Pith. sign in

REVIEW 4 major objections 4 minor 65 references

MAT: Multi-Range Attention Transformer for Efficient Image Super-Resolution

T0 review · 4 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a Transformer for image super-resolution can get a larger effective receptive field and richer features by mixing regional attention over several neighborhood sizes with sparse, dilated global attention—reaching…

desk verdict New architecture, strong ablations, but the SOTA claim is contradicted by their own Table VII on Manga109 x4. read the letter →

arxiv 2411.17214 v3 pith:O3B2QII6 submitted 2024-11-26 cs.CV

Chengxing Xie , Xiaoming Zhang , Linze Li , Yuqian Fu , Biao Gong , Tianrui Li , Kai Zhang This is my paper · ORCID
classification cs.CV
keywords imagesuper-resolutionmulti-rangeattentionsparsedilatedlightweighttransformerhierarchicalfeaturerepresentation
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 the fixed-size attention window used by most Transformer super-resolution models is the wrong unit of efficiency. Its claim is that attending over several neighborhood sizes at once—regional attention in $7\times7$, $9\times9$, and $11\times11$ blocks—plus a sparsely sampled, dilated global attention, captures hierarchical and self-similar image structure better than any single window, and at lower cost. The lightweight model MAT-light reaches $26.83$ dB on Urban100 at $\times4$ with 714K parameters and reports the best PSNR/SSIM among lightweight methods trained on DIV2K across five benchmarks; the classical-scale model matches larger state-of-the-art transformers with about 9.6M parameters. If the claim is right, the efficiency bottleneck in SR transformers is not attention itself but how window size is chosen.

What carries the argument

The load-bearing object is sparse multi-range attention (SMA), defined in Eqs. (8)–(10): for range size $k$ and dilation rate $\delta$, the key-value set for pixel $(i,j)$ contains only positions $(i+x\delta,\ j+y\delta)$ with $-k/2 \le x,y \le k/2$, so the number of attended keys stays roughly $k^2$ while the covered area grows by about $\delta$. Multi-range attention (MA) is the non-sparse companion: several heads compute regional attention over different $k$ values and their outputs are concatenated and fused. A Local Aggregation Block uses depth-wise convolutions and channel attention for local features, and the MSConvStar module replaces the standard feed-forward network with parallel multi-scale depth-wise convolutions plus a star operation. The dilation schedule in the paper sets $\delta$ as the floor division of the $64\times64$ patch size by the range size, i.e., $\{9,7,5\}$ for ranges $\{7,9,11\}$, which lets SMA reach across the whole training patch at roughly constant cost.

What would settle it

Train MAT-light with the same parameter and compute budget but with dense global attention in place of SMA, or evaluate the model on inputs whose patch size is $128\times128$ while recomputing the dilation schedule; if the dense version matches or beats MAT, or if performance drops sharply when the schedule is re-derived for a different patch size, the multi-range sparse design is not doing the claimed work. A complementary test is to construct images whose repeating texture has a period not commensurate with the dilation step and check whether SMA misses those matches.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that dilation can be transplanted from convolutions into self-attention to make a Transformer's receptive field flexible without raising its compute. Multi-range attention applies regional attention at several range sizes across different heads, and sparse multi-range attention samples keys on a grid stepped by a dilation rate, so a $k\times k$ set of keys spreads across a far larger spatial footprint. Combined with a local aggregation block and the MSConvStar feed-forward module, this yields a model whose effective receptive field, measured by LAM diffusion index, covers nearly the whole input image. The paper reports that this design outperforms fixed-window transformers such as SwinIR-light and SRFormer-light on all five benchmarks at $\times2$, $\times3$, and $\times4$, with the largest gains on Urban100 and Manga109, while using fewer parameters, Multi-Adds, inference time, and GPU memory.

Load-bearing premise

The argument hinges on the idea that a sparse, evenly spaced sample of distant pixels preserves the useful non-local information that dense attention would find, at the specific dilation ratios chosen for $64\times64$ training patches; if informative matches fall between the sampled positions, the efficiency claim loses its performance basis.

Editorial extensions

If this is right

  • Lightweight super-resolution can match or beat fixed-window Transformer accuracy with much less compute: MAT-light reports $26.83$ dB on Urban100 at $\times4$ with 714K parameters and 48.5G Multi-Adds.
  • Scaling the same multi-range recipe works: the $9.6$M-parameter classical MAT outperforms or ties SwinIR, HAT-S, ART-S, DAT-S, SRFormer and MambaIR on most benchmark and scale combinations.
  • The receptive-field gain is attributable to sparse global attention: ablations show that removing SMA lowers the LAM diffusion index and PSNR more than removing MA or LAB.
  • The MSConvStar feed-forward module improves feature diversity while reducing parameters relative to a plain MLP or ConvFFN baseline.
  • Because the design avoids large padded windows, it is faster in practice: MAT-light runs about $3.3\times$ faster than SRFormer-light and uses about 24% less GPU memory at $1536^2$ output.

