Pith. sign in

REVIEW 5 major objections 4 minor 50 references

WaveHiT-SR: Hierarchical Wavelet Network for Efficient Image Super-Resolution

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read A wavelet-based attention mechanism inside a hierarchical transformer is claimed to make window self-attention scale linearly with window size, and the authors report it improves three SR baselines while cutting parameters, FLOPs, and runti

desk verdict Promising wavelet-hierarchical SR architecture undermined by an internally contradictory complexity claim. read the letter →

arxiv 2508.19927 v1 pith:5PD5DPIW submitted 2025-08-27 cs.CV cs.AI

classification cs.CVcs.AI
keywords imagesuper-resolutionwavelettransformhierarchicaltransformerWaveAttentionwindowself-attentionefficientsingle-image
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

The paper claims that embedding a discrete wavelet transform inside window self-attention removes the quadratic cost that usually forces transformer super-resolution models to use small fixed windows. It introduces WaveAttention, which downsamples value features via DWT before computing attention, making complexity linear in window size and enabling windows such as 64×64. Reframing three existing transformer SR models with this mechanism, plus expanding hierarchical windows and channel attention, is reported to yield higher PSNR/SSIM on standard benchmarks with fewer parameters, lower FLOPs, and faster inference. If the claim holds, it offers a general drop-in way to make transformer image restoration both cheaper and more accurate.

What carries the argument

WaveAttention: a replacement for spatial self-attention in which value features are passed through a discrete wavelet transform (DWT), yielding low- and high-frequency subbands and halving spatial resolution, so the attention correlation is computed against quarter-resolution values. Combined with hierarchical windows whose size expands across transformer blocks and a channel-attention branch, it forms the Hybrid Attention Block that carries the efficiency and quality gains.

What would settle it

A controlled experiment: train the same hierarchical transformer with identical windows and FLOPs but replace the DWT value-downsampling in Eq. (4) with average pooling or a strided 3×3 convolution. If the DWT variant does not beat both on Urban100 and Manga109 PSNR/SSIM across ×2, ×3, and ×4, the wavelet-specific claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a wavelet transform can be embedded inside window self-attention so that transformer super-resolution no longer has to choose between small windows and quadratic cost. Concretely, WaveHiT-SR replaces spatial self-attention with WaveAttention: value features are passed through a discrete wavelet transform, decomposed into four subbands, and spatially downsampled by half before the attention correlation is computed, cutting the multiply-add count for attention by roughly a factor of four and making complexity linear in window size. On top of this, hierarchical windows that expand across transformer blocks give the network multi-scale receptive fields. The authors repo

Load-bearing premise

The efficiency claim rests on the assumption that applying a discrete wavelet transform to the value features and keeping only the quarter-resolution approximation preserves whatever information the attention needs, and the paper does not include the ablation that would test this directly.

Editorial extensions

If this is right

  • Window sizes such as 64×64 become practical for transformer SR, letting a single model combine local and global context without the quadratic penalty that forces small windows.
  • The same recipe — DWT-downsampled values plus expanding windows — transfers across at least three different attention designs, so it behaves as a general efficiency module rather than a fix tied to one architecture.
  • Lower FLOPs and parameter counts at equal or better PSNR imply reduced memory and faster inference, which matters for deploying SR on edge devices or in video pipelines.
  • The reported 6× inference speedup for WaveHiT-SRF over its base model suggests the main bottleneck in the baseline is the attention correlation itself rather than parameter count.

Reading between the lines

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

  • The paper defers the ablation of the DWT-downsampling choice to a supplementary file; a direct comparison against average pooling or strided convolution at matched complexity would isolate whether DWT specifically, rather than any downsampling, drives the gains.
  • Because DWT is invertible and the low-frequency subband carries most of the energy, a natural next step is multi-level DWT (downsampling by 4 or 8) to push attention complexity even lower, though texture quality at those depths is untested.
  • The same wavelet substitution could benefit other Swin-style restoration tasks such as denoising or deblurring, and other permutation-based attention designs in video or medical imaging SR, by replacing their value-downsampling path.
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

