Pith. sign in

REVIEW 4 major objections 5 minor 35 references

Assessing the Use of Face Swapping Methods as Face Anonymizers in Videos

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Face swapping models can act as video face anonymizers, hiding identity while keeping videos smooth and realistic.

desk verdict A useful empirical benchmark of face-swapping-as-anonymization, but the temporal consistency metrics in Table I can be fooled by a static-face baseline, so the central claim is weaker than the abstract suggests. read the letter →

arxiv 2505.20985 v1 pith:WPXQ2JXD submitted 2025-05-27 cs.CV

Mustafa İzzet Muştu , Haz{i}m Kemal Ekenel This is my paper
classification cs.CV
keywords faceswappinganonymizationvideoprivacytemporalconsistencyidentityretrievalsyntheticfacesSimSwapDancer
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 asks whether face swapping models, originally built for entertainment, can serve as video face anonymizers by replacing each person's face with a synthetic source face. Using one fixed synthetic face per video from the FaceForensics++ dataset, the authors measure temporal consistency, anonymization strength, and content quality across four face swapping and two dedicated anonymization models. They find that face swapping models, especially SimSwap and FaceDancer, produce smoother and more natural anonymized videos than dedicated image anonymization models, while still reducing identity match. This matters because existing anonymization methods often distort video data or flicker across frames, whereas face swapping could preserve usable video quality and still hide identity.

What carries the argument

The pipeline pairs a single synthetic source face, selected from a 5,000-image set, with an existing face swapping model that transfers that identity onto every video frame. Anonymization strength is measured by identity retrieval against the original 889 identities enrolled with ArcFace embeddings, while temporal consistency is measured by SSIM between consecutive frames and by L2 distances between consecutive ArcFace embeddings. The central mechanism is this synthetic-source-plus-swap pairing, evaluated with metrics that compare anonymized videos to original videos; the choice of source face and the face swapping architecture together determine whether the output stays smooth and hides the original identity.

What would settle it

A stress test that feeds the original video and a frame-averaged or heavily temporally smoothed version of the same face through the same metrics; if the frozen-face baseline matches or beats SimSwap and FaceDancer on SSIM and embedding distance while clearly destroying expressions in human review, the temporal-consistency measure does not establish the claimed smoothness.

Watch

Extended reading notes

Core claim

The central claim is that off-the-shelf face swapping, driven by a single synthetic source face, works as a functional video anonymizer. In the authors' measurements, SimSwap and FaceDancer maintain frame-to-frame facial consistency close to or better than the original video, while REFace and G2Face achieve stronger identity hiding at the cost of realism and stability. Face swapping models also score better on Fréchet Video Distance than image-focused anonymization models, meaning their outputs look more like real video. The paper concludes that the face anonymization task can be performed using a face swapping approach, with a tradeoff between temporal consistency and anonymization strength across methods.

Load-bearing premise

The paper assumes that high SSIM and low ArcFace embedding distances between consecutive frames really capture good temporal behavior, but a model that simply froze or heavily smoothed the face would score well on both without preserving the original movement and expression.

Editorial extensions

If this is right

  • Face swapping models can serve as the backbone of practical video anonymization pipelines without additional training.
  • Models that already excel at temporal consistency, such as SimSwap and FaceDancer, could be retrained with a stronger anonymization objective to improve identity hiding while keeping smoothness.
  • Because SimSwap has a relatively low number of parameters, real-time video anonymization with face swapping is a plausible next step.
  • Dedicated image anonymizers such as G2Face and FAMS are less suited to video because they either flicker or fail to obscure identity, suggesting future anonymizer design should optimize temporal coherence directly.

Reading between the lines

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

  • Editorial inference: the reported temporal-consistency metrics would also reward a trivial solution, such as a face that is frozen or heavily smoothed across frames, so the evaluation would be stronger if it checked that the original facial dynamics and expressions survive anonymization.
  • Editorial inference: a natural stress test is to use the same synthetic source face across many videos of different people and verify that identity retrieval stays low, since a fixed source identity could leak information across anonymized videos.
  • Editorial inference: the observed tradeoff between temporal consistency and anonymization suggests a possible decoupled design where identity is swapped in a latent space that preserves expression trajectories, rather than generating every frame independently.
  • Editorial inference: because only one synthetic source face was used per video, varying the source face by age or gender, as the authors list for future work, could shift both realism and anonymization strength in ways the current single-source setup does not reveal.
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 using existing face-swapping models with a synthetic source face as a video face anonymization pipeline. It evaluates four face-swapping models (SimSwap, REFace, FaceDancer, E4S) and two dedicated anonymization models (G2Face, FAMS) on FaceForensics++ using temporal-consistency metrics (consecutive-frame SSIM and ArcFace embedding L2 distance), face/identity metrics (SPIGA landmarks and pose, ArcFace ID similarity, identity retrieval), and FVD for content quality. The reported results indicate that SimSwap and FaceDancer offer the best trade-off between temporal consistency and visual quality, that REFace and G2Face anonymize more strongly but with lower realism and consistency, and that FAMS fails to anonymize effectively. The paper concludes that face-swapping methods with synthetic source faces are suitable for privacy-preserving video applications.

