Pith. sign in

REVIEW 4 major objections 7 minor 76 references

Enhancing Image Restoration Transformer via Adaptive Translation Equivariance

T0 review · 4 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper claims that image-restoration transformers can be made exactly translation-equivariant with an adaptive sliding key-value self-attention module, and that this yields state-of-the-art restoration quality at linear complexity.

desk verdict Strong empirical restoration paper whose exact translation-equivariance claim does not hold up under finite-domain details; still worth serious refereeing. read the letter →

arxiv 2506.18520 v1 pith:JHQYJR64 submitted 2025-06-23 cs.CV

JiaKui Hu , Zhengjian Yao , Lujia Jin , Hangzhou He , Yanye Lu This is my paper · ORCID
classification cs.CV
keywords translationequivarianceimagerestorationself-attentionadaptiveslidingindexingdownsampledTEAFormersuper-resolutioninductivebias
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

This paper tries to establish that translation equivariance—shifting the input shifts the output the same way—is essential for image-restoration transformers, and that it can be restored without losing global receptive field or linear cost. It gives two construction rules: slide indexing, which keeps operator responses at fixed positions, and component stacking, which combines equivariant operators in parallel or in series. On top of these it builds TEA, an attention module that gathers each query's keys and values from an adaptive sliding window plus a downsampled global branch. The full network, TEAFormer, is claimed to satisfy exact translation equivariance and to reach state-of-the-art restoration quality with faster convergence and better generalization than prior restoration transformers.

What carries the argument

The load-bearing object is the TEA attention module, an additive combination of Adaptive Sliding key-value Self-Attention (ASkvSA) and Downsampled Self-Attention (DSA). ASkvSA reshapes keys and values into 2D arrays, applies a depthwise convolution to produce a coordinate map $F \in \mathbb{R}^{H\times W\times 2}$, and shuffles pixels at those coordinates into each query's fixed sliding window; DSA pools keys and values down to a small token count and attends to them globally. The identity doing the argument's work is that the depthwise-convolution-generated index map is shift-equivariant, so a translated input produces translated indices and translated outputs, and Theorem 3.3 transfers equivariance to the stacked and summed architecture.

What would settle it

Take a test image, shift it by one pixel, run TEAFormer, shift the output back, and compare the result with the output on the unshifted image; a nonzero difference larger than floating-point tolerances, especially near image borders, would falsify the exact-equivariance claim. The boundary-blocking rule in Eq. 4 and the average pooling in DSA are the components most likely to show such a difference.

Watch

Extended reading notes

Core claim

The central discovery is that the failure of transformers to be translation-equivariant can be traced to how key-value pairs are indexed, and that this indexing can be made equivariant. In TEA, a depthwise convolution generates a per-pixel offset map; the convolution is shift-equivariant, so when the input shifts, the offsets shift with it, and the gathered keys and values move correspondingly. That branch (ASkvSA) is summed with a downsampled self-attention branch (DSA) that supplies coarse global context. Because equivariant operators stay equivariant under serial and parallel combination (Theorem 3.3), the whole TEA module is declared translation-equivariant; the paper reports that this property yields linear complexity, faster convergence, and higher restoration accuracy than the baseline and previous state-of-the-art methods.

Load-bearing premise

The argument collapses if the complete TEA module is only approximately translation-equivariant; that requires the depthwise-convolution-generated index map, the boundary-blocking rule at finite edges, and DSA's average pooling all to be shift-equivariant, and the paper itself notes that the average pooling is not strictly equivariant.

