Pith. sign in

REVIEW 3 major objections 61 references

CURE: Controllable Unified Image Restoration for Complex Degradations

T0 review · 3 major / 0 minor · reviewed 2026-07-12 · grok-4.5

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

Pith's one-line read Four training losses let existing image restorers remove mixed degradations one factor at a time, at any intensity, without caring about order.

desk verdict Practical training-only recipe that adds usable selective and continuous control to existing prompt restorers; evidence is consistent but mostly synthetic and half-supervised. read the letter →

arxiv 2607.03044 v1 pith:RNA2EOQH submitted 2026-07-03 cs.CV

Boseong Kim , Donghyeon Cho This is my paper
classification cs.CV
keywords all-in-oneimagerestorationcontrollablecompositedegradationdisentangledembeddingsratiocontrolpermutationinvariancetext-guided
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

When an image is ruined by several degradations at once (haze plus rain plus low light, for example), ordinary restorers either fix everything or nothing and often leave artifacts that depend on the order of the fixes. CURE shows that you can keep any existing text-guided restorer and simply retrain it with four extra objectives: an identity path that leaves the image untouched, a ratio path that mixes that identity with a degradation prompt so intensity can be dialed continuously, intermediate supervision that forces each step to remove only one factor, and a permutation loss that makes the final result independent of order. The result is higher restoration scores on composite benchmarks and, more importantly, the ability for a user to keep some degradations, partially remove others, or fuse them all in one pass by changing only the mixing weights of the embeddings. Because the network architecture itself never changes, the same recipe can be dropped onto several published models and immediately improves both quality and control.

What carries the argument

The ratio-control embedding formed by linear interpolation between a fixed identity vector and a degradation-specific text embedding, supervised only at the half-intensity point and reinforced by intermediate and permutation-invariant losses.

What would settle it

Measure whether PSNR/SSIM and residual-degradation classification accuracy remain linear when the mixing weight is swept from 0 to 1 on held-out triple-composite images that were never used in ratio-control training; any sharp non-monotonicity or order dependence would falsify the claim.

Watch

Extended reading notes

Core claim

A model-agnostic training recipe consisting of an all-ones identity embedding plus four losses (identity reconstruction, half-intensity ratio consistency, single-factor intermediate supervision, and order-invariant final consistency) is sufficient to turn existing text-guided image restorers into systems that can selectively and continuously control each factor inside composite degradations.

Load-bearing premise

That mixing the identity embedding and a degradation embedding only at the halfway point is enough to produce smooth, linear intensity control for every other strength and for combinations the model never saw during training.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 0 minor

Summary. CURE proposes a model-agnostic training strategy for text-guided all-in-one restorers that targets composite degradations with explicit controllability. It introduces a fixed all-ones identity embedding, a linear ratio-control embedding (Eq. 2) that mixes identity and degradation embeddings via a scalar w, and four losses: identity (Eq. 3), ratio-control supervised only at half intensity (Eq. 4), intermediate single-factor supervision (Eq. 5), and permutation-invariant consistency (Eq. 6). The method is plugged into OneRestore, TextPromptIR, and AM-PromptIR without architectural change. On the authors’ CCDD-11 weather composites and a Blur–Noise–JPEG set, CURE reports modest PSNR/SSIM gains over the same baselines (Table 1), strong identity preservation (Table 3), improved selective removal (Tables 5–6), reduced order dependence (Table 7), and monotonic residual-degradation classifier accuracy as w varies (Table 4). Code and the synthetic dataset are released.

Significance. If the continuous selective-control claim holds, CURE fills a genuine gap: existing composite restorers (notably OneRestore) lack reliable identity bypass, soft intensity control, and order-invariant sequential removal. The contribution is primarily a training recipe rather than a new architecture, which is a strength for adoption. Positive aspects include multi-baseline integration, ablations isolating each loss (Table 2), selective-restoration and order-invariance metrics with explicit GTs, extension to digital degradations, real-world NIQE checks, and public code/dataset. The absolute PSNR gains are small (often <1 dB), so the main value is controllability rather than raw restoration quality. The work is of clear interest to the all-in-one restoration community provided the continuous-control evidence is tightened.

