change across the image.
  • Lower or comparable parameter count, FLOPs, and runtime than static expert-pool methods.
  • Avoidance of homogeneous updates and unstable sparse routing that can occur with discrete expert selection.
  • A transparent linear-attention view of the routing process that follows directly from the dense affinity rule.

Where Pith is reading between the lines

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

  • The same per-token continuous routing could be tested on video sequences where degradations also vary across both space and time.
  • Replacing discrete expert banks with synthesized low-rank components may reduce memory footprint in other conditional vision models.
  • The linear-attention equivalence suggests the method could be accelerated further with existing efficient attention kernels without changing the learned behavior.

Load-bearing premise

A lightweight cross-attention module can reliably extract effective low-rank bases and residual directions directly from the image's own intermediate features and that the resulting dense signed assembly stays stable and non-homogeneous.

What would settle it

No measurable quality gain, or outright worse results, on test sets containing spatially varying or compositional degradations when compared with strong prompt- and expert-based baselines under matched compute budgets.

Figures

Figures reproduced from arXiv: 2605.06127 by Haisen He, Heng Li, SongLin Dong, Xiangyu Zou, Yihong Gong, Zhiheng Ma.

Figure 1
Figure 1. Figure 1: Conceptual comparison of adaptation paradigms for All-in-One Image Restoration. (a) Representative global or discrete expert adaptation. A compact image-level signal or sparse router selects or modulates a fixed set of parameters, which can become coarse when degradations vary spatially or appear in composition. (b) Continuous Expert Assembly (CEA). A lightweight hyper-adapter probes spatial features and g… view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of the proposed method. (a) Backbone. We build upon a U-shaped Transformer backbone. To dynamically adapt to unknown degradations, our proposed modules are integrated into the decoder blocks to modulate the intermediate spatial features. (b) Cross-Attention Hyper-Adapter. Spatial features are condensed via depthwise-pointwise (DW/PW) convolutions to provide local context. A compact set… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on the AIO-3 benchmark. We compare CEA-IR-S against AirNet, PromptIR, and MoCE-IR-S across dehazing, deraining, and denoising. The bottom rows show zoom-in crops and absolute error maps, where darker regions indicate lower reconstruction error. CEA reduces residual degradation artifacts and preserves fine structures across the three restoration tasks. F.2 Qualitative Comparison on AIO-5 view at source ↗
Figure 4
Figure 4. Figure 4: Additional qualitative comparison on the AIO-5 benchmark. We compare our method with representative baselines across dehazing, deraining, denoising, deblurring, and low-light en￾hancement. F.3 Qualitative Comparison on CDD-11 Figures 5–7 provide additional CDD-11 comparisons. The advantage of CEA is most visible when multiple degradation operators interact, especially in haze-related compositions where bot… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on CDD-11 single degradations. We compare OneRestore, MoCE-IR-S, and CEA-IR-S on low-light, haze, rain, and snow. 22 view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison on CDD-11 double degradations. Representative examples include low-light+haze, low-light+rain, low-light+snow, haze+rain, and haze+snow view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison on CDD-11 triple degradations. We show representative examples with low-light+haze+rain and low-light+haze+snow. 23 view at source ↗
read the original abstract