Editorial extensions

If this is right

  • Translation equivariance can be built into high-performance restoration transformers without quadratic attention: TEA's FLOPs grow linearly as $O(N)$ with fixed hyperparameters.
  • An exactly equivariant module can be used as a drop-in building block: stacking TEA blocks, residual groups, and convolutions preserves equivariance by Theorem 3.3.
  • Restoration models trained with this inductive bias should converge faster and transfer better to out-of-distribution degradations; the paper measures this with NTK convergence, SRGA generalization, and feature-consistency statistics.
  • If the reported results hold, TEAFormer becomes a new state-of-the-art backbone across classical super-resolution, defocus deblurring, Gaussian denoising, and all-in-one restoration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper, one way to test whether equivariance is the actual cause of the gains is to ablate the equivariance of the index map $F$ while keeping parameter count fixed—for instance, replacing the depthwise convolution with a non-equivariant offset predictor—and checking whether the reported convergence and accuracy benefits disappear.
  • A practical strictness fix suggested by the paper's own caveat is to replace DSA's average pooling with a learnable polyphase downsampler; if exact equivariance is the mechanism, this should improve border fidelity and further sharpen generalization.
  • The adaptive slide-indexing principle should transfer to other dense prediction tasks that require pixel-wise fidelity, such as video restoration or medical-image denoising, provided the boundary-blocking rule is adapted to their edge conditions.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper proposes TEAFormer, a transformer for image restoration that combines an adaptive sliding key-value self-attention (ASkvSA) with a downsampled self-attention (DSA) global branch, assembled to preserve translation equivariance (TE). The central theoretical claim is that the resulting TEA attention satisfies exact TE (Definition 3.1 and Eq. 7), and the paper reports state-of-the-art results across image super-resolution, defocus deblurring, Gaussian denoising, and all-in-one restoration. The paper also proposes two general design strategies, slide indexing and component stacking, and provides theoretical statements (Theorems 3.2 and 3.3).

Significance. If the exact-equivalence claim held, the paper would offer a simple and broadly applicable way to restore a principled inductive bias in restoration transformers while keeping global receptive fields and linear complexity. The empirical validation is unusually broad (four tasks, many benchmarks, comparisons with recent transformers such as HAT, IPG, GRL, and InstructIR), and the reported gains are consistent. The manuscript also makes the theoretical framework explicit and includes extensive ablations. However, the exactness claim is not currently supported: the proof of Theorem 3.2 assumes an infinite domain, and Section 5.6 concedes that average pooling in DSA is not strictly TE; moreover, the ablations do not isolate TE as the cause of the observed gains. Thus the paper's main conceptual message requires revision before the result can be accepted.

