Pith. sign in

REVIEW 4 major objections 5 minor 44 references

D^3-Talker: Dual-Branch Decoupled Deformation Fields for Few-Shot 3D Talking Head Synthesis

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read D3-Talker splits lip motion into a shared facial-motion field and a personal audio field, so a 3D talking head trained on seconds of video lip-syncs to unseen voices.

desk verdict Believable incremental extension of InsTaG with a strong lip-sync signal borrowed from a pretrained prior; the central claim needs a Real3D-Portrait baseline before it fully lands. read the letter →

arxiv 2508.14449 v1 pith:JDOIEFQN submitted 2025-08-20 cs.CV

Yuhang Guo , Kaijun Deng , Siyang Song , Jindong Xie , Wenhui Ma , Linlin Shen This is my paper
classification cs.CV
keywords 3Dtalkingheadsynthesisfew-shotlearningGaussianSplattingaudio-drivenlipsynchronizationdeformationfieldscontrastiveFacialMotionpriorcoarse-to-finerendering
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

Few-shot 3D talking head synthesis has a chicken-and-egg problem: the model must learn how a person's mouth moves from very little video, while the audio that drives it is full of speaker-specific cues like pitch and timbre that have nothing to do with lip shape. D3-Talker's proposal is to stop asking one network to learn everything from audio. It builds a static 3D Gaussian head and adds two separate deformation fields: a General Field, fed by an identity-agnostic facial-motion prior extracted from audio, predicts speaker-independent lip movements; an Individual Field, fed by raw audio features, predicts the target person's personalized motion. A similarity contrastive loss during multi-identity pre-training pushes the General Field to encode only shared motion. The paper reports that with as little as ten seconds of training video, this dual-branch design produces sharper frames and markedly better audio-lip synchronization than previous NeRF- and 3DGS-based methods, especially when the driving audio comes from a speaker never seen in training.

What carries the argument

The load-bearing mechanism is a pair of decoupled deformation fields over a static 3D Gaussian head: a General Field driven by Facial Motion (a canonical-subtracted PNCC feature map from a pretrained audio-to-motion model) and a per-identity Individual Field driven by audio features. They predict additive Gaussian attribute deformations, and a similarity contrastive loss trained during the multi-identity pre-training stage pulls the General Field's position deformation toward the matching individual's deformation and away from other identities', forcing the General Field to encode only speaker-independent lip motion. A tri-plane hash encoder with region attention supplies spatial and audio c

What would settle it

Take the same transcribed sentence spoken by two different voices, drive D3-Talker with each, and compare the generated lip shapes with a landmark distance metric; the identity-agnostic claim predicts the two mouth shapes should be nearly identical. If same-content, different-voice pairs produce substantially different lip landmarks, the Facial Motion prior still carries speaker identity. A second test: replace the pretrained audio-to-motion feature extractor with a different one and rerun the cross-speaker sync evaluation, since the architecture's claim requires the decoupling to transfer acr

Watch

Extended reading notes

Core claim

The central claim is that decoupling what every mouth does from how this particular mouth does it is enough to make few-shot 3D talking head synthesis work. Concretely, D3-Talker constructs a static 3D Gaussian attribute field and two additive deformation fields over it. The General Field takes the Facial Motion signal, defined as the difference between a pretrained audio-to-motion model's PNCC feature map for a frame and its canonical expression, and predicts identity-agnostic deformations of Gaussian position, scale, and rotation. The Individual Field takes speech-recognition audio features and predicts identity-specific deformations of the same attributes. The two predictions are added to

Load-bearing premise

The method assumes the Facial Motion signal, a pretrained model's face-prior feature map minus a neutral expression, is a true speaker-independent description of lip motion; if that external prior is noisy or speaker-specific, the claimed decoupling between general and personal deformations fails, and the paper does not test that assumption independently.

Editorial extensions

If this is right

  • With ten seconds of training video, D3-Talker reports the best visual-quality scores (PSNR 31.518, SSIM 0.927) and the best lip-sync confidence score (Sync-C 6.015) among the six methods compared.
  • It keeps that lead across 5, 10, 15, and 20 seconds of training data, with consistent gains over the two closest few-shot baselines at every length.
  • On audio from speakers never seen in training, its lip-sync confidence is about twice that of the prior few-shot 3DGS method being extended (Sync-C 6.966 versus 3.510).
  • Ablations show each component is doing work: removing the facial-motion prior, the similarity contrastive loss, or the coarse-to-fine renderer each hurts either lip sync or image quality.
  • The paper also finds that deforming only center position, scale, and rotation, rather than all Gaussian attributes, gives the best lip synchronization.