5 major / 4 minor

Summary. The paper proposes WaveHiT-SR, an image super-resolution architecture that embeds discrete wavelet transforms into hierarchical transformer blocks. Three variants are built by adapting SwinIR-Light, SwinIR-NG, and SRFormer-Light to use a proposed WaveAttention/channel-correlation mechanism with hierarchical windows. The authors claim that this yields linear (or near-linear) complexity with respect to window size, enabling large 64×64 windows, and report improved PSNR/SSIM on Set5, Set14, B100, Urban100, and Manga109 at ×2, ×3, and ×4 with lower parameter counts and FLOPs. The main text refers to an accompanying supplementary file and ablation study, but no supplementary material is included in the submission.

Significance. If the complexity and efficiency claims held, the idea of wavelet-domain downsampled values inside hierarchical window attention could be a useful contribution to lightweight SR. The paper has a concrete strength: it evaluates the proposed modifications against three recent transformer-based lightweight baselines on standard benchmarks and reports parameter/FLOP reductions. However, the central derivation is internally inconsistent, and the empirical claims are not independently verifiable because no code, error bars, or promised ablation are supplied. The main contribution's mathematical basis is not sound as written, so the paper cannot be accepted in its current form.

major comments (5)
  1. [Eq. (5)/(6); Table 1] The central claim of linear complexity is contradicted by the paper's own equations. In Eq. (5), the left multiplicand Q_i is full-resolution, h_i w_i × C/2, and the right multiplicand W A(V^T_{i,↓}) is defined from the half-resolution wavelet downsampling of V^T_i, i.e., (h_i w_i/4) × C/2 (up to transpose). Forming the correlation Q_i · W A(V^T_{i,↓}) therefore costs O(C (h_i w_i)^2/4) multiply-adds per window, which remains quadratic in the window pixel count. Eq. (6)'s '2N·Ch·w/2·h/2' is ambiguous and, under the natural reading 2N C (h/2)(w/2), counts only a subset of the multiply-adds and does not follow from Eq. (5). Table 1 gives a third, conflicting complexity, O(h w log(h w)). Even if all frequency subbands were retained, the stated operation is quadratic; the abstract's 'linear complexity' and the 64×64-window claim are unsupported.
  2. [Eq. (1)] The wavelet subband labels are reversed. Eq. (1) lists {F_HH, F_HL, F_LH, F_LL}, and the text states that F_HH and F_HL are low-frequency components capturing global structure while F_LH and F_LL are high-frequency components. In standard DWT notation, F_LL is the low-low approximation and F_HH, F_HL, F_LH are the high-frequency detail subbands. Since the paper's frequency-based motivation is used to justify preserving 'low-frequency components' and 'high-frequency textures,' the text should be corrected or the design re-explained; as written, the frequency semantics are inverted.
  3. [Eq. (3) and Eq. (5)] The attention computation is under-specified. The text defines queries, keys, and values, but Eq. (3) splits only [Q,V] = DFE(X), and the next sentence refers to 'queries and keys' while providing Q_i, V_i with no K_i. Eq. (5) uses only Q_i and V^T_{i,↓}. If the design intentionally removes keys, that should be stated and justified; otherwise the missing K_i makes the module non-reproducible. This is a load-bearing omission because the claimed efficiency and the reported results depend on the exact operation being evaluated.
  4. [Note after Eq. (6); Experimental Settings] The main text explicitly promises supporting material that is absent: it states that 'in-depth information on Spatial and Channel Self-Correlation and Hierarchical Information Aggregation' is provided 'in the supplementary material along with Ablation Study,' and figure captions refer to a 'DWT: Supplementary file.' No supplementary file or ablation study is included in this submission. The missing ablation is precisely the experiment needed to test whether the DWT-downsampling in Eq. (4) preserves the information required by attention; without it, the efficiency/quality trade-off claimed in Table 2 is unsubstantiated.
  5. [Table 2] The empirical claims are single-run numbers without error bars or code. Several reported gains are large for lightweight SR—e.g., WaveHiT-SNG improves Set5 ×3 from 34.64 to 35.12 dB and B100 ×3 SSIM from .8090 to .8222 relative to SwinIR-NG. Without standard deviations, baseline evaluation protocol details, or released code, it is not possible to distinguish genuine gains from training or metric discrepancies. The authors should provide code, random-seed error bars, or at least a detailed comparison protocol.
minor comments (4)
  1. [Notation] The notation 'Ch' in Eq. (6) is ambiguous; use C·h or C_h explicitly. Also, γ=2 is introduced for DWT downsampling but never appears in the equations; either use it or remove it.
  2. [References] Several citations are imprecise: 'Alexey 2020' should cite Dosovitskiy et al.; the Adam optimizer is cited to Lai et al. 2017 instead of Kingma and Ba; and the RCAN/VDSR citations appear interchanged.
  3. [Figures and captions] Figure 1 uses both 'WaveHit' and 'WaveHiT'; unify the spelling. Figure 5's axes and labels are difficult to read and should be redrawn.
  4. [Experimental Settings, Table 2] For a given model, parameter counts differ across scales (e.g., WaveHiT-SNG: 1038K, 1043K, 1054K for ×2, ×3, ×4). The reason for these differences should be explained.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular reduction: SR results are empirical benchmarks; the linear-complexity claim is internally inconsistent (a correctness issue, not circularity).