major comments (4)
  1. [3, Supplement B, Eq. (4)] Theorem 3.2 and its proof in Supplement B assume a one-dimensional infinite sequence when translating indices x_j to x_{j+delta}. In the actual finite-input implementation, the boundary-blocking rule in Eq. (4) changes the indexing window for boundary queries, and a translation of the input can move a query from the blocked boundary regime into the interior regime or vice versa. Consequently the set of key/value indices selected at the translated position is not in general the translation of the set selected at the original position, so Phi(T(x)) = T(Phi(x)) fails at image boundaries. The same boundary issue affects the depthwise convolution used to generate adaptive indices, which is not exactly equivariant with zero-padded finite inputs. Definition 3.1 is therefore not established for ASkvSA or TEA as implemented; the authors should either prove equivariance on finite domains under explicit boundary conventions, or characterize TEA as approximately translation-equivariant and quantify the boundary error.
  2. [5.6, Eq. (6)-(7)] The paper explicitly concedes that "average pooling employed in DSA does not strictly adhere to the property of TE." Since TEA is the parallel combination in Eq. (7), Theorem 3.3's equivariance of the sum requires both branches to be exactly equivariant; a single non-equivariant branch is enough to make the sum non-equivariant. The statement in Section 5.6 that Table 12 "experimentally demonstrate[s] that the use of average pooling does not compromise the model's performance or its TE" is not supported by the reported evidence, which consists of PSNR values and an undefined S-Conv percentage rather than a direct equivariance-error measurement. Please either replace average pooling with a strictly equivariant downsampler, or rephrase the paper's central claim to approximate equivariance and report a quantitative equivariance error.
  3. [5.1, Tables 1-2] The ablations in Tables 1 and 2 compare SwinIR with successive additions of SkvSA, ASkvSA, and DSA, but these additions also change the attention mechanism (adaptive indexing, global branch) independently of translation equivariance. The paper's causal claim that TE drives the observed faster convergence and better generalization is therefore underdetermined; a non-equivariant control that keeps the adaptive indexing and global branch while breaking the equivariance condition (e.g., position-dependent offsets or non-equivariant boundary handling) is needed to support that claim.
  4. [Table 3, TEAFormer-L rows] The TEAFormer-L SSIM entries in Table 3 are inconsistent with the corresponding PSNR values (e.g., 38.27 dB PSNR but 0.8618 SSIM on Set5 x2, where all comparable methods report approximately 0.96; and 0.7442 SSIM on Set5 x4). If accurate, these values contradict the paper's claim that the lightweight model outperforms HAN at 1.3% of the parameters; if typographical errors, they must be corrected. Please verify all TEAFormer-L rows and the text in Section 5.2.
minor comments (7)
  1. [Eq. (3)] The summation bounds are garbled; please rewrite the sliding-window index set as j = i-b1, i-b1+s, ..., i+b2 with explicit boundary handling.
  2. [Table 12, Section 5.6] Define S-Conv and state what it measures; also indicate whether the reported percentages are computed on the same dataset and shift schedule as the rest of the ablation.
  3. [Tables 1-2] The NTK "Condition" and SRGA "Value" metrics are used to support the convergence and generalization claims but are never defined in the main text; please add a sentence describing each metric and the interpretation of lower values.
  4. [4.2 (ASkvSA)] Clarify whether the adaptive index map F is computed separately for K and V, and describe the shuffle operation with explicit tensor shapes (e.g., gather/scatter from 2D coordinates to 1D token indices).
  5. [4.3] The claim that TEA's cost is "slightly lower" than window attention with window size 16 appears inconsistent with Table 2, where the full TEAFormer costs 1.035 T FLOPs versus 0.897 T for SwinIR-Large; please clarify the comparison basis.
  6. [1, 5.2] "In4×image super-resolution" should read "In 4× image super-resolution," and the lightweight comparison says "the CNN method HAN," but HAN uses attention; please correct the description.
  7. [Supplement B, Proof of Theorem 3.3] The proof labels composition as "stacking in parallel"; parallel combination is the sum, and series combination is composition, so the terminology should be made consistent with the main text.

Circularity Check

2 steps flagged · score 2.0 of 10

No load-bearing circularity; the TE claim is definitional and partly unproven at boundaries, and the Urban100 headline is mildly fitted by hyperparameter tuning, but the main empirical comparisons are external.

  1. self definitional [Section 3 (Theorem 3.2), Section 4.2 (boundary processing), Supplement B (Proof of Theorem 3.2, Eqs. 8–9)]
    "Given a function Φ(x)_i is transformed from x_j=[i−b,i+b], where b is the sliding boundary, Φ(x)_i can be rewritten as follows. ... T(Φ(x)_i) = Φ(x)_{i+δ} = Φ(x_{j=[i+δ−b,i+δ+b]}) = Φ(T(x)_{j=[i−b,i+b]}) = Φ(T(x))_i, which completes the proof."

    Theorem 3.2 is the definition of translation equivariance rewritten in index notation: it says that a function whose i-th output is computed from the input window [i−b,i+b] satisfies Definition 3.1. The proof is exactly the definition, so the equivariance of SkvSA is true by construction of the sliding-index rule, not derived from an independent principle. Moreover, the proof assumes the infinite-sequence translation T(x_j)=x_{j+δ}; the finite boundary-blocking rule in Eq. 4 is not part of the theorem's premise, so exact TE at image edges is asserted rather than proven.

  2. fitted input called prediction [Section 5.6 (hyperparameter analysis, Tables 8–11) and Section 5.2 (Table 3)]
    "The performance under various hyperparameters is reported in Table 8 (w), Table 9 (s), Table 10 (k) and Table 11 (Nd). Our settings (w=15, s=4, k=3, Nd=16) optimally balanced performance and computation cost."

