Pith. sign in

REVIEW 3 major objections 5 minor 39 references

CDG-MAE: Cross-view Masked Modeling using Diffusion Generated Views

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CDG-MAE trains masked autoencoders on diffusion-generated views of static images, matching video-based pretraining on label propagation tasks.

desk verdict Diffusion-generated views plus multi-anchor masking give a real but modest win for cross-view MAE; the empirical result is credible, but the consistency metric behind the mechanism claim is unvalidated. read the letter →

arxiv 2506.18164 v2 pith:4PEALFSQ submitted 2025-06-22 cs.CV

Varun Belagali , Pierre Marza , Srikar Yellapragada , Zilinghan Li , Tarak Nath Nandi , Ravi K Madduri , Joel Saltz , Stergios Christodoulidis
show 2 more authors
Maria Vakalopoulou Dimitris Samaras
This is my paper
classification cs.CV
keywords self-supervisedlearningmaskedautoencoderdiffusion-generatedviewsdensecorrespondencevideolabelpropagationmulti-anchormaskingcross-viewpretrainingimage-only
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

CDG-MAE asks whether static images can substitute for videos when training a masked autoencoder to learn dense correspondences. The answer proposed is yes: for each ImageNet image, a self-supervised diffusion model generates a small bag of synthetic views with pose and perspective changes, and a Siamese MAE learns to reconstruct a heavily masked target view from one or more masked anchor views drawn from that bag. The paper also gives a quantitative way to score view pairs on global and local consistency, and shows that synthetic views score close to real video frames. With three anchors masked at 25%, CDG-MAE reaches 62.6 J&Fm on DAVIS and 38.1 mIoU on VIP, above the CropMAE crop-based reproduction and essentially on par with SiamMAE trained on Kinetics-400, despite never seeing video.

What carries the argument

The machinery is a bag of views plus a multi-anchor cross-view MAE. Each real ImageNet image is paired offline with M=4 views sampled from S-LDM, a latent diffusion model conditioned on a frozen self-supervised image encoder; these synthetic views are meant to vary pose, motion, and perspective while staying globally consistent. Training uses a weight-shared ViT encoder: the target is masked at 90%, each of N anchors is independently masked at ratio ra, and a decoder reconstructs the masked target by cross-attending to the concatenated anchor tokens. The paper's view-evaluation tool (global similarity, local similarity, nearest-patch similarity) is what justifies that S-LDM views behave like video frames rather than k-NN images or random pairs, and it is what guided the choice of generation model.

What would settle it

A decisive experiment would measure ground-truth correspondence accuracy between real images and their S-LDM views (for example, using a dataset with known camera poses or optical flow): if patches that the proxy local-similarity score calls matched are often not the same physical point, the pretext task would be teaching image-level co-occurrence rather than correspondence, and the central claim would not hold.

Watch

Extended reading notes

Core claim

The central claim is that diffusion-generated views carry the same correspondence-relevant signal as video frames. Concretely, when the target view is masked at 90% and the anchor is one of the synthetic views, reconstructing the target forces the encoder to match patches across views, so the learned features support training-free label propagation. The paper reports CDG-MAE-a3 (three anchors, 25% anchor masking) at 62.6 J&Fm on DAVIS, 38.1 mIoU on VIP, and 47.8 PCK@0.1 on JHMDB, outperforming the CropMAE reproduction (59.7, 33.8, 43.9) and matching or slightly exceeding SiamMAE (62.0, 37.3, 47.0), with a ViT-S/16 trained for only 100 epochs on ImageNet-1K. The authors interpret this as narrowing the gap to video-based MAE methods while keeping the data advantages of image-only pretraining.

Load-bearing premise

The method assumes the diffusion-generated views are alternative views of the same scene, so that reconstructing a masked synthetic view from the real image teaches dense correspondences; that assumption is supported only by proxy similarities from a pretrained MAE, not by ground-truth correspondences, and the authors note they cannot control which pose or motion changes occur.

Editorial extensions

If this is right

  • Image-only pretraining can close much of the gap to video-based correspondence learning, since synthetic views supply the pose and motion variation that crops lack.
  • The GS/LS/NPS evaluation can be used to choose which generative model, and which generated samples, are worth using for cross-view pretraining.
  • Going from one anchor to three anchors with 25% anchor masking improves all three downstream propagation tasks, while anchor masking keeps the added compute in check.
  • Choosing the target randomly between real and synthetic views performs as well as always using the real image, so synthetic views are usable as targets, not just anchors.
  • Halving the patch size from 16 to 8 gives CDG-MAE further gains on DAVIS, VIP, and JHMDB, and lets it surpass the conditioning encoder of the diffusion model that produced its data.