full rationale

The paper's super-resolution results are empirical benchmark measurements, not derivations from a theory, so they do not reduce to their inputs by construction. The method builds on HiT-SR (Aslahishahri et al. 2024) and replaces pooling/self-correlation with DWT-based wavelet attention; this is an architectural transplant, transparently labeled as an adaptation, not a renaming. Self-citations (Dharejo et al. 2021/2022 for wavelet SR motivation; Liang et al. 2021 for SwinIR baselines; Agustsson & Timofte 2017 for DIV2K) are present but are not used to justify the paper's central claims or to forbid alternatives; they are background and baselines. I checked the central efficiency claim because it is the closest thing to a first-principles statement. The paper asserts in Eq. 6 that WA-SC has Mult-Add = 2NCh/2w/2, but by the paper's own Eq. 5, WA-SC(Q_i, V_down) multiplies a full-resolution Q_i ∈ R^{h_iw_i × C/2} by a half-resolution V_down ∈ R^{(h_i/2)(w_i/2) × C/2}, producing a correlation matrix of size h_iw_i × h_iw_i/4 and hence O(C(h_iw_i)^2/4) multiply-adds per window. Table 1's O(hw log hw) also conflicts with Eq. 6. This is an internal-consistency/correctness defect in the 'linear complexity' contribution, not circularity: the conclusion is not equivalent to the input by definition; it is simply not supported by the written equations. The note promising a supplementary derivation and ablation ('We have provided in-depth information on Spatial and Channel Self-Correlation and Hierarchical Information Aggregation in the supplementary material along with Ablation Study') is unfulfilled in the arXiv record, so that missing support should be weighed as a completeness problem, not as evidence of circular reasoning. No parameter is fitted and then renamed a prediction, and no external 'uniqueness theorem' from the authors is invoked. Therefore the circularity score is low.

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

The central claims rest on hand-chosen architecture hyperparameters and the unverified assumption that DWT downsampling preserves attention-relevant information. No new physical entities are introduced.

free parameters (5)
  • Channel number (C=60) = 60
    Set to match SwinIR-Light settings; chosen by hand, not derived.
  • Number of heads (Nh=6) = 6
    Chosen by hand, consistent with SwinIR-Light.
  • Base window size (8x8) = 8x8
    Adopted from SwinIR-Light; not justified by analysis.
  • Hierarchical window ratios = HiT-SR ratios
    Borrowed without modification from HiT-SR; these determine receptive field growth.
  • DWT downsampling factor gamma = 2
    Chosen to halve spatial dimensions, making the complexity claim linear; no ablation of other factors.
assumptions (4)
  • standard math Discrete Wavelet Transform preserves and separates information into low and high frequency subbands as assumed.
    The method relies on DWT properties, e.g., invertibility and frequency separation, without explicitly proving them for learned features.
  • domain assumption Training on DIV2K with bicubic degradation transfers to the five evaluation benchmarks.
    Standard practice in SR, but the paper does not test distribution shift.
  • ad hoc to paper Hierarchical window ratios from HiT-SR are appropriate for the three adapted baselines.
    The ratios are taken from HiT-SR without justification or ablation, yet they are central to the hierarchical design.
  • ad hoc to paper The DWT coefficients, after downsampling, retain enough information for the attention mechanism to produce high-quality SR.
    This is the core design assumption behind WaveAttention; the paper does not ablate it and the promised supplementary ablation is missing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WaveHiT-SR: Hierarchical Wavelet Network for Efficient Image Super-Resolution." pith.science (2026). https://pith.science/paper/5PD5DPIW