    All four TEA hyperparameters (w, s, k, Nd) are selected by comparing Urban100 PSNR in Tables 8–11, and the same Urban100 column is then reported as the headline super-resolution result (28.67 dB in Table 3). The Urban100 number is therefore a selected optimum of a tuning sweep on the evaluation benchmark rather than an independent prediction. This is a partial, mild circularity; other datasets and restoration tasks remain external evidence.

full rationale

The paper's central empirical claims are benchmarked against external methods and datasets (Set5, Set14, B100, Urban100, Manga109, DPDD, etc.), so the main SR, deblurring, denoising, and all-in-one results are not generated by a self-citation chain or by renaming a fitted parameter as a prediction. No load-bearing argument reduces to the authors' own prior work: the DCPT citation [23] only supplies the all-in-one protocol, and the learnable-polyphase citation [45] is an external result. The TE derivation itself is definitional rather than predictive: Theorem 3.2 restates Definition 3.1, and the exactness claim in Eq. 7 is not fully established because the finite boundary-blocking rule of Eq. 4 falls outside the theorem's infinite-sequence proof, and Section 5.6 admits that 'average pooling employed in DSA does not strictly adhere to the property of TE.' These are rigor gaps in the exact-equivariance claim, not instances where an output is circularly reused as an input. The only genuine circularity is mild: hyperparameters are tuned on Urban100 and the same Urban100 figure is featured as the headline SR result, which slightly inflates that specific number. Overall, the architecture is tested against external benchmarks, so the circularity score remains low at 2.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's central claim depends on the equivariance of sliding-window indexing, composition of equivariant operators, equivariant adaptive-index generation, and the coarse equivariance of average pooling. The fourth premise is explicitly acknowledged by the authors as not strictly true, and the first two are proven only for infinite or circular translation without finite-boundary effects.

free parameters (4)
  • sliding window size w = 15
    Chosen via ablation on Urban100 (Table 8); trades FLOPs and PSNR.
  • sliding stride s = 4
    Chosen via ablation on Urban100 (Table 9); does not change FLOPs.
  • adaptive index kernel size k = 3
    Chosen via ablation on Urban100 (Table 10); k=None hurts PSNR.
  • DSA token count N_d = 16
    Chosen via ablation on Urban100 (Table 11); N_d=N gives 28.80 but higher cost.
assumptions (4)
  • domain assumption Sliding window extraction with fixed boundaries is translation equivariant (Theorem 3.2).
    Proof in Supplement B assumes an infinite sequence T(x_j)=x_{j+delta}; at finite image edges the boundary-blocking rule (Eq. 4) is not equivariant.
  • standard math Composition and sum of translation-equivariant operators are translation-equivariant (Theorem 3.3).
    True for any group-equivariant maps; used to stack SkvSA, ASkvSA, and DSA.
  • domain assumption Depthwise convolution generating adaptive indices F is translation-equivariant.
    Invoked in Section 4.2; holds for no-padding or circular convolution but not with zero padding at finite boundaries.
  • ad hoc to paper Average pooling in DSA provides a coarse but acceptable form of translation equivariance.
    Paper admits avg pooling is not strictly TE (Section 5.6) but keeps it; the exact TE claim in Eq. 7 relies on this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Image Restoration Transformer via Adaptive Translation Equivariance." pith.science (2026). https://pith.science/paper/JHQYJR64

@misc{pith2026250618520,
  author       = {Pith},
  title        = {Pith review of: Enhancing Image Restoration Transformer via Adaptive Translation Equivariance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JHQYJR64}},
  note         = {Machine review of arXiv:2506.18520}
}
read the original abstract

