Pith. sign in

REVIEW 4 major objections 7 minor 1 cited by

An Empirical Study of Autoregressive Pre-training from Videos

T0 review · 4 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Autoregressive next-token prediction over a trillion visual tokens yields visual representations competitive with self-supervised baselines across image, video, tracking, and robot tasks.

desk verdict A large, honest empirical study of autoregressive video pre-training with real scaling data, but the headline comparison is not controlled and the tokenizer table is misleading; still worth referee time. read the letter →

arxiv 2501.05453 v1 pith:AXLK4TV7 submitted 2025-01-09 cs.CV cs.AI

Jathushan Rajasegaran , Ilija Radosavovic , Rahul Ravishankar , Yossi Gandelsman , Christoph Feichtenhofer , Jitendra Malik This is my paper · ORCID
classification cs.CVcs.AI
keywords autoregressivepre-trainingvideorepresentationlearningnext-tokenpredictionvisualtokensself-supervisedscalinglawsactionrecognitionrobotmanipulation
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 attempts to establish that the language-model recipe—autoregressive next-token prediction over a long sequence of discrete tokens—transfers to vision when the tokens come from videos and images. The authors build a family of causal transformers, called Toto, train them on over one trillion visual tokens from ImageNet, Kinetics-600, Ego4D, and HowTo100M, and show the resulting features transfer to image classification, video action recognition, action forecasting, tracking, object-permanence reasoning, and robot manipulation with little or no task-specific design. If these results hold, video becomes a large, unlabeled data source that can be mined with the same scaling recipe that drove language models, rather than requiring handcrafted inductive biases or labels. The paper also claims a compute-scaling law for visual next-token prediction, $L(C)=7.32 \cdot C^{-0.0378}$, which increases with compute but at a slower rate than the GPT-3 language-model exponent.

What carries the argument

The load-bearing object is the Toto model family: a causal transformer with LLaMA-style blocks—RMSNorm, SwiGLU activations, and RoPE positional embeddings—trained on sequences of 4096 discrete visual tokens produced by a fixed dVAE tokenizer with an 8K vocabulary. Videos (16 frames) and images (16 random images) are flattened into the same token sequence by raster-scanning each frame, with start and end tokens marking modality, and the training loss is the negative log-likelihood of each next token. This setup unifies image and video data, removes labels and almost all inductive bias, and makes the model a drop-in autoregressive learner. For evaluation, the paper uses attention pooling—a learned query token cross-attending to intermediate tokens—rather than average pooling, because causal attention gives later tokens more context; probing peaks at about half the model depth across tasks and model sizes.

What would settle it

Retrain the Toto-1b recipe on the same one-trillion-token mixture but shuffle the token order within each video, or replace causal next-token prediction with a bidirectional predictive objective, keeping compute fixed; if ImageNet and Kinetics-400 probe accuracies stay within one or two points, the temporal autoregressive order is not the source of the representations. Alternatively, train 3B- and 5B-parameter Toto models with the same mixture and check whether validation loss follows $L(C)=7.32 \cdot C^{-0.0378}$; a clear bend at larger compute would falsify the claimed scaling exponent.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a causal transformer trained solely to predict the next visual token, in raster-scan order over independently dVAE-tokenized frames, learns general visual representations. On ImageNet, attention-probing the middle layer of the 1.1B-parameter model reaches 75.3% top-1, the best reported for autoregressive generative pretraining at comparable size and above iGPT-XL's 72.0% despite using one-sixth of the parameters; on Kinetics-400 it reaches 74.4%, on DAVIS semi-supervised tracking the J&F score reaches 62.4% at 512 resolution, and on CATER object localization it beats task-specific baselines at both 16 and 32 temporal resolutions. The same frozen features make robot manipulation policies learn faster in simulation than MAE-pretrained baselines and reach 63% success on a real Franka cube-picking task. The authors interpret these results as evidence that minimal-inductive-bias next-token prediction is a competitive paradigm for visual representation learning, and that in decoder-only vision models the middle layers act as an implicit encoder while the later layers reconstruct token space.

Load-bearing premise

The central claim assumes the fixed dVAE tokenizer is a fair, label-free front end whose information loss does not determine the downstream gap; if the tokenizer—rather than the next-token objective—is what sets representation quality, the comparisons against masked and contrastive baselines would not isolate autoregressive pretraining.