Real-world image degradation is often unknown, spatially non-uniform, and compositional, requiring all-in-one restoration models to adapt a single set of weights to diverse local corruption patterns without test-time degradation labels. Existing methods typically modulate a shared backbone with global prompts or degradation descriptors, or route features through predefined expert pools. However, compact global conditioning can bottleneck localized degradation evidence, while static expert routing may produce homogeneous updates or rely on unstable sparse assignments. We propose \textbf{Continuous Expert Assembly} (CEA), a token-wise dynamic parameterization framework for all-in-one image restoration. CEA employs a lightweight \textbf{Cross-Attention Hyper-Adapter} to probe intermediate spatial features and synthesize instance-conditioned low-rank routing bases and residual directions. Each spatial token then assembles its own residual update via dense signed dot-product affinities over the generated rank-wise components, avoiding external prompts, static expert banks, and discrete Top- selection. The resulting assembly rule also admits a linear-attention perspective, making its dense token-wise routing behavior transparent. Experiments on AIO-3, AIO-5, and CDD-11 show that CEA improves average restoration quality over strong prompt-, descriptor-, and expert-based baselines, with the clearest gains on spatially varying and compositional degradations, while maintaining favorable parameter, FLOP, and runtime efficiency.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces Continuous Expert Assembly (CEA) for all-in-one image restoration, addressing unknown, spatially non-uniform, and compositional degradations. It proposes a token-wise dynamic parameterization using a lightweight Cross-Attention Hyper-Adapter that synthesizes instance-conditioned low-rank routing bases and residual directions from intermediate spatial features. Each token then computes its residual update via dense signed dot-product affinities over these components, avoiding external prompts, static expert banks, and discrete Top-k selection. The assembly is also interpreted through a linear-attention lens. Experiments on AIO-3, AIO-5, and CDD-11 report average quality improvements over prompt-, descriptor-, and expert-based baselines, with stronger gains on spatially varying degradations, while claiming favorable parameter, FLOP, and runtime efficiency.

Significance. If the empirical gains and the mechanism's instance-specific adaptation hold under scrutiny, CEA could meaningfully advance all-in-one restoration by enabling more localized, non-homogeneous updates without the bottlenecks of global conditioning or fixed expert pools. The linear-attention view provides a useful transparency angle. The approach builds on low-rank adaptation ideas but applies them continuously and densely at the token level, which may generalize to other adaptive vision tasks if the synthesized bases prove meaningfully diverse.

major comments (3)
  1. [Experiments] Experiments (assumed §4): The abstract and introduction assert average restoration quality gains over strong baselines on AIO-3, AIO-5, and CDD-11 with clearest benefits on spatially varying degradations, yet no specific PSNR/SSIM numbers, per-degradation breakdowns, or statistical significance tests are referenced in the provided text. This makes it impossible to verify the magnitude of improvement or rule out that gains derive primarily from the shared backbone rather than the proposed assembly.
  2. [Method] Method (§3, Cross-Attention Hyper-Adapter description): The central claim requires that the adapter produces distinct, instance-conditioned low-rank bases and residual directions that vary meaningfully with local spatial features. No ablation or analysis (e.g., cosine similarity of generated bases across degradation types or spatial maps of assembled residuals) is described to confirm non-homogeneous updates; without this, the performance advantage over global prompt methods remains unproven.
  3. [Method] Method (dense signed dot-product assembly): The paper states that the dense assembly avoids unstable sparse assignments and admits a linear-attention perspective. However, no analysis of numerical stability, gradient flow, or comparison to sparse alternatives is provided, which is load-bearing for the claim that this routing is reliably superior for compositional degradations.
minor comments (2)
  1. [Abstract] Abstract: The phrase 'dense signed dot-product affinities over the generated rank-wise components' is introduced without a brief equation or notation preview, reducing immediate clarity for readers unfamiliar with the low-rank construction.
  2. [Method] Notation: The manuscript should explicitly define the dimensions of the synthesized low-rank bases (e.g., rank r, feature dimension d) and the exact form of the signed dot-product in the first method subsection to aid reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment below and have revised the paper to strengthen the presentation of results and supporting analyses.

read point-by-point responses

  1. Referee: [Experiments] Experiments (assumed §4): The abstract and introduction assert average restoration quality gains over strong baselines on AIO-3, AIO-5, and CDD-11 with clearest benefits on spatially varying degradations, yet no specific PSNR/SSIM numbers, per-degradation breakdowns, or statistical significance tests are referenced in the provided text. This makes it impossible to verify the magnitude of improvement or rule out that gains derive primarily from the shared backbone rather than the proposed assembly.

    Authors: We appreciate this observation. The full manuscript in §4 and the associated tables report concrete PSNR/SSIM values, per-degradation breakdowns on AIO-3/AIO-5/CDD-11, and comparisons isolating the contribution of the assembly module from the shared backbone. In the revised version we will add explicit numerical references (e.g., average PSNR gains) and direct citations to these tables in both the abstract and introduction, along with a brief note on the consistency of improvements across datasets. This will make the magnitude and source of the gains immediately verifiable without altering the experimental claims. revision: yes

  2. Referee: [Method] Method (§3, Cross-Attention Hyper-Adapter description): The central claim requires that the adapter produces distinct, instance-conditioned low-rank bases and residual directions that vary meaningfully with local spatial features. No ablation or analysis (e.g., cosine similarity of generated bases across degradation types or spatial maps of assembled residuals) is described to confirm non-homogeneous updates; without this, the performance advantage over global prompt methods remains unproven.