major comments (3)
  1. §3.2–3.3, Eqs. (2) and (4), Table 4: The central continuous-intensity claim rests on linear interpolation between a fixed all-ones identity embedding and a degradation embedding, supervised only at the single half-intensity point w=0.5 (plus a second half-step to clean). Table 4 and the Blur–Noise–JPEG counterpart report only residual-degradation classifier accuracy versus w; they do not report pixel-level PSNR/SSIM of intermediate outputs against true partial-degradation GTs for untrained w values (e.g., 0.2, 0.7). Because half-intensity GTs are themselves generated by simply halving the same synthesis parameters used at test time, the monotonic classifier curve is a weak proxy for a continuous, approximately linear control manifold. Either provide pixel-level fidelity curves against partial GTs for a dense set of w, or clearly limit the claim to the trained half-intensity regime and qu
  2. §3.4 and Supp. §1: CCDD-11 (and the Blur–Noise–JPEG set) are synthesized with the same operators and parameter ranges used to define half-intensity and selective GTs. Training and evaluation therefore share the degradation model, creating mild circularity for the ratio-control and intermediate objectives. The paper already notes rain-mask diversity relative to CDD-11; it should also quantify how sensitive the reported control metrics are to changes in the synthesis parameters (different β, γ, α ranges, or non-halved intensity schedules) and, where possible, include at least one real partial-degradation protocol or cross-operator test so that controllability is not solely an artifact of the training distribution.
  3. §3.4 / Supp. Table 2: Triple-composite images are excluded from CURE training “due to combinatorial complexity” yet appear in the test set and in selective-restoration tables. The paper asserts generalization from dual to triple composites, but the intermediate and permutation losses are never applied to three-factor mixtures during training. A short controlled experiment (or explicit negative result) showing whether the same losses remain stable when a modest number of triple samples are included—or a clearer statement that triple performance is zero-shot extrapolation—would make the scope of the claim precise.

Circularity Check

0 steps flagged · score 1.0 of 10

Empirical training objectives on author-synthesized data; no derivation reduces to its inputs by construction.

full rationale

CURE is a supervised training recipe (identity embedding + four losses) that is plugged into existing text-guided restorers. The ratio-control embedding (Eq. 2) and half-intensity loss (Eq. 4) are design choices, not claimed first-principles derivations; continuous control for other w and for triple composites is an empirical extrapolation tested only via residual-degradation classifier accuracy (Tab. 4) and selective-restoration PSNR against the same synthetic pipeline. Dataset synthesis (CCDD-11, Blur-Noise-JPEG) follows the public CDD-11 protocol with transparent half-parameter GTs; this is ordinary for the field and does not make any reported metric tautological. No self-citation is load-bearing, no uniqueness theorem is imported, and no fitted parameter is renamed a prediction. Real-world NIQE numbers and external baselines further ground the claims. The single mild note is that half-intensity supervision plus linear mixing is never pixel-verified for untrained intensities, but that is a validity gap, not circularity under the stated criteria.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The central claim rests on the empirical effectiveness of four new losses and a fixed identity embedding, plus the assumption that synthetic half-intensity and selective pairs generated by the authors' pipeline are faithful proxies for controllable restoration. No free physical constants; the free parameters are ordinary training hyper-parameters and the hand-chosen all-ones identity vector. Invented entities are the identity embedding and the ratio-control embedding construction.