Reading between the lines

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

  • If generated views are filtered by the proposed consistency metrics rather than used indiscriminately, the same pipeline may yield larger gains; the paper does not filter views before training.
  • Because the pretext task is correspondence-centric, CDG-MAE features could transfer to optical flow, stereo matching, or point tracking, which the paper does not evaluate.
  • A controllable diffusion model that lets the user specify the amount of pose change could turn the bag of views into a curriculum of task difficulty, directly addressing the limitation the authors state.
  • The near-parity with a video-trained model suggests that the remaining gap may be due less to the data source and more to the lack of temporal consistency across generated views; adding a temporal or geometric constraint to the generation could be the next step.
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 / 5 minor

Summary. The paper proposes CDG-MAE, a self-supervised cross-view masked autoencoding method that trains on synthetic views generated from static ImageNet images by an image-conditioned self-supervised latent diffusion model (S-LDM from Gen-SIS). The method first creates an offline 'bag of views' for each image, introduces a quantitative view-consistency evaluation (GS, LS, NPS) to select the view generator, and extends the cross-view MAE to a multi-anchor setting with per-anchor masking. Downstream evaluation on video label propagation tasks (DAVIS, VIP, JHMDB) shows that CDG-MAE outperforms a reproduced CropMAE baseline and often matches or slightly exceeds SiamMAE trained on Kinetics-400, despite using only static images and training for 100 epochs.

Significance. If the central mechanism claim is established, this is a useful contribution: it offers an image-only pretraining route for video correspondence tasks, narrows the gap to video-based MAEs, and introduces a simple multi-anchor masking strategy that is well ablated (Table 4 reports mean±std over three seeds). The view-evaluation metrics could help the community select generative models for correspondence pretraining, and the inclusion of a comparison between Gen-SIS and RCG diffusion models (Tables 9-10) is a welcome partial validation of the metric's practical value. The paper is generally clearly written and the experiments are extensive for a 100-epoch ViT-S setting. However, the main empirical comparison lacks error bars, the view-consistency metrics are never validated against ground-truth correspondences, and the paper's own Limitations section acknowledges that the generated view variations are uncontrolled. These issues leave the core claim—that diffusion-generated views provide a genuine dense-correspondence signal—in need of stronger support.

major comments (3)
  1. [§3.2, Eqs. (1)-(3) and Table 1] The view-consistency metrics are presented as evidence that S-LDM views 'mimic video frames,' but they are never calibrated against ground-truth correspondences. LS compares features at identical grid positions, which conflates genuine local motion with global misalignment, and NPS is an average of maximum patch similarities that can be high for any two images sharing class-level texture, since MAE patch features are not instance-specific. The paper should validate these metrics on pairs with known geometric correspondences (e.g., homography-warped images, rendered views, or video frames with optical flow) and report, for example, how well the metrics discriminate true correspondence from unrelated pairs. Additionally, the description says a pretrained ViT-B/16 MAE is used to extract CLS tokens, but the MAE architecture in [16] does not include a [CLS] token; please clarify the exact global pooling procedure.
  2. [§5.3, Table 5] The main comparison table reports a single point per method without error bars, while the ablations in Table 4 include mean±std over three seeds. The headline claim that CDG-MAE-a3 outperforms the CropMAE reproduction (62.6 vs 59.7 J&Fm, 38.1 vs 33.8 mIoU, 47.8 vs 43.9 PCK0.1) needs the same treatment: at least the CDG-MAE variants and the CropMAE† baseline should report seed-averaged means and standard deviations, and the number of seeds for the reproduction should be stated. As written, the reader cannot assess whether the improvements are statistically reliable, especially since CropMAE† is a reproduction rather than a published number.
  3. [§3.1 and Limitations] The paper claims that S-LDM generates views 'mimicking the changes between video frames,' but also acknowledges in the Limitations that 'we cannot control which specific variations occur between generated views.' If a substantial fraction of generated views introduce new objects or rearrange parts rather than applying a true viewpoint/pose change, the reconstruction pretext can be solved by image-level co-occurrence statistics instead of dense instance-level correspondences. The authors should provide direct evidence of correspondence quality of the generated views, e.g., by manually annotating or automatically matching a subset of generated view pairs with ground-truth keypoints or flow, and report the fraction of views that preserve instance identity. This is necessary to support the claim that the method learns correspondences rather than merely benefiting from diverse augmentation.
minor comments (5)
  1. [§4] The text says 'We use a target masking ratio ra = 90%' but ra denotes the anchor masking ratio in §3.3; this should be rt = 90%. Also, in §5.1, 'such high ration performs poorly' should read 'such a high ratio performs poorly.'
  2. [Table 5] 'ImagNet-1K' should be 'ImageNet-1K', and the dataset name 'DAVIS' is inconsistently typeset as 'DA VIS' in the main text and captions.
  3. [Tables 2 and 3] These ablation tables do not report standard deviations or the number of seeds; please state whether the numbers are single runs or averaged over multiple seeds, as in Table 4.
  4. [Appendix A.3.1, Table 8] The N=4 rows are from a single seed, as stated in the text, but this is not marked in the table; add a footnote or symbol so the reader is not misled.
  5. [Eq. (10)] The loss expression contains typesetting artifacts in the PDF (the norm bars and the subtraction sign appear corrupted); ensure the final rendering is correct.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's downstream gains are measured on external benchmarks, and the diffusion-view choice is validated against an independent baseline (RCG).