Reading between the lines

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

  • Because the gain is tied to a dilation schedule fixed by the training patch size, a natural testable extension is to make dilation rates learned or resolution-adaptive; the paper does not report such a variant.
  • The same sparse-sampling logic should transfer to other dense prediction tasks such as denoising, deblurring, or segmentation, since the underlying redundancy argument concerns natural image statistics rather than super-resolution specifically.
  • If the benefit comes from covering a larger footprint with the same number of attended keys, then very high resolutions will thin the sampled grid and may require more range levels; the paper's experiments stop at $1536^2$ output, so this scaling behavior remains unshown.
  • The paper's own 'not all global dependencies are equally valuable' framing suggests that a learned gating mechanism over ranges could outperform the fixed two-heads-per-range allocation, a variant the authors leave untested.
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

4 major / 4 minor

Summary. The paper proposes MAT, a transformer architecture for single-image super-resolution that replaces fixed-window self-attention with multi-range attention (MA) and sparse multi-range attention (SMA), and introduces the MSConvStar module for multi-range representation learning. The authors evaluate lightweight and classical variants of MAT on five benchmarks, reporting state-of-the-art PSNR/SSIM with lower parameter counts and Multi-Adds, and support the design with component ablations, LAM visualizations, ERF measurements, and runtime comparisons.

Significance. If the empirical results are reproducible, MAT is a genuinely useful efficient SR architecture: the sparse dilated attention in SMA provides a larger effective receptive field at the same computational cost, and the reported 714K-parameter lightweight model reaches 26.83 dB on Urban100 x4, outperforming SRFormer-light with lower Multi-Adds. The paper's strengths are its component-by-component ablations (Tables I-V), the LAM/ERF analyses, the public code release, and the breadth of benchmark comparisons. However, the classical-SR state-of-the-art claim is not fully established because one baseline entry in Table VII is better in PSNR, and the reported gains are single-run values without variance; the precise magnitude of the claimed advantages is therefore uncertain.

major comments (4)
  1. [§IV-D, Table VII] The statement in §IV-D that "MAT achieves superior performance across all five benchmark datasets and scale factors" is contradicted by Table VII: on the x4 Manga109 row, MambaIR has PSNR 33.32 dB while MAT+ has 32.49 dB, a 0.83 dB deficit for the proposed method on the primary metric. Because PSNR is the standard comparison metric in SR, this sentence is false as written. The claim should be corrected or qualified (e.g., per-dataset exceptions or parameter-efficiency framing), and the MambaIR value should be verified against the original MambaIR publication because it is considerably higher than neighboring entries such as HAT-S (32.35 dB) and SwinIR (32.03 dB).
  2. [§IV-A, Tables VI-VII] The benchmark tables report single-run PSNR/SSIM without variance. Several of the headline gains are in the 0.08-0.16 dB range, e.g., Urban100 x4: MAT-light 26.83 vs. SRFormer-light 26.67, and Manga109 x4: MAT-light 31.38 vs. SRFormer-light 31.17; in this regime training noise of 0.1-0.2 dB is common. Please provide mean plus/minus standard deviation over at least three training runs for the main lightweight comparisons, or otherwise report seed variability, so that the claimed improvements can be distinguished from training noise.
  3. [§IV-G, Fig. 16] The abstract's "3.3x faster" claim is taken from the largest resolution shown in Fig. 16 (1536x1536), but at the 1280x720 setting used in Table VIII the speedup is only 162.9/71.1 about 2.29x, and the Multi-Adds ratio is 62.8/48.5 about 1.29x. The paper should state the speedup at the standard benchmark resolution and clearly qualify that the factor depends on image resolution; otherwise the headline efficiency claim is misleading.
  4. [Eq. (9), §IV-A] The dilation schedule delta = {9,7,5} is defined as the floor division of the 64x64 training patch size by the range sizes, and Eq. (9) samples keys at positions offset by multiples of delta. The paper does not specify how SMA is applied at test-time resolutions that are not 64x64, nor how boundary positions are handled (padding versus masking). Because the model is evaluated on arbitrary image sizes, an explicit boundary description and an experiment at a different resolution or patch size are needed to support the generalization claim for the multi-range design.
minor comments (4)
  1. [Figures 3, 5, 15 and Table IV] There are several typographical errors: "MSCvonStar" in Fig. 3 should be "MSConvStar", "MSC ONV STAR" in the Table IV heading should be "MSConvStar", and "efective" in the Fig. 15 caption should be "effective".
  2. [§IV-A] The text says "the the number of RMAG, MAB and channel increase" with a duplicated article; also "Flicker2K" should read "Flickr2K".
  3. [Table VII caption] The caption says Multi-Adds are calculated for "upscaling one image to 2562 resolution", which appears to be a typo; the intended resolution should be stated consistently with Table VIII, where 1280x720 is used.
  4. [Tables VI-VII, References] Some method entries in the tables have compound venue labels (e.g., CRAFT as "ICCV23&PAMI25" and SRFormer cited with both ICCV and journal years); the reference list should identify the primary venue unambiguously.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MAT's performance and efficiency claims are empirical benchmark comparisons and direct measurements, not derivations from fitted targets or self-cited premises.