    Authors: We agree that direct evidence of instance- and spatially-conditioned variation is important for substantiating the advantage over global prompt baselines. While the current manuscript emphasizes end-to-end performance, the revised version will incorporate an additional analysis subsection that includes (i) cosine-similarity statistics of the synthesized low-rank bases across different degradation types and (ii) qualitative spatial maps of the assembled residual updates. These additions will demonstrate that the Cross-Attention Hyper-Adapter generates meaningfully distinct, non-homogeneous components conditioned on local features. revision: yes

  3. Referee: [Method] Method (dense signed dot-product assembly): The paper states that the dense assembly avoids unstable sparse assignments and admits a linear-attention perspective. However, no analysis of numerical stability, gradient flow, or comparison to sparse alternatives is provided, which is load-bearing for the claim that this routing is reliably superior for compositional degradations.

    Authors: The dense signed dot-product formulation is presented precisely to circumvent the instability of discrete Top-k selection while admitting a linear-attention interpretation. We concur that explicit supporting analysis would reinforce this design choice. The revised manuscript will add a short discussion and comparison that reports training stability metrics, gradient-norm behavior, and performance differences versus a sparse Top-k variant, with particular attention to compositional degradation cases. This will provide concrete evidence for the reliability of the dense assembly. revision: yes

Circularity Check

0 steps flagged

No significant circularity; new architectural proposal with empirical validation

full rationale

The paper proposes a novel Continuous Expert Assembly (CEA) framework that introduces new components including a Cross-Attention Hyper-Adapter for synthesizing instance-conditioned low-rank routing bases and residual directions, followed by dense signed dot-product assembly. No equations, fitted parameters, or derivations are presented that reduce by construction to prior inputs, self-citations, or renamed known results. Claims rest on empirical gains over baselines on AIO-3, AIO-5, and CDD-11 rather than any load-bearing mathematical chain that loops back to the method's own definitions. This is a standard case of an independent architectural contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The central claim rests on the effectiveness of the newly introduced Cross-Attention Hyper-Adapter and the dense signed dot-product assembly rule for producing instance-specific updates. No explicit free parameters, standard mathematical axioms, or independently evidenced invented entities are detailed in the abstract.

invented entities (1)
  • Cross-Attention Hyper-Adapter no independent evidence
    purpose: Probe intermediate spatial features to synthesize instance-conditioned low-rank routing bases and residual directions
    Core new module introduced to enable the token-wise dynamic parameterization; no independent evidence supplied in abstract

pith-pipeline@v0.9.0 · 5553 in / 1296 out tokens · 29999 ms · 2026-05-08T13:53:48.210497+00:00 · methodology

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

59 extracted references · 45 canonical work pages · 4 internal anchors

  1. [1]

    Lora-ir: taming low-rank experts for efficient all-in-one image restoration,

    Yuang Ai, Huaibo Huang, and Ran He. Lora-ir: Taming low-rank experts for efficient all-in-one image restoration.arXiv preprint arXiv:2410.15385, 2024. doi: 10.48550/arXiv.2410.15385

    work page doi:10.48550/arxiv.2410.15385 2024

  2. [3]

    Taming Transformers for High-Resolution Image Synthesis , booktitle =

    Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12294– 12305, 2021. doi: 10.1109/CVPR46437.2021.01212

    work page doi:10.1109/cvpr46437.2021.01212 2021

  3. [4]

    Hinet: Half instance normalization network for image restoration

    Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 182–192, June 2021. doi: 10.1109/ CVPRW53098.2021.00027

    work page arXiv 2021

  4. [5]

    Dvs-voltmeter: Stochastic process-based event simulator for dynamic vision sensors,

