Pith. sign in

REVIEW 5 major objections 5 minor 38 references

Navigating Large-Pose Challenge for High-Fidelity Face Reenactment with Video Diffusion Model

T0 review · 5 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 A frozen video diffusion model's latent space can repair warped facial regions, enabling high-fidelity face reenactment at extreme poses.

desk verdict Solid engineering integration for large-pose face reenactment, but the headline claim outruns the evidence; needs a stronger evaluation before acceptance. read the letter →

arxiv 2507.16341 v1 pith:VVT2DP5T submitted 2025-07-22 cs.CV

Mingtao Guo , Guanyu Xing , Yanci Zhang , Yanli Liu This is my paper
classification cs.CV
keywords facereenactmentvideodiffusionmodellargeposevariationwarpingcorrectionimplicitkeypointsStabletalkingheadgenerationidentitypreservation
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 claims that warping artifacts under large pose changes can be repaired inside the latent space of a pretrained image-to-video diffusion model, rather than with a dedicated inpainting network. The proposed Face Reenactment Video Diffusion model (FRVD) extracts implicit facial keypoints from the source image and each driving frame, warps the source features toward the driving pose, and feeds the warped features through a Warping Feature Mapper into a frozen Stable Video Diffusion model. Inside that frozen latent space the distorted regions are reconstructed from the source identity and appearance priors, and a rectified guidance term shifts the predicted noise to preserve global appearance. Experiments on HDTF and CelebV-HQ report lower reconstruction and perceptual error, higher pose accuracy, and better human-rated identity and expression preservation than seven prior methods, with the largest gains in extreme-pose cases.

What carries the argument

The load-bearing mechanism is the Warping Feature Mapper (WFM) plus its Internal Feature Modulators (IFM), which adapt the warped source features so they can be added into the SVD backbone's per-layer feature maps by element-wise addition. The motion extractor supplies implicit keypoints (canonical keypoints plus estimated rotation, translation, and expression coefficients), and the warping module aligns the source feature to the driving frame; the WFM then projects this warped feature into the motion-aware latent space of Stable Video Diffusion. The rectified guidance $r$ acts as a mean-shift term in the denoising objective, and the cross-identity Motion Alignment Module re-references driving motion to the source's first-frame pose so identity does not leak.

What would settle it

Run FRVD on driving poses outside its training distribution (for example, near back-facing yaw) and compare the corrected outputs against a control where the fusion is replaced by channel-wise concatenation or where SVD is fine-tuned: visible geometric distortions or identity drift in those poses would indicate the frozen-latent-space compatibility does not hold where it is claimed to matter.

Watch

Extended reading notes

Core claim

The central discovery is that element-wise injection of motion-conditioned warped features into a frozen video diffusion latent space converts a degraded warping output into a high-fidelity reenactment. Formally, given the warped source features $F_s^{(i)}$, the Warping Feature Mapper produces modulated features $F_{s}^{(i)m}$; at layer $j$ of the SVD backbone these are fused as $F^{(j)}_{fuse} = F^{(j)} + F_{s}^{(i)m}$ and fed onward, while the final layer predicts a rectified guidance $r$ that shifts the Gaussian mean of the SVD output. With training loss Eq. (2) and classifier-free guidance Eq. (5), the method reconstructs regions lost to warping while keeping identity and temporal coherence. On the reported benchmarks it outperforms seven state-of-the-art reenactment methods on L1, PSNR, SSIM, LPIPS, POSE, EXP, FID, FVD and VQ, and wins a user study on all four perceptual axes; the ablation shows the WFM alone accounts for large improvements in every metric.

Load-bearing premise

The paper assumes that element-wise addition of WFM-modulated warped features into the frozen SVD feature maps yields a latent-space combination that SVD can interpret well enough to restore warped regions; this compatibility is supported chiefly by the ablation rather than by analysis of the two feature distributions.