Editorial extensions

If this is right

  • Autoregressive pre-training over visual tokens is a viable label-free alternative to masked autoencoding and contrastive methods; at 1.1B parameters it outperforms much larger iGPT and comes within range of MAE and DINO on image benchmarks.
  • The unified image-and-video token-sequence format means one pretrained backbone can serve both image and video benchmarks, including egocentric forecasting, without separate video-specific pretraining.
  • Video-derived features are useful beyond recognition: frozen Toto representations support label propagation for tracking and accelerate reinforcement-learning policies, so generative video pretraining can feed control stacks.
  • Scaling behavior, while slower than language modeling, is monotone and predictable, so the practical bottleneck for these models is compute and tokenizer quality rather than the next-token objective itself.

Reading between the lines

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

  • The paper leaves implicit that its own limitation section—representation quality is bounded by the tokenizer—cuts against the strong reading that the next-token objective alone is responsible; a clean test is to repeat the same suite with an end-to-end learned tokenizer and see whether the gap to discriminative baselines closes.
  • The consistent middle-layer peak hints that decoder-only visual transformers organize themselves into an encoder-like first half and a reconstruction-oriented second half; practitioners could exploit this by attaching task heads at half depth, and a testable hypothesis is that this split sharpens with scale.
  • Because video frames are redundant, and the appendix's loss-per-token curve shows the first frame carries the bulk of the loss, removing redundancy through frame subsampling, adaptive token rates, or temporal compression might yield a steeper scaling exponent than the reported $L(C)=7.32 C^{-0.0378}$; that is an untested extension of the paper's own observation.
  • The robot results, obtained with a small frozen backbone, suggest physical priors emerge from generative video pretraining; if true, larger autoregressive video models should improve manipulation without any robot-specific data.
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 / 7 minor

Summary. The paper presents Toto, a family of decoder-only transformers pre-trained autoregressively on discrete visual tokens from images and videos, with roughly one trillion tokens of mixed ImageNet/Kinetics/Ego4D/HowTo100M data. The authors study design choices (tokenizer, resolution, pooling, architecture) and evaluate the learned representations on image classification, video action recognition, action forecasting, DAVIS tracking, CATER object permanence, and robotic manipulation. They report that, despite minimal inductive biases, the models achieve competitive accuracy across tasks, and they fit a power-law relationship between validation loss and compute, claiming a slower scaling rate than GPT-3-style language modeling.

Significance. If the claims were fully supported, this would be a useful empirical data point for the field: it would show that large-scale autoregressive next-token prediction on video can yield transferable visual representations across a broad task suite, and it would characterize how such models scale. The paper is unusually broad in its evaluation and includes honest limitations (Section 5) and several controlled comparisons in the appendix, notably the linear-probe comparison with iGPT in A.4 and the same-protocol K400 comparison in A.7. These are genuine strengths. However, the central 'competitive across all benchmarks' claim is weakened by evaluation-protocol mismatches and by the informal scaling-law analysis, so the paper requires substantial additional experimentation or careful reframing before it can be accepted.