full rationale

The paper's only notable self-reference is the use of Gen-SIS' S-LDM [4] as the view generator, but this is not a load-bearing circular step. The choice of S-LDM is supported by Table 9, which compares Gen-SIS S-LDM to RCG S-LDM, an independently developed method, using the proposed consistency metrics; more decisively, Table 10 shows that CDG-MAE trained with Gen-SIS views outperforms CDG-MAE trained with RCG views under identical settings. This is an external, empirical validation rather than a reduction to the paper's own assumptions. The view-consistency metrics in Table 1 (GS, LS, NPS) are uncalibrated proxies and do not by themselves prove correspondence, but they are used only as a screening tool, not as the source of the downstream predictions. The main results are obtained on external benchmarks (DAVIS, VIP, JHMDB) with a fixed evaluation protocol, so no fitted parameter is renamed as a prediction. The reconstruction loss in Eq. (10) is standard MSE, and the multi-anchor formulation in Eqs. (4)-(9) is a direct extension of SiamMAE and CropMAE; it does not smuggle the generated views back into the objective by construction. Controlled ablations (Tables 2-4) determine masking ratios and anchor counts empirically, not by definition. The acknowledged limitation that the authors cannot control which variations S-LDM produces weakens the causal interpretation of the pretext task, but this is a validity concern about the assumed correspondence signal, not circular reasoning. Thus, no part of the derivation chain reduces to its own inputs, and the circularity score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The main tunable knobs are M, rt, ra, and N; each is set by ablation or hand choice. The method borrows S-LDM from prior work, and the proxy metrics used to justify it are not externally validated. No new physical or theoretical entities are posited.

free parameters (4)
  • Number of generated views M = 4
    Chosen by the authors; no ablation over M is reported, though it directly controls training data size and the 48-hour generation cost.
  • Target masking ratio rt = 90%
    Selected by ablation in Table 3; 75% and 98.5% perform worse.
  • Anchor masking ratio ra = 25%
    Selected by ablation in Table 4; 0%, 25%, and 50% are tested, with 25% best for N=3.
  • Number of anchors N = 3
    Selected by ablation in Table 4; N=4 shows no further improvement.
assumptions (3)
  • domain assumption The S-LDM from Gen-SIS generates views that preserve the object identities and global scene layout of the conditioning image.
    The entire pretraining signal depends on generated views being usable as corresponding views; Table 1 supports this via proxy metrics but there is no ground-truth correspondence check. Stated in Section 3.1 and Figure 5.
  • domain assumption Local and global similarity measured with a pretrained ViT-B/16 MAE reflect the correspondence quality needed for label propagation.
    Section 3.2 defines GS, LS, and NPS on MAE features and uses them to select views, but the metric is never validated against human or geometric correspondences.
  • domain assumption The training-free label propagation protocol on DAVIS, VIP, and JHMDB is a valid proxy for dense correspondence quality.
    Used as the sole downstream evaluation, inherited from SiamMAE and CropMAE; no linear probing or other SSL evaluation is reported. Sections 4 and A.1.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CDG-MAE: Cross-view Masked Modeling using Diffusion Generated Views." pith.science (2026). https://pith.science/paper/4PEALFSQ

@misc{pith2026250618164,
  author       = {Pith},
  title        = {Pith review of: CDG-MAE: Cross-view Masked Modeling using Diffusion Generated Views},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4PEALFSQ}},
  note         = {Machine review of arXiv:2506.18164}
}
read the original abstract