free parameters (3)
  • identity embedding vector = constant vector of ones
    Fixed to the constant vector of ones; no training, but the choice of the all-ones vector is arbitrary and load-bearing for the identity and ratio-control losses.
  • ratio-control training intensity w = 0.5
    Only w=0.5 is supervised; all other intensities are assumed to interpolate linearly at inference.
  • loss-balancing coefficients and learning-rate schedule for CURE stage
    Inherited or slightly lowered from each baseline; not ablated exhaustively.
assumptions (3)
  • ad hoc to paper Linear interpolation between identity and degradation embeddings yields continuous, approximately linear restoration intensity.
    Stated in Eq. (2) and justified only by the half-intensity loss (Eq. 4) and post-hoc classification curves.
  • domain assumption Synthetic degradations generated by the CCDD-11 / Blur-Noise-JPEG pipelines (halved parameters, same masks) are valid ground truth for selective and partial restoration.
    Dataset construction section; all quantitative claims rest on these pairs.
  • domain assumption Existing text-guided restorers already provide a usable cross-attention or feature-injection interface for arbitrary embeddings.
    Baseline framework section; CURE never modifies architecture.
invented entities (2)
  • identity embedding (constant ones vector)
    purpose: Represents the do-nothing operation so that ratio control and identity loss can be defined.
    Introduced in Sec. 3.2; no independent evidence outside the paper's own identity-PSNR table.
  • ratio-control embedding
    purpose: Continuous intensity dial obtained by linear mix of identity and degradation embeddings.
    Eq. (2); validated only inside the paper's classification and visual experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CURE: Controllable Unified Image Restoration for Complex Degradations." pith.science (2026). https://pith.science/paper/RNA2EOQH

@misc{pith2026260703044,
  author       = {Pith},
  title        = {Pith review of: CURE: Controllable Unified Image Restoration for Complex Degradations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RNA2EOQH}},
  note         = {Machine review of arXiv:2607.03044}
}
read the original abstract

The presence of composite degradations poses a significant challenge, since the underlying corruption factors exhibit complex and interdependent interactions. Even when the degradation types are known, accurately restoring the image remains difficult due to the intertwined nature of their effects and the need for selective control during the recovery process. To address this, we introduce CURE, a unified framework that enables controllable restoration in complex degradation settings by learning disentangled and adjustable representations. CURE is driven by four complementary objectives. First, an identity embedding is incorporated, along with a reconstruction constraint, to ensure that the model can reproduce the input image when restoration is unnecessary. Second, the ratio control mechanism blends the identity embedding with degradation-specific embeddings using user-regulated mixing ratios, allowing continuous control over restoration intensity. Third, an intermediate loss is applied to supervise stepwise outputs, each encouraged to tackle the removal of only a single degradation factor within a composite mixture. Finally, a permutation-invariant loss ensures that the model achieves consistent restoration quality regardless of the order in which multiple degradations are addressed. Since CURE modifies only the training strategy and not the underlying network architecture, it can be seamlessly integrated into existing controllable restoration models. Experiments demonstrate that CURE delivers state-of-the-art performance on composite degradation benchmarks, while enabling both selective and jointly fused restoration through flexible modulation of embedding ratios. The code and dataset are available at https://github.com/bo-oseng/CURE.

Figures

Figures reproduced from arXiv: 2607.03044 by the authors.

