Pith. sign in

REVIEW 4 major objections 5 minor 47 references

CASIAL: Geometric Distortion Robust Image Watermarking

T0 review · 4 major / 5 minor · reviewed 2026-07-30 · grok-4.5

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

Pith's one-line read Geometric watermark robustness needs global cover-aware message spread plus attention that realigns desynchronized features.

desk verdict Solid systems paper: cover-gated spreading plus attention really does move the geometric numbers, but the headline margin is still partly a matched-noise story. read the letter →

arxiv 2607.26729 v1 pith:6I2GUR6V submitted 2026-07-29 cs.CV

Yupeng Qiu , Han Fang , Ee-Chien Chang This is my paper · ORCID
classification cs.CV
keywords imagewatermarkinggeometricrobustnesscover-awaremessagespreadinginvariancealignmentspatialattentiondesynchronizationregionremovaldeeplearning
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

Deep image watermarks already survive compression and noise, but geometric attacks still break them in two distinct ways: cropping or masking erases the pixels that carried the bits, and rotation, scaling, shear, or scrambling misaligns the remaining evidence so a fixed decoder cannot read it. This paper argues that fixing those failures requires two properties at once—message evidence must be spread across the whole image in a way that depends on the cover content, and the decoder must form geometry-invariant features that can re-aggregate displaced evidence. CASIAL implements the first property with cover-aware message spreading (CAS), which turns each bit into a selector over candidate features derived from the cover, and the second with invariance alignment learning (IAL), spatial attention that reweights reliable regions after distortion. Trained only on differentiable white-box distortions, the system reports near-perfect bit accuracy on six geometric attacks while keeping higher visual quality than eleven prior methods, and it transfers to unseen black-box appearance changes and non-square image shapes.

What carries the argument

CASIAL’s paired mechanisms: cover image-aware message spreading (CAS), which generates bit-0/bit-1 candidate features from the cover latent and selects them by the message bits before attention fusion, and invariance alignment learning (IAL), spatial attention that maps distorted features into a shared geometry-invariant space.

What would settle it

Retrain or re-evaluate under stronger or held-out geometric attacks (for example heavier crop ratios, larger rotations, or adaptive desynchronization) and check whether bit accuracy collapses toward chance while baselines do not, or whether a steganalytic/adaptive attacker can strip the CAS residual without destroying the image.

Watch

Extended reading notes

Core claim

Geometric robustness in deep watermarking is not achieved by noise layers alone; it requires explicit global, cover-conditioned message spreading to survive region removal and spatial-attention-based alignment to recover from desynchronization. CASIAL realizes both and, on the authors’ benchmarks, substantially outperforms eleven baselines on six geometric transformations while preserving the best reported visual quality and retaining strong signal, photometric, and black-box transfer performance.

Load-bearing premise

Training with a fixed menu of differentiable geometric noise layers plus latent attention is assumed sufficient to produce truly geometry-invariant, transferable watermarks without explicit sync marks or adaptive attackers.

Editorial extensions

If this is right

  • Message-processing blocks that ignore cover content will remain brittle under heavy crop or erase even if they upsample or broadcast bits.
  • Spatial attention (or equivalent long-range aggregation) becomes a practical requirement for decoders that must survive rotation, shear, elastic warp, and jigsaw permutation.
  • JND residual attenuation can raise perceptual quality at matched PSNR by moving energy out of smooth regions, complementary to the robustness gains from CAS and IAL.
  • A single architecture conditioned on cover latents can serve variable resolutions and aspect ratios without a separate shape-conversion stage.

Reading between the lines

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

  • The same cover-conditioned spreading idea may transfer to video or audio watermarking where temporal or spectral ‘region removal’ is common.
  • If IAL’s attention maps are stable under permutation, they could double as a lightweight desynchronization diagnostic or soft sync signal.
  • Capacity–robustness curves under extreme crop ratios would test whether per-pixel information density from CAS scales beyond the 64-bit setting used here.
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 / 5 minor

Summary. The paper proposes CASIAL, an encoder–noise–decoder image watermarking framework aimed at geometric robustness. It attributes geometric failures to region removal and desynchronization, and addresses them with two components: CAS, which generates cover-conditioned candidate features and selects them via message bits to spread watermark evidence globally, and IAL, which uses spatial attention to aggregate and realign features under spatial warps. Encoding/decoding operate in VAE latent space with JND residual attenuation for imperceptibility. Empirically, against eleven retrained baselines under matched resolution, message length, and white-box noise settings, CASIAL reports the best fidelity (Table 1: 40.82 dB PSNR) and near-ceiling bit accuracy on six geometric transforms plus competitive signal/photometric results (Table 2: 99.40% average), with ablations (Table 4), residual/attention visualizations (Figs. 4–5), multi-shape tests (Table S1), and transfer to five appearance-style black-box filters (Table 3).