Editorial extensions

If this is right

  • Large-pose reenactment no longer needs explicit landmark reliability: implicit keypoints carry motion, and the I2V prior repairs what warping destroys.
  • Because SVD is frozen and only WFM and IFM are trained, the approach inherits the temporal coherence and appearance priors of large-scale video diffusion without retraining a video generator.
  • The rectified guidance term can be read as a learned appearance-conditioned mean shift; the ablations tie it to color-fidelity gains in skin tone, lip color, and hair color.
  • On the reported HDTF benchmark, the method lowers L1 by 11.7 percent and LPIPS by 25.5 percent over the second-best baseline, and improves POSE and FID by 4.0 percent and 6.5 percent in the cross-identity setting.

Reading between the lines

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

  • If the latent-injection recipe transfers, the same WFM-style mapper could repair warping artifacts in other warping-based animation tasks, such as full-body pose transfer, where the frozen video prior already contains plausible motion.
  • The paper does not measure distribution alignment between $F^{(j)}$ and $F_{s}^{(i)m}$; a diagnostic such as feature covariance distances across poses would test whether the range of viable poses is bounded by the coverage of the SVD prior.
  • The reported 4 minutes per 100 frames suggests the same architecture would reach real-time use only after distillation, a limitation the paper itself concedes.
  • The ablation without WFM degenerates into a random I2V generator, implying the motion constraint carried by the warped features is doing most of the pose steering; a natural extension would expose the motion signal as an explicit conditioning channel and measure how much of the gain comes from motion versus appearance restoration.
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

5 major / 5 minor

Summary. The paper proposes FRVD, a face reenactment framework that combines implicit-keypoint-based motion warping with a frozen Stable Video Diffusion (SVD) backbone. A Motion Extractor estimates pose, expression, and canonical keypoints; a warping module aligns the source image features to the driving frame; and a Warping Feature Mapper (WFM) with Internal Feature Modulators (IFM) injects the warped features into SVD's latent space by element-wise addition, supported by a rectified guidance term. Training is self-supervised on VFHQ, freezing the Motion Extractor and SVD while updating only WFM/IFM. The authors report quantitative comparisons on HDTF and CelebV-HQ against seven baselines, qualitative comparisons, ablations of WFM, rectified guidance, and classifier-free guidance strength, plus a user study.

Significance. If the reported results are reliable, FRVD is a useful contribution: it shows how a frozen video diffusion prior can be adapted for large-pose face reenactment while training only a lightweight warping-feature mapper, and it compares against a broad set of modern baselines. The paper is clearly written, the framework is well described, and the ablation isolating WFM shows a very large effect. However, the central claim of superiority 'especially in challenging scenarios with extreme pose variations' is not yet established by the evidence: the quantitative evaluation is aggregate over HDTF and CelebV-HQ with no pose stratification, several leading margins are small point estimates without error bars, the guidance scale is tuned on the same benchmark, and the paper's own conclusion admits that extreme non-frontal views were not evaluated.