Translation equivariance is a fundamental inductive bias in image restoration, ensuring that translated inputs produce translated outputs. Attention mechanisms in modern restoration transformers undermine this property, adversely impacting both training convergence and generalization. To alleviate this issue, we propose two key strategies for incorporating translation equivariance: slide indexing and component stacking. Slide indexing maintains operator responses at fixed positions, with sliding window attention being a notable example, while component stacking enables the arrangement of translation-equivariant operators in parallel or sequentially, thereby building complex architectures while preserving translation equivariance. However, these strategies still create a dilemma in model design between the high computational cost of self-attention and the fixed receptive field associated with sliding window attention. To address this, we develop an adaptive sliding indexing mechanism to efficiently select key-value pairs for each query, which are then concatenated in parallel with globally aggregated key-value pairs. The designed network, called the Translation Equivariance Adaptive Transformer (TEAFormer), is assessed across a variety of image restoration tasks. The results highlight its superiority in terms of effectiveness, training convergence, and generalization.

Figures

Figures reproduced from arXiv: 2506.18520 by the authors.

Figure 1
Figure 1. T (·) means the translation function. ”SA” and ”WA” are commonly used for self-attention and window attention, respec￾tively, but disrupt translation equivariance (TE) due to position en￾coding and feature shifting. ”TEA” is our proposed translation equivariance adaptive attention, which satisfies TE. TE promotes faster convergence and better generalization. Attentions TE Performance NTK [25] SRGA [33] Complexity PS… view at source ↗
Figure 2
Figure 2. The overall architecture of our proposed TEAFormer. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The inner structure of TEA with two components. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison with recent SOTA methods on the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison with recent SOTA methods on the [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison on low-light+haze+rain mixed degradation. TEAFormer can restore the correct lightness. Method Params low-light+haze+rain low-light+haze+snow (M) PSNR↑ / SSIM↑ PSNR↑ / SSIM↑ AirNet [29] 8.90 21.80 / 0.708 22.23 / 0.725 TransWeather [50] 31.0 21.55 / 0.…
Figure 7
Figure 7. Figure 7: Comparison on convergence (left) and generalization (right). 6. Conclusion In this paper, we start by examining the translation equiv￾ariance, which serves as an inherent inductive bias in im￾age restoration. We then define the translation equivariance and offer two st…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 60 canonical work pages

  1. [46]

    Making vision transformers truly shift- equivariant

    Renan A Rojas-Gomez, Teck-Yian Lim, Minh N Do, and Raymond A Yeh. Making vision transformers truly shift- equivariant. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5568– 5577, 2024. 2

    work page 2024

  2. [45]

    Learnable polyphase sam- pling for shift invariant and equivariant convolutional net- works.Advances in Neural Information Processing Systems, 35:35755–35768, 2022

    Renan A Rojas-Gomez, Teck-Yian Lim, Alex Schwing, Minh Do, and Raymond A Yeh. Learnable polyphase sam- pling for shift invariant and equivariant convolutional net- works.Advances in Neural Information Processing Systems, 35:35755–35768, 2022. 4

    work page 2022

  3. [1]

    Defocus deblur- ring using dual-pixel data

    Abdullah Abuolaim and Michael S Brown. Defocus deblur- ring using dual-pixel data. InEuropean Conference on Com- puter Vision, pages 111–126. Springer, 2020. 6

    work page 2020

  4. [2]

    Learning to re- duce defocus blur by realistically modeling dual-pixel data

    Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to re- duce defocus blur by realistically modeling dual-pixel data. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6

    work page 2021

  5. [3]

    Ntire 2017 challenge on single image super-resolution: Dataset and study

    Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. InThe IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR) Workshops, 2017. 6

    2017

  6. [4]

    Attention augmented convolutional net- works

    Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional net- works. InProceedings of the IEEE/CVF international con- ference on computer vision, pages 3286–3295, 2019. 2

    work page 2019

  7. [5]

    Low-complexity single-image super-resolution based on nonnegative neighbor embedding

    Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. InBritish Machine Vision Conference. BMV A press, 2012. 5, 6

    work page 2012

  8. [6]

    Masked image training for generalizable deep image denois- ing

    Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denois- ing. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 1692–1703,

Show all 76 references
  1. [7]