Significance. Geometric robustness remains a genuine bottleneck for deep watermarking, and the paper’s two-failure-mode framing is clear and useful. The empirical package is comparatively strong for this area: matched retraining of eleven baselines, component ablations, content-coupling visualizations, decoder attention under jigsaw, distortion-strength curves, and multi-aspect-ratio evaluation. If the gains hold under fairer geometric augmentation of baselines and under geometric operators outside the training menu, CASIAL would be a meaningful practical advance for copyright/provenance pipelines that must survive crop, warp, and layout disruption. Strengths to credit explicitly include the systematic baseline reimplementation under shared noise settings, the CAS residual analysis showing cover-dependent spreading, and the complementary CAS/IAL ablation.

major comments (4)
  1. [Experimental Settings; Table 2] Experimental Settings / Table 2: The headline claim of substantially stronger geometric robustness than eleven baselines is load-bearing on “all models are trained under the same noise settings.” Several END/invertible baselines were originally designed around JPEG/signal layers; under the shared Kornia geometric menu they fall near chance on C&R, shear, rotate, elastic, and jigsaw while CASIAL stays ≥98.44%. Please document capacity/schedule/hyperparameter parity (epochs, LR, noise sampling probabilities, whether spreading variants received equal geometric exposure) and, ideally, report a stronger geometric-augmented baseline protocol (or architecture-appropriate sync auxiliaries). Without this, the margin may partly reflect unequal fitness to this exact attack family rather than a general geometric solution.
  2. [Abstract; Motivation and Methods; Noise layer] Motivation & Methods / Abstract: IAL is described as aligning features into a “shared geometry-invariant representation space,” but the evidence is recovery under the same white-box geometric family used in training (crop-resize, erase, shear, rotate, elastic, jigsaw) plus attention realignment on jigsaw (Fig. 5). There is no explicit geometric normalization, group-equivariant design, or test on held-out geometric operators (e.g., perspective/homography, similarity with continuous scale outside the trained range, free-form mesh warps, or print-cam geometry). Please either (i) add such out-of-menu geometric evaluations, or (ii) substantially temper “geometry-invariant” language to “robust under the trained geometric attack family,” and discuss limits of attention-based realignment.
  3. [Table 3; Abstract] Table 3 / black-box claim: Transfer is demonstrated only on appearance filters (Crayon, Film, Heavy, Layering, Sketch), not on unseen geometric transforms or adaptive/steganalytic attackers. The abstract’s “strong transfer robustness to unseen black-box distortions” is easy to over-read as geometric generalization. Restrict or split the claim (appearance transfer vs. geometric transfer), and add at least one geometric black-box or composition setting (e.g., random perspective + JPEG, or screen-cam-style geometry) so the transfer statement matches the paper’s geometric thesis.
  4. [CAS Strategy; Algorithm S1] CAS formulation (Algorithm S1 / Fig. 2d): Each bit selects a full-resolution candidate F^(0)/F^(1), then L candidates are channel-concatenated to shape (L·C)×H/4×W/4 before IAL fusion. For L=64 this is a large intermediate tensor and a strong inductive bias toward global bit evidence, but the paper does not analyze bit interference, payload scaling, or complexity versus MPB baselines. Please report accuracy vs. L (and wall-clock/memory), and clarify whether performance at L=64 depends on this expensive concatenation rather than on cover-aware selection alone—an ablation with shared low-rank bit embeddings or per-bit lightweight modulators would isolate the claimed mechanism.
minor comments (5)
  1. [Figure 1] Fig. 1 is referenced for robustness comparison but the main text does not fully specify what each bar aggregates; align caption with Table 2 protocol (exact strengths: erase 80%, C&R 20%, etc.).
  2. [Encoder; Supplementary Table S2] JND module is important for Table 1 quality yet only lightly specified (“perceptual visibility mask”); cite the exact JND formulation used and whether it is frozen or learned.
  3. [Loss Function] Loss uses MSE on bits (Eq. 2) rather than BCE; a one-sentence justification would help, given binary messages.
  4. [Introduction; Motivation] Typos/spacing artifacts appear throughout (e.g., “asasscalingorrotation”, “re-gion removal”, “noiselayersalone”); a careful copy-edit pass is needed.
  5. [Related Work] Related Work on transformers is fine but could briefly position against classical sync-template / RST-invariant watermarking to clarify what is new versus classical desynchronization defenses.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical END watermarking with held-out evaluation, not a derivation that equals its inputs by construction.

