Pith. sign in

REVIEW 3 major objections 4 minor 32 references

Fake & Square: Training Self-Supervised Vision Transformers with Synthetic Data and Synthetic Hard Negatives

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Synthetic clones plus fake hard negatives lift DeiT-S by ~3 points.

desk verdict An honest, incremental empirical study whose headline DeiT-S gain is confounded by doubled optimizer steps; the paper's own matched-iteration rows show Syn2Co below MoBY. read the letter →

arxiv 2509.02029 v1 pith:OCBE7EYV submitted 2025-09-02 cs.CV cs.AI

Nikos Giakoumoglou , Andreas Floros , Kleanthis Marios Papadopoulos , Tania Stathaki This is my paper
classification cs.CVcs.AI
keywords self-supervisedlearningcontrastivevisiontransformerssyntheticdatadiffusionmodelshardnegativesImageNet-100linearprobing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tests whether a self-supervised vision transformer can be trained with synthetic images and synthetic hard negatives in place of large amounts of carefully curated real data. It builds on the MoBY contrastive framework and the SynCo hard-negative idea, adding a diffusion-generated clone of ImageNet-100 and representation-space synthetic negatives to the training queue. The paper's claim is that these two "fake" ingredients help DeiT-S learn more discriminative features: at 300 epochs, Syn2Co reaches 82.12% top-1 linear-probing accuracy on ImageNet-100, compared with 79.36% for the MoBY baseline, although Syn2Co runs twice as many iterations over those 300 epochs. The authors conclude that synthetic data can complement rather than replace real data, and that the value of each synthetic ingredient depends strongly on the transformer architecture.

What carries the argument

The load-bearing mechanism is the InfoNCE contrastive objective over a momentum-updated queue of negatives, extended with two synthetic inputs. A diffusion model generates roughly 130,000 synthetic clones of ImageNet-100. A synthesis function F picks the top-N hardest real negatives for each query by cosine similarity and creates new L2-normalized contrasts using six strategies—interpolation, extrapolation, mixing, noise jittering, perturbation, and adversarial synthesis. These synthetic negatives are inserted into the queue, so each InfoNCE update has to separate the query from a richer set of difficult examples.

What would settle it

Run the MoBY baseline for 152,400 iterations with the same augmentations and batch size and compare ImageNet-100 top-1 linear-probing accuracy to Syn2Co's 82.12; if MoBY matches or exceeds it at matched iterations, the reported DeiT-S gain is explained by compute rather than by synthetic data and negatives.

Watch

Extended reading notes

Core claim

The central discovery, as the authors state it, is that vision transformers can productively use synthetic contrasts in self-supervised learning, and that the two forms of "faking it" work differently. Syn2Co combines a diffusion-generated clone of the ImageNet-100 training set with synthetic hard negatives built from the top-N hardest entries in the momentum queue; the negatives are produced by interpolation, extrapolation, mixing, jittering, perturbation, or adversarial strategies and then normalized. On the paper's main comparison, DeiT-S improves from 79.36% to 82.12% top-1 with both ingredients under extended training, while Swin-T's best result comes from synthetic negatives alone (84.

Load-bearing premise

The result depends on treating MoBY's 76,200-iteration run as the fair control for Syn2Co's 152,400-iteration run; at equal iterations Syn2Co actually trails the baseline, so the extra training length must not be what causes the gain.

Editorial extensions

If this is right

  • Pretraining purely on synthetic images remains close to, but below, real-data pretraining on ImageNet-100, so diffusion clones can stand in for real data when real data is scarce.
  • Synthetic hard negatives alone improve Swin-T at fixed training length (84.04 vs 83.90), making them a low-cost way to sharpen representations.
  • For DeiT-S, the best configuration needs both synthetic data and synthetic negatives over extended training, pointing to architecture-specific design rules.
  • The result implies self-supervised pipelines can be bootstrapped without large real datasets, which matters for low-resource domains.