    Simple baselines for image restoration

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. InComputer Vision – ECCV 2022, pages 17–33, Cham, 2022. Springer Nature Switzerland. 8

    2022

  2. [8]

    When vision transformers outperform resnets without pre-training or strong data augmentations

    Xiangning Chen, Cho-Jui Hsieh, and Boqing Gong. When vision transformers outperform resnets without pre-training or strong data augmentations. InInternational Conference on Learning Representations, 2022. 1

    2022

  3. [9]

    Activating more pixels in image super- resolution transformer

    Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image super- resolution transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22367–22377, 2023. 2, 5, 6, 8

    2023

  4. [10]

    Dual aggregation transformer for image super-resolution

    Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xi- aokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. InICCV, 2023. 2, 5, 6

    2023

  5. [11]

    In- structir: High-quality image restoration following human in- structions

    Marcos V Conde, Gregor Geigle, and Radu Timofte. In- structir: High-quality image restoration following human in- structions. InProceedings of the European Conference on Computer Vision (ECCV), 2024. 7, 8

    2024

  6. [12]

    Selective frequency network for image restoration

    Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. InThe Eleventh In- ternational Conference on Learning Representations, 2022. 6

    2022

  7. [13]

    Irnext: Rethinking convolutional network de- sign for image restoration

    Yuning Cui, Wenqi Ren, Sining Yang, Xiaochun Cao, and Alois Knoll. Irnext: Rethinking convolutional network de- sign for image restoration. InProceedings of the 40th Inter- national Conference on Machine Learning, 2023. 6

    2023

  8. [14]

    Selective frequency network for image restoration

    Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. InThe Eleventh In- ternational Conference on Learning Representations, 2023. 6, 7

    2023

  9. [15]

    Cswin transformer: A general vision transformer backbone with cross-shaped windows

    Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pag...

    2022

  10. [16]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition ...

    2021

  11. [17]

    Kodak lossless true color image suite, 2013

    Rich Franzen. Kodak lossless true color image suite, 2013. 7

    2013

  12. [18]

    Bruinsma, Andrew Y

    Jonathan Gordon, Wessel P. Bruinsma, Andrew Y . K. Foong, James Requeima, Yann Dubois, and Richard E. Turner. Con- volutional conditional neural processes. InInternational Conference on Learning Representations, 2020. 1

    2020

  13. [19]

    Onerestore: A universal restoration framework for com- posite degradation

    Yu Guo, Yuan Gao, Yuxu Lu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for com- posite degradation. InEuropean Conference on Computer Vision, 2024. 8

    2024

  14. [20]

    Neighborhood attention transformer

    Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6185–6194, 2023. 2

    2023

  15. [21]

    Fastervit: Fast vision transformers with hierarchical attention

    Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. InThe Twelfth International Conference on Learning Representa- tions, 2024. 2

    2024

  16. [22]

    Squeeze-and-excitation net- works

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation net- works. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7132–7141, 2018. 2

    2018

  17. [23]

    Univer- sal image restoration pre-training via degradation classifica- tion.The Thirteenth International Conference on Learning Representations, 2025

    JiaKui Hu, Lujia Jin, Zhengjian Yao, and Yanye Lu. Univer- sal image restoration pre-training via degradation classifica- tion.The Thirteenth International Conference on Learning Representations, 2025. 7 9

    2025

  18. [24]

    Sin- gle image super-resolution from transformed self-exemplars

    Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Sin- gle image super-resolution from transformed self-exemplars. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5197–5206, 2015. 5, 6, 7, 8

    2015

  19. [25]

    Neu- ral tangent kernel: Convergence and generalization in neural networks.Advances in neural information processing sys- tems, 31, 2018

    Arthur Jacot, Franck Gabriel, and Cl ´ement Hongler. Neu- ral tangent kernel: Convergence and generalization in neural networks.Advances in neural information processing sys- tems, 31, 2018. 1, 5

    2018

  20. [26]

    Adam: A method for stochastic optimization

    Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInternational Conference on Learning Representations (ICLR), San Diega, CA, USA,