full rationale

CASIAL is an empirical deep-watermarking systems paper. Its load-bearing claims are measured bit accuracies and fidelity metrics (Tables 1–4, S1–S2) on held-out USC-SIPI images, including black-box appearance filters never inserted in the noise layer. The architectural story (region removal → CAS; desynchronization → IAL) motivates design choices; it does not algebraically define the reported accuracies. Training with a shared white-box Kornia geometric menu is standard END practice and creates train–test kinship for those distortions, but baselines are retrained under the same noise settings and still collapse near chance on several geometric attacks while CASIAL does not—so the headline margins are not forced by construction. Ablations (Table 4) further show CAS/IAL are necessary inside the pipeline rather than tautological renamings of the loss. Self-citations (LightweightMark, etc.) appear only as baselines, not as uniqueness theorems or smuggled ansatze that forbid alternatives. No fitted scalar is relabeled as a prediction; no equation reduces target robustness to an input identity. Circularity score is therefore 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The work is empirical deep learning. Load-bearing commitments are the END training paradigm, the sufficiency of the chosen differentiable geometric noise family, VAE latent features as the embedding space, spatial attention as a proxy for geometry invariance, MSE bit supervision, and hand-chosen loss weights plus JND masking. No formal invariance theorem is claimed.

free parameters (4)
  • λ1, λ2 loss weights = 20 and 1
    Total loss is λ1 L_image + λ2 L_message; set to 20 and 1 without a reported sensitivity sweep, directly governing the quality-robustness tradeoff.
  • Message length L and input resolution = L=64, 128x128 default
    Default L=64 on 128x128 images fixes payload and spatial budget for all main tables.
  • Geometric noise hyperparameters used in training/eval = e.g. erase 80%, C&R retain 20%, shear ±60°, rotate ±45°, elastic α=2.0, jigsaw 8x8
    Erase ratio, crop retain ratio, shear/rotation ranges, elastic alpha, jigsaw grid, etc., define the robustness claim surface.
  • Adam learning rate and train budget = 1e-5
    Optimizer hyperparameters affect convergence of encoder/decoder and are stated as 1e-5 with defaults, not derived.
assumptions (5)
  • domain assumption Differentiable white-box noise-layer training yields decoders that remain accurate under corresponding real geometric transforms and can transfer to unseen appearance distortions.
    Stated throughout Motivation & Methods and Experimental Settings; central to both white-box and black-box claims.
  • ad hoc to paper VAE latent features plus spatial attention can implement a shared geometry-invariant representation sufficient for desynchronization recovery without explicit geometric normalization.
    IAL design paragraph and decoder analysis (Fig. 5) treat attention aggregation as the synchronization mechanism.
  • domain assumption MSE on images and on continuous message predictions is an adequate surrogate for imperceptibility and bit accuracy.
    Loss Function section defines L_image and L_message as MSE.
  • domain assumption JND-style perceptual masks identify regions where residuals are less visible and can be applied as a multiplicative attenuation on the encoder residual.
    Encoder forward path and Related Work on perceptual masking; ablated in Table S2/Fig. S4.
  • domain assumption Matched retraining of prior official implementations under the same noise settings is a fair robustness comparison.
    Experimental Settings benchmarks paragraph.
invented entities (2)
  • CAS (cover image-aware message spreading) with binary gated feature selection
    purpose: Generate bit-dependent features from cover latents so each bit is globally spread and content-coupled.
    Core proposed module in Fig. 2(d) and Algorithm S1; not a physical entity but a new architectural operator the claims depend on.
  • IAL (invariance alignment learning) module
    purpose: Use spatial attention to align distorted features into a geometry-invariant space for encoding/decoding.
    Second named pillar of CASIAL; implemented as attention blocks rather than a proven invariant feature theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CASIAL: Geometric Distortion Robust Image Watermarking." pith.science (2026). https://pith.science/paper/6I2GUR6V

@misc{pith2026260726729,
  author       = {Pith},
  title        = {Pith review of: CASIAL: Geometric Distortion Robust Image Watermarking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6I2GUR6V}},
  note         = {Machine review of arXiv:2607.26729}
}
read the original abstract