Reading between the lines

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

  • If the Facial Motion prior is genuinely identity-agnostic, the same dual-field architecture could be reused with any dense 3D face prior (expression coefficients, landmarks, blendshapes), turning it into a general few-shot avatar personalizer rather than a lip-sync-only method.
  • The General Field's ceiling is set by the external pretrained audio-to-motion model; a fair test of the architecture's claim would be to retrain the pipeline with two different priors and compare cross-speaker Sync-C, isolating how much of the gain is decoupling versus the prior.
  • The contrastive loss supervises only center-position deformations, leaving rotation and scale unconstrained; extending it to those attributes might stabilize decoupling under head motion and would be a cheap further ablation.
  • Because the coarse-to-fine renderer works at input resolution and is trained as an inpainter, it is likely correcting general splatting artifacts rather than just head-motion blur, which suggests it could help other 3DGS avatar pipelines beyond talking heads.
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 D3-Talker, a few-shot 3D talking head synthesis method built on 3D Gaussian Splatting. The key idea is to decouple general, identity-agnostic lip deformation from personalized deformation using two control signals: audio (for the Individual Field) and a Facial Motion prior (for the General Field). The Facial Motion signal is obtained by taking PNCC features from a pretrained Real3D-Portrait Audio-to-Motion model and subtracting a canonical expression (Eq. 3). A Similarity Contrastive Loss (Eq. 10) is introduced during pretraining to encourage the General Field to learn speaker-invariant features, and a Coarse-to-Fine neural renderer refines the 3DGS output. The method is evaluated on six speaking videos under self-driven and cross-driven settings, reporting improved PSNR/SSIM and large SyncNet gains over prior art, with ablations on the main components.

Significance. If the reported results hold, D3-Talker would be a meaningful step for few-shot 3D talking head synthesis: it explicitly separates audio-driven personalized deformation from a learned identity-agnostic motion prior, and the proposed similarity contrastive loss is a sensible mechanism for this decoupling. The paper also contributes a Coarse-to-Fine refinement that visibly improves rendering quality. However, the main empirical evidence is not yet conclusive. The method relies on a strong pretrained audio-to-motion model (Real3D-Portrait) as the source of the Facial Motion prior, yet this model is never included in quantitative comparisons. The ablation table contains duplicate rows and an inconsistency with Table 2, and no error bars or code are provided. These issues leave the central claim—that the dual-branch decoupling, rather than the external prior, drives the lip-sync gains—unsupported. The architectural ideas are promising, but the evaluation must be strengthened before the claims can be accepted.

major comments (4)
  1. [§3.2, Eq. (3); §4.2, Tables 1–3] The General Field is driven exclusively by Facial Motion, which is extracted from the pretrained Real3D-Portrait Audio-to-Motion model. Real3D-Portrait itself is a complete one-shot talking head generator, and its audio-to-motion module already converts audio into realistic lip movements. This means a substantial portion of the reported Sync-C gains could be inherited from that external prior. Yet Real3D-Portrait appears only in the qualitative Figure 4, not in any quantitative comparison in Tables 1–3. To attribute the gains to the proposed decoupled deformation fields, the authors must add Real3D-Portrait (or an equivalent motion-prior baseline) to the quantitative experiments, and/or ablate by replacing the PNCC-difference features with another motion representation. Without this, the central claim of the paper is not testable.
  2. [§4.4, Table 4] Table 4 is internally inconsistent and undermines the ablation claims. The row 'w/o SCLoss' (32.539/0.936/3.228/6.021) is numerically identical to the row 'NCLoss (InsTaG) Only', and the row 'w/o FM' (32.176/0.932/3.251/5.012) is identical to 'Audio Only'. Additionally, the full 'D3-Talker' row (32.566/0.936/3.060/6.453) matches exactly the 20s row of Table 2, while the text in §4.4 refers to Figure 5 as using 10s training data. These duplications suggest the ablation rows are not independent, and the specific contribution of the Similarity Contrastive Loss is not cleanly established. The authors must clarify which training duration Table 4 uses and re-run the ablations so that each row removes exactly one component.
  3. [§3.2, Eq. (3); §4.2, Table 3] The claim that PNCC difference features are 'appearance-agnostic' and 'identity-agnostic' is assumed, not tested. If the Real3D-Portrait Audio-to-Motion model was trained on identities similar to the test speakers (Obama, May, or the AD-NeRF/HDTF identities), or if PNCC encodes source-speaker articulation style, then the cross-driven Sync-C improvements in Table 3 are inflated by leakage from the pretrained prior. The paper should report whether Real3D-Portrait's training data overlaps with the test speakers and should evaluate the General Field on truly unseen identities whose PNCC features are generated by a model never fine-tuned on those identities. Without this test, the identity-agnosticism claim that underpins the cross-driven evaluation is not validated.
  4. [§4.2, Tables 1–3] The quantitative comparisons report no error bars or variance across runs/random seeds. Given that the test set is small (6 videos) and the SyncNet metric is known to be noisy, the large claimed margins (e.g., Sync-C 6.966 vs 3.510 in Table 3) could be partly due to train/test variability. The authors should either report standard deviations over multiple runs or over per-video scores, and ideally provide confidence intervals for the main metric. This is not a fatal flaw, but it is load-bearing because the paper's strongest claim is a quantitative superiority in lip synchronization.
minor comments (5)
  1. [General] The paper does not release code or pretrained models, despite the abstract stating 'code will be provided upon acceptance'. For reproducibility, the authors should commit to a public release or at least provide detailed training configurations and hyperparameters.
  2. [§3.2] Typo: 'Real3D-Potrait' should be 'Real3D-Portrait'.
  3. [Table 1] The column header 'Training FPS↑Time' is ambiguous; it should be separated into 'Training Time ↓' and 'FPS ↑' for clarity.
  4. [§4.4] The sentence 'We also visualize the results with 10s training data in Figure 5' conflicts with Table 4's numbers, which appear to be from 20s training. Please reconcile the training duration used for ablations.
  5. [§4.2, Table 3] The cross-driven table reports only SyncNet metrics; it would be helpful to also report LMD or a perceptual metric to show that the improved synchronization does not come at the cost of visual fidelity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is not definitionally self-referential; the Real3D-Portrait prior is an external input, not a fitted/predicted quantity.