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restora- tion. InComputer Vision – ECCV 2022, pages 17–33, 2022. doi: 10.1007/978-3-031-20071-7\ _2

    work page doi:10.1007/978-3-031-20071-7 2022

  5. [6]

    nuScenes: A multimodal dataset for autonomous driving,

    Yinpeng Chen, Xiyang Dai, Mengchen Liu, Dongdong Chen, Lu Yuan, and Zicheng Liu. Dynamic convolution: Attention over convolution kernels. In2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11027–11036, 2020. doi: 10.1109/ CVPR42600.2020.01104

    work page arXiv 2020

  6. [7]

    Cross aggregation transformer for image restoration

    Zheng Chen, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Cross aggregation transformer for image restoration. InAdvances in Neural Information Processing Systems (NeurIPS), volume 35, pages 25478–25490, 2022

    2022

  7. [8]

    Colwell, and Adrian Weller

    Krzysztof Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamás Sarlós, Peter Hawkins, Jared Davis, Afroz Mohiuddin, Lukasz Kaiser, David Belanger, Lucy J. Colwell, and Adrian Weller. Rethinking attention with performers. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  8. [9]

    Conde, Gregor Geigle, and Radu Timofte

    Marcos V . Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. InComputer Vision – ECCV 2024, pages 1–21, 2024. doi: 10.1007/978-3-031-72764-1\_1

    work page doi:10.1007/978-3-031-72764-1 2024

  9. [10]

    Dynamic filter net- works

    Bert De Brabandere, Xu Jia, Tinne Tuytelaars, and Luc Van Gool. Dynamic filter net- works. InAdvances in Neural Information Processing Systems (NeurIPS), page 667–675,

  10. [11]

    10.48550/arXiv.1605.09673

    work page doi:10.48550/arxiv.1605.09673

  11. [12]

    FD-GAN: Generative adversarial networks with fusion-discriminator for single image dehazing

    Yu Dong, Yihao Liu, He Zhang, Shifeng Chen, and Yu Qiao. FD-GAN: Generative adversarial networks with fusion-discriminator for single image dehazing. InProceedings of the AAAI Conference on Artificial Intelligence, 2020. doi: 10.1609/aaai.v34i07.6701

    work page doi:10.1609/aaai.v34i07.6701 2020

  12. [13]

    Uniprocessor: A text-induced unified low-level image processor

    Huiyu Duan, Xiongkuo Min, Sijing Wu, Wei Shen, and Guangtao Zhai. Uniprocessor: A text-induced unified low-level image processor. InComputer Vision – ECCV 2024, pages 180–199, 2024. doi: 10.1007/978-3-031-72855-6\_11. 10

    work page doi:10.1007/978-3-031-72855-6 2024

  13. [14]

    A general decoupled learning framework for parameterized image operators.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:33–47, 2021

    Qingnan Fan, Dongdong Chen, Lu Yuan, Gang Hua, Nenghai Yu, and Baoquan Chen. A general decoupled learning framework for parameterized image operators.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43:33–47, 2021. doi: 10.1109/TPAMI.2019.2925793

    work page doi:10.1109/tpami.2019.2925793 2021

  14. [15]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research, 23 (120):1–39, 2022

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity.Journal of Machine Learning Research, 23 (120):1–39, 2022

    2022

  15. [16]

    Dynamic scene deblurring with parameter selective sharing and nested skip connections

    Hongyun Gao, Xin Tao, Xiaoyong Shen, and Jiaya Jia. Dynamic scene deblurring with parameter selective sharing and nested skip connections. In2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3843–3851, 2019. URL 10.1109/ CVPR.2019.00397

    work page arXiv 2019

  16. [17]

    Parameter efficient adaptation for image restoration with heterogeneous mixture-of-experts

    Hang Guo, Tao Dai, Yuanchao Bai, Bin Chen, Xudong Ren, Zexuan Zhu, and Shu-Tao Xia. Parameter efficient adaptation for image restoration with heterogeneous mixture-of-experts. In Advances in Neural Information Processing Systems (NeurIPS), pages 13522–13547, 2024. doi: 10.52202/079017-0432

    work page doi:10.52202/079017-0432 2024

  17. [18]