Deep learning-based watermarking has shown strong robustness against non-geometric distortions, yet its performance under geometric transformations remains limited. Such transformations induce two fundamental failure modes: region removal, such as cropping or masking, which eliminates the information carried by removed pixels, and desynchronization, such as scaling or rotation, which misaligns pixel positions and disrupts decoding. We argue that achieving geometric robustness requires two essential properties: (1) global spread of the watermark message, ensuring resilience even when large regions are removed, and (2) geometry-invariant representations, enabling decoding to remain synchronized despite spatial transformations. Building on these insights, we propose CASIAL, a geometric distortion-robust watermarking framework with cover image-aware message spreading (CAS) strategy and invariance alignment learning (IAL) module. CAS tightly couples watermark bits with cover image features and distributes them adaptively across the entire image, enhancing per-pixel information capacity and robustness to region removal. IAL leverages spatial attention to capture cross-pixel dependencies and align perturbed features into a shared geometry-invariant representation space, mitigating failures due to desynchronization. Across six challenging geometric transformations, CASIAL achieves substantially stronger robustness than eleven prior baselines while preserving high visual quality. It also maintains competitive performance under six signal distortions and four photometric transformations. Notably, although trained only with white-box distortions, CASIAL also exhibits strong transfer robustness to unseen black-box distortions. Comprehensive experiments demonstrate the broad robustness and superior visual quality of our method.

Figures

Figures reproduced from arXiv: 2607.26729 by the authors.

Figure 1
Figure 1. Robustness comparison under different distortions. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Representative message processing blocks (MPBs) in prior deep watermarking methods and the proposed CAS strategy. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of CASIAL. In the encoder, a VAE produces cover-image features, CAS produces a cover-aware message [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of message spreading and cover-image coupling. For each method and cover image, we encode (i) an [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Decoder attention under jigsaw distortion. Rows from top to bottom show the watermarked image, the attention map on [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 3 linked inside Pith

  1. [1]

    Morgan Kaufmann Publishers , volume=

    Digital watermarking , author=. Morgan Kaufmann Publishers , volume=. 2008 , publisher=

    2008

  2. [2]

    Forty-third International Conference on Machine Learning , year=

    Revisiting Coding-Based Approaches to Overcome the Curse of Dimensionality in Learning-Based Watermarking , author=. Forty-third International Conference on Machine Learning , year=

  3. [3]

    Proceedings of 1st international conference on image processing , volume=

    A digital watermark , author=. Proceedings of 1st international conference on image processing , volume=. 1994 , organization=

    1994

  4. [4]

    IEEE transactions on image processing , volume=

    Secure spread spectrum watermarking for multimedia , author=. IEEE transactions on image processing , volume=. 1997 , publisher=

    1997

  5. [5]

    Signal processing , volume=

    A DCT-domain system for robust image watermarking , author=. Signal processing , volume=. 1998 , publisher=

    1998

  6. [6]

    Proceedings of the IEEE , volume=

    Multimedia watermarking techniques , author=. Proceedings of the IEEE , volume=. 1999 , publisher=

    1999

  7. [7]

    Proceedings of the IEEE , volume=

    Information hiding-a survey , author=. Proceedings of the IEEE , volume=. 1999 , publisher=

    1999

  8. [8]

    Technologies , volume=

    An authorship protection technology for electronic documents based on image watermarking , author=. Technologies , volume=. 2020 , publisher=

    2020

Show all 47 references
  1. [9]

    Multimedia Tools and Applications , volume=

    A review of image watermarking for identity protection and verification , author=. Multimedia Tools and Applications , volume=. 2024 , publisher=

    2024

  2. [10]

    IEEE Transactions on circuits and systems for video technology , volume=

    A perceptually tuned subband image coder based on the measure of just-noticeable-distortion profile , author=. IEEE Transactions on circuits and systems for video technology , volume=. 1995 , publisher=

    1995

  3. [11]

    IEEE Transactions on Image Processing , volume=

    Enhanced just noticeable difference model for images with pattern complexity , author=. IEEE Transactions on Image Processing , volume=. 2017 , publisher=

    2017

  4. [12]

    Forty-second International Conference on Machine Learning , year=

    XAttnMark: Learning Robust Audio Watermarking with Cross-Attention , author=. Forty-second International Conference on Machine Learning , year=

  5. [13]

    Proceedings of the European conference on computer vision (ECCV) , pages=

    Hidden: Hiding data with deep networks , author=. Proceedings of the European conference on computer vision (ECCV) , pages=

  6. [14]

    , author=

    A Dwt-Based Image Watermarking Algorithm. , author=. ICME , year=

  7. [15]

    International Conference on Information Technology: Coding and Computing, 2004

    A novel DCT-based approach for secure color image watermarking , author=. International Conference on Information Technology: Coding and Computing, 2004. Proceedings. ITCC 2004. , volume=. 2004 , organization=