Figure 1
Figure 1. Overview of our CURE. (a) Embedding Generation constructs degradation-aware embeddings via a pre-trained text embedder and our ratio￾control mechanism. This enables continuous intensity control by interpolating between identity and degradation embeddings. (b) The identity loss is a loss func￾tion designed specifically for the identity input, which preserves the input without any modifications. It serves to suppress … view at source ↗
Figure 2
Figure 2. Qualitative comparisons on the CCDD-11 dataset. Ours refers to the [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. (a) Ratio-control embedding. (b) Selective restoration. (c) Restoration [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (16 more)
Figure 1
Figure 1. Figure 1: Qualitative Restoration Results on Real-World Datasets. [PITH_FULL_IMAGE:figures/full_fig_p023_1.png]
Figure 2
Figure 2. Figure 2: Qualitative comparison of identity operation for Snow image, where the [PITH_FULL_IMAGE:figures/full_fig_p025_2.png]
Figure 3
Figure 3. Figure 3: Qualitative comparison of identity operation for Low [PITH_FULL_IMAGE:figures/full_fig_p026_3.png]
Figure 4
Figure 4. Figure 4: Qualitative results of Ratio Control Embedding with varying [PITH_FULL_IMAGE:figures/full_fig_p027_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results of Ratio Control Embedding with varying [PITH_FULL_IMAGE:figures/full_fig_p028_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results of Ratio Control Embedding with varying [PITH_FULL_IMAGE:figures/full_fig_p029_6.png]
Figure 7
Figure 7. Figure 7: (a) Qualitative results of selective restoration for the Haze component in [PITH_FULL_IMAGE:figures/full_fig_p030_7.png]
Figure 8
Figure 8. Figure 8: (a) Qualitative results of selective restoration for the Low component in [PITH_FULL_IMAGE:figures/full_fig_p031_8.png]
Figure 9
Figure 9. Figure 9: (a) Qualitative results of selective restoration for the Low component in [PITH_FULL_IMAGE:figures/full_fig_p032_9.png]
Figure 10
Figure 10. Figure 10: Example results of simultaneous Ratio Control Restoration and Selective [PITH_FULL_IMAGE:figures/full_fig_p033_10.png]
Figure 11
Figure 11. Figure 11: Comparison of one-stage and two-stage restoration results for (a) OneRe [PITH_FULL_IMAGE:figures/full_fig_p034_11.png]
Figure 12
Figure 12. Figure 12: Comparison of one-stage and two-stage restoration results for (a) OneRe [PITH_FULL_IMAGE:figures/full_fig_p035_12.png]
Figure 13
Figure 13. Figure 13: Comparison of one-stage and two-stage restoration results for (a) OneRe [PITH_FULL_IMAGE:figures/full_fig_p036_13.png]
Figure 14
Figure 14. Figure 14: Qualitative comparison of Haze+Snow image restoration results between OneRestore and Ours [PITH_FULL_IMAGE:figures/full_fig_p037_14.png]
Figure 15
Figure 15. Figure 15: Qualitative comparison of Low+Haze+Snow image restoration results between OneRestore and Ours [PITH_FULL_IMAGE:figures/full_fig_p038_15.png]
Figure 16
Figure 16. Figure 16: Qualitative comparison of Blur image restoration results between OneRe [PITH_FULL_IMAGE:figures/full_fig_p039_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

61 extracted references · 1 linked inside Pith

  1. [1]

    IJCV93(3), 348–367 (2011)

    Bossu, J., Hautière, N., Tarel, J.P.: Rain or snow detection in image sequences through use of a histogram of orientation of streaks. IJCV93(3), 348–367 (2011)

    2011

  2. [2]

    In: CVPR

    Chen, H., Wang, Y., Guo, T., Xu, C., Deng, Y., Liu, Z., Ma, S., Xu, C., Xu, C., Gao, W.: Pre-trained image processing transformer. In: CVPR. pp. 12299–12310 (2021)

    2021

  3. [3]

    In: CVPR

    Chen, I., Chen, W.T., Liu, Y.W., Chiang, Y.C., Kuo, S.Y., Yang, M.H., et al.: Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior. In: CVPR. pp. 17969–17979 (2025)

    2025

  4. [4]

    In: ECCV

    Chen, L., Chu, X., Zhang, X., Sun, J.: Simple baselines for image restoration. In: ECCV. pp. 17–33. Springer (2022)

    2022

  5. [5]

    In: ICCV

    Chen, W.T., Fang, H.Y., Hsieh, C.L., Tsai, C.C., Chen, I., Ding, J.J., Kuo, S.Y., et al.: All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In: ICCV. pp. 4196–4205 (2021)

    2021

  6. [7]

    In: CVPR

    Chen, W.T., Huang, Z.K., Tsai, C.C., Yang, H.H., Ding, J.J., Kuo, S.Y.: Learning multiple adverse weather removal via two-stage knowledge learning and multi- contrastive regularization: Toward a unified model. In: CVPR. pp. 17653–17662 (2022)

    2022

  7. [8]

    In: CVPR

    Chen, X., Pan, J., Dong, J.: Bidirectional multi-scale implicit neural representations for image deraining. In: CVPR. pp. 25627–25636 (2024)

    2024

  8. [9]

    In: ECCV

    Conde, M.V., Geigle, G., Timofte, R.: Instructir: High-quality image restoration following human instructions. In: ECCV. pp. 1–21. Springer (2024)

    2024

Show all 61 references
  1. [10]

    In: AAAI

    Cui, Y., Ren, W., Knoll, A.: Omni-kernel network for image restoration. In: AAAI. vol. 38, pp. 1426–1434 (2024)

    2024

  2. [11]