full rationale

The paper is an empirical architecture study. Its central claims are PSNR/SSIM values on standard benchmarks obtained by training fixed architectures with standard L1 loss and comparing against published baselines. No parameter is fitted to the target benchmark numbers, and the reported gains are not constructed from those numbers. The design choices, such as range sizes 7x7/9x9/11x11 and dilation rates set to floor division of the 64x64 patch by the range sizes, are motivated by ablations on Urban100 x2 (Tables I-V), which is a model-selection procedure rather than a circular derivation of the final SOTA claim. Equation (9) defines the sparse sampling pattern, and its effectiveness is tested empirically against alternatives in Table III; it is not a renaming of a known result. The efficiency claim of ~3.3x faster than SRFormer-light is a direct runtime measurement (Fig. 16). Self-citations [2] and [15] appear only in related work as examples of CNN-based SR approaches and are not load-bearing; no uniqueness theorem is invoked to forbid alternatives. The apparent conflict in Table VII, where MambaIR reaches 33.32 dB while MAT+ reaches 32.49 dB on Manga109 x4, is a correctness or comparison issue, not a circularity issue, because the claim does not reduce to its inputs by definition. Overall, the derivation chain is self-contained and externally benchmarked.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical, not theoretical. It rests on a set of tuned architectural hyperparameters (range sizes, dilation rates, head allocation, kernel sizes, network depth) and on the domain assumptions that images are self-similar and that sparse dilated sampling captures useful global structure. There are no invented physical entities; the newly introduced modules MA, SMA and MSConvStar are network components, not independent evidential entities.

free parameters (5)
  • MA and SMA range sizes = 7x7, 9x9, 11x11 (lightweight); 13x13, 15x15, 17x17 (classical)
    Chosen by hand and validated through ablations; directly controls the attention receptive field and the number of key-value pairs.
  • SMA dilation rates = 5, 7, 9 (lightweight), derived as floor(64 / range)
    The 'Maximum' setting is selected because Table II shows it performs best; setting it higher with padding degrades performance substantially, so this is a tuned design choice.
  • Attention head allocation per range = 2 heads per range, 6 heads total
    Work allocation among the three ranges is arbitrary; no analysis is provided for why two heads per range is optimal.
  • MSConvStar depth-wise kernel sizes = 1x1, 3x3, 5x5, 7x7
    Multi-scale kernel sizes chosen by hand for feature diversity; ablation only compares including or excluding multi-scale, not the specific kernel set.
  • Network depth and width = RMAG/MAB/channel = 4/2/60 (lightweight), 6/3/156 (classical)
    Model capacity hyperparameters set following prior work; not derived from first principles.
assumptions (4)
  • domain assumption Natural images contain hierarchical features at multiple spatial scales and exhibit self-similarity/redundancy.
    Motivates MA and SMA; supported by cited prior work [3]-[5], [25], but assumed to translate directly to SR performance.
  • domain assumption Sparse sampled global attention at stride delta retains the essential non-local dependencies.
    Used in constructing SMA, Eq. (8)-(9); no proof is given that sampling one pixel per delta-th position preserves enough useful information.
  • standard math Scaled dot-product attention with softmax and learnable relative position bias provides effective feature aggregation.
    Equation (5) is the standard transformer attention mechanism, treated as a building block.
  • domain assumption Training on DIV2K/DF2K with L1 loss transfers to the five benchmark test sets.
    Standard SR protocol inherited from SwinIR [21] and SRFormer [23]; no explicit validation of cross-dataset generalization beyond test metrics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MAT: Multi-Range Attention Transformer for Efficient Image Super-Resolution." pith.science (2026). https://pith.science/paper/O3B2QII6