major comments (5)
  1. [Abstract; Sec. 1; Sec. 4.3; Sec. 6] The headline claim that FRVD is superior 'especially in challenging scenarios with extreme pose variations' is not quantitatively supported. Tables 1 and 2 report aggregate metrics over HDTF without any stratification by pose angle or pose discrepancy, and the qualitative evidence in Figs. 2–3 is a small set of selected frames. The conclusion explicitly states that 'further evaluation on extreme non-frontal views (e.g., back-facing poses) is warranted,' which is an admission that the regime named in the abstract was not systematically tested. The authors should either add an evaluation bucketed by pose magnitude or a dedicated large-pose benchmark, or temper the claim to match the evidence.
  2. [Sec. 4.3, Table 1] All objective metrics are single-run point estimates with no error bars, confidence intervals, or significance tests. Several headline margins are extremely small (SSIM 0.8702 vs. 0.8692; PSNR 27.708 vs. 27.582), so it is unclear whether these differences are meaningful. The authors should report standard deviations across multiple runs or across video subsets, and perform a significance test where appropriate.
  3. [Sec. 5.3, Table 5; Sec. 4.1] The classifier-free guidance scale w is selected by ablating on the same HDTF benchmark used for the main comparison, which inflates the apparent gains. Moreover, there is an internal inconsistency: Table 5 reports SSIM 0.8730 at w=2, but Table 1 reports SSIM 0.8702 for the same configuration; Sec. 4.1 states that inference uses a guidance scale of 2.5, a value that is not tested in Table 5. The guidance scale should be selected on a validation set, and the reported numbers for the exact inference setting should be reconciled.
  4. [Sec. 4.2, Table 2] The user study uses only 11 raters and reports no inter-rater reliability, per-item variance, or statistical significance. The claim of superiority in identity preservation (ID-User 4.375) also conflicts with the objective ID results in Table 1, where FRVD trails LivePortrait in both self-reenactment (0.8570 vs. 0.8967) and cross-identity reenactment (0.8975 vs. 0.9294). The authors should substantially expand the user study or report distributions and significance tests, and they should temper the identity-preservation claim in the abstract.
  5. [Sec. 3.3, Eqs. (2) and (5)] The core mechanism—that element-wise addition of WFM-modulated features into SVD's latent feature maps restores warping-degraded regions—is asserted rather than analyzed. The notation in Eq. (2) predicts ϵ with a minus sign on the noised latent, and Eq. (5) adds the rectified term r to the classifier-free guidance combination; the relationship between these two expressions and the standard DDIM/CFG sampling used in Sec. 3.4 should be clarified. More importantly, the paper provides no evidence that the WFM feature statistics are compatible with SVD's internal feature distribution, beyond the internal ablation of Table 3, where removing WFM also removes the motion condition entirely. A control experiment replacing the WFM features with random or independent features of the same shape would strengthen the claim that the mechanism is specifically about feature-space alignment.
minor comments (5)
  1. [Sec. 3.4, Eq. (3)] In Eq. (3), the expression alignment formula writes δi_s, which should be δ_s (the source expression coefficient) rather than a per-frame superscript; the same typo appears in the surrounding text.
  2. [Sec. 3.4] The index ranges are written as 'i=i' in the text (e.g., {R_i_d}_{i=i}^m); these should be i=1.
  3. [Sec. 5.3] The text says 'when the guidance strength exceeds 1, the performance metrics consistently degrade as w increases,' but Table 5 shows w=2 improving over w=1 on all four metrics; the statement should say 'exceeds 2.'
  4. [Sec. 4.1] The implementation details mention '8-bit Adam [16],' but reference [16] is the original Adam paper; the 8-bit optimizer is a specific implementation (e.g., bitsandbytes) and should be cited separately or described without the misleading reference.
  5. [Sec. 4.1; Sec. 5.3] Please state clearly whether the main experiments use w=2 or w=2.5, and ensure Table 1, Table 5, and the inference description are consistent.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation; the central claim is an empirical comparison against external benchmarks, with only a non-load-bearing self-citation.

full rationale

FRVD's core mechanism injects WFM-modulated warped features into the frozen SVD latent space via F_fuse = F_j + F_s^m and a learned rectified guidance r. This is not defined in terms of the quantity it predicts: the training objective in Eq. (2) targets the driving-frame latent z0, and the reported results are evaluated against independent HDTF and CelebV-HQ benchmarks and external baselines. The WFM ablation in Table 3 is an internal control rather than circular inference, because removing WFM makes the SVD backbone degenerate into unconstrained image-to-video generation, and the comparison directly measures the module's contribution. The only self-citation, reference [9] in the related-work list of explicit-keypoint methods, is not load-bearing: no central premise or uniqueness claim is justified by that citation. The guidance-strength sweep in Table 5 is performed on HDTF, and the paper itself concedes that 'further evaluation on extreme non-frontal views (e.g., back-facing poses) is warranted' (Sec. 6). These are external-validity concerns about the large-pose headline, not circularity: they weaken the strength of the empirical claim but do not make any prediction equivalent to its input. No fitted parameter is renamed as a prediction, no ansatz is smuggled in via self-citation, and no known result is merely renamed.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No physical or standalone entities are introduced. WFM and IFM are architectural components whose only evidence is the paper's internal ablation; I have not listed them as invented entities because they are not independent postulates.