  21. [27]

    Iterative filter adaptive network for single image defocus deblurring

    Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. InProceedings of the IEEE Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2021. 6

    2021

  22. [28]

    Benchmarking single- image dehazing and beyond.TIP, 2018

    Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single- image dehazing and beyond.TIP, 2018. 8

    2018

  23. [29]

    All-in-one image restoration for unknown cor- ruption

    Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown cor- ruption. InCVPR, 2022. 8

    2022

  24. [30]

    On effi- cient transformer-based image pre-training for low-level vi- sion

    Wenbo Li, Xin Lu, Shengju Qian, and Jiangbo Lu. On effi- cient transformer-based image pre-training for low-level vi- sion. InProceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1089–

  25. [31]

    Swinir: Image restoration us- ing swin transformer

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration us- ing swin transformer. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 1833–1844,

  26. [32]

    Enhanced deep residual networks for single image super-resolution

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. Enhanced deep residual networks for single image super-resolution. InProceedings of the IEEE confer- ence on computer vision and pattern recognition workshops, pages 136–144, 2017. 1, 5, 6

    2017

  27. [33]

    Evaluating the generalization ability of super- resolution networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14497–14513, 2023

    Yihao Liu, Hengyuan Zhao, Jinjin Gu, Yu Qiao, and Chao Dong. Evaluating the generalization ability of super- resolution networks.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14497–14513, 2023. 1, 5, 8

    2023

  28. [34]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 4

    2021

  29. [35]

    Decoupled weight de- cay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight de- cay regularization. InInternational Conference on Learning Representations, 2019. 6

    2019

  30. [36]

    Controlling vision-language models for universal image restoration

    Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sj¨olund, and Thomas B Sch ¨on. Controlling vision-language models for universal image restoration. InThe Twelfth International Conference on Learning Representations, 2023. 7, 8

    2023

  31. [37]

    Martin, C

    D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecolog- ical statistics. InProceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 41...

    2001

  32. [38]

    A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics

    David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. InICCV, 2001. 7, 8

    2001

  33. [39]

    A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics

    David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. InProceedings Eighth IEEE International Conference on Computer Vision. ICC...

    2001

  34. [40]

    Sketch-based manga retrieval using manga109 dataset.Mul- timedia Tools Appl., 76(20):21811–21838, 2017

    Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset.Mul- timedia Tools Appl., 76(20):21811–21838, 2017. 5, 6

    2017

  35. [41]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. InCVPR, 2017. 8

    2017

  36. [42]

    Single image super-resolution via a holistic attention network

    Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Pro...

    2020

  37. [43]

    Promptir: Prompting for all-in- one image restoration.NeurIPS, 2023

    Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Promptir: Prompting for all-in- one image restoration.NeurIPS, 2023. 7, 8

    2023

  38. [44]

    Stand-alone self- attention in vision models.Advances in neural information processing systems, 32, 2019

    Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya, and Jon Shlens. Stand-alone self- attention in vision models.Advances in neural information processing systems, 32, 2019. 2

    2019

  39. [47]

    Single image defocus deblurring using kernel- sharing parallel atrous convolutions

    Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungy- ong Lee. Single image defocus deblurring using kernel- sharing parallel atrous convolutions. InProc. ICCV, 2021. 6

    2021

  40. [48]

    Image processing gnn: Breaking rigidity in super-resolution

    Yuchuan Tian, Hanting Chen, Chao Xu, and Yunhe Wang. Image processing gnn: Breaking rigidity in super-resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24108–24117, 2024. 2, 4, 5, 6

    2024

  41. [49]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. InInternational conference on machine learning, pages 10347–10357. PMLR, 2021. 2 10

    2021

  42. [50]

    Transweather: Transformer-based restoration of im- ages degraded by adverse weather conditions

    Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of im- ages degraded by adverse weather conditions. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353–2363, 2022. 8

    2022

  43. [51]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017. 2

    2017

  44. [52]

    Pyramid vision transformer: A versatile backbone for dense prediction without convolutions

    Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. InProceedings of the IEEE/CVF international conference on computer vision, pag...

