Pith. sign in

REVIEW 6 minor 62 references

CUST restores high-resolution images by clustering similar patches across windows and guiding attention with multi-scale high-frequency residuals, cutting memory while matching or beating recent lightweight super-resolution models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review

2026-07-14 07:08 UTC pith:5MIFXWXS

load-bearing objection Solid empirical architecture paper that advances the lightweight SISR Pareto front on memory/latency/PSNR with two concrete modules and thorough ablations; the clustering heuristic is the main soft spot but is already acknowledged and not load-bearing.

arxiv 2607.11088 v1 pith:5MIFXWXS submitted 2026-07-13 cs.CV

CUST: Clustered Unit-level Similarity Transformer for Lightweight Image Super-Resolution

Jeongsoo Kim This is my paper
classification cs.CV
keywords image super-resolutionvision transformerwindow attentionlightweight SRpatch clusteringhigh-frequency residualmemory efficiency
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Vision Transformers deliver strong single-image super-resolution by modelling long-range dependencies, yet their quadratic cost forces most practical designs into fixed local windows that sacrifice that very advantage. CUST claims that a better trade-off is possible: group windows into larger search regions, average-pool each window into a representative token, then re-cluster every patch by its affinity to those tokens so that semantically similar patches become neighbours and can attend to one another (with a modest key/value expansion and mask to keep the clusters coherent). Immediately afterwards a multi-frequency residual—original features minus their own down-sampled/up-sampled version—explicitly highlights edges and textures, which are refined by dilated convolutions and a spatial gate before ordinary overlapping-window attention finishes the local cleanup. The resulting network runs with substantially lower peak GPU memory and competitive or better latency than recent global-context and lightweight baselines while posting higher average PSNR on standard benchmarks, especially at ×3 and ×4. The practical upshot is a hardware-aware transformer that keeps the global modelling power of ViTs without the memory and latency tax that has kept them out of resource-constrained devices.

Core claim

A lightweight transformer that first re-clusters patches by similarity to pooled window tokens inside broadened search regions (CANA) and then steers overlapping-window attention with explicit multi-scale high-frequency residuals (MEDA) simultaneously recovers long-range context and fine detail, yielding lower memory footprint, faster inference and equal or better restoration metrics than prior window-based and super-token SR models under realistic hardware constraints.

What carries the argument

Cross-window Affinity Neighbor Attention (CANA) plus Multi-frequency Error-driven Dense Attention (MEDA): patches are sorted and clustered by affinity to average-pooled window tokens so similar patches can attend across original window boundaries; high-frequency residuals extracted by down-up sampling then gate and refine the subsequent overlapping attention.

Load-bearing premise

That ranking every patch by its similarity to a few average-pooled window tokens inside a fixed search grid will form clusters coherent enough to capture true long-range structure without losing critical information at the artificial cluster boundaries, even on irregular or stochastic textures.

What would settle it

On a held-out set of highly irregular textures (or a controlled synthetic pattern with known long-range but non-repeating structure), measure whether CANA’s clusters still produce a measurably larger Local Attribution Map diffusion index and higher PSNR than a pure window baseline of identical parameter count; if the gap collapses or reverses, the clustering premise fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

0 major / 6 minor

Summary. The paper proposes CUST, a lightweight Vision Transformer for single-image super-resolution that combines two modules: Cross-window Affinity Neighbor Attention (CANA) and Multi-frequency Error-driven Dense Attention (MEDA). CANA partitions features into search regions, forms representative tokens by average-pooling windows (Eq. 1), clusters patches by affinity via argmax/argsort, expands Key/Value sets from adjacent clusters (Eq. 2), and applies masked attention with a learned gate (Eq. 3). MEDA extracts high-frequency residuals via down/up-sampling, refines them with dilated convolutions and a spatial gate (Eq. 4, Algorithm 1), then applies overlapping-window self-attention. Four model variants (Base/Base+/Small/Small+) are trained on DIV2K and evaluated on Set5/14, B100, Urban100, Manga109 (and Test2K in the appendix). The central claim is an empirical Pareto improvement: competitive or superior PSNR/SSIM (e.g., average +0.094 dB at ×4 vs. CATANet) together with substantially lower peak GPU memory (~83% reduction vs. CATANet) and competitive latency under fixed-resolution measurements on an RTX 3090.