Reading between the lines

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

  • Beyond the paper: a matched-iteration comparison—both at 76,200 or both at 152,400 iterations—would cleanly separate the contribution of the synthetic components from the contribution of longer training; the paper's table includes enough numbers to set this test up.
  • Beyond the paper: the synthetic-negative mechanism should transfer to any queue-based contrastive learner; testing it with a larger queue or a harder selection policy could reveal where gains saturate.
  • Beyond the paper: the DeiT/Swin difference suggests attention locality changes how models use extra synthetic contrasts; varying the window size in Swin is a direct way to probe this.
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

3 major / 4 minor

Summary. The paper studies two forms of synthetic augmentation for self-supervised contrastive learning with vision transformers: diffusion-generated synthetic images that complement ImageNet-100, and representation-space synthetic hard negatives built on top of MoBY, following the authors' earlier SynCo method. The resulting framework, Syn2Co, is evaluated with DeiT-S and Swin-T on ImageNet-100 linear probing. The main reported result is that Syn2Co reaches 82.12% top-1 with DeiT-S at 300 epochs/152,400 iterations, versus 79.36% for MoBY at 300 epochs/76,200 iterations; for Swin-T the authors report 83.70% versus 83.90%. The paper also studies the effect of the real/synthetic mixing ratio and of synthetic-negative hyperparameters. The abstract explicitly frames the work as an empirical study rather than a fundamentally new method.

Significance. If the headline gains held under controlled training budgets, the paper would provide useful evidence that diffusion-generated data and representation-space hard negatives can reduce the reliance on real data and on curated hard negatives in transformer self-supervised learning. The manuscript reproduces the baselines, implements all six SynCo synthesis strategies, and includes a useful architecture-level sensitivity analysis. Its primary limitation is experimental control: the strongest claim rests on an unmatched compute comparison, and the matched-iteration rows in Table 1 do not show a benefit. The paper is honest about its exploratory framing, which is appropriate for an empirical study, but the quantitative support for the central claims currently needs substantially stronger controls before the conclusions can be accepted.

major comments (3)
  1. [Table 1, Sections 4.2 and 5] The headline DeiT-S result is confounded by optimizer steps and sample budget. Syn2Co reaches 82.12 at 300 epochs/152,400 iterations, versus MoBY's 79.36 at 300 epochs/76,200; these runs differ by a factor of two in total batches and total samples seen. The matched-iteration row in the same table is Syn2Co at 150 epochs/76,200: 78.12, below MoBY's 79.36. Thus the advertised gain appears only under an extended-training regime, and no real-only MoBY baseline trained for 152,400 iterations is reported. Section 4.2's 'during extended training' caveat does not separate the synthetic components from the extra training budget. A compute-matched MoBY baseline is required before the claim that DeiT-S leverages synthetic contrasts is supported.
  2. [Table 1, Section 4.2] For Swin-T, the same-compute rows do not support the statement that 'Swin benefits primarily from synthetic negatives.' With synthetic negatives only at 300 epochs/76,200 iterations, Swin-T improves by 0.14 points (84.04 vs 83.90); Syn2Co at matched iterations is 83.55, below MoBY. No seeds, error bars, or significance tests are reported, so these differences are within plausible run-to-run noise. The paper should either provide multiple-seed results or substantially soften the architecture-specific claim.
  3. [Section 4.4, Figure 4] The hyperparameters for the reported runs (N and the synthetic-negative ratio) are selected using the same ImageNet-100 linear-probe metric on which the final numbers are reported, with no validation split or tuning on a separate criterion. This can overstate the benefit of the synthetic components. Report the selection procedure and, ideally, evaluate on downstream transfer tasks to distinguish genuine representation quality from tuning to the probe.
minor comments (4)
  1. [Figure 3] The caption does not specify the architecture, training length, or whether the percentage is over the combined real+synthetic dataset. Define the x-axis and state the evaluation protocol clearly.
  2. [Abstract and Section 1] The abstract says the paper 'does not introduce a new method per se,' while the contributions claim a framework that 'adapts and extends' synthetic negatives to transformers. Align these statements or clarify that Syn2Co is an integration of existing components rather than a novel method.
  3. [Equation (4), Section 3.4] The synthesis function F is normalized after application, but the paper does not specify how each of the six strategies behaves under that normalization (e.g., whether interpolation weights are applied before or after normalization). A short description or pseudocode for the six strategies would improve reproducibility.
  4. [Section 3.4] Hard negatives selected from the queue may include false negatives (samples that are semantically similar to the query). Since these candidates are used to synthesize contrasts, the paper should discuss this issue and any safeguards, as prior work on hard-negative mining typically does.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; minor self-citation to the authors' prior SynCo method is not load-bearing. The main weakness is an experimental-control confound (doubled optimizer steps), which is a validity issue, not circularity.