    2021

  45. [53]

    Pvt v2: Improved baselines with pyramid vision transformer

    Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415–424, 2022. 2

    2022

  46. [54]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 17683–17693, 2022. 6

    2022

  47. [55]

    Deep retinex decomposition for low-light enhancement

    Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retinex decomposition for low-light enhancement. In British Machine Vision Conference, 2018. 8

    2018

  48. [56]

    Cvt: Introduc- ing convolutions to vision transformers

    Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introduc- ing convolutions to vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 22–31, 2021. 2

    2021

  49. [57]

    Vision transformer with deformable attention

    Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 4794–4803, 2022. 2

    2022

  50. [58]

    InUncertainty in Artificial Intelligence, pages 2356–2366

    Renjun Xu, Kaifan Yang, Ke Liu, and Fengxiang He.e(2)- equivariant vision transformer. InUncertainty in Artificial Intelligence, pages 2356–2366. PMLR, 2023. 2

    2023

  51. [59]

    Vi- tae: Vision transformer advanced by exploring intrinsic in- ductive bias.Advances in Neural Information Processing Systems, 34, 2021

    Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vi- tae: Vision transformer advanced by exploring intrinsic in- ductive bias.Advances in Neural Information Processing Systems, 34, 2021. 2

    2021

  52. [60]

    Joint rain detection and removal from a single image with contextualized deep net- works.TPAMI, 2019

    Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep net- works.TPAMI, 2019. 8

    2019

  53. [61]

    Effi- cient and explicit modelling of image hierarchies for image restoration

    Xiaoyu Xiang Yawei Li, Yuchen Fan, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Effi- cient and explicit modelling of image hierarchies for image restoration. InProceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition, 2023. 2, 5, 6, 7

    2023

  54. [62]

    Multi-stage progressive image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. InCVPR,

  55. [63]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. InCVPR, 2022. 2, 3, 6, 7, 8

    2022

  56. [64]

    On single image scale-up using sparse-representations

    Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. InCurves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711–

    2010

  57. [65]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825–5835, 2023. 7, 8

    2023

  58. [66]

    Accurate image restoration with attention retractable transformer

    Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. InICLR, 2023. 7

    2023

  59. [67]

    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. 7

    2017

  60. [68]

    Plug-and-play image restora- tion with deep denoiser prior.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 44(10):6360–6376,

    Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restora- tion with deep denoiser prior.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 44(10):6360–6376,

  61. [69]

    Color demosaicking by local directional interpolation and nonlocal adaptive thresholding.Journal of Electronic imaging, 20(2): 023016–023016, 2011

    Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding.Journal of Electronic imaging, 20(2): 023016–023016, 2011. 7

    2011

  62. [70]

    Vitaev2: Vision transformer advanced by exploring induc- tive bias for image recognition and beyond.arXiv preprint arXiv:2202.10108, 2022

    Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vitaev2: Vision transformer advanced by exploring induc- tive bias for image recognition and beyond.arXiv preprint arXiv:2202.10108, 2022. 2

    2022 arXiv

  63. [71]

    Image super-resolution using very deep residual channel attention networks

    Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. InECCV, 2018. 1, 5, 6

    2018

  64. [72]

    Selective hourglass mapping for universal image restoration based on diffusion model

    Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7, 8

    2024

  65. [73]

    Srformer: Permuted self-attention for single image super-resolution

    Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 12780–12791, 2023. 2, 5, 6

    2023

  66. [74]

    Biformer: Vision transformer with bi-level routing attention

    Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Biformer: Vision transformer with bi-level routing attention. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 10323–10333, 2023. 2 11 Enhancing Image Restoration T...

    2023

  67. [730]

    Springer, 2012. 5, 6

    2012

  68. [1097]

    Main Track

    International Joint Conferences on Artificial Intelli- gence Organization, 2023. Main Track. 5, 6, 7

    2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.