Significance. If the reported numbers hold, CUST supplies a practical, hardware-aware advance for resource-constrained SISR. The work is valuable for emphasizing peak memory and measured latency rather than FLOPs alone, for providing extensive ablations that isolate CANA, MEDA and their sub-components (Tables 4–6, A5), and for releasing code. LAM visualizations, clustering maps, and residual-alignment analyses further support that the modules deliver broader context and high-frequency guidance on the evaluated data. The contribution is architectural and empirical rather than theoretical; its significance lies in the demonstrated efficiency–accuracy trade-off under realistic constraints.

minor comments (6)
  1. Abstract and introduction contain several missing spaces (“Itachievesalowermemoryfootprint”, “inputlow-resolution”, etc.). A careful proof-reading pass is needed.
  2. Section 3.2 / Eq. 3: the attention-mask definition (value 0 when window IDs match, -∞ otherwise) is clear in prose but would benefit from an explicit mathematical definition of M_ij for reproducibility.
  3. Figure 1 caption and axis labels are readable, yet the precise measurement protocol (batch size, precision, warm-up) for the plotted latency/memory points is only fully specified later; a short note in the caption would help.
  4. Tables 1–2 and A1: occasional inconsistency in model naming (HPI-Net vs. HPINet) and missing FLOPs for ESRT; standardizing these would improve readability.
  5. Section 5 already notes the limitation on irregular textures; a brief quantitative stress test (e.g., on a textured subset of Urban100 or a synthetic noise pattern) would strengthen the discussion, though it is not required for acceptance.
  6. Appendix C (Table A4) is useful for explaining the FLOPs–latency discrepancy; a one-sentence pointer in the main text (near Table 3) would make the hardware-efficiency argument more self-contained.

Circularity Check

0 steps flagged

No significant circularity; purely empirical architecture paper whose claims rest on measured PSNR/SSIM, latency and peak-memory numbers on held-out standard benchmarks.

full rationale

CUST is an engineering paper that defines two modules (CANA via affinity clustering of patches to average-pooled window tokens inside a search region, MEDA via residual high-frequency error maps feeding overlap-window attention) and then trains the resulting network on DIV2K. All quantitative claims (Tables 1-3, A1-A3, ablations 4-6) are direct measurements of reconstruction metrics and hardware counters on the conventional held-out sets Set5/14, B100, Urban100, Manga109 and Test2K. No equation equates a predicted quantity to a fitted parameter by construction; no uniqueness theorem or ansatz is imported from prior work by the same author; the single independent-researcher citation list contains no load-bearing self-reference. Hyper-parameters (channel counts, block counts, cyclic window schedules) are design choices, not quantities that redefine the evaluation metrics. The acknowledged limitation on irregular textures is orthogonal to circularity. Consequently the derivation chain is empty of circular steps and the score is zero.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

Method paper whose claims rest on standard deep-learning training assumptions plus a handful of architectural hyper-parameters and the unproved but empirically tested premise that affinity-to-pooled-window clustering yields useful long-range groups. No new physical entities or free constants fitted to the evaluation metrics themselves.