@misc{pith2026250819927,
  author       = {Pith},
  title        = {Pith review of: WaveHiT-SR: Hierarchical Wavelet Network for Efficient Image Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5PD5DPIW}},
  note         = {Machine review of arXiv:2508.19927}
}
read the original abstract

Transformers have demonstrated promising performance in computer vision tasks, including image super-resolution (SR). The quadratic computational complexity of window self-attention mechanisms in many transformer-based SR methods forces the use of small, fixed windows, limiting the receptive field. In this paper, we propose a new approach by embedding the wavelet transform within a hierarchical transformer framework, called (WaveHiT-SR). First, using adaptive hierarchical windows instead of static small windows allows to capture features across different levels and greatly improve the ability to model long-range dependencies. Secondly, the proposed model utilizes wavelet transforms to decompose images into multiple frequency subbands, allowing the network to focus on both global and local features while preserving structural details. By progressively reconstructing high-resolution images through hierarchical processing, the network reduces computational complexity without sacrificing performance. The multi-level decomposition strategy enables the network to capture fine-grained information in lowfrequency components while enhancing high-frequency textures. Through extensive experimentation, we confirm the effectiveness and efficiency of our WaveHiT-SR. Our refined versions of SwinIR-Light, SwinIR-NG, and SRFormer-Light deliver cutting-edge SR results, achieving higher efficiency with fewer parameters, lower FLOPs, and faster speeds.

Figures

Figures reproduced from arXiv: 2508.19927 by the authors.

Figure 1
Figure 1. Comparison of efficient SR transformers: SwinIR [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed SR framework, with specific block- and layer-level enhancements from our WaveHiT-SR [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The WaveHiT-SR framework’s layer-level design includes dual feature extraction (DFE) and the use of wave attention [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visual comparisons with state-of-the-art SISR methods on the Urban100 Dataset on Scale ( [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: To compare convergence rates, we evaluated [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 39 canonical work pages

  1. [1]

    Agustsson, E.; and Timofte, R. 2017. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, 126--135

  2. [2]

    Ahn, N.; Kang, B.; and Sohn, K.-A. 2018. Fast, accurate, and lightweight super-resolution with cascading residual network. In Proceedings of the European conference on computer vision (ECCV), 252--268

  3. [3]

    Alexey, D. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929

  4. [4]

    Aslahishahri, M.; Ubbens, J.; and Stavness, I. 2024. HiTSR: A Hierarchical Transformer for Reference-based Super-Resolution. arXiv preprint arXiv:2408.16959

  5. [5]

    Baek, S.; and Lee, C. 2020. Single image super-resolution using frequency-dependent convolutional neural networks. In 2020 IEEE International Conference on Industrial Technology (ICIT), 692--695. IEEE

  6. [6]

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

  7. [7]

    Bittner, M.; Hobeichi, S.; Zawish, M.; Diatta, S.; Ozioko, R.; Xu, S.; and Jantsch, A. 2023. An LSTM-based Downscaling Framework for Australian Precipitation Projections. In NeurIPS 2023 Workshop on Tackling Climate Change with Machine Learning

  8. [8]

    Cai, H.; Li, J.; Hu, M.; Gan, C.; and Han, S. 2023. Efficientvit: Lightweight multi-scale attention for high-resolution dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 17302--17313

Show all 50 references
  1. [9]