free parameters (1)
  • Classifier-free guidance scale w = 2.0
    Selected from the ablation in Sec. 5.3 (Table 5) on the HDTF benchmark, then used in all final reported comparisons on the same benchmark. This is a hand-tuned value, not a verified held-out choice.
assumptions (4)
  • domain assumption The frozen SVD feature space is compatible with element-wise fusion of WFM features and can repair warped face regions while preserving identity.
    Core premise of Sec. 3.3, operationalized by F_fuse = F_j + F_s^m; the paper provides no alignment analysis or alternative fusion comparison.
  • domain assumption OSFV-style implicit keypoints and the warping module remain reliable enough under large pose changes to guide the diffusion model.
    Adopted in Sec. 3.2 from [26]; the large-pose regime is exactly where keypoint predictions can collapse, so this premise is load-bearing.
  • domain assumption Self-supervised training on VFHQ transfers to HDTF, CelebV-HQ, and unseen cross-identity pairs.
    Stated in Sec. 4.1; no domain-shift analysis or per-identity generalization study is provided.
  • standard math Standard diffusion epsilon-prediction and DDIM sampling assumptions hold for Eq. 2 and Eq. 5.
    The paper builds on SVD and DDIM; however the formulas as printed are nonstandard and appear to contain typos.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Navigating Large-Pose Challenge for High-Fidelity Face Reenactment with Video Diffusion Model." pith.science (2026). https://pith.science/paper/VVT2DP5T

@misc{pith2026250716341,
  author       = {Pith},
  title        = {Pith review of: Navigating Large-Pose Challenge for High-Fidelity Face Reenactment with Video Diffusion Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VVT2DP5T}},
  note         = {Machine review of arXiv:2507.16341}
}
read the original abstract

Face reenactment aims to generate realistic talking head videos by transferring motion from a driving video to a static source image while preserving the source identity. Although existing methods based on either implicit or explicit keypoints have shown promise, they struggle with large pose variations due to warping artifacts or the limitations of coarse facial landmarks. In this paper, we present the Face Reenactment Video Diffusion model (FRVD), a novel framework for high-fidelity face reenactment under large pose changes. Our method first employs a motion extractor to extract implicit facial keypoints from the source and driving images to represent fine-grained motion and to perform motion alignment through a warping module. To address the degradation introduced by warping, we introduce a Warping Feature Mapper (WFM) that maps the warped source image into the motion-aware latent space of a pretrained image-to-video (I2V) model. This latent space encodes rich priors of facial dynamics learned from large-scale video data, enabling effective warping correction and enhancing temporal coherence. Extensive experiments show that FRVD achieves superior performance over existing methods in terms of pose accuracy, identity preservation, and visual quality, especially in challenging scenarios with extreme pose variations.

Figures

Figures reproduced from arXiv: 2507.16341 by the authors.

Figure 1
Figure 1. Our face reenactment framework comprises two stages: (1) Training stage: We begin by employing the Motion Extractor to extract pose and [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Self-reenactment qualitative comparison with state-of-the-art methods including OSFV [26], TPSMM [37], LivePortrait [8], FADM [33], AniPortrait [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Cross-identity reenactment qualitative comparison with state-of-the-art methods, including OSFV [26], TPSMM [37], LivePortrait [8], FADM [33], [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Qualitative results of the ablation study with and without the Warp [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison of the ablation study on classifier-free guid [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 26 canonical work pages

  1. [1]