major comments (4)
  1. [Section 4.2 / Table 7 / Appendix A.4] The ImageNet comparison is not controlled. The Toto numbers in Table 7 are obtained with attention pooling plus self-supervised next-token fine-tuning on ImageNet at 32x32 token resolution with retuned RoPE, while the cited baselines (MAE, DINOv2, iGPT) are mostly frozen-feature linear probes or their original supervised protocols. The paper's own controlled linear-probe experiment in Table 14 gives Toto-1b only 66.2%, which is 1.0 point above iGPT-L and far below the MAE/DINOv2 numbers quoted in Table 7. To support the claim that autoregressive video pre-training yields competitive ImageNet representations, the authors should either evaluate the baselines under the same fine-tuning/attention-pooling protocol or restrict the claim to the controlled comparison.
  2. [Section 4.5 / Table 10] The DAVIS tracking comparison confounds model quality with input resolution. Toto-large at 512/8 achieves 62.4 J&F, but the DINO and MAE baselines are shown only at 224 resolution. At a more comparable setting, Toto-large at 256/8 is 44.8, which is below DINO-base at 224/8 (54.3). The statement that 'at large resolution (512), it outperforms all methods' is therefore an artifact of the resolution difference rather than a demonstrated advantage of the learned representation. A same-resolution baseline should be added, or the claim should be qualified accordingly.
  3. [Section 4.9 / Figure 9 / Table 15] The scaling-law exponent is not sufficiently supported. The power law L(C) = 7.32 * C^-0.0378 is fit to six models trained on a single fixed data mixture with one learning rate, and no confidence intervals, fit residuals, or sensitivity analysis are reported. The comparison to the GPT-3 exponent is also informal because the two losses are on different tokenizers and data distributions. Please report the fitting procedure, the number of points, uncertainty estimates, and a check of whether the exponent is stable across data mixtures or learning rates; otherwise the result should be described as a qualitative trend, not a quantitative scaling law.
  4. [Section 4.6 / Table 11] The robotics result is presented as 'on-par' and 'comparable', but Table 11 reports 63% success for Toto-base versus 75% for MVP. With 16 trials, the difference is within binomial noise, but the wording overstates the evidence. The authors should either report confidence intervals or phrase this as 'within the resolution of this small trial count', and they should avoid using the word 'on-par' without statistical support.
minor comments (7)
  1. [Section 3] The phrase 'casual transformer' should be 'causal transformer'.
  2. [Section 2] 'Mamaba Gu & Dao (2023)' should be 'Mamba (Gu & Dao, 2023)'.
  3. [Section 4.7] The text says Toto-large achieves '62.8% and 70.9%' for 16 and 32 frames, but Table 12 reports 72.9% for 32 frames. The table and text should be reconciled.
  4. [Section 4.7] The unrendered reference 'Table??shows' should be replaced with the actual table number.
  5. [Section 3.4] 'per-trained' should be 'pre-trained'.
  6. [Section 4.1 / Table 3] It would be helpful to state explicitly which layer is used for the linear-probing comparison in Table 3, since the text says 'optimal intermediate layer' but does not give the layer number for each tokenizer.
  7. [Section 4.9 / Figure 9] The caption of Figure 9 says 'with optimal learning rates', but the text and Figure 11 state that a single optimal learning rate of 2^-7 is used for all widths. Please clarify whether the learning rate was per-model or shared.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the pretraining-to-evaluation chain is externally benchmarked, and the scaling curve is a fitted empirical summary rather than a derivation from its own inputs.

full rationale

The paper's claimed derivation is an empirical pipeline: minimize the next-token negative log-likelihood in Eq. (2) on tokenized image/video data, then transfer the learned representations to external benchmarks (ImageNet, Kinetics-400, Ego4D, DAVIS, CATER, robot manipulation). None of Tables 7-12 feeds a downstream result back into the pretraining objective, so there is no self-definitional or fitted-input-called-prediction step. The dVAE tokenizer is an upstream fixed preprocessing choice, but the paper does not define its target metric in terms of the tokenizer and explicitly lists the tokenizer bound as a limitation (Section 5); this is an external dependency, not circularity. The Section 4.9 scaling law L(C) = 7.32*C^-0.0378 is a power-law fit to the authors' own validation-loss-versus-compute data. It is presented as an observed empirical curve ('Toto shows a power law'), not as a prediction of an external quantity; fitting a curve to one's own measurements is not circular by this rubric. The self-citations (e.g., MVP protocols and robot demonstrations in Section 4.6) are used as baselines/protocol references, not as load-bearing justification for the central claim, and no uniqueness theorem is imported from the authors' prior work. The skeptic's point that Toto's ImageNet/K400 numbers use additional fine-tuning and attention pooling while some baselines use frozen linear probes is an evaluation-protocol validity concern, not a circularity of the derivation chain; it does not raise the circularity score.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The ledger contains two fitted quantities: the scaling-law exponent and the data mixture ratios. The axioms are standard domain assumptions for self-supervised vision pre-training. The paper introduces no new theoretical entities.