    In: ICLR

    Cui, Y., Zamir, S.W., Khan, S., Knoll, A., Shah, M., Khan, F.S.: Adair: Adaptive all-in-one image restoration via frequency mining and modulation. In: ICLR. pp. 57335–57356. International Conference on Learning Representations, ICLR (2025)

    2025

  3. [12]

    In: Proceedings of the 6th ACM multimedia systems conference

    Dang-Nguyen, D.T., Pasquini, C., Conotter, V., Boato, G.: Raise: A raw images dataset for digital image forensics. In: Proceedings of the 6th ACM multimedia systems conference. pp. 219–224 (2015)

    2015

  4. [13]

    In: CVPR

    Dong, H., Pan, J., Xiang, L., Hu, Z., Zhang, X., Wang, F., Yang, M.H.: Multi-scale boosted dehazing network with dense feature fusion. In: CVPR. pp. 2157–2167 (2020)

    2020

  5. [14]

    In: CVPR (2017)

    Fu, X., Huang, J., Zeng, D., Huang, Y., Ding, X., Paisley, J.: Removing rain from single images via a deep detail network. In: CVPR (2017)

    2017

  6. [15]

    In: AAAI

    Fu, X., Qi, Q., Zha, Z.J., Zhu, Y., Ding, X.: Rain streak removal via dual graph convolutional network. In: AAAI. vol. 35, pp. 1352–1360 (2021)

    2021

  7. [16]

    ACM Transactions on Graphics (TOG)25(3), 996–1002 (2006)

    Garg, K., Nayar, S.K.: Photorealistic rendering of rain streaks. ACM Transactions on Graphics (TOG)25(3), 996–1002 (2006)

    2006

  8. [17]

    In: CVPR

    Guo, C.L., Yan, Q., Anwar, S., Cong, R., Ren, W., Li, C.: Image dehazing trans- former with transmission-aware 3d position embedding. In: CVPR. pp. 5812–5820 (2022)

    2022

  9. [18]

    In: ECCV

    Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple baseline for image restoration with state-space model. In: ECCV. pp. 222–241. Springer (2024) CURE: Controllable Unified Image Restoration for Complex Degradations 41

    2024

  10. [19]

    IEEE TIP26(2), 982–993 (2016)

    Guo, X., Li, Y., Ling, H.: Lime: Low-light image enhancement via illumination map estimation. IEEE TIP26(2), 982–993 (2016)

    2016

  11. [20]

    In: ECCV

    Guo, Y., Gao, Y., Lu, Y., Zhu, H., Liu, R.W., He, S.: Onerestore: A universal restoration framework for composite degradation. In: ECCV. pp. 255–272. Springer (2024)

    2024

  12. [21]

    In: CVPR

    He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. In: CVPR. pp. 1956–1963 (2009)

    1956

  13. [22]

    IEEE TPAMI33(12), 2341–2353 (2010)

    He, K., Sun, J., Tang, X.: Single image haze removal using dark channel prior. IEEE TPAMI33(12), 2341–2353 (2010)

    2010

  14. [23]

    arXiv preprint arXiv:2501.15510 (2025)

    Hu, J., Jin, L., Yao, Z., Lu, Y.: Universal image restoration pre-training via degradation classification. arXiv preprint arXiv:2501.15510 (2025)

    2025 arXiv

  15. [24]

    In: ECCV

    Kulkarni, A., Phutke, S.S., Murala, S.: Unified transformer network for multi- weather image restoration. In: ECCV. pp. 344–360 (2022)

    2022

  16. [25]

    Scientific American237(6), 108–129 (1977)

    Land, E.H.: The retinex theory of color vision. Scientific American237(6), 108–129 (1977)

    1977

  17. [26]

    IEEE TIP28(1), 492–505 (2018)

    Li, B., Ren, W., Fu, D., Tao, D., Feng, D., Zeng, W., Wang, Z.: Benchmarking single-image dehazing and beyond. IEEE TIP28(1), 492–505 (2018)

    2018

  18. [27]

    In: CVPR

    Li, B., Liu, X., Hu, P., Wu, Z., Lv, J., Peng, X.: All-in-one image restoration for unknown corruption. In: CVPR. pp. 17452–17462 (2022)

    2022

  19. [28]