    Cao, J.; Li, Y.; Zhang, K.; and Van Gool, L. 2021. Video super-resolution transformer. arXiv preprint arXiv:2106.06847

  2. [10]

    Chen, X.; Wang, X.; Zhou, J.; Qiao, Y.; and Dong, C. 2023. Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 22367--22377

  3. [11]

    Choi, H.; Lee, J.; and Yang, J. 2023. N-gram in swin transformers for efficient lightweight image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2071--2081

  4. [12]

    A.; Deeba, F.; Zhou, Y.; Das, B.; Jatoi, M

    Dharejo, F. A.; Deeba, F.; Zhou, Y.; Das, B.; Jatoi, M. A.; Zawish, M.; Du, Y.; and Wang, X. 2021. TWIST-GAN: Towards wavelet transform and transferred GAN for spatio-temporal single image super resolution. ACM Transactions on Intelligent Systems and Technology (TIST), 12(6): 1--20

  5. [13]

    A.; Zawish, M.; Deeba, F.; Zhou, Y.; Dev, K.; Khowaja, S

    Dharejo, F. A.; Zawish, M.; Deeba, F.; Zhou, Y.; Dev, K.; Khowaja, S. A.; and Qureshi, N. M. F. 2022. Multimodal-boost: Multimodal medical image super-resolution using multi-attention network with wavelet transform. IEEE/ACM Transactions on Computational Biology and Bioinforma...

  6. [14]

    C.; He, K.; and Tang, X

    Dong, C.; Loy, C. C.; He, K.; and Tang, X. 2014. Learning a deep convolutional network for image super-resolution. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part IV 13, 184--199. Springer

  7. [15]

    Fan, H.; Xiong, B.; Mangalam, K.; Li, Y.; Yan, Z.; Malik, J.; and Feichtenhofer, C. 2021. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 6824--6835

  8. [16]

    Guo, T.; Seyed Mousavi, H.; Huu Vu, T.; and Monga, V. 2017. Deep wavelet prediction for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, 104--113

  9. [17]

    Huang, H.; He, R.; Sun, Z.; and Tan, T. 2017. Wavelet-srnet: A wavelet-based cnn for multi-scale face super resolution. In Proceedings of the IEEE international conference on computer vision, 1689--1697

  10. [18]

    Huang, J.-B.; Singh, A.; and Ahuja, N. 2015. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, 5197--5206

  11. [19]

    Hui, Z.; Gao, X.; Yang, Y.; and Wang, X. 2019. Lightweight image super-resolution with information multi-distillation network. In Proceedings of the 27th acm international conference on multimedia, 2024--2032

  12. [20]

    K.; and Lee, K

    Kim, J.; Lee, J. K.; and Lee, K. M. 2016. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 1646--1654

  13. [21]

    M.; and Dogan, Z

    Korkmaz, C.; Tekalp, A. M.; and Dogan, Z. 2024. Training Generative Image Super-Resolution Models by Wavelet-Domain Losses Enables Better Control of Artifacts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 5926--5936

  14. [22]

    Lai, W.-S.; Huang, J.-B.; Ahuja, N.; and Yang, M.-H. 2017. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, 624--632

  15. [23]

    Lee, A.; Tsekouras, K.; Calderon, C.; Bustamante, C.; and Press \'e , S. 2017. Unraveling the thousand word picture: an introduction to super-resolution data analysis. Chemical reviews, 117(11): 7276--7330

  16. [24]

    Li, K.; Wang, Y.; Zhang, J.; Gao, P.; Song, G.; Liu, Y.; Li, H.; and Qiao, Y. 2023. Uniformer: Unifying convolution and self-attention for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10): 12581--12600

  17. [25]

    Li, W.; Lu, X.; Qian, S.; Lu, J.; Zhang, X.; and Jia, J. 2021. On efficient transformer-based image pre-training for low-level vision. arXiv preprint arXiv:2112.10175

  18. [26]

    Liang, J.; Cao, J.; Sun, G.; Zhang, K.; Van Gool, L.; and Timofte, R. 2021. Swinir: Image restoration using swin transformer. In Proceedings of the IEEE/CVF international conference on computer vision, 1833--1844

  19. [27]

    Lim, B.; Son, S.; Kim, H.; Nah, S.; and Mu Lee, K. 2017. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, 136--144

  20. [28]

    Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, 10012--10022

  21. [29]

    Luo, X.; Xie, Y.; Zhang, Y.; Qu, Y.; Li, C.; and Fu, Y. 2020. Latticenet: Towards lightweight image super-resolution with lattice block. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XXII 16, 272--289. Springer

  22. [30]

    Martin, D.; Fowlkes, C.; Tal, D.; and Malik, J. 2001. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, ...