    Mambair: A simple baseline for image restoration with state-space model

    Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. InComputer Vision – ECCV 2024, 2024. doi: 10.1007/978-3-031-72649-1\_13

    work page doi:10.1007/978-3-031-72649-1 2024

  18. [19]

    Onerestore: A universal restoration framework for composite degradation

    Yu Guo, Yuan Gao, Yuxu Lu, Huilin Zhu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. InComputer Vision – ECCV 2024, pages 255–272, 2025. doi: 10.1007/978-3-031-72655-2\_15

    work page doi:10.1007/978-3-031-72655-2 2024

  19. [20]

    David Ha, Andrew Dai, and Quoc V . Le. Hypernetworks. InInternational Conference on Learning Representations (ICLR), 2017

    2017

  20. [21]

    Parameter-efficient transfer learning for NLP

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. InProceedings of the 36th International Conference on Machine Learning (ICML), volume 97, pages 2790–2799, 2019

    2019

  21. [22]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations (ICLR), 2022. doi: 10.48550/arXiv.2106. 09685

    work page doi:10.48550/arxiv.2106 2022

  22. [23]

    In: Proceedings of the IEEE/CVF ICCV, pp

    Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 13899–13909, October 2021. doi: 10.1109/ICCV48922.2021.01366

    work page doi:10.1109/iccv48922.2021.01366 2021

  23. [24]

    Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M. Patel. Transweather: Transformer- based restoration of images degraded by adverse weather conditions. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2343–2353, 2022. doi: 10.1109/CVPR52688.2022.00239

    work page doi:10.1109/cvpr52688.2022.00239 2022

  24. [25]

    Transformers are RNNs: Fast autoregressive transformers with linear attention

    Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are RNNs: Fast autoregressive transformers with linear attention. InProceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 5156–5165, 2020

    2020

  25. [26]

    Adam: A Method for Stochastic Optimization

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014. doi: 10.48550/arXiv.1412.6980

    work page internal anchor Pith review doi:10.48550/arxiv.1412.6980 2014

  26. [27]

    Gshard: Scaling giant models with condi- tional computation and automatic sharding

    Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with condi- tional computation and automatic sharding. InInternational Conference on Learning Represen- tations (ICLR), 2021. 11

    2021

  27. [28]

    Parmar, R

    Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17431–17441, 2022. doi: 10.1109/CVPR52688.2022. 01693

    work page doi:10.1109/cvpr52688.2022 2022

  28. [29]

    arXiv preprint arXiv:2312.05038 (2023) MMFE-IR 17

    Zilong Li, Yiming Lei, Chenglong Ma, Junping Zhang, and Hongming Shan. Prompt-in- prompt learning for universal image restoration.arXiv preprint arXiv:2312.05038, 2023. doi: 10.48550/arXiv.2312.05038

    work page doi:10.48550/arxiv.2312.05038 2023

  29. [30]

    Swinir: Image restoration using swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swin transformer. In2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW), pages 1833–1844, 2021. doi: 10.1109/ICCVW54120. 2021.00210

    work page doi:10.1109/iccvw54120 2021

  30. [31]

    TAPE: Task-agnostic prior embedding for image restoration

    Lin Liu, Lingxi Xie, Xiaopeng Zhang, Shanxin Yuan, Xiangyu Chen, Wengang Zhou, Houqiang Li, and Qi Tian. TAPE: Task-agnostic prior embedding for image restoration. InComputer Vision – ECCV 2022, 2022. doi: 10.1007/978-3-031-19797-0\_26

    work page doi:10.1007/978-3-031-19797-0 2022

  31. [32]

    F., Cheng, K.-T., and Chen, M.-H

    Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. In Proceedings of the 41st International Conference on Machine Learning (ICML), volume 235, pages 32100–32121, 2024. doi: 10.48550/arXiv.2402.09353

    work page doi:10.48550/arxiv.2402.09353 2024

  32. [33]

    Gustafsson, Zheng Zhao, Jens Sjölund, and Thomas B

    Ziwei Luo, Fredrik K. Gustafsson, Zheng Zhao, Jens Sjölund, and Thomas B. Schön. Controlling vision-language models for multi-task image restoration. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  33. [35]

    doi:10.1109/CVPR.2018

    Ben Mildenhall, Jonathan T. Barron, Jiawen Chen, Dillon Sharlet, Ren Ng, and Robert Carroll. Burst denoising with kernel prediction networks. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2502–2510, 2018. doi: 10.1109/CVPR.2018. 00265

    work page doi:10.1109/cvpr.2018 2018

  34. [36]