full rationale

This is an empirical paper that explicitly disclaims a new derivation ('This paper does not introduce a new method per se'), so there is no derivation chain whose conclusion is equivalent to its premises. The synthetic-data component uses an external diffusion model [22]; the synthetic-negative component reuses the six strategies from the authors' prior SynCo paper [12] ('Following [12], we implement six different synthesis strategies for F'), but this is method reuse, not a load-bearing justification: the central comparison in Table 1 is against externally reproduced baselines (MoBY, DINO, MoCo-v3, supervised), so the claim that DeiT/Swin can use synthetic contrasts is tested rather than assumed. The tuning of N and the synthetic-negative ratio on ImageNet-100 linear probing (Figure 4) is benchmark selection and can inflate reported numbers, but it is not a case of fitting a parameter and then renaming the fit a prediction; no equation in the paper reduces a reported quantity to an input by construction. The most serious weakness is experimental control, not circularity: in Table 1, Syn2Co at 300 epochs runs 152,400 iterations while MoBY runs 76,200, and the paper's own matched-iteration row (Syn2Co 150 epochs, 76,200 iterations) gives DeiT-S 78.12 vs MoBY 79.36, so the headline 82.12 gain is confounded by doubled optimizer steps. Section 4.2's caveat 'during extended training' acknowledges this dependence. That confound is a correctness/validity concern and does not make the argument circular. Score 2 reflects the minor self-citation to SynCo [12] and the benchmark-tuned hyperparameters; neither is load-bearing.

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

No new theoretical entities are introduced; Syn2Co and Syn1Co are framework labels for combining existing components, and synthetic negatives are data constructions rather than postulated objects. The load-bearing external assumptions are the fidelity of the diffusion-generated clone and the validity of the synthesized hard negatives, both domains where the paper provides only indirect evidence. The free parameters N, the synthetic-negative ratio, and the real/synthetic mixing ratio are all tuned on the reported evaluation metric.

free parameters (4)
  • N, number of selected hardest negatives = 256 / 512 / 1024 (swept)
    Eq. 3 TopK selection; Figure 4 shows accuracy varies with N; the best configuration is chosen with the same linear-probe metric used for the final results.
  • synthetic negative ratio L/(K+L) = 0 to 40%, optimum around 10-15% (Figure 4)
    Swept on the evaluation metric; no held-out validation split is described; the reported gains are maxima over this grid.
  • real/synthetic data mixing ratio = 0 to 100% real samples (Figure 3)
    Explored as a percentage of real samples; final Syn2Co uses the union X and Xs, but the curve itself is fit on the eval metric.
  • synthesis function F parameters (interpolation weight, jitter scale, etc.) = inherited from SynCo [12], not re-derived here
    All six synthesis strategies are implemented from the authors' own prior paper; their internal constants are not stated in this preprint.
assumptions (5)
  • standard math InfoNCE objective (Eq. 1) is a valid surrogate for learning discriminative representations
    Adopted from Oord et al. [26]; standard contrastive learning objective.
  • domain assumption Synthesized negatives s = F(q,n)/||F(q,n)|| (Eq. 4) are valid hard negatives and not false positives
    Core of the method from [12]; interpolation and extrapolation of normalized features can land near actual positives. The paper gives no validation of the synthesized points beyond final accuracy.
  • domain assumption Relay Diffusion [22] clones of ImageNet-100 preserve the semantic class structure needed for transfer
    The 130k synthetic set is used as pretraining data; Figure 3 shows fully synthetic pretraining drops 10+ points, so fidelity is imperfect.
  • domain assumption ImageNet-100 linear probing measures representation quality and transferability
    This is the only evaluation protocol; no downstream transfer tasks are run despite the abstract claiming 'transferable' representations.
  • standard math MoBY/BYOL momentum-updated target encoder (Eq. 2) behaves as expected
    Standard MoBY/BYOL machinery, inherited unchanged from [29, 13].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fake & Square: Training Self-Supervised Vision Transformers with Synthetic Data and Synthetic Hard Negatives." pith.science (2026). https://pith.science/paper/OCBE7EYV