free parameters (2)
  • Scaling-law coefficient and exponent = 7.32 and -0.0378
    Fit to validation loss versus compute for six mu-Parametrized models in Figure 9. The exponent is the central quantitative claim of Section 4.9 and is presented without confidence intervals.
  • Data mixture ratios = 20% ImageNet, 10% Ego4D, 10% Kinetics, 60% HowTo100M
    Chosen by the authors without ablations over mixing ratios. The downstream results depend on this choice, and the paper does not test alternatives.
assumptions (4)
  • domain assumption The dVAE tokenizer provides a faithful, label-free discretization of images and videos.
    The whole pre-training pipeline operates on dVAE tokens. The paper acknowledges in Section 5 that representation quality is bounded by the tokenizer, but the downstream comparisons assume the tokenizer is not the decisive factor.
  • domain assumption Validation loss on the pre-training mixture is a meaningful proxy for downstream representation quality.
    Section 4.9 uses validation loss to fit the scaling law and to compare with language models, but the paper does not show that this loss correlates with ImageNet or Kinetics accuracy across the six scaling models.
  • domain assumption Raster-scan ordering of patches is a reasonable way to impose a sequence on images and videos.
    The paper follows iGPT and PixelCNN in Section 3.1. The redundancy issues with this ordering are discussed in Appendix A.1.
  • domain assumption Attention pooling at a middle layer is a fair way to extract representations.
    Section 3.5 and Figure 4 justify this for decoder-only models, but the choice of probe layer is selected on the validation set for each task, which can inflate reported accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Empirical Study of Autoregressive Pre-training from Videos." pith.science (2026). https://pith.science/paper/AXLK4TV7

@misc{pith2026250105453,
  author       = {Pith},
  title        = {Pith review of: An Empirical Study of Autoregressive Pre-training from Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AXLK4TV7}},
  note         = {Machine review of arXiv:2501.05453}
}
read the original abstract

We empirically study autoregressive pre-training from videos. To perform our study, we construct a series of autoregressive video models, called Toto. We treat videos as sequences of visual tokens and train transformer models to autoregressively predict future tokens. Our models are pre-trained on a diverse dataset of videos and images comprising over 1 trillion visual tokens. We explore different architectural, training, and inference design choices. We evaluate the learned visual representations on a range of downstream tasks including image recognition, video classification, object tracking, and robotics. Our results demonstrate that, despite minimal inductive biases, autoregressive pre-training leads to competitive performance across all benchmarks. Finally, we find that scaling our video models results in similar scaling curves to those seen in language models, albeit with a different rate. More details at https://brjathu.github.io/toto/

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Poly-Autoregressive Prediction for Modeling Interactions

    cs.CV 2025-02 conditional novelty 4.0 of 10

    A single transformer training recipe, poly-autoregressive prediction, improves multi-agent ego forecasting over autoregressive baselines on three distinct tasks.

Reference graph

Works this paper leans on

29 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [2]

    Figure 11 µ-Parameterization Learning Rate: We show that µ-Parameterization (Yang et al., 2022), we can train all width Toto models, with an single optimal learning rate of 2−7

    Model Params Dimension Heads Layers a1 14.8M 256 16 12 a2 77.2M 512 16 16 a3 215M 768 16 20 a4 458M 1024 16 24 a5 1.2B 1536 16 28 a6 1.9B 1792 16 32 Table 15 Toto Varients: We scaleToto models by increasing hidden dimension and number of layers linearly while keeping number of heads constant following (Yang et al., 2022; Touvron et al., 2023). Figure 11 µ...

    work page 2022

  2. [3]

    A short note on the kinetics-700 human action dataset

    Joao Carreira, Eric Noland, Chloe Hillier, and Andrew Zisserman. A short note on the kinetics-700 human action dataset. arXiv preprint arXiv:1907.06987,

    arXiv 1907

  3. [9]

    Mamba: Linear-time sequence modeling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752,

    arXiv

  4. [10]

    Scaling laws for autoregressive generative modeling.arXiv preprint arXiv:2010.14701,

    Tom Henighan, Jared Kaplan, Mor Katz, Mark Chen, Christopher Hesse, Jacob Jackson, Heewoo Jun, Tom B Brown, Prafulla Dhariwal, Scott Gray, et al. Scaling laws for autoregressive generative modeling.arXiv preprint arXiv:2010.14701,

    arXiv 2010

  5. [11]

    Perceptual losses for real-time style transfer and super-resolution

    13 Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pp. 694–711. Springer,

    work page 2016

  6. [13]