    Parmar, R

    Chong Mou, Qian Wang, and Jian Zhang. Deep generalized unfolding networks for image restoration. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. doi: 10.1109/CVPR52688.2022.01688

    work page doi:10.1109/cvpr52688.2022.01688 2022

  35. [37]

    Restoring vision in adverse weather conditions with patch-based denoising diffusion models.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10346–10357, 2023

    Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10346–10357, 2023. doi: 10.1109/TPAMI.2023.3238179

    work page doi:10.1109/tpami.2023.3238179 2023

  36. [38]

    Khan, and Fahad Shahbaz Khan

    Vaishnav Potlapalli, Syed Waqas Zamir, Salman H. Khan, and Fahad Shahbaz Khan. Promptir: prompting for all-in-one blind image restoration. InAdvances in Neural Information Processing Systems (NeurIPS), pages 71275–71293, 2023. doi: 10.52202/075280-3121

    work page doi:10.52202/075280-3121 2023

  37. [39]

    From sparse to soft mixtures of experts

    Joan Puigcerver, Carlos Riquelme Ruiz, Basil Mustafa, and Neil Houlsby. From sparse to soft mixtures of experts. InInternational Conference on Learning Representations, pages 28435–28445, 2024

    2024

  38. [40]

    Manifold-aware representation learning for degradation-agnostic image restoration.arXiv preprint arXiv:2505.18679, 2025

    Bin Ren, Yawei Li, Xu Zheng, Yuqian Fu, Danda Pani Paudel, Hong Liu, Ming-Hsuan Yang, Luc Van Gool, and Nicu Sebe. Efficient degradation-agnostic image restoration via channel-wise functional decomposition and manifold regularization.arXiv preprint arXiv:2505.18679, 2026. doi: 10.48550/arXiv.2505.18679

    work page doi:10.48550/arxiv.2505.18679 2026

  39. [41]

    Scaling vision with sparse mixture of experts

    Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, An- dré Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. InAdvances in Neural Information Processing Systems (NeurIPS), pages 8583–8595,

  40. [42]

    10.48550/arXiv.2106.05974. 12

    work page doi:10.48550/arxiv.2106.05974

  41. [43]

    chinchilla optimal

    David Serrano-Lozano, Luis Herranz, Shaolin Su, and Javier Vazquez-Corral. Adaptive blind all-in-one image restoration.arXiv preprint arXiv:2411.18412, 2025. doi: 10.48550/arXiv. 2411.18412

    work page internal anchor Pith review doi:10.48550/arxiv 2025

  42. [44]

    Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V . Le, Geoffrey E. Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture- of-experts layer. InInternational Conference on Learning Representations (ICLR), 2017. 10.48550/arXiv.1701.06538

    work page internal anchor Pith review doi:10.48550/arxiv.1701.06538 2017

  43. [45]

    Localized gaussian splatting editing with contextual awareness

    Zhuoran Shen, Mingyuan Zhang, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Efficient attention: Attention with linear complexities. In2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 3530–3538, 2021. doi: 10.1109/W ACV48630.2021.00357

    work page doi:10.1109/w 2021

  44. [46]

    Image denoising using deep CNN with batch renormalization.Neural Networks, 121:461–473, 2020

    Chunwei Tian, Yong Xu, and Wangmeng Zuo. Image denoising using deep CNN with batch renormalization.Neural Networks, 121:461–473, 2020. doi: 10.1016/j.neunet.2019.08.022

    work page doi:10.1016/j.neunet.2019.08.022 2020

  45. [47]

    Shadows can be

    Zhengzhong Tu, Hossein Talebi, Han Zhang, Feng Yang, Peyman Milanfar, Alan Bovik, and Yinxiao Li. MAXIM: Multi-axis MLP for image processing. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5759–5770, 2022. doi: 10.1109/ CVPR52688.2022.00568

    work page arXiv 2022

  46. [48]

    Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions.International Journal of Computer Vision, 132:4541–4563, 2024