@misc{pith2026250902029,
  author       = {Pith},
  title        = {Pith review of: Fake & Square: Training Self-Supervised Vision Transformers with Synthetic Data and Synthetic Hard Negatives},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OCBE7EYV}},
  note         = {Machine review of arXiv:2509.02029}
}
read the original abstract

This paper does not introduce a new method per se. Instead, we build on existing self-supervised learning approaches for vision, drawing inspiration from the adage "fake it till you make it". While contrastive self-supervised learning has achieved remarkable success, it typically relies on vast amounts of real-world data and carefully curated hard negatives. To explore alternatives to these requirements, we investigate two forms of "faking it" in vision transformers. First, we study the potential of generative models for unsupervised representation learning, leveraging synthetic data to augment sample diversity. Second, we examine the feasibility of generating synthetic hard negatives in the representation space, creating diverse and challenging contrasts. Our framework - dubbed Syn2Co - combines both approaches and evaluates whether synthetically enhanced training can lead to more robust and transferable visual representations on DeiT-S and Swin-T architectures. Our findings highlight the promise and limitations of synthetic data in self-supervised learning, offering insights for future work in this direction.

Figures

Figures reproduced from arXiv: 2509.02029 by the authors.

Figure 1
Figure 1. Overview of our pipeline for contrastive learning. It [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. DeiT Grad-CAMs. Left: input image. Middle: pretraining [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. Top-1 linear probing accuracy on ImageNet-100. We pretrain for 100 epochs on real data ( [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 26 canonical work pages

  1. [1]

    A theoretical analysis of contrastive unsupervised representation learning, 2019

    Sanjeev Arora, Hrishikesh Khandeparkar, Mikhail Khodak, Orestis Plevrakis, and Nikunj Saunshi. A theoretical analysis of contrastive unsupervised representation learning, 2019. 2

    2019

  2. [2]

    This dataset does not exist: training models from generated images

    Victor Besnier, Himalaya Jain, Andrei Bursuc, Matthieu Cord, and Patrick P´erez. This dataset does not exist: training models from generated images. In ICASSP, 2020. 2

    work page 2020

  3. [3]

    Deep clustering for unsupervised learning of visual features, 2019

    Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features, 2019. 2

    work page 2019

  4. [4]

    Unsupervised learn- ing of visual features by contrasting cluster assignments

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learn- ing of visual features by contrasting cluster assignments. In Advances in Neural Information Processing Systems, pages 9912–9924. Curran Associates, Inc., 2020. 2

    work page 2020

  5. [5]

    Emerg- ing properties in self-supervised vision transformers, 2021

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´e J´egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerg- ing properties in self-supervised vision transformers, 2021. 2, 4

    work page 2021

  6. [6]

    A simple framework for contrastive learning of visual representations, 2020

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geof- frey Hinton. A simple framework for contrastive learning of visual representations, 2020. 1, 2, 3

    2020

  7. [7]

    An empirical study of training self-supervised vision transformers, 2021

    Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers, 2021. 2, 4

    work page 2021

  8. [8]

    When vision transformers outperform resnets without pre-training or strong data augmentations, 2022

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

    work page 2022

Show all 32 references
  1. [9]

    Li, and Li Fei-Fei

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, K. Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 4

    2009

  2. [10]

    An image is worth 16x16 words: Transformers for image recognition at scale, 2021

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

    2021

  3. [11]

    A review on discriminative self-supervised learning methods, 2024

    Nikolaos Giakoumoglou and Tania Stathaki. A review on discriminative self-supervised learning methods, 2024. 1, 2

    2024

  4. [12]

    Synco: Synthetic hard negatives in contrastive learning for better unsupervised visual representations, 2024

    Nikolaos Giakoumoglou and Tania Stathaki. Synco: Synthetic hard negatives in contrastive learning for better unsupervised visual representations, 2024. 2, 3, 4

    2024

  5. [13]

    Richemond, Elena Buchatskaya, Carl Do- ersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Moham- mad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, R´emi Munos, and Michal Valko

    Jean-Bastien Grill, Florian Strub, Florent Altch ´e, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Do- ersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Moham- mad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, R´emi Munos, and Michal Valko. Bootstrap your own ...