@misc{pith2026241117214,
  author       = {Pith},
  title        = {Pith review of: MAT: Multi-Range Attention Transformer for Efficient Image Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O3B2QII6}},
  note         = {Machine review of arXiv:2411.17214}
}
read the original abstract

Image super-resolution (SR) has significantly advanced through the adoption of Transformer architectures. However, conventional techniques aimed at enlarging the self-attention window to capture broader contexts come with inherent drawbacks, especially the significantly increased computational demands. Moreover, the feature perception within a fixed-size window of existing models restricts the effective receptive field (ERF) and the intermediate feature diversity. We demonstrate that a flexible integration of attention across diverse spatial extents can yield significant performance enhancements. In line with this insight, we introduce Multi-Range Attention Transformer (MAT) for SR tasks. MAT leverages the computational advantages inherent in dilation operation, in conjunction with self-attention mechanism, to facilitate both multi-range attention (MA) and sparse multi-range attention (SMA), enabling efficient capture of both regional and sparse global features. Combined with local feature extraction, MAT adeptly capture dependencies across various spatial ranges, improving the diversity and efficacy of its feature representations. We also introduce the MSConvStar module, which augments the model's ability for multi-range representation learning. Comprehensive experiments show that our MAT exhibits superior performance to existing state-of-the-art SR models with remarkable efficiency (~3.3 faster than SRFormer-light).

Figures

Figures reproduced from arXiv: 2411.17214 by the authors.

Figure 1
Figure 1. Comparison of trade-offs between model performance and overheads [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. (a) Illustration of image redundancy in natural images with self [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The overall architecture of Multi-Range Attention Transformer (MAT). B. Transformer-based Image SR Vision Transformers have demonstrated remarkable success across diverse visual tasks [34]–[43]. Their superior ability to capture long-range dependencies and extract features from extensive regions makes them particularly effective for SR tasks. SwinIR [21] successfully adapts Swin-Transformer [35] architecture for ima… view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Illustration of the window self-attention (WSA), multi-range attention (MA) and sparse multi-range attention (SMA). MA and SMA set different [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Illustration of FFN [21], ConvFFN [23], ConvStar and MSConvStar. [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: LAM [20] results of SwinIR [21], SRFormer [23] and our MAT with various dilation rates. The experiments is conducted using lightweight models. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: PSNR (dB) comparison of different attention mechanisms on Ur [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Visual comparison of features and reconstruction results on [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 9
Figure 9. Figure 9: Visual comparison on ×2 lightweight image SR. The patches for comparison are marked with red boxes in the original images. TABLE IV ABLATION STUDY ON MSCONVSTAR. THE SCHEMATIC DIAGRAM OF FOUR NETWORK STRUCTURAL COMBINATIONS IS SHOWN IN FIG. 5 Conv Star Multi-Scale Para…
Figure 11
Figure 11. Figure 11: Visualization of feature maps for multi-range dependencies. LAB and MA emphasize more on low-frequency structural information, while SMA focuses more on high-frequency edge details. while MA or LAB removal yields more modest decrements. This pattern suggests SMA’s pri…
Figure 10
Figure 10. Figure 10: LAM [20] results under different range dependency configurations. larger-range features showing particularly significant impact. We further analyze component contributions through LAM visualization across four model configurations ( [PITH_FULL_IMAGE:figures/full_fig_…
Figure 12
Figure 12. Figure 12: Visual comparison on ×4 lightweight image SR. The patches for comparison are marked with red boxes in the original images. Quantitative Comparison. Table VI presents a quantitative comparison of lightweight image SR models. We report both model parameters and Multi-Ad…
Figure 13
Figure 13. Figure 13: Visual comparison on ×4 classical image SR. The patches for comparison are marked with red boxes in the original images. TABLE VIII THE MODEL DEPTH AND PERFORMANCE COMPARISONS ON ×4 SR. ’MULTI-ADDS’, ’RUNNING TIME’ AND ’MEMORY’ IS CALCULATED UNDER THE SETTING OF UPSCA…
Figure 14
Figure 14. Figure 14: Comparison of LAM results of EDSR-baseline [10], SwinIR-light [21], SRFormer-light [23], and the proposed MAT-light. [PITH_FULL_IMAGE:figures/full_fig_p010_14.png]
Figure 15
Figure 15. Figure 15: The efective receptive field (ERF) [24] visualization and comparison [PITH_FULL_IMAGE:figures/full_fig_p011_15.png]
Figure 16
Figure 16. Figure 16: Running time and memory comparisons on ×4 SR. F. Analyses of LAM and ERF We use local attribution maps (LAM) [20] to visualize the spatial range of information used in target area reconstruction. As shown in [PITH_FULL_IMAGE:figures/full_fig_p011_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