free parameters (5)
  • channel count / block count (Base: 40 ch, 12 blocks; Small: 30 ch, 8 blocks)
    Chosen by authors to define model capacity; directly affect all reported PSNR and efficiency numbers.
  • MEDA window-size schedule ([12,14,16,18] cyclic or fixed 18)
    Hand-selected; ablation Table 6 shows performance is sensitive to these values.
  • search-region scale factor gs = GS/ws (default 10)
    Controls clustering scope; reduced to 5 in ablation and hurts DI/PSNR.
  • learnable residual scaler α and gate parameters inside MEDA
    Trained end-to-end; their presence is required for the claimed high-frequency guidance.
  • learning-rate schedule and warm-up (5e-4, MultiStepLR milestones)
    Standard but specific choices that affect final checkpoint quality.
axioms (4)
  • ad hoc to paper Average-pooled window tokens are sufficiently representative of local window content for subsequent patch-to-token affinity clustering to recover long-range semantic groups.
    Core premise of CANA (§3.2, Eq. 1); not derived, only ablated.
  • domain assumption The residual X − Up(Down(X)) isolates the high-frequency components most useful for SR reconstruction.
    Standard multi-scale residual idea in SR; used as the explicit guide in MEDA (Alg. 1).
  • domain assumption Training on DIV2K with standard flips/rotations and evaluating PSNR/SSIM on the five classic benchmarks is a valid proxy for practical SR quality.
    Community convention followed in §4.1; not re-validated here.
  • domain assumption Peak GPU memory via torch.cuda.max_memory_allocated and average latency over 50 fixed-resolution images are faithful measures of real-world efficiency.
    Stated measurement protocol in §4.1; hardware-dependent but standard.
invented entities (2)
  • Cross-window Affinity Neighbor Attention (CANA) module no independent evidence
    purpose: Cluster patches by affinity to pooled window tokens inside a search region and perform masked attention with KV expansion from adjacent clusters.
    New architectural block introduced in §3.2; independent evidence is only the paper’s own ablations and visualizations.
  • Multi-frequency Error-driven Dense Attention (MEDA) module no independent evidence
    purpose: Extract high-frequency residual via down/up-sampling, refine it with dilated conv + gate, then apply overlapping self-attention.
    New architectural block introduced in §3.3 / Alg. 1; evidence limited to internal ablations.

reviewed 2026-07-14 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CUST: Clustered Unit-level Similarity Transformer for Lightweight Image Super-Resolution." pith.science (2026). https://pith.science/paper/5MIFXWXS

@misc{pith2026260711088,
  author       = {Pith},
  title        = {Pith review of: CUST: Clustered Unit-level Similarity Transformer for Lightweight Image Super-Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5MIFXWXS}},
  note         = {Machine review of arXiv:2607.11088}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recently, Vision Transformer (ViT)-based models have exhibited remarkable performance in image super-resolution. However, the quadratic computational complexity of ViTs with respect to spatial resolution severely constrains their efficiency, leading to high latency and massive memory consumption. To alleviate this, various window-based attention mechanisms have been proposed; yet, they inherently compromise the long-range dependency modeling that is the primary advantage of ViTs. To overcome these limitations, we propose the Clustered Unit-level Similarity Transformer (CUST), a novel architecture that efficiently integrates global and local information. Specifically, CUST enables each patch to aggregate and attend to similar patches within a broadened regional scope outside its local window, thereby capturing extensive contextual understanding. Furthermore, it employs overlapping attention windows to capture local dependencies, while explicitly extracting high-frequency details by computing the residual difference between the original features and their downsampled-upsampled counterparts. Comprehensive experiments demonstrate that our proposed model achieves a practical balance between computational efficiency and restoration performance. It achieves a lower memory footprint and faster inference speed compared to recent global context or lightweight models under realistic constraints. Code is available at [https://github.com/jwgdmkj/CUST].

Figures

Figures reproduced from arXiv: 2607.11088 by JeongSoo Kim.

Figure 1
Figure 1. Figure 1: Efficiency-Performance trade-off analysis on Urban100 (×4). been proposed [7, 20, 28, 58]. However, CNNs suffer from a fundamental limita￾tion regarding their restricted receptive fields, making it difficult to capture long￾range dependencies effectively. Although various approaches have attempted to overcome this issue by stacking deeper layers [26,31,59], this strategy often leads to increased computatio… view at source ↗
Figure 2
Figure 2. Figure 2: The architec￾ture of CUST. (a) The overall framework of CUST. (b) The struc￾ture of Cross-window Affinity Neighbor At￾tention (CANA) mod￾ule. (c) The structure of Multi-frequency Error￾driven Dense Attention (MEDA) module. inference speed. ELAN [57] utilized group-wise multi-scale attention. SPIN [52] adopted superpixel-based interactions. CATANet [34] efficiently captured broad￾range attention by generati… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the proposed Cross-window Affinity Neighbor Attention (CANA) mechanism. (a) Patches within a search region are clustered based on their similarity to pooled window tokens. (b) To prevent information disconnection at cluster boundaries, Key and Value sets are expanded by incorporating patches from adjacent clusters that share the same window token affinity. ILR. Detailed mechanisms of the CA… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of state-of-the-art SR models at Urban100 ×4 scale. Red boxes indicate the regions selected for detailed comparison. is increased to 12 to match the parameter count of CUST-Small (rows 4 and 5). Our experimental results show that simultaneously utilizing both modules yields a significant synergetic effect compared to using either one in isolation. Specifically, the CANA-only variant … view at source ↗
Figure 5
Figure 5. Figure 5: LAM (Local Attribution Maps) comparison between the proposed model and other ViT-based lightweight SR models. As illustrated, the proposed model utilizes a broader spatial range for image reconstruction compared to other models [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: LAM comparison between CUST-Small and its variants using only the CANA module (b) or only the MEDA module (c). The results demonstrate the synergistic effect achieved when CANA and MEDA are utilized together [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visual analysis using Local Attribution Map (LAM). (a) A query patch from Urban100. (b)-(d) LAM results for different configurations. Our baseline (gs=10) shows a significantly higher Diffusion Index (DI = 13.92) compared to the Swin-based coun￾terpart (DI = 8.11), indicating that our model effectively utilizes a broader range of spatial information for high-quality image reconstruction [PITH_FULL_IMAGE:f… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 3 linked inside Pith