Cross-view masked autoencoding has emerged as a powerful pretext task for learning dense correspondences, which are essential for applications such as video label propagation. The cross-view pretext task is modeled with a masked autoencoder, where a masked target view is reconstructed from an anchor view. However, acquiring effective training data remains a challenge - collecting diverse video datasets is costly, while simple image crops lack the necessary pose variations, underperforming video-based methods. This paper introduces CDG-MAE, a novel MAE-based self-supervised method that uses diverse synthetic views generated from static images via an image-conditioned diffusion model. We present a quantitative method to evaluate the local and global consistency of the generated views to choose the right diffusion model for cross-view self-supervised pretraining. These generated views exhibit substantial changes in pose and perspective, providing a rich training signal that overcomes the limitations of video and crop-based anchors. Furthermore, we enhance the standard single-anchor MAE setting to a multi-anchor masking strategy to increase the difficulty of the pretext task. CDG-MAE substantially narrows the gap to video-based MAE methods, while maintaining the data advantages of image-only MAEs.

Figures

Figures reproduced from arXiv: 2506.18164 by the authors.

Figure 1
Figure 1. CDG-MAE: We train a vision encoder in a self-supervised manner by finding correspon￾dences between real and synthetic views generated by a diffusion model. These synthetic views preserve important scene information while introducing diverse dynamics. Given the large availability of 2D images, a natural question arises: can we generate dynamic varia￾tions from images equivalent to those found in videos for correspond… view at source ↗
Figure 2
Figure 2. Overview of CDG-MAE: (a) For every real image, we generate M views using an off￾the-shelf S-LDM [4]. (b) We develop a quantitative tool to evaluate local and global consistencies between view pairs. (c) We develop a multi-anchor framework to train cross-view MAE. Having multiple anchors allows for anchor masking, which creates a more challenging pretext task. 3 Method Our approach, CDG-MAE, consists of three stages:… view at source ↗
Figure 3
Figure 3. Visualization of pose propagation across frames (frame:1 is provided with ground-truth [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Performance of SiamMAE, CropMAE and CDG-MAE with ViT-S/16 and ViT-S/8 on [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Bag of views visualization. Real denotes an image in ImageNet dataset. The views represent [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Visualization of label propagation using CDG-MAE ViT-S/16 on DAVIS [ [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 7
Figure 7. Figure 7: Visualization of label propagation using CDG-MAE ViT-S/16 on VIP [ [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Visualization of label propagation using CDG-MAE ViT-S/16 on JHMDB [ [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Performance of CropMAE when trained with 400 epochs schedule. We report the evaluation [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 24 canonical work pages

  1. [16]

    K. He, X. Chen, S. Xie, Y . Li, P. Dollár, and R. Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000–16009, 2022

    2022

  2. [1]

    Assran, Q

    M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y . LeCun, and N. Bal- las. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 15619–15629, 2023

    work page 2023

  3. [2]

    H. Bao, L. Dong, S. Piao, and F. Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021

    arXiv 2021

  4. [3]

    Bardes, J

    A. Bardes, J. Ponce, and Y . LeCun. Vicregl: Self-supervised learning of local visual features. Advances in Neural Information Processing Systems , 35:8799–8810, 2022

    work page 2022

  5. [4]

    Belagali, S

    V . Belagali, S. Yellapragada, A. Graikos, S. Kapse, Z. Li, T. N. Nandi, R. K. Madduri, P. Prasanna, J. Saltz, and D. Samaras. Gen-sis: Generative self-augmentation improves self- supervised learning, 2024

    work page 2024

  6. [5]

    Caron, H

    M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, and A. Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650–9660, 2021

    2021

  7. [6]

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning , pages 1597–1607. PmLR, 2020

    work page 2020

  8. [7]

    X. Chen, S. Xie, and K. He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640–9649, 2021

    work page 2021

Show all 39 references
  1. [8]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition , pages 248–255. Ieee, 2009

    2009

  2. [9]

    Dhariwal and A

    P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  3. [10]

    Esser, S

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, Z. English, K. Lacey, A. Goodwin, Y . Marek, and R. Rom- bach. Scaling rectified flow transformers for high-resolution image synthesis, 2024

    2024

  4. [11]

    Eymaël, R

    A. Eymaël, R. Vandeghen, A. Cioppa, S. Giancola, B. Ghanem, and M. Van Droogenbroeck. Efficient image pre-training with siamese cropped masked autoencoders. In ECCV, 2024

    2024

  5. [12]

    Feichtenhofer, Y

    C. Feichtenhofer, Y . Li, K. He, et al. Masked autoencoders as spatiotemporal learners.Advances in neural information processing systems , 35:35946–35958, 2022

    2022

  6. [13]

    Graikos, S

    A. Graikos, S. Yellapragada, M.-Q. Le, S. Kapse, P. Prasanna, J. Saltz, and D. Samaras. Learned representation-guided diffusion models for large-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 8532–8542, Jun...