65 extracted references · 63 canonical work pages

  1. [1]

    Reconstructing an image from its local descriptors,

    P. Weinzaepfel, H. J ´egou, and P. P ´erez, “Reconstructing an image from its local descriptors,” in CVPR. IEEE, 2011, pp. 337–344. 1

    work page 2011

  2. [2]

    Boosting single image super-resolution via partial channel shifting,

    X. Zhang, T. Li, and X. Zhao, “Boosting single image super-resolution via partial channel shifting,” in ICCV, 2023, pp. 13 223–13 232. 1, 2

    work page 2023

  3. [3]

    Matching local self-similarities across images and videos,

    E. Shechtman and M. Irani, “Matching local self-similarities across images and videos,” in CVPR, 2007, pp. 1–8. 1

    work page 2007

  4. [4]

    Super-resolution from a single image,

    D. Glasner, S. Bagon, and M. Irani, “Super-resolution from a single image,” in ICCV, 2009, pp. 349–356. 1

    work page 2009

  5. [5]

    Image super-resolution with self- similarity prior guided network and sample-discriminating learning,

    Y . Hu, J. Li, Y . Huang, and X. Gao, “Image super-resolution with self- similarity prior guided network and sample-discriminating learning,” TCSVT, vol. 32, no. 4, pp. 1966–1985, Apr. 2022. 1

    work page 1966

  6. [6]

    Single image super-resolution from transformed self-exemplars,

    J.-B. Huang, A. Singh, and N. Ahuja, “Single image super-resolution from transformed self-exemplars,” in CVPR, 2015, pp. 5197–5206. 1, 5

    work page 2015

  7. [7]

    Image super-resolution using deep convolutional networks,

    C. Dong, C. C. Loy, K. He, and X. Tang, “Image super-resolution using deep convolutional networks,” IEEE TPAMI, vol. 38, no. 2, pp. 295–307,

    work page

  8. [8]

    Accurate image super-resolution using very deep convolutional networks,

    J. Kim, J. K. Lee, and K. M. Lee, “Accurate image super-resolution using very deep convolutional networks,” in CVPR, 2016, pp. 1646–1654. 1, 2

    work page 2016

Show all 65 references
  1. [9]

    Photo-realistic single im- age super-resolution using a generative adversarial network,

    C. Ledig, L. Theis, F. Husz ´ar, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang et al., “Photo-realistic single im- age super-resolution using a generative adversarial network,” in CVPR, 2017, pp. 4681–4690. 1

    2017

  2. [10]

    Enhanced deep resid- ual networks for single image super-resolution,

    B. Lim, S. Son, H. Kim, S. Nah, and K. Mu Lee, “Enhanced deep resid- ual networks for single image super-resolution,” in CVPR Workshops, 2017, pp. 136–144. 1, 2, 5, 7, 8, 9, 10, 11

    2017

  3. [11]

    Residual dense network for image super-resolution,

    Y . Zhang, Y . Tian, Y . Kong, B. Zhong, and Y . Fu, “Residual dense network for image super-resolution,” in CVPR, 2018, pp. 2472–2481. 2

    2018

  4. [12]

    Multi-grained attention networks for single image super-resolution,

    H. Wu, Z. Zou, J. Gui, W.-J. Zeng, J. Ye, J. Zhang, H. Liu, and Z. Wei, “Multi-grained attention networks for single image super-resolution,” IEEE TCSVT, vol. 31, no. 2, pp. 512–522, 2021. 2

    2021

  5. [13]

    Lightweight image super- resolution with information multi-distillation network,

    Z. Hui, X. Gao, Y . Yang, and X. Wang, “Lightweight image super- resolution with information multi-distillation network,” in ACM MM , 2019, pp. 2024–2032. 2, 3, 7, 8

    2019

  6. [14]

    Blueprint separable residual network for efficient image super-resolution,

    Z. Li, Y . Liu, X. Chen, H. Cai, J. Gu, Y . Qiao, and C. Dong, “Blueprint separable residual network for efficient image super-resolution,” in CVPR Workshops, 2022, pp. 833–843. 2

    2022

  7. [15]

    Large kernel distillation network for efficient single image super-resolution,

    C. Xie, X. Zhang, L. Li, H. Meng, T. Zhang, T. Li, and X. Zhao, “Large kernel distillation network for efficient single image super-resolution,” in CVPR Workshops, 2023, pp. 1283–1292. 2