  1. [1]

    arXiv preprint arXiv:1607.06450 (2016)

    Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016)

    Pith/arXiv arXiv 2016

  2. [2]

    In: British ma- chine vision conference (BMVC) (2012)

    Bevilacqua, M., Roumy, A., Guillemot, C., Morel, M.L.A.: Low-complexity single- image super-resolution based on nonnegative neighbor embedding. In: British ma- chine vision conference (BMVC) (2012)

    2012

  3. [3]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Chen,H.,Wang,Y.,Guo,T.,Xu,C.,Deng,Y.,Liu,Z.,Ma,S.,Xu,C.,Xu,C.,Gao, W.: Pre-trained image processing transformer. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 12299–12310 (2021)

    2021

  4. [4]

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

    Chen, X., Wang, X., Zhou, J., Qiao, Y., Dong, C.: Activating more pixels in image super-resolution transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 22367–22377 (June 2023)

    2023

  5. [5]

    Advances in Neural Information Processing Systems 35, 25478–25490 (2022)

    Chen, Z., Zhang, Y., Gu, J., Kong, L., Yuan, X., et al.: Cross aggregation trans- former for image restoration. Advances in Neural Information Processing Systems 35, 25478–25490 (2022)

    2022

  6. [6]

    Choi, H., Lee, J., Yang, J.: N-gram in swin transformers for efficient lightweight imagesuper-resolution.In:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition (CVPR). pp. 2071–2081 (2023)

    2071

  7. [7]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Dai, T., Cai, J., Zhang, Y., Xia, S.T., Zhang, L.: Second-order attention network for single image super-resolution. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11065–11074 (2019)

    2019

  8. [8]

    In: European conference on computer vision

    Dong, C., Loy, C.C., He, K., Tang, X.: Learning a deep convolutional network for image super-resolution. In: European conference on computer vision. pp. 184–199. Springer (2014)

    2014

  9. [9]

    In: European conference on computer vision

    Dong, C., Loy, C.C., Tang, X.: Accelerating the super-resolution convolutional neural network. In: European conference on computer vision. pp. 391–407. Springer (2016)

    2016

  10. [10]

    In Proceedings of the Computer Vision and Pattern Recognition (CVPR) pp

    Dong, X., Bao, J., Chen, D., Zhang, W., Yu, N., Yuan, L., Chen, D., Guo, B.: Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the Computer Vision and Pattern Recognition (CVPR) pp. 12114– 12124 (2021)

    2021

  11. [11]

    arXiv preprint arXiv:2010.11929 (2020)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)

    Pith/arXiv arXiv 2010

  12. [12]

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

    Du, Z., Liu, D., Liu, J., Tang, J., Wu, G., Fu, L.: Fast and memory-efficient net- work towards efficient image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 853–862 (2022)

    2022

  13. [13]

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

    Fang, J., Lin, H., Chen, X., Zeng, K.: A hybrid network of cnn and transformer for lightweight image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1103–1112 (2022)

    2022

  14. [14]