Significance. If the empirical claims hold, the paper provides a useful benchmark and a practical recipe for video face anonymization, with an actionable ordering among six open-source methods. The choice to include both face-swapping and dedicated anonymization models, and to measure temporal, identity, and quality aspects separately, is valuable. The strongest aspect is that the evaluation is grounded in publicly available models and a standard dataset, which makes the comparison broadly reproducible in principle. However, the quantitative evidence is currently not strong enough to support the headline trade-off conclusion: the temporal-consistency metrics admit a trivial static-face solution, no variance or statistical comparisons are reported, and the retrieval and FVD protocols are underspecified. With additional baselines, per-video statistics, and protocol details, the contribution would be solid.

major comments (4)
  1. [§III-C1, Table I] The two temporal-consistency metrics—consecutive-frame SSIM and consecutive-frame ArcFace embedding L2 distance—are not sufficient to establish "consistent facial transitions" as claimed in the abstract. A baseline that outputs a single static swapped face pasted into every frame would score near 1.0 on SSIM and near 0 on embedding distance, thereby beating every reported method in Table I without preserving any expression, gaze, or pose dynamics. The paper does not report such a trivial baseline, nor does it report any dynamics-sensitive measure (e.g., per-frame landmark or pose trajectories, optical flow, or temporal derivative correlation) that would distinguish genuine transitions from freezing. FVD is a distribution-level statistic and would not identify per-video freezing either. Since the abstract and Section IV explicitly rely on Table I to recommend SimSwap and FaceDancer, this is a load-bearing gap. Please add a static-face baseline and per-video temporal-dynamics metrics, or explicitly reframe the claim as one about appearance stability only.
  2. [§IV, Tables I–IV] All conclusions are based on aggregate means over videos with no error bars, per-video variance, number of videos or frames, or statistical significance tests. For example, the recommendation that SimSwap and FaceDancer offer "the best trade-off" relies on differences in Table I (SSIM 0.9203 vs 0.9216; embedding distance 0.1667 vs 0.1954) that could be within noise when evaluated over 1000 videos or a subset thereof. Please report standard deviations or confidence intervals, the exact number of videos and sampled frames, and paired tests (e.g., Wilcoxon signed-rank) for the key comparisons; otherwise the ordering of the close methods is not statistically established.
  3. [§III-C2, Table III] The identity retrieval experiment is underspecified and lacks a calibrating baseline. It is not stated what constitutes retrieval (e.g., top-1 nearest neighbor in an ArcFace embedding space), whether the gallery contains one enrollment per identity or per video, which distance metric is used, or whether the same random synthetic source per video is fixed across methods. More importantly, there is no baseline retrieval rate when the original videos themselves are used as queries; without that upper bound, a score of 0.0228 for REFace cannot be interpreted as "strong anonymization" relative to the protocol. Please specify the retrieval protocol and add an original-vs-original baseline.
  4. [§III-B, §III-C3] The evaluation resizes all inputs to 224×224 to satisfy face-swapping models, but the six methods have different native operating resolutions, and this uniform resizing may disadvantage some methods; the paper itself speculates that FAMS's weak anonymization is caused by input resolution. This is a potential fairness confound for the headline comparison. Additionally, the FVD calculation is not reproducible as reported: the feature extractor, embedding layer, number of clips, and exact comparison protocol are not specified. Please justify the uniform resizing or report native-resolution results, and provide the FVD protocol details.
minor comments (5)
  1. [Table I] The heading "Embeddings µ" uses an undefined symbol; the text should define µ as the mean consecutive-frame L2 distance.
  2. [Figure 2] The caption and text should label each result row by method and clarify that all four face-swapping methods use the same synthetic source image; the current wording "the following four rows" is easy to misread.
  3. [§III-A6] The method name is typeset as "F AMS" in Section III-A6 but as "FAMS" elsewhere; please unify the notation.
  4. [§IV] The statement that FAMS's weak anonymization "is caused by the resolution of the input images" is presented without supporting ablation; please mark it explicitly as a hypothesis or remove it.
  5. [References] The reference for the synthetic face dataset "5k AI generated faces" should include a version/date and license information if available.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical evaluation against external benchmarks, and its conclusions are measurements rather than derivations from self-referential assumptions.