    2023

  8. [16]

    Image super- resolution using very deep residual channel attention networks,

    Y . Zhang, K. Li, K. Li, L. Wang, B. Zhong, and Y . Fu, “Image super- resolution using very deep residual channel attention networks,” in ECCV, 2018, pp. 286–301. 2, 3, 5, 9

    2018

  9. [17]

    Residual feature aggregation network for image super-resolution,

    J. Liu, W. Zhang, Y . Tang, J. Tang, and G. Wu, “Residual feature aggregation network for image super-resolution,” in CVPR, 2020, pp. 2359–2368. 2

    2020

  10. [18]

    Image super-resolution with non-local sparse attention,

    Y . Mei, Y . Fan, and Y . Zhou, “Image super-resolution with non-local sparse attention,” in CVPR, 2021, pp. 3517–3526. 2, 4, 9

    2021

  11. [19]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” NeurIPS, vol. 30, 2017. 2, 4, 5

    2017

  12. [20]

    Interpreting super-resolution networks with local attribution maps,

    J. Gu and C. Dong, “Interpreting super-resolution networks with local attribution maps,” in CVPR, 2021, pp. 9199–9208. 2, 6, 7, 11

    2021

  13. [21]

    Swinir: Image restoration using swin transformer,

    J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, “Swinir: Image restoration using swin transformer,” in ICCV Workshops, 2021, pp. 1833–1844. 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY 12

    2021

  14. [22]

    Activating more pixels in image super-resolution transformer,

    X. Chen, X. Wang, J. Zhou, Y . Qiao, and C. Dong, “Activating more pixels in image super-resolution transformer,” in CVPR, 2023, pp. 22 367–22 377. 2, 3, 9

    2023

  15. [23]

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

    Y . Zhou, Z. Li, C.-L. Guo, S. Bai, M.-M. Cheng, and Q. Hou, “Srformer: Permuted self-attention for single image super-resolution,” in ICCV, 2023, pp. 12 780–12 791. 2, 3, 4, 5, 6, 7, 8, 9, 10, 11

    2023

  16. [24]

    Scaling up your kernels to 31x31: Revisiting large kernel design in cnns,

    X. Ding, X. Zhang, J. Han, and G. Ding, “Scaling up your kernels to 31x31: Revisiting large kernel design in cnns,” in CVPR, 2022, pp. 11 963–11 975. 2, 11

    2022

  17. [25]

    Efficient and explicit modelling of image hierarchies for image restoration,

    Y . Li, Y . Fan, X. Xiang, D. Demandolx, R. Ranjan, R. Timofte, and L. Van Gool, “Efficient and explicit modelling of image hierarchies for image restoration,” in CVPR, 2023, pp. 18 278–18 289. 2, 3

    2023

  18. [26]

    Rethinking atrous convolution for semantic image segmentation,

    L.-C. Chen, G. Papandreou, F. Schroff, and H. Adam, “Rethinking atrous convolution for semantic image segmentation,” arXiv preprint arXiv:1706.05587, 2017. 2

    2017 arXiv

  19. [27]

    Understanding convolution for semantic segmentation,

    P. Wang, P. Chen, Y . Yuan, D. Liu, Z. Huang, X. Hou, and G. Cottrell, “Understanding convolution for semantic segmentation,” in WACV, 2018, pp. 1451–1460. 2

    2018

  20. [28]

    Rewrite the stars,

    X. Ma, X. Dai, Y . Bai, Y . Wang, and Y . Fu, “Rewrite the stars,” in CVPR, Jun. 2024, pp. 5694–5703. 2, 5

    2024

  21. [29]

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

    E. Agustsson and R. Timofte, “Ntire 2017 challenge on single image super-resolution: Dataset and study,” in CVPR Workshops , 2017, pp. 126–135. 2, 5

    2017

  22. [30]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in ICML, vol. 37, Jul. 2015, pp. 448–456. 2

    2015

  23. [31]

    Cross-scale internal graph neural network for image super-resolution,

    S. Zhou, J. Zhang, W. Zuo, and C. C. Loy, “Cross-scale internal graph neural network for image super-resolution,” NeurIPS, vol. 33, pp. 3499– 3509, 2020. 2, 9

    2020

  24. [32]

    Feedback pyramid attention networks for single image super-resolution,

    H. Wu, J. Gui, J. Zhang, J. T. Kwok, and Z. Wei, “Feedback pyramid attention networks for single image super-resolution,” IEEE TCSVT , vol. 33, no. 9, pp. 4881–4892, Sep. 2023. 2

    2023

  25. [33]