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

    Fang, W., Tang, Y., Guo, H., Yuan, M., Mok, T.C.W., Yan, K., Yao, J., Chen, X., Liu, Z., Lu, L., Zhang, L., Xu, M.: Cycleinr: Cycle implicit neural representation for arbitrary-scale volumetric super-resolution of medical data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11631–11641 (June 2024)

    2024

  15. [15]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Gu, J., Dong, C.: Interpreting super-resolution networks with local attribution maps. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 9199–9208 (2021) CUST for Lightweight Image Super-Resolution 17

    2021

  16. [16]

    In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW)

    Gu, S., Lugmayr, A., Danelljan, M., Fritsche, M., Lamour, J., Timofte, R.: Div8k: Diverse 8k resolution image dataset. In: 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW). pp. 3512–3516. IEEE (2019)

    2019

  17. [17]

    arXiv preprint arXiv:2411.15269 (2024)

    Guo, H.,Guo,Y.,Zha,Y.,Zhang,Y.,Li,W.,Dai,T., Xia,S.T.,Li,Y.:Mambairv2: Attentive state space restoration. arXiv preprint arXiv:2411.15269 (2024)

    Pith/arXiv arXiv 2024

  18. [18]

    In: European Conference on Computer Vision

    Guo, H., Li, J., Dai, T., Ouyang, Z., Ren, X., Xia, S.T.: Mambair: A simple base- line for image restoration with state-space model. In: European Conference on Computer Vision. pp. 222–241. Springer (2024)

    2024

  19. [19]

    In: International Conference on Learning Representations

    Hatamizadeh, A., Heinrich, G., Yin, H., Tao, A., Alvarez, J.M., Kautz, J., Molchanov, P.: Fastervit: Fast vision transformers with hierarchical attention. In: International Conference on Learning Representations. vol. 2024, pp. 29368–29391 (2024)

    2024

  20. [20]

    IEEE Transactions on Circuits and Sys- tems for Video Technology30(11), 3911–3927 (2019)

    Hu, Y., Li, J., Huang, Y., Gao, X.: Channel-wise and spatial feature modulation network for single image super-resolution. IEEE Transactions on Circuits and Sys- tems for Video Technology30(11), 3911–3927 (2019)

    2019

  21. [21]

    Advances in Neural Information Processing Systems35, 14596–14607 (2022)

    Huang, H., Zhou, X., He, R.: Orthogonal transformer: An efficient vision trans- former backbone with token orthogonalization. Advances in Neural Information Processing Systems35, 14596–14607 (2022)

    2022

  22. [22]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

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

    2015

  23. [23]

    In: ACM MM (2019)

    Hui, Z., Gao, X., Yang, Y., Wang, X.: Lightweight image super-resolution with information multi-distillation network. In: ACM MM (2019)

    2019

  24. [24]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Ignatov, A., Perevozchikov, G., Timofte, R., Zhang, Z., Gao, T., Yang, Y., Zhu, S., Wang, S., Yoon, K., Gankhuyag, G., et al.: Quantized image super-resolution on mobile npus, mobile ai 2025 challenge: Report. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 1908–1921 (2025)

    2025

  25. [25]

    Kim, J., Nang, J., Choe, J.: Lmlt : Low-to-high multi-level vision transformer for lightweightimagesuper-resolution.In:ProceedingsoftheIEEE/CVFInternational ConferenceonComputerVision(ICCV)Workshops.pp.5509–5519(October2025)

  26. [26]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)

    Kim,J.,Lee,J.K.,Lee,K.M.:Accurateimagesuper-resolutionusingverydeepcon- volutional networks. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2016)

    2016

  27. [27]

    In: ICLR (2015)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. In: ICLR (2015)

    2015

  28. [28]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition

    Ledig, C., Theis, L., Huszár, F., Caballero, J., Cunningham, A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang, Z., et al.: Photo-realistic single image super- resolution using a generative adversarial network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 4681–4690 (2017)

    2017

  29. [29]

    In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2023)

    Li, Y., Fan, Y., Xiang, X., Demandolx, D., Ranjan, R., Timofte, R., Gool, L.V.: Efficient and explicit modelling of image hierarchies for image restoration. In: Pro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition (2023)

    2023

  30. [30]

    Proceedings Eighth IEEE International Con- ference on Computer Vision (ICCV) (2021)