full rationale

The paper makes no fitted-parameter prediction and contains no derivation chain in which an output is defined in terms of an input. Its central claim, that face swapping with a synthetic source can act as a video face anonymizer, is supported by measured metrics: SSIM between consecutive frames, ArcFace embedding distances between consecutive frames, landmark and pose L2 errors against original frames, identity similarity and retrieval rates, and FVD. These are all external benchmarks or standard evaluation measures; none of them is constructed from the paper's own conclusions. The only self-citation is reference [18], the authors' prior work on text-guided face anonymization, mentioned in Related Work as one example of a diffusion-based pipeline. That citation is background context and is not used as evidence for the central claim, so it is not load-bearing. The reviewer's concern that high SSIM and low consecutive-frame embedding distance could also be achieved by a static or frozen face is a validity or failure-mode concern about the temporal consistency metric, not a circularity: the metric does not assume the paper's conclusion, and the paper's claim is not equivalent to the metric by construction. No equation in the paper reduces to its own input, no parameter is fitted to a subset and then renamed as a prediction, and no uniqueness theorem or ansatz is imported from the authors' own prior work to force a choice. The empirical evaluation is self-contained against external benchmarks, so the appropriate finding is no significant circularity.

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

The paper introduces no new entities or fitted models. Its central claims rest on experimental design choices (frame count, resolution, source sampling) and on assumptions that the chosen metrics and datasets measure privacy and quality. No parameter is fitted to data in the paper itself.

free parameters (3)
  • frame_count = 64
    Only the first 64 frames of each video are used; this is a hand-chosen limit that prevents assessment of long-term temporal consistency.
  • input_resolution = 224x224
    All frames are resized to 224x224 for compatibility with face swapping models. The authors attribute FAMS's poor anonymization to this resolution, so comparative conclusions are tied to this preprocessing choice.
  • source_faces_per_video = 1
    A single synthetic source face is randomly sampled per video. Anonymization strength may vary with source choice, and no multiple-seed averaging is reported.
assumptions (4)
  • domain assumption FaceForensics++ videos are representative of video anonymization use cases.
    The entire evaluation is performed on this dataset without discussion of domain shift to surveillance or meeting recordings.
  • domain assumption ArcFace embeddings are valid for measuring identity similarity and for identity retrieval in anonymized videos.
    Used in Section III-C for both temporal consistency and identity difference, but no validation that ArcFace is reliable on heavily edited or synthetic faces.
  • domain assumption SSIM between consecutive frames and FVD capture temporal consistency and visual quality as intended.
    Section III-C1 and III-C3 define these as quality metrics without showing they discriminate trivial solutions such as static faces.
  • domain assumption The synthetic faces from the Kaggle '5k AI generated faces' dataset do not correspond to real people.
    Source images in Section III-B are assumed to be safe replacement identities; the paper provides no verification that these faces are non-identifiable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Assessing the Use of Face Swapping Methods as Face Anonymizers in Videos." pith.science (2026). https://pith.science/paper/WPXQ2JXD

@misc{pith2026250520985,
  author       = {Pith},
  title        = {Pith review of: Assessing the Use of Face Swapping Methods as Face Anonymizers in Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WPXQ2JXD}},
  note         = {Machine review of arXiv:2505.20985}
}
read the original abstract

The increasing demand for large-scale visual data, coupled with strict privacy regulations, has driven research into anonymization methods that hide personal identities without seriously degrading data quality. In this paper, we explore the potential of face swapping methods to preserve privacy in video data. Through extensive evaluations focusing on temporal consistency, anonymity strength, and visual fidelity, we find that face swapping techniques can produce consistent facial transitions and effectively hide identities. These results underscore the suitability of face swapping for privacy-preserving video applications and lay the groundwork for future advancements in anonymization focused face-swapping models.

Figures

Figures reproduced from arXiv: 2505.20985 by the authors.