    Non-local neural net- works,

    X. Wang, R. Girshick, A. Gupta, and K. He, “Non-local neural net- works,” in CVPR, 2018, pp. 7794–7803. 2

    2018

  26. [34]

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

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” in ICLR, 2021. 3, 4

    2021

  27. [35]

    Swin transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin transformer: Hierarchical vision transformer using shifted windows,” in ICCV, 2021, pp. 10 012–10 022. 3, 4

    2021

  28. [36]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,” NeurIPS, vol. 34, pp. 12 077–12 090, 2021. 3

    2021

  29. [37]

    End-to-end object detection with transformers,

    N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, “End-to-end object detection with transformers,” in ECCV, 2020, pp. 213–229. 3

    2020

  30. [38]

    Restormer: Efficient transformer for high-resolution image restoration,

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M.-H. Yang, “Restormer: Efficient transformer for high-resolution image restoration,” in CVPR, 2022, pp. 5728–5739. 3, 5

    2022

  31. [39]

    Egocentric early action prediction via multimodal transformer-based dual action prediction,

    W. Guan, X. Song, K. Wang, H. Wen, H. Ni, Y . Wang, and X. Chang, “Egocentric early action prediction via multimodal transformer-based dual action prediction,” IEEE TCSVT , vol. 33, no. 9, pp. 4472–4483,

  32. [40]

    Msvt: Multiple spatiotemporal views transformer for deepfake video detection,

    Y . Yu, R. Ni, Y . Zhao, S. Yang, F. Xia, N. Jiang, and G. Zhao, “Msvt: Multiple spatiotemporal views transformer for deepfake video detection,” IEEE TCSVT, vol. 33, no. 9, pp. 4462–4471, Sep. 2023. 3

    2023

  33. [41]

    Lightweight image super-resolution with pyramid clustering transformer,

    M. Li, B. Ma, and Y . Zhang, “Lightweight image super-resolution with pyramid clustering transformer,” IEEE TCSVT, 2023. 3

    2023

  34. [42]

    Hybrid attention- based u-shaped network for remote sensing image super-resolution,

    J. Wang, B. Wang, X. Wang, Y . Zhao, and T. Long, “Hybrid attention- based u-shaped network for remote sensing image super-resolution,” IEEE TGRS, vol. 61, pp. 1–15, 2023. 3

    2023

  35. [43]

    Cross-spatial pixel integration and cross-stage feature fusion- based transformer network for remote sensing image super-resolution,

    Y . Lu, L. Min, B. Wang, L. Zheng, X. Wang, Y . Zhao, L. Yang, and T. Long, “Cross-spatial pixel integration and cross-stage feature fusion- based transformer network for remote sensing image super-resolution,” IEEE TGRS, vol. 61, pp. 1–16, 2023. 3

    2023

  36. [44]

    Omni aggregation networks for lightweight image super-resolution,

    H. Wang, X. Chen, B. Ni, Y . Liu, and J. Liu, “Omni aggregation networks for lightweight image super-resolution,” in CVPR, 2023, pp. 22 378–22 387. 3, 7, 8

    2023

  37. [45]

    Accurate image restoration with attention retractable transformer,

    J. Zhang, Y . Zhang, J. Gu, Y . Zhang, L. Kong, and X. Yuan, “Accurate image restoration with attention retractable transformer,” in ICLR, 2023. 3, 6, 9

    2023

  38. [46]

    Transformer-based dual-branch multiscale fusion network for pan-sharpening remote sensing images,

    Z. Li, J. Li, L. Ren, and Z. Chen, “Transformer-based dual-branch multiscale fusion network for pan-sharpening remote sensing images,” IEEE JSTARS, vol. 17, pp. 614–632, 2023. 3

    2023

  39. [47]

    Slide-transformer: Hierarchical vision transformer with local self-attention,

    X. Pan, T. Ye, Z. Xia, S. Song, and G. Huang, “Slide-transformer: Hierarchical vision transformer with local self-attention,” in CVPR, 2023, pp. 2082–2091. 3

    2023

  40. [48]

    Dilateformer: Multi-scale dilated transformer for visual recognition,

    J. Jiao, Y .-M. Tang, K.-Y . Lin, Y . Gao, A. J. Ma, Y . Wang, and W.-S. Zheng, “Dilateformer: Multi-scale dilated transformer for visual recognition,” TMM, vol. 25, pp. 8906–8919, 2023. 3

    2023

  41. [49]

    Real-time single image and video super- resolution using an efficient sub-pixel convolutional neural network,