    In: CVPR

    Li, R., Cheong, L.F., Tan, R.T.: Heavy rain image restoration: Integrating physics model and conditional adversarial learning. In: CVPR. pp. 1633–1642 (2019)

    2019

  20. [29]

    In: CVPR

    Li, R., Tan, R.T., Cheong, L.F.: All in one bad weather removal using architectural search. In: CVPR. pp. 3175–3185 (2020)

    2020

  21. [30]

    In: CVPR

    Li, Z., Snavely, N.: Megadepth: Learning single-view depth prediction from internet photos. In: CVPR. pp. 2041–2050 (2018)

  22. [31]

    In: Proc

    Liang, Y., Anwar, S., Liu, Y.: Drt: A lightweight single image deraining recursive transformer. In: Proc. IEEE CVPR. pp. 589–598 (2022)

    2022

  23. [32]

    In: CVPR

    Lin, Y., Lin, Z., Chen, H., Pan, P., Li, C., Chen, S., Wen, K., Jin, Y., Li, W., Ding, X.: Jarvisir: Elevating autonomous driving perception with intelligent image restoration. In: CVPR. pp. 22369–22380 (2025)

    2025

  24. [33]

    In: CVPR

    Liu, R., Ma, L., Zhang, J., Fan, X., Luo, Z.: Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In: CVPR. pp. 10561–10570 (2021)

    2021

  25. [34]

    IEEE TIP27(6), 3064–3073 (2018)

    Liu, X., Zhang, M., Fu, X., Huang, Y., Ding, X., Paisley, J.: Desnownet: Context- aware deep network for snow removal. IEEE TIP27(6), 3064–3073 (2018)

    2018

  26. [35]

    In: CVPR

    Ma, L., Ma, T., Liu, R., Fan, X., Luo, Z.: Toward fast, flexible, and robust low-light image enhancement. In: CVPR. pp. 5637–5646 (2022)

    2022

  27. [36]

    In: CVPR

    Mou, C., Wang, Q., Zhang, J.: Deep generalized unfolding networks for image restoration. In: CVPR. pp. 17399–17410 (2022)

    2022

  28. [37]

    In: ICRA

    Porav, H., Bruls, T., Newman, P.: I can see clearly now: Image restoration via de-raining. In: ICRA. pp. 7087–7093. IEEE (2019)

    2019

  29. [38]

    In: NeurIPS

    Potlapalli, V., Zamir, S.W., Khan, S., Khan, F.S.: Promptir: Prompting for all-in-one blind image restoration. In: NeurIPS. pp. 71275–71293 (2023)

    2023

  30. [39]

    In: CVPR

    Qu, L., Tian, J., He, S., Tang, Y., Lau, R.W.: Deshadownet: A multi-context embedding deep network for shadow removal. In: CVPR. pp. 4067–4075 (2017)

    2017

  31. [40]

    IEEE TCSVT33(7), 3133–3144 (2023)

    Quan, Y., Tan, X., Huang, Y., Xu, Y., Ji, H.: Image desnowing via deep invertible separation. IEEE TCSVT33(7), 3133–3144 (2023)

    2023

  32. [41]

    In: CVPR

    Tian, X., Liao, X., Liu, X., Li, M., Ren, C.: Degradation-aware feature perturbation for all-in-one image restoration. In: CVPR. pp. 28165–28175 (2025)

    2025

  33. [42]

    NeurIPS30(2017) 42 Boseong Kim and Donghyeon Cho

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. NeurIPS30(2017) 42 Boseong Kim and Donghyeon Cho

    2017

  34. [43]

    IEEE TIP33, 191–204 (2023)

    Wang, Q., Jiang, K., Wang, Z., Ren, W., Zhang, J., Lin, C.W.: Multi-scale fusion and decomposition network for single image deraining. IEEE TIP33, 191–204 (2023)

    2023

  35. [44]

    IEEE TIP22(9), 3538–3548 (2013)

    Wang, S., Zheng, J., Hu, H.M., Li, B.: Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE TIP22(9), 3538–3548 (2013)

    2013

  36. [45]

    In: ICCV

    Wang, X., Xie, L., Dong, C., Shan, Y.: Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In: ICCV. pp. 1905–1914 (2021)