Figure 1
Figure 1. Proposed face anonymization pipeline. Our proposed pipeline can be seen in [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Qualitative anonymization results. The first row shows consecutive frames from an original video in FaceForensics++ dataset. The following four rows [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 25 canonical work pages

  1. [1]

    What is GDPR, the EU’s new data protection law? - GDPR.eu — gdpr.eu,

    “What is GDPR, the EU’s new data protection law? - GDPR.eu — gdpr.eu,” https://gdpr.eu/what-is-gdpr, [Accessed 12-02-2025]

    work page 2025

  2. [2]

    Face deidentification with generative deep neural networks,

    B. Meden, R. C. Mallı, S. Fabijan, H. K. Ekenel, V . ˇStruc, and P. Peer, “Face deidentification with generative deep neural networks,” IET Signal Processing , vol. 11, no. 9, pp. 1046–1054, 2017

    work page 2017

  3. [3]

    Anonymousnet: Natural face de-identification with measurable privacy,

    T. Li and L. Lin, “Anonymousnet: Natural face de-identification with measurable privacy,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , 2019, pp. 56– 65

    work page 2019

  4. [4]

    Deepprivacy: A generative adversarial network for face anonymization,

    H. Hukkel ˚as, R. Mester, and F. Lindseth, “Deepprivacy: A generative adversarial network for face anonymization,” in Advances in Visual Computing. Springer International Publishing, 2019, pp. 565–578

    work page 2019

  5. [5]

    Ciagan: Conditional identity anonymization generative adversarial networks,

    M. Maximov, I. Elezi, and L. Leal-Taixe, “Ciagan: Conditional identity anonymization generative adversarial networks,” in IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR) , June 2020

    work page 2020

  6. [6]

    G2face: High-fidelity reversible face anonymization via generative and geometric priors,

    H. Yang, X. Xu, C. Xu, H. Zhang, J. Qin, Y . Wang, P.-A. Heng, and S. He, “G2face: High-fidelity reversible face anonymization via generative and geometric priors,” IEEE Transactions on Information F orensics and Security, 2024

    work page 2024

  7. [7]

    Latent Diffusion Models for Attribute-Preserving Image Anonymization

    L. Piano, P. Basci, F. Lamberti, and L. Morra, “Latent diffusion models for attribute-preserving image anonymization,” arXiv preprint arXiv:2403.14790, 2024

    work page Pith review arXiv 2024

  8. [8]

    Face deidentification with controllable privacy protection,

    B. Meden, M. Gonzalez-Hernandez, P. Peer, and V . ˇStruc, “Face deidentification with controllable privacy protection,” Image and Vision Computing, vol. 134, p. 104678, 2023

    work page 2023

Show all 35 references
  1. [9]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in Neural Information Processing Systems , vol. 27, 2014

    2014

  2. [10]

    Deepprivacy2: Towards realistic full- body anonymization,

    H. Hukkel ˚as and F. Lindseth, “Deepprivacy2: Towards realistic full- body anonymization,” in 2023 IEEE/CVF Winter Conference on Appli- cations of Computer Vision (WACV) , 2023, pp. 1329–1338

    2023

  3. [11]

    U-net: Convolutional net- works for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional net- works for biomedical image segmentation,” in Medical Image Comput- ing and Computer-Assisted Intervention – MICCAI 2015 , N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi, Eds. Cham: Springer Internationa...

    2015

  4. [12]

    Analyzing and improving the image quality of stylegan,

    T. Karras, S. Laine, M. Aittala, J. Hellsten, J. Lehtinen, and T. Aila, “Analyzing and improving the image quality of stylegan,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 8107–8116

    2020

  5. [13]

    Cfa- net: Controllable face anonymization network with iden- tity representation manipulation,

    T. Ma, D. Li, W. Wang, and J. Dong, “Cfa- net: Controllable face anonymization network with iden- tity representation manipulation,” 2021. [Online]. Available: https://api.semanticscholar.org/CorpusID:238583157

    2021

  6. [14]

    Ganonymization: A gan-based face anonymization framework for preserving emotional expressions,

    F. Hellmann, S. Mertes, M. Benouis, A. Hustinx, T.-C. Hsieh, C. Conati, P. Krawitz, and E. Andr ´e, “Ganonymization: A gan-based face anonymization framework for preserving emotional expressions,” ACM Trans. Multimedia Comput. Commun. Appl. , vol. 21, no. 1, Dec

  7. [15]

    Ldfa: Latent diffusion face anonymization for self-driving applications,

    M. Klemp, K. R ¨osch, R. Wagner, J. Quehl, and M. Lauer, “Ldfa: Latent diffusion face anonymization for self-driving applications,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 3199–3205

    2023

  8. [16]

    High-resolution image synthesis with latent diffusion models,

    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, 2022, pp. 10 684–10 695

    2022

  9. [17]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” in IEEE International Conference on Computer Vision (ICCV) , 2023, pp. 3836–3847

    2023

  10. [18]

    Facial attribute based text guided face anonymization,

    M. ˙I. Mus ¸tu and H. K. Ekenel, “Facial attribute based text guided face anonymization,” in Proceedings of the Joint visuAAL- GoodBrother Conference on trustworthy video- and audio-based assistive technologies . Zenodo, Jun. 2024, pp. 50–55. [Online]. Available: https://doi.o...