    2024

  7. [14]

    Grill, F

    J.-B. Grill, F. Strub, F. Altché, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems , 33:21271–21284, 2020

    2020

  8. [15]

    Gupta, J

    A. Gupta, J. Wu, J. Deng, and F.-F. Li. Siamese masked autoencoders. NeurIPS, 2023

    2023

  9. [17]

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729–9738, 2020

    2020

  10. [18]

    Ho and T

    J. Ho and T. Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  11. [20]

    Jhuang, J

    H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. In Proceedings of the IEEE international conference on computer vision , pages 3192–3199, 2013

    2013

  12. [21]

    W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017

    2017 arXiv

  13. [22]

    B. F. Labs. Flux. https://github.com/black-forest-labs/flux , 2024

    2024

  14. [23]

    T. Li, D. Katabi, and K. He. Return of unconditional generation: A self-supervised representation generation method. Advances in Neural Information Processing Systems , 37:125441–125468, 2024

    2024

  15. [24]

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models, 2023

    2023

  16. [25]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  17. [26]

    Peebles and S

    W. Peebles and S. Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 4195–4205, 2023

    2023

  18. [28]

    Pont-Tuset, F

    J. Pont-Tuset, F. Perazzi, S. Caelles, P. Arbeláez, A. Sorkine-Hornung, and L. Van Gool. The 2017 davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017

    2017 arXiv

  19. [29]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  20. [30]

    J. Song, C. Meng, and S. Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  21. [31]

    Y . Tian, L. Fan, K. Chen, D. Katabi, D. Krishnan, and P. Isola. Learning vision from models rivals learning vision from data. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15887–15898, 2024. 11

    2024

  22. [32]

    Y . Tian, L. Fan, P. Isola, H. Chang, and D. Krishnan. Stablerep: Synthetic images from text- to-image models make strong visual representation learners. Advances in Neural Information Processing Systems, 36, 2024

    2024

  23. [33]

    Z. Tong, Y . Song, J. Wang, and L. Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35:10078–10093, 2022

    2022

  24. [34]

    C. Wei, H. Fan, S. Xie, C.-Y . Wu, A. Yuille, and C. Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14668–14678, 2022

    2022

  25. [35]

    Weinzaepfel, V

    P. Weinzaepfel, V . Leroy, T. Lucas, R. Brégier, Y . Cabon, V . Arora, L. Antsfeld, B. Chidlovskii, G. Csurka, and J. Revaud. Croco: Self-supervised pre-training for 3d vision tasks by cross-view completion. NeurIPS, 2022

    2022

  26. [36]

    Weinzaepfel, T

    P. Weinzaepfel, T. Lucas, V . Leroy, Y . Cabon, V . Arora, R. Brégier, G. Csurka, L. Antsfeld, B. Chidlovskii, and J. Revaud. Croco v2: Improved cross-view completion pre-training for stereo matching and optical flow. In ICCV, 2023

    2023

  27. [37]

    Z. Xie, Z. Zhang, Y . Cao, Y . Lin, J. Bao, Z. Yao, Q. Dai, and H. Hu. Simmim: A simple framework for masked image modeling. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9653–9663, 2022

    2022

  28. [38]

    Zhang, A

    L. Zhang, A. Rao, and M. Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF international conference on computer vision , pages 3836–3847, 2023

    2023

  29. [39]

    J. Zhou, C. Wei, H. Wang, W. Shen, C. Xie, A. Yuille, and T. Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021

    2021 arXiv

  30. [40]

    Q. Zhou, X. Liang, K. Gong, and L. Lin. Adaptive temporal encoding network for video instance-level human parsing. In Proceedings of the 26th ACM international conference on Multimedia, pages 1527–1535, 2018

    2018

  31. [41]

    Q. Zhou, X. Liang, K. Gong, and L. Lin. Adaptive temporal encoding network for video instance-level human parsing. In Proceedings of the 26th ACM international conference on Multimedia, pages 1527–1535, 2018. A Appendix A.1 Implementation details A.1.1 Training Our implementat...

    2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.