    1905

  37. [46]

    In: CVPR

    Wang, Y., Ma, C., Liu, J.: Smartassign: Learning a smart knowledge assignment strategy for deraining and desnowing. In: CVPR. pp. 3677–3686 (2023)

    2023

  38. [47]

    In: CVPR

    Wang, Z., Cun, X., Bao, J., Zhou, W., Liu, J., Li, H.: Uformer: A general u-shaped transformer for image restoration. In: CVPR. pp. 17683–17693 (2022)

    2022

  39. [48]

    In: CVPR

    Xu, X., Wang, R., Fu, C.W., Jia, J.: Snr-aware low-light image enhancement. In: CVPR. pp. 17714–17724 (2022)

    2022

  40. [49]

    EAAI155, 110981 (2025)

    Yan, Q., Jiang, A., Chen, K., Peng, L., Yi, Q., Zhang, C.: Textual prompt guided image restoration. EAAI155, 110981 (2025)

    2025

  41. [50]

    In: CVPR

    Yang, W., Tan, R.T., Feng, J., Liu, J., Guo, Z., Yan, S.: Deep joint rain detection and removal from a single image. In: CVPR. pp. 1357–1366 (2017)

    2017

  42. [51]

    In: CVPR

    Zamfir, E., Wu, Z., Mehta, N., Tan, Y., Paudel, D.P., Zhang, Y., Timofte, R.: Complexity experts are task-discriminative learners for any image restoration. In: CVPR. pp. 12753–12763 (2025)

    2025

  43. [52]

    In: CVPR

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H.: Restormer: Efficient transformer for high-resolution image restoration. In: CVPR. pp. 5728–5739 (2022)

    2022

  44. [53]

    In: ECCV

    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. pp. 492–511. Springer (2020)

    2020

  45. [55]

    In: CVPR

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Multi-stage progressive image restoration. In: CVPR. pp. 14821–14831 (2021)

    2021

  46. [56]

    IEEE TPAMI45(2), 1934–1948 (2022)

    Zamir, S.W., Arora, A., Khan, S., Hayat, M., Khan, F.S., Yang, M.H., Shao, L.: Learning enriched features for fast image restoration and enhancement. IEEE TPAMI45(2), 1934–1948 (2022)

    1934

  47. [57]

    In: CVPR

    Zeng, H., Wang, X., Chen, Y., Su, J., Liu, J.: Vision-language gradient descent- driven all-in-one deep unfolding networks. In: CVPR. pp. 7524–7533 (2025)

    2025

  48. [58]

    In: CVPR

    Zhang, H., Patel, V.M.: Density-aware single image de-raining using a multi-stream dense network. In: CVPR. pp. 695–704 (2018)

    2018

  49. [59]

    zhang et al

    Zhang, S., Wang, J., He, N., Xu, S., Liu, S., Yu, P., Ma, X.: Adaptive multi-modal prompting for universal image restoration amidst diverse degradations: S. zhang et al. Vis. Comput. pp. 1–13 (2025)

    2025

  50. [60]

    In: CVPR

    Zheng, Y., Zhan, J., He, S., Dong, J., Du, Y.: Curricular contrastive regularization for physics-aware single image dehazing. In: CVPR. pp. 5785–5794 (2023)

    2023

  51. [61]

    In: ICML

    Zhou, M., Huang, J., Guo, C.L., Li, C.: Fourmer: An efficient global modeling paradigm for image restoration. In: ICML. pp. 42589–42601. PMLR (2023)

    2023

  52. [62]

    IEEE TMM26, 650–662 (2023)

    Zhou, M., Wu, X., Wei, X., Xiang, T., Fang, B., Kwong, S.: Low-light enhancement method based on a retinex model for structure preservation. IEEE TMM26, 650–662 (2023)

    2023

  53. [63]

    In: CVPR

    Zhu, Y., Wang, T., Fu, X., Yang, X., Guo, X., Dai, J., Qiao, Y., Hu, X.: Learning weather-general and weather-specific features for image restoration under multiple adverse weather conditions. In: CVPR. pp. 21747–21758 (2023)

    2023

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.