    Tao Wang, Kaihao Zhang, Ziqian Shao, Wenhan Luo, Bjorn Stenger, Tong Lu, Tae-Kyun Kim, Wei Liu, and Hongdong Li. Gridformer: Residual dense transformer with grid structure for image restoration in adverse weather conditions.International Journal of Computer Vision, 132:4541–4563, 2024. doi: 10.1007/s11263-024-02056-0

    work page doi:10.1007/s11263-024-02056-0 2024

  47. [49]

    Uformer: A general u-shaped transformer for image restoration

    Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17662–17672, June

  48. [50]

    doi: 10.1109/CVPR52688.2022.01716

    work page doi:10.1109/cvpr52688.2022.01716 2022

  49. [51]

    Harmony in diversity: Improving all-in-one image restoration via multi-task collaboration

    Gang Wu, Junjun Jiang, Kui Jiang, and Xianming Liu. Harmony in diversity: Improving all-in-one image restoration via multi-task collaboration. InProceedings of the 32nd ACM International Conference on Multimedia, 2024. doi: 10.1145/3664647.3680762

    work page doi:10.1145/3664647.3680762 2024

  50. [52]

    Le, and Jiquan Ngiam

    Brandon Yang, Gabriel Bender, Quoc V . Le, and Jiquan Ngiam. Condconv: conditionally parameterized convolutions for efficient inference. In Hanna M. Wallach, Hugo Larochelle, Alina Beygelzimer, Florence d’Alché Buc, Emily B. Fox, and Roman Garnett, editors,Advances in Neural Information Processing Systems (NeurIPS), pages 1307–1318, 2019

    2019

  51. [53]

    In: IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)

    Eduard Zamfir, Zongwei Wu, Nancy Mehta, Yuedong Tan, Danda Pani Paudel, Yulun Zhang, and Radu Timofte. Complexity experts are task-discriminative learners for any image restoration. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12753–12763, 2025. doi: 10.1109/CVPR52734.2025.01190

    work page doi:10.1109/cvpr52734.2025.01190 2025

  52. [54]

    Taming Transformers for High-Resolution Image Synthesis , booktitle =

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming- Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14816–14826, 2021. doi: 10.1109/CVPR46437.2021.01458

    work page doi:10.1109/cvpr46437.2021.01458 2021

  53. [55]

    Parmar, R

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5728–5739, 2022. doi: 10.1109/CVPR52688.2022.00564

    work page doi:10.1109/cvpr52688.2022.00564 2022

  54. [56]

    In: IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR)

    Haijin Zeng, Xiangming Wang, Yongyong Chen, Jingyong Su, and Jie Liu. Vision-language gradient descent-driven all-in-one deep unfolding networks. In2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. doi: 10.1109/CVPR52734.2025. 00705

    work page doi:10.1109/cvpr52734.2025 2025

  55. [57]

    Ingredient-oriented multi-degradation learning for image restoration

    Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multi-degradation learning for image restoration. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5825–5835. IEEE,

  56. [58]

    doi: 10.1109/CVPR52729.2023.00564. 13

    work page doi:10.1109/cvpr52729.2023.00564 2023

  57. [59]

    Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.IEEE Transactions on Image Processing, 26(7):3142–3155, 2017

    Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising.IEEE Transactions on Image Processing, 26(7):3142–3155, 2017. doi: 10.1109/TIP.2017.2662206

    work page doi:10.1109/tip.2017.2662206 2017

  58. [60]

    AdaLoRA: Adaptive Budget Allocation for Parameter-Efficient Fine-Tuning

    Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adaptive budget allocation for parameter- efficient fine-tuning. InInternational Conference on Learning Representations (ICLR), 2023. 10.48550/arXiv.2303.10512

    work page internal anchor Pith review doi:10.48550/arxiv.2303.10512 2023

  59. [61]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yurui Zhu, Tianyu Wang, Xueyang Fu, Xuanyu Yang, Xin Guo, Jifeng Dai, Yu Qiao, and Xiaowei Hu. Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21747–21758, 2023. doi: 10.1109/CVPR52729.2023. 02083. A A...

    work page doi:10.1109/cvpr52729.2023 2023