    2024 doi

  11. [19]

    Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion,

    X. Ju, X. Liu, X. Wang, Y . Bian, Y . Shan, and Q. Xu, “Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion,” in European Conference on Computer Vision . Springer, 2024, pp. 150–168

    2024

  12. [20]

    Face anonymization made simple,

    H.-W. Kung, T. Varanka, S. Saha, T. Sim, and N. Sebe, “Face anonymization made simple,” in Proceedings of the Winter Conference on Applications of Computer Vision (WACV) , February 2025, pp. 1040– 1050

    2025

  13. [21]

    Simswap: An efficient framework for high fidelity face swapping,

    R. Chen, X. Chen, B. Ni, and Y . Ge, “Simswap: An efficient framework for high fidelity face swapping,” in MM ’20: The 28th ACM Interna- tional Conference on Multimedia , 2020

    2020

  14. [22]

    Realistic and efficient face swapping: A unified approach with diffusion models,

    S. Baliah, Q. Lin, S. Liao, X. Liang, and M. H. Khan, “Realistic and efficient face swapping: A unified approach with diffusion models,” in 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025, pp. 1062–1071

    2025

  15. [23]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” in International Conference on Learning Representations , 2021

    2021

  16. [24]

    Learning transferable visual models from natural language supervi- sion,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervi- sion,” in International Conference on Machine Learning , 2021

    2021

  17. [25]

    Facedancer: Pose- and occlusion-aware high fidelity face swapping,

    F. Rosberg, E. E. Aksoy, F. Alonso-Fernandez, and C. Englund, “Facedancer: Pose- and occlusion-aware high fidelity face swapping,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , January 2023, pp. 3454–3463

    2023

  18. [26]

    E4s: Fine-grained face swapping via editing with regional gan inversion,

    M. Li, G. Yuan, C. Wang, Z. Liu, Y . Zhang, Y . Nie, J. Wang, and D. Xu, “E4s: Fine-grained face swapping via editing with regional gan inversion,” arXiv preprint arXiv:2310.15081 , 2023

    2023 arXiv

  19. [27]

    Simswap++: Towards faster and high-quality identity swapping,

    X. Chen, B. Ni, Y . Liu, N. Liu, Z. Zeng, and H. Wang, “Simswap++: Towards faster and high-quality identity swapping,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 46, no. 1, pp. 576– 592, 2024

    2024

  20. [28]

    FaceForensics++: Learning to detect manipulated facial images,

    A. R ¨ossler, D. Cozzolino, L. Verdoliva, C. Riess, J. Thies, and M. Nießner, “FaceForensics++: Learning to detect manipulated facial images,” in International Conference on Computer Vision (ICCV), 2019

    2019

  21. [29]

    Sample and com- putation redistribution for efficient face detection,

    J. Guo, J. Deng, A. Lattas, and S. Zafeiriou, “Sample and com- putation redistribution for efficient face detection,” arXiv preprint arXiv:2105.04714, 2021

    2021 arXiv

  22. [30]

    Arcface: Additive angular margin loss for deep face recognition,

    J. Deng, J. Guo, N. Xue, and S. Zafeiriou, “Arcface: Additive angular margin loss for deep face recognition,” in 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2019, pp. 4685– 4694

    2019

  23. [31]

    5k ai generated faces,

    “5k ai generated faces,” https://www.kaggle.com/datasets/chelove4draste/5k- ai-generated-faces, 2022, [Accessed 13-02-2025]

    2022

  24. [32]

    Image quality assessment: from error visibility to structural similarity,

    Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Trans- actions on Image Processing , vol. 13, no. 4, pp. 600–612, 2004

    2004

  25. [33]

    Shape preserving facial landmarks with graph attention networks,

    A. Prados-Torreblanca, J. M. Buenaposada, and L. Baumela, “Shape preserving facial landmarks with graph attention networks,” in 33rd British Machine Vision Conference 2022, BMVC 2022, London, UK, November 21-24, 2022 . BMV A Press, 2022. [Online]. Available: https://bmvc2022.m...

    2022

  26. [34]

    Towards accurate generative models of video: A new metric & challenges,

    T. Unterthiner, S. Van Steenkiste, K. Kurach, R. Marinier, M. Michalski, and S. Gelly, “Towards accurate generative models of video: A new metric & challenges,” arXiv preprint arXiv:1812.01717 , 2018

    2018 arXiv

  27. [2024]

    Available: https://doi.org/10.1145/3641107

    [Online]. Available: https://doi.org/10.1145/3641107

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.