    2020

  6. [14]

    Momentum contrast for unsupervised visual repre- sentation learning, 2020

    Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual repre- sentation learning, 2020. 1, 2, 3

    2020

  7. [15]

    Masked autoencoders are scalable vision learners, 2021

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners, 2021. 2

    2021

  8. [16]

    Hard negative mixing for contrastive learning, 2020

    Yannis Kalantidis, Mert Bulent Sariyildiz, Noe Pion, Philippe Weinzaepfel, and Diane Larlus. Hard negative mixing for contrastive learning, 2020. 1, 2, 3

    2020

  9. [17]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410, 2019. 2

    2019

  10. [18]

    Self-supervised learning: Generative or contrastive

    Xiao Liu, Fanjin Zhang, Zhenyu Hou, Liming Mian, Zhaoyu Wang, Jing Zhang, and Jie Tang. Self-supervised learning: Generative or contrastive. IEEE Transactions on Knowledge and Data Engineering, 35(1):21–40, 2021. 1

    2021

  11. [19]

    Swin transformer: Hierarchical vision transformer using shifted windows, 2021

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows, 2021. 2, 4

    2021

  12. [20]

    Synthetic data for deep learning

    Sergey I Nikolenko. Synthetic data for deep learning. arXiv preprint arXiv:1909.11512, 2021. 2

    1909 arXiv

  13. [21]

    Fake it till you make it: Learning trans- ferable representations from synthetic imagenet clones

    Mert Bulent Sariyildiz, Karteek Alahari, Diane Larlus, and Yannis Kalantidis. Fake it till you make it: Learning trans- ferable representations from synthetic imagenet clones. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition (CVPR), 2023. 2

    2023

  14. [22]

    Relay diffusion: Unifying diffusion process across resolutions for image syn- thesis

    Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jian- qiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image syn- thesis. In The Twelfth International Conference on Learning Representations, 2024. 3, 4

    2024

  15. [23]

    Contrastive multiview coding, 2020

    Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding, 2020. 4

    2020

  16. [24]

    Learning vision from models rivals learning vision from data

    Yonglong Tian, Lijie Fan, Kaifeng Chen, Dina Katabi, Dilip Krishnan, and Phillip Isola. Learning vision from models rivals learning vision from data. arXiv preprint arXiv:2312.17742, 2023. 2

    2023 arXiv

  17. [25]

    Training data-efficient image transformers & distillation through atten- tion, 2021

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv´e J´egou. Training data-efficient image transformers & distillation through atten- tion, 2021. 2, 4

    2021

  18. [26]

    Repre- sentation learning with contrastive predictive coding, 2019

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Repre- sentation learning with contrastive predictive coding, 2019. 3

    2019

  19. [27]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998–6008, 2017. 2

    2017

  20. [28]

    Un- supervised feature learning via non-parametric instance-level discrimination, 2018

    Zhirong Wu, Yuanjun Xiong, Stella Yu, and Dahua Lin. Un- supervised feature learning via non-parametric instance-level discrimination, 2018. 3

    2018

  21. [29]

    Self-supervised learning with swin transformers, 2021

    Zhenda Xie, Yutong Lin, Zhuliang Yao, Zheng Zhang, Qi Dai, Yue Cao, and Han Hu. Self-supervised learning with swin transformers, 2021. 2, 4

    2021

  22. [30]

    Decoupled contrastive learning, 2022

    Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. Decoupled contrastive learning, 2022. 3

    2022

  23. [31]

    Understanding hard nega- tives in noise contrastive estimation, 2021

    Wenzheng Zhang and Karl Stratos. Understanding hard nega- tives in noise contrastive estimation, 2021. 1

    2021

  24. [32]

    ibot: Image bert pre-training with online tokenizer, 2022

    Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer, 2022. 2

    2022

Pith tools

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