    Liang, J., Cao, J., Sun, G., Zhang, K., Van Gool, L., Timofte, R.: Swinir: Image restoration using swin transformer. Proceedings Eighth IEEE International Con- ference on Computer Vision (ICCV) (2021)

    2021

  31. [31]

    In: Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) Workshops (July 2017) 18 J

    Lim, B., Son, S., Kim, H., Nah, S., Mu Lee, K.: Enhanced deep residual networks for single image super-resolution. In: Proceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition (CVPR) Workshops (July 2017) 18 J. Kim

    2017

  32. [32]

    In: Proceedings of the AAAI Con- ference on Artificial Intelligence (2023)

    Liu, J., Chen, C., Tang, J., Wu, G.: From coarse to fine: Hierarchical pixel inte- gration for lightweight image super-resolution. In: Proceedings of the AAAI Con- ference on Artificial Intelligence (2023)

    2023

  33. [33]

    In: European Conference on Computer Vision

    Liu, J., Tang, J., Wu, G.: Residual feature distillation network for lightweight image super-resolution. In: European Conference on Computer Vision. pp. 41–55. Springer (2020)

    2020

  34. [34]

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

    Liu, X., Liu, J., Tang, J., Wu, G.: Catanet: Efficient content-aware token aggre- gation for lightweight image super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 17902– 17912 (June 2025)

    2025

  35. [35]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin trans- former: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 10012– 10022 (2021)

    2021

  36. [36]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Long, W., Zhou, X., Zhang, L., Gu, S.: Progressive focused transformer for sin- gle image super-resolution. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 2279–2288 (2025)

    2025

  37. [37]

    In: European Conference on Computer Vision

    Luo, X., Xie, Y., Zhang, Y., Qu, Y., Li, C., Fu, Y.: Latticenet: Towards lightweight image super-resolution with lattice block. In: European Conference on Computer Vision. pp. 272–289. Springer (2020)

    2020

  38. [38]

    In: Proceedings eighth IEEE international conference on com- puter vision

    Martin, D., Fowlkes, C., Tal, D., Malik, J.: 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 com- puter vision. ICCV 2001. vol. 2, pp. 416–423. Ieee (2001)

    2001

  39. [39]

    Multimedia tools and applications76(20), 21811–21838 (2017)

    Matsui, Y., Ito, K., Aramaki, Y., Fujimoto, A., Ogawa, T., Yamasaki, T., Aizawa, K.: Sketch-based manga retrieval using manga109 dataset. Multimedia tools and applications76(20), 21811–21838 (2017)

    2017

  40. [40]

    Advances in Neural Information Processing Systems 38, 100092–100118 (2026)

    Qiu,Z.,Wang,Z.,Zheng,B.,Huang,Z.,Wen,K.,Yang,S.,Men,R.,Yu,L.,Huang, F., Huang, S., et al.: Gated attention for large language models: Non-linearity, spar- sity, and attention-sink-free. Advances in Neural Information Processing Systems 38, 100092–100118 (2026)

    2026

  41. [41]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)

    Shi, W., Caballero, J., Huszar, F., Totz, J., Aitken, A.P., Bishop, R., Rueckert, D., Wang, Z.: Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2016)

    2016

  42. [42]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Sun, L., Dong, J., Tang, J., Pan, J.: Spatially-adaptive feature modulation for efficient image super-resolution. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 13190–13199 (October 2023)

    2023

  43. [43]

    Advances in Neural Information Processing Systems35, 17314–17326 (2022)

    Sun, L., Pan, J., Tang, J.: Shufflemixer: An efficient convnet for image super- resolution. Advances in Neural Information Processing Systems35, 17314–17326 (2022)

    2022

  44. [44]

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

    Tian, Y., Chen, H., Xu, C., Wang, Y.: Image processing gnn: Breaking rigidity in super-resolution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 24108–24117 (June 2024)

    2024

  45. [45]

    In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops

    Timofte, R., Agustsson, E., Van Gool, L., Yang, M.H., Zhang, L.: Ntire 2017 chal- lenge on single image super-resolution: Methods and results. In: Proceedings of the IEEE conference on computer vision and pattern recognition workshops. pp. 114–125 (2017)