    Network in network.arXiv preprint arXiv:1312.4400,

    Min Lin, Qiang Chen, and Shuicheng Yan. Network in network.arXiv preprint arXiv:1312.4400,

    arXiv

  7. [15]

    In-context learning and induction heads.arXiv preprint arXiv:2209.11895,

    Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. In-context learning and induction heads.arXiv preprint arXiv:2209.11895,

    arXiv

  8. [16]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193,

    arXiv

Show all 29 references
  1. [18]

    Video (language) modeling: a baseline for generative models of natural videos.arXiv preprint arXiv:1412.6604,

    MarcAurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos.arXiv preprint arXiv:1412.6604,

    arXiv

  2. [21]

    Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556,

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556,

    arXiv

  3. [22]

    Llama: Open and efficient foundation language models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971,

    arXiv

  4. [24]

    Scaling autoregressive video models.arXiv preprint arXiv:1906.02634,

    Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models.arXiv preprint arXiv:1906.02634,

    1906 arXiv

  5. [25]

    Masked visual pre-training for motor control.arXiv preprint arXiv:2203.06173,

    Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control.arXiv preprint arXiv:2203.06173,

    arXiv

  6. [26]

    Tfcnet: Temporal fully connected networks for static unbiased temporal reasoning.arXiv preprint arXiv:2203.05928,

    Shiwen Zhang. Tfcnet: Temporal fully connected networks for static unbiased temporal reasoning.arXiv preprint arXiv:2203.05928,

    arXiv

  7. [27]

    Vision mamba: Efficient visual representation learning with bidirectional state space model.arXiv preprint arXiv:2401.09417,

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model.arXiv preprint arXiv:2401.09417,

    arXiv

  8. [28]

    to make the images and videos into a 1D sequence by scanning the patches in raster order. While this ordering allows for example to learn to predict the bottom half of the image from the top part of the image, in many places, the tokens can be predicted by interpolating rather...

    2017

  9. [1951]

    Glu variants improve transformer.arXiv preprint arXiv:2002.05202,

    Noam Shazeer. Glu variants improve transformer.arXiv preprint arXiv:2002.05202,

    2002 arXiv

  10. [1954]

    Beit: Bert pre-training of image transformers.arXiv preprint arXiv:2106.08254,

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers.arXiv preprint arXiv:2106.08254,

    arXiv

  11. [2013]

    Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101,

    arXiv

  12. [2015]

    Hiera: A hierarchical vision transformer without the bells-and- whistles

    Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and- whistles. arXiv preprint arXiv:2306.00989,

    arXiv

  13. [2016]

    The kinetics human action video dataset.arXiv preprint arXiv:1705.06950,

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset.arXiv preprint arXiv:1705.06950,

    arXiv

  14. [2017]

    Videomae v2: Scaling video masked autoencoders with dual masking

    Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14549–14560, 2023a. Rui Wang,...

    arXiv

  15. [2018]

    The 2017 davis challenge on video object segmentation.arXiv preprint arXiv:1704.00675,

    Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation.arXiv preprint arXiv:1704.00675,

    2017 arXiv

  16. [2019]

    Scalable pre-training of large autoregressive image models.arXiv preprint arXiv:2401.08541,

    Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Alexander Toshev, Vaishaal Shankar, Joshua M Susskind, and Armand Joulin. Scalable pre-training of large autoregressive image models.arXiv preprint arXiv:2401.08541,

    arXiv

  17. [2020]

    Data filtering networks.arXiv preprint arXiv:2309.17425,

    Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks.arXiv preprint arXiv:2309.17425,

    arXiv

  18. [2021]

    Language models are few-shot learners.arXiv preprint arXiv:2005.14165,

    Tom B Brown. Language models are few-shot learners.arXiv preprint arXiv:2005.14165,

    2005 arXiv

  19. [2022]

    Cater: A diagnostic dataset for compositional actions and temporal reasoning

    Rohit Girdhar and Deva Ramanan. Cater: A diagnostic dataset for compositional actions and temporal reasoning. arXiv preprint arXiv:1910.04744,

    1910 arXiv

  20. [2023]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pp. 248–255. Ieee,

    2009

  21. [2024]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 ieee. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12868–12878,

    2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.