    2004

  8. [16]

    Proceedings of the 29th ACM international conference on multimedia , pages=

    Towards robust deep hiding under non-differentiable distortions for practical blind watermarking , author=. Proceedings of the 29th ACM international conference on multimedia , pages=

  9. [17]

    Proceedings of the 30th ACM International Conference on Multimedia , pages=

    Towards blind watermarking: Combining invertible and non-invertible mechanisms , author=. Proceedings of the 30th ACM International Conference on Multimedia , pages=

  10. [18]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Flow-based robust watermarking with invertible noise layer for black-box distortions , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  11. [19]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  12. [20]

    2008 , publisher=

    Digital watermarking and steganography , author=. 2008 , publisher=

    2008

  13. [21]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Stegastamp: Invisible hyperlinks in physical photographs , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  14. [22]

    Proceedings of the 30th ACM international conference on multimedia , pages=

    PIMoG: An effective screen-shooting noise-layer simulation for deep-learning-based watermarking network , author=. Proceedings of the 30th ACM international conference on multimedia , pages=

  15. [23]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  16. [24]

    The eleventh international conference on learning representations , year=

    Learning iterative neural optimizers for image steganography , author=. The eleventh international conference on learning representations , year=

  17. [25]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  18. [26]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    The stable signature: Rooting watermarks in latent diffusion models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  19. [27]

    International Conference on Learning Representations-ICLR 2025 , year=

    Watermark anything with localized messages , author=. International Conference on Learning Representations-ICLR 2025 , year=

    2025

  20. [28]

    arXiv preprint arXiv:2512.16874 , year=

    Pixel Seal: Adversarial-only training for invisible image and video watermarking , author=. arXiv preprint arXiv:2512.16874 , year=

  21. [29]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Gaussian shading: Provable performance-lossless image watermarking for diffusion models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  22. [30]

    International Conference on Machine Learning , pages=

    GaussMarker: Robust Dual-Domain Watermark for Diffusion Models , author=. International Conference on Machine Learning , pages=. 2025 , organization=

    2025

  23. [31]

    Proceedings of the 29th ACM international conference on multimedia , pages=

    Mbrs: Enhancing robustness of dnn-based watermarking by mini-batch of real and simulated jpeg compression , author=. Proceedings of the 29th ACM international conference on multimedia , pages=

  24. [32]

    European conference on computer vision , pages=

    Microsoft coco: Common objects in context , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  25. [33]

    The USC-SIPI Image Database , howpublished =

    USC Viterbi , year=. The USC-SIPI Image Database , howpublished =

  26. [34]

    arXiv preprint arXiv:1412.6980 , year=

    Adam: A method for stochastic optimization , author=. arXiv preprint arXiv:1412.6980 , year=

    arXiv

  27. [35]

    Advances in neural information processing systems , volume=

    Pytorch: An imperative style, high-performance deep learning library , author=. Advances in neural information processing systems , volume=

  28. [36]

    ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Watermarking images in self-supervised latent spaces , author=. ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2022 , organization=

    2022

  29. [37]

    Forty-second International Conference on Machine Learning , year=

    Lightweight-Mark: Rethinking Deep Learning-Based Watermarking , author=. Forty-second International Conference on Machine Learning , year=

  30. [38]

    arXiv preprint arXiv:2412.09492 , year=

    Video seal: Open and efficient video watermarking , author=. arXiv preprint arXiv:2412.09492 , year=

    arXiv

  31. [39]

    arXiv preprint arXiv:2510.12812 , year=

    We can hide more bits: The unused watermarking capacity in theory and in practice , author=. arXiv preprint arXiv:2510.12812 , year=

  32. [40]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

    2004

  33. [41]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  34. [42]

    Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

    Kornia: an open source differentiable computer vision library for pytorch , author=. Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

  35. [43]

    arXiv preprint arXiv:2010.11929 , year=

    An image is worth 16x16 words: Transformers for image recognition at scale , author=. arXiv preprint arXiv:2010.11929 , year=

    2010 arXiv

  36. [44]

    International conference on machine learning , pages=

    Training data-efficient image transformers & distillation through attention , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  37. [45]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Swin transformer: Hierarchical vision transformer using shifted windows , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  38. [46]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Restormer: Efficient transformer for high-resolution image restoration , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  39. [47]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Uformer: A general u-shaped transformer for image restoration , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

Pith tools

Reviewed July 30, 2026 · model on record in the stance chip above.