  23. [31]

    Matsui, Y.; Ito, K.; Aramaki, Y.; Fujimoto, A.; Ogawa, T.; Yamasaki, T.; and Aizawa, K. 2017. Sketch-based manga retrieval using manga109 dataset. Multimedia tools and applications, 76: 21811--21838

  24. [32]

    Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; et al. 2019. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32

  25. [33]

    Tai, Y.; Yang, J.; Liu, X.; and Xu, C. 2017. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, 4539--4547

  26. [34]

    N.; Kaiser, .; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems. Advances in neural information processing systems, 30(2017)

  27. [35]

    Wang, C.; Jiang, J.; Jiang, K.; and Liu, X. 2024. Structure Prior-Aware Dynamic Network for Face Super-Resolution. IEEE Transactions on Biometrics, Behavior, and Identity Science

  28. [36]

    Wang, H.; Chen, X.; Ni, B.; Liu, Y.; and Liu, J. 2023. Omni aggregation networks for lightweight image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 22378--22387

  29. [37]

    Wang, W.; Xie, E.; Li, X.; Fan, D.-P.; Song, K.; Liang, D.; Lu, T.; Luo, P.; and Shao, L. 2021. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, 568--578

  30. [38]

    Wang, W.; Xie, E.; Li, X.; Fan, D.-P.; Song, K.; Liang, D.; Lu, T.; Luo, P.; and Shao, L. 2022 a . Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3): 415--424

  31. [39]

    Wang, Z.; Cun, X.; Bao, J.; Zhou, W.; Liu, J.; and Li, H. 2022 b . Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 17683--17693

  32. [40]

    Yuan, Y.; Fu, R.; Huang, L.; Lin, W.; Zhang, C.; Chen, X.; and Wang, J. 2021. Hrformer: High-resolution vision transformer for dense predict. Advances in neural information processing systems, 34: 7281--7293

  33. [41]

    Zeyde, R.; Elad, M.; and Protter, M. 2012. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, 711--730. Springer

  34. [42]

    Zhang, D.; Huang, F.; Liu, S.; Wang, X.; and Jin, Z. 2022 a . Swinfir: Revisiting the swinir with fast fourier convolution and improved training for image super-resolution. arXiv preprint arXiv:2208.11247

  35. [43]

    Zhang, Y.; Ji, B.; Hao, J.; and Yao, A. 2022 b . Perception-distortion balanced ADMM optimization for single-image super-resolution. In European Conference on Computer Vision, 108--125. Springer

  36. [44]

    Zhang, Y.; Li, K.; Li, K.; Zhong, B.; and Fu, Y. 2019. Residual non-local attention networks for image restoration. arXiv preprint arXiv:1903.10082

  37. [45]

    Zhang, Y.; Tian, Y.; Kong, Y.; Zhong, B.; and Fu, Y. 2018. Residual dense network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2472--2481

  38. [46]

    H.; et al

    Zheng, S.; Lu, J.; Zhao, H.; Zhu, X.; Luo, Z.; Wang, Y.; Fu, Y.; Feng, J.; Xiang, T.; Torr, P. H.; et al. 2021. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern ...

  39. [47]

    Zhou, Y.; Li, Z.; Guo, C.-L.; Bai, S.; Cheng, M.-M.; and Hou, Q. 2023. Srformer: Permuted self-attention for single image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 12780--12791

  40. [48]

    Zhu, X.; Su, W.; Lu, L.; Li, B.; Wang, X.; and Dai, J. 2020. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159

  41. [49]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  42. [50]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.