    2017

  46. [46]

    Advances in neural information pro- cessing systems30(2017) CUST for Lightweight Image Super-Resolution 19

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017) CUST for Lightweight Image Super-Resolution 19

    2017

  47. [47]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pyramid vision transformer: A versatile backbone for dense prediction with- out convolutions. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 568–578 (October 2021)

    2021

  48. [48]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Wang, Y., Zhang, T.: OSFFNet: Omni-Stage Feature Fusion Network for Lightweight Image Super-Resolution. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 5660–5668 (2024)

    2024

  49. [49]

    IEEE transactions on image processing 13(4), 600–612 (2004)

    Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13(4), 600–612 (2004)

    2004

  50. [50]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Yu,W.,Luo,M.,Zhou,P.,Si,C.,Zhou,Y.,Wang,X.,Feng,J.,Yan,S.:Metaformer is actually what you need for vision. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10819–10829 (2022)

    2022

  51. [51]

    In: Curves and Surfaces

    Zeyde, R., Elad, M., Protter, M.: On single image scale-up using sparse- representations. In: Curves and Surfaces. pp. 711–730. Springer Berlin Heidelberg, Berlin, Heidelberg (2012)

    2012

  52. [52]

    In: International Conference on Computer Vision (2023)

    Zhang, A., Ren, W., Liu, Y., Cao, X.: Lightweight image super-resolution with su- perpixel token interaction. In: International Conference on Computer Vision (2023)

    2023

  53. [53]

    In: IEEE Conference on Computer Vision and Pattern Recognition

    Zhang, K., Zuo, W., Gu, S., Zhang, L.: Learning deep cnn denoiser prior for image restoration. In: IEEE Conference on Computer Vision and Pattern Recognition. pp. 3929–3938 (2017)

    2017

  54. [54]

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

    Zhang, L., Li, Y., Zhou, X., Zhao, X., Gu, S.: Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 2856–2865 (June 2024)

    2024

  55. [55]

    IEEE Transactions on Pattern Analysis and Machine Intelligence46(5), 3608–3624 (2024)

    Zhang, Q., Zhang, J., Xu, Y., Tao, D.: Vision transformer with quadrangle at- tention. IEEE Transactions on Pattern Analysis and Machine Intelligence46(5), 3608–3624 (2024)

    2024

  56. [56]

    In: ECCV (2024)

    Zhang, X., Zhang, Y., Yu, F.: Hit-sr: Hierarchical transformer for efficient image super-resolution. In: ECCV (2024)

    2024

  57. [57]

    Proceedings of the IEEE Conference on European Confer- ence on Computer Vision (ECCV) (2022)

    Zhang, X., Zeng, H., Guo, S., Zhang, L.: Efficient long-range attention network for image super-resolution. Proceedings of the IEEE Conference on European Confer- ence on Computer Vision (ECCV) (2022)

    2022

  58. [58]

    In: ECCV (2018)

    Zhang, Y., Li, K., Li, K., Wang, L., Zhong, B., Fu, Y.: Image super-resolution using very deep residual channel attention networks. In: ECCV (2018)

    2018

  59. [59]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    Zhang, Y., Tian, Y., Kong, Y., Zhong, B., Fu, Y.: Residual dense network for image super-resolution. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (2018)

    2018

  60. [60]

    In: International Conference on Learning Representations (2021)

    Zhang, Y., Wang, H., Qin, C., Fu, Y.: Learning efficient image super-resolution net- works via structure-regularized pruning. In: International Conference on Learning Representations (2021)

    2021

  61. [61]

    In: ECCV (2024)

    Zheng, M., Sun, L., Dong, J., Pan, J.: Smfanet: A lightweight self-modulation feature aggregation network for efficient image super-resolution. In: ECCV (2024)

    2024

  62. [62]

    Oth- ers

    Zhou, Y., Li, Z., Guo, C.L., Bai, S., Cheng, M.M., Hou, Q.: Srformer: Permuted self-attention for single image super-resolution. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 12780–12791 (2023) 20 J. Kim Table A1:Quantitative comparison of CUST-Base and CUST-Base+ with state-of- the-art lightweight SISR models. The best ...

    2023

This paper was first reviewed by grok-4.5 on July 14, 2026.