    Stable video di ffusion: Scaling latent video di ffusion models to large datasets

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video di ffusion: Scaling latent video di ffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023

    arXiv 2023

  2. [2]

    DiffusionAct: Controllable Diffusion Autoencoder for One-shot Face Reenactment

    Stella Bounareli, Christos Tzelepis, Vasileios Argyriou, Ioannis Pa- tras, and Georgios Tzimiropoulos. Di ffusionact: Controllable dif- fusion autoencoder for one-shot face reenactment. arXiv preprint arXiv:2403.17217, 2024

    work page Pith review arXiv 2024

  3. [3]

    How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks)

    Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In International Conference on Computer Vision, 2017

    work page 2017

  4. [4]

    Echomimic: Lifelike audio-driven portrait animations through ed- itable landmark conditions

    Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through ed- itable landmark conditions. arXiv preprint arXiv:2407.08136, 2024

    arXiv 2024

  5. [5]

    Arcface: Additive angular margin loss for deep face recognition

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690–4699, 2019

    2019

  6. [6]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. InProceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, pages 12873–12883, 2021

    work page 2021

  7. [7]

    Gener- ative adversarial nets.Advances in neural information processing systems, 27, 2014

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gener- ative adversarial nets.Advances in neural information processing systems, 27, 2014

    2014

  8. [8]

    Liveportrait: E fficient por- trait animation with stitching and retargeting control

    Jianzhu Guo, Dingyun Zhang, Xiaoqiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Liveportrait: E fficient por- trait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024

    arXiv 2024

Show all 38 references
  1. [9]

    High-fidelity relightable monocular portrait animation with lighting-controllable video di ffusion model

    Mingtao Guo, Guanyu Xing, and Yanli Liu. High-fidelity relightable monocular portrait animation with lighting-controllable video di ffusion model. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 228–238, 2025

    2025

  2. [10]

    Animatedi ff: An- imate your personalized text-to-image di ffusion models without specific tuning

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatedi ff: An- imate your personalized text-to-image di ffusion models without specific tuning. In The Twelfth International Conference on Learning Represen- tations

  3. [11]

    Gans trained by a two time-scale update rule con- verge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule con- verge to a local nash equilibrium. Advances in neural information pro- cessing systems, 30, 2017

    2017

  4. [12]

    Depth-aware gen- erative adversarial network for talking head video generation

    Fa-Ting Hong, Longhao Zhang, Li Shen, and Dan Xu. Depth-aware gen- erative adversarial network for talking head video generation. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recog- nition, pages 3397–3406, 2022

    2022

  5. [13]

    Animate anyone: Consistent and controllable image-to-video syn- thesis for character animation

    Li Hu. Animate anyone: Consistent and controllable image-to-video syn- thesis for character animation. In Proceedings of the IEEE /CVF Con- ference on Computer Vision and Pattern Recognition, pages 8153–8163, 2024

    2024

  6. [14]

    Alias-free generative adversarial net- works

    Tero Karras, Miika Aittala, Samuli Laine, Erik H ¨ark¨onen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial net- works. Advances in Neural Information Processing Systems, 34:852–863, 2021

    2021

  7. [15]

    Analyzing and improving the image quality of style- gan

    Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehti- nen, and Timo Aila. Analyzing and improving the image quality of style- gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110–8119, 2020

    2020

  8. [16]

    Adam: A method for stochastic opti- mization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  9. [17]

    Anchored di ffusion for video face reenactment

    Idan Kligvasser, Regev Cohen, George Leifman, Ehud Rivlin, and Michael Elad. Anchored di ffusion for video face reenactment. arXiv preprint arXiv:2407.15153, 2024

    2024 arXiv

  10. [18]

    Imagenet clas- sification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geo ffrey E Hinton. Imagenet clas- sification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012

    2012

  11. [19]

    Mediapipe: A framework for building perception pipelines

    Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019

    1906 arXiv

  12. [20]