    W. Shi, J. Caballero, F. Husz ´ar, J. Totz, A. P. Aitken, R. Bishop, D. Rueckert, and Z. Wang, “Real-time single image and video super- resolution using an efficient sub-pixel convolutional neural network,” in CVPR, 2016, pp. 1874–1883. 3

    2016

  42. [50]

    Efficient non- local contrastive attention for image super-resolution,

    B. Xia, Y . Hang, Y . Tian, W. Yang, Q. Liao, and J. Zhou, “Efficient non- local contrastive attention for image super-resolution,” in AAAI, vol. 36, 2022, pp. 2759–2767. 4

    2022

  43. [51]

    Stand-alone self-attention in vision models,

    P. Ramachandran, N. Parmar, A. Vaswani, I. Bello, A. Levskaya, and J. Shlens, “Stand-alone self-attention in vision models,” NeurIPS, vol. 32, 2019. 4

    2019

  44. [52]

    Neighborhood attention transformer,

    A. Hassani, S. Walton, J. Li, S. Li, and H. Shi, “Neighborhood attention transformer,” in CVPR, 2023, pp. 6185–6194. 4

    2023

  45. [53]

    Scaling local self-attention for parameter efficient visual backbones,

    A. Vaswani, P. Ramachandran, A. Srinivas, N. Parmar, B. Hechtman, and J. Shlens, “Scaling local self-attention for parameter efficient visual backbones,” in CVPR, 2021, pp. 12 894–12 904. 4

    2021

  46. [54]

    Exploring frequency-inspired optimization in transformer for efficient single image super-resolution,

    A. Li, L. Zhang, Y . Liu, and C. Zhu, “Exploring frequency-inspired optimization in transformer for efficient single image super-resolution,” IEEE TPAMI, pp. 1–18, 2025. 5, 7, 8

    2025

  47. [55]

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

    M. Bevilacqua, A. Roumy, C. Guillemot, and M. L. Alberi-Morel, “Low- complexity single-image super-resolution based on nonnegative neighbor embedding,” in BMVC, 2012. 5

    2012

  48. [56]

    On single image scale-up using sparse-representations,

    R. Zeyde, M. Elad, and M. Protter, “On single image scale-up using sparse-representations,” in Curves and Surfaces , 2012, pp. 711–730. 5

    2012

  49. [57]

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

    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 ecological statistics,” in ICCV, vol. 2, 2001, pp. 416–423. 5

    2001

  50. [58]

    Sketch-based manga retrieval using manga109 dataset,

    Y . Matsui, K. Ito, Y . Aramaki, A. Fujimoto, T. Ogawa, T. Yamasaki, and K. Aizawa, “Sketch-based manga retrieval using manga109 dataset,” Multimed Tools Appl, vol. 76, pp. 21 811–21 838, 2017. 5

    2017

  51. [59]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014. 5

    2014 arXiv

  52. [60]

    Latticenet: Towards lightweight image super-resolution with lattice block,

    X. Luo, Y . Xie, Y . Zhang, Y . Qu, C. Li, and Y . Fu, “Latticenet: Towards lightweight image super-resolution with lattice block,” in ECCV, 2020, pp. 272–289. 7, 8

    2020

  53. [61]

    N-gram in swin transformers for efficient lightweight image super-resolution,

    H. Choi, J. Lee, and J. Yang, “N-gram in swin transformers for efficient lightweight image super-resolution,” in CVPR, 2023, pp. 2071–2081. 7, 8

    2023

  54. [62]

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

    H. Guo, J. Li, T. Dai, Z. Ouyang, X. Ren, and S.-T. Xia, “Mambair: A simple baseline for image restoration with state-space model,” in ECCV,

  55. [63]

    Srconvnet: A transformer-style convnet for lightweight image super-resolution,

    F. Li, R. Cong, J. Wu, H. Bai, M. Wang, and Y . Zhao, “Srconvnet: A transformer-style convnet for lightweight image super-resolution,” IJCV, vol. 133, no. 1, pp. 173–189, 2025. 7, 8

    2025

  56. [64]

    Dual aggregation transformer for image super-resolution,

    Z. Chen, Y . Zhang, J. Gu, L. Kong, X. Yang, and F. Yu, “Dual aggregation transformer for image super-resolution,” in ICCV, 2023, pp. 12 312–12 321. 9

    2023

  57. [65]

    Image super- resolution via efficient transformer embedding frequency decomposition with restart,

    Y . Zuo, W. Yao, Y . Hu, Y . Fang, W. Liu, and Y . Peng, “Image super- resolution via efficient transformer embedding frequency decomposition with restart,” IEEE TIP, vol. 33, pp. 4670–4685, 2024. 9

    2024

Pith tools

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