    Follow-your-emoji: Fine-controllable and expressive freestyle portrait an- imation

    Yue Ma, Hongyu Liu, Hongfa Wang, Heng Pan, Yingqing He, Junkun Yuan, Ailing Zeng, Chengfei Cai, Heung-Yeung Shum, Wei Liu, et al. Follow-your-emoji: Fine-controllable and expressive freestyle portrait an- imation. In SIGGRAPH Asia 2024 Conference Papers, pages 1–12, 2024

    2024

  13. [21]

    High-resolution image synthesis with latent diffusion mod- els

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj¨orn Ommer. High-resolution image synthesis with latent diffusion mod- els. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  14. [22]

    First order motion model for image animation

    Aliaksandr Siarohin, St ´ephane Lathuili`ere, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in neural information processing systems, 32, 2019

    2019

  15. [23]

    Denoising di ffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising di ffusion implicit models. In International Conference on Learning Representa- tions, 2020

    2020

  16. [24]

    Towards accurate gen- erative models of video: A new metric & challenges

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate gen- erative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018

    2018 arXiv

  17. [25]

    Wan: Open and advanced large-scale video generative models

    Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pan- deng Li, P...

    2025 arXiv

  18. [26]

    One-shot free-view neural talking-head synthesis for video conferencing

    Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021

    2021

  19. [27]

    Image quality assessment: from error visibility to structural similarity

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600–612, 2004

    2004

  20. [28]

    Aniportrait: Audio-driven synthesis of photorealistic portrait animations, 2024

    Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animations, 2024

    2024

  21. [29]

    Neighbourhood representative sampling for efficient end-to-end video quality assessment, 2022

    Haoning Wu, Chaofeng Chen, Liang Liao, Jingwen Hou, Wenxiu Sun, Qiong Yan, Jinwei Gu, and Weisi Lin. Neighbourhood representative sampling for efficient end-to-end video quality assessment, 2022

    2022

  22. [30]

    Vfhq: A high-quality dataset and benchmark for video face super- resolution

    Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super- resolution. In Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pages 657–666, 2022. / Computers & Graphics (2025) 11

    2025

  23. [31]

    X-portrait: Expressive portrait animation with hierarchical motion attention

    You Xie, Hongyi Xu, Guoxian Song, Chao Wang, Yichun Shi, and Linjie Luo. X-portrait: Expressive portrait animation with hierarchical motion attention. In ACM SIGGRAPH 2024 Conference Papers , pages 1–11, 2024

    2024

  24. [32]

    Cogvideox: Text-to-video di ffusion models with an expert trans- former

    Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Ji- azheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video di ffusion models with an expert trans- former. arXiv preprint arXiv:2408.06072, 2024

    2024 arXiv

  25. [33]

    Face animation with an attribute-guided dif- fusion model

    Bohan Zeng, Xuhui Liu, Sicheng Gao, Boyu Liu, Hong Li, Jianzhuang Liu, and Baochang Zhang. Face animation with an attribute-guided dif- fusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 628–637, 2023

    2023

  26. [34]

    The unreasonable e ffectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable e ffectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018

    2018

  27. [35]

    Mimicmotion: High-quality human motion video generation with confidence-aware pose guidance

    Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yue- feng Zhu, and Fangyuan Zou. Mimicmotion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024

    2024 arXiv

  28. [36]

    Flow- guided one-shot talking face generation with a high-resolution audio- visual dataset

    Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow- guided one-shot talking face generation with a high-resolution audio- visual dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3661–3670, 2021

    2021

  29. [37]

    Thin-plate spline motion model for image animation

    Jian Zhao and Hui Zhang. Thin-plate spline motion model for image animation. In Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pages 3657–3666, 2022

    2022

  30. [38]

    Celebv-hq: A large-scale video fa- cial attributes dataset

    Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebv-hq: A large-scale video fa- cial attributes dataset. In European conference on computer vision, pages 650–667. Springer, 2022

    2022

Pith tools

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