full rationale

I walked the claimed derivation chain: static Gaussian field (Eq. 1), audio and Facial Motion control signals (Eqs. 3-5), dual deformation fields (Eqs. 6-8), rendering (Eq. 9), and the coarse-to-fine refinement. The Facial Motion signal is explicitly defined as a difference of PNCC features extracted by a pretrained, external Audio-to-Motion model from Real3D-Portrait (Section 3.2, Eq. 3), not by the paper's own architecture or its own fitted outputs. The General Field consumes this external prior (Eq. 6), while the Individual Field consumes audio (Eq. 7); the final deformation is their sum (Eq. 8). No target quantity is defined in terms of the model's own prediction, and no fitted parameter is renamed as a prediction. The Similarity Contrastive Loss (Eq. 10) is a standard supervised/contrastive objective over the fields' outputs; it does not smuggle the result into the input. The two-stage training and losses (Eqs. 11-13) are conventional. The only self-citation, DegsTalk [8], appears in a related-work enumeration and is not load-bearing. The absence of a quantitative Real3D-Portrait baseline in Tables 1-3 and the assumption that PNCC is identity-agnostic are evaluation/robustness concerns, not circularity: even if the lip-sync improvement is partly inherited from the external prior, the paper's contribution is the decoupled Gaussian deformation architecture trained on top of that prior, which is an independent empirical claim. The paper itself flags only the lack of ground truth in cross-driven evaluation, which is not a circular-step admission. Therefore, no circular step meets the required evidentiary bar.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central result rests on standard 3DGS machinery plus two external pretrained feature extractors (DeepSpeech, Real3D-Portrait). The paper's own contribution is architectural and empirical; no new theoretical entities are introduced.

free parameters (3)
  • temperature tau in LSC
    Eq. (10), chosen, no value reported.
  • loss weights lambda_D-SSIM, lambda_SC, lambda_C2F, lambda_C, lambda_Geo
    Section 3.5, hand-set, no values reported.
  • latent dimension 32 for FM and FA
    Section 3.2, chosen for feature alignment.
assumptions (5)
  • standard math 3D Gaussian Splatting is a valid rendering representation for static heads
    Taken from Kerbl et al.; used in Section 3.1.
  • domain assumption DeepSpeech audio features encode speech content relevant to lip motion
    Used in Section 3.2 as audio branch input.
  • domain assumption The Real3D-Portrait Audio-to-Motion model provides identity-agnostic, accurate lip motion
    Section 3.2; critical for the General Field. If false, the decoupling fails.
  • domain assumption SyncNet confidence/distance is a valid lip-sync evaluation
    Used as primary lip-sync metric in Section 4.1.
  • domain assumption 5 pre-training identities suffice for generalizable deformations
    Section 4.1, dataset choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of D^3-Talker: Dual-Branch Decoupled Deformation Fields for Few-Shot 3D Talking Head Synthesis." pith.science (2026). https://pith.science/paper/JDOIEFQN

@misc{pith2026250814449,
  author       = {Pith},
  title        = {Pith review of: D^3-Talker: Dual-Branch Decoupled Deformation Fields for Few-Shot 3D Talking Head Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JDOIEFQN}},
  note         = {Machine review of arXiv:2508.14449}
}
read the original abstract

A key challenge in 3D talking head synthesis lies in the reliance on a long-duration talking head video to train a new model for each target identity from scratch. Recent methods have attempted to address this issue by extracting general features from audio through pre-training models. However, since audio contains information irrelevant to lip motion, existing approaches typically struggle to map the given audio to realistic lip behaviors in the target face when trained on only a few frames, causing poor lip synchronization and talking head image quality. This paper proposes D^3-Talker, a novel approach that constructs a static 3D Gaussian attribute field and employs audio and Facial Motion signals to independently control two distinct Gaussian attribute deformation fields, effectively decoupling the predictions of general and personalized deformations. We design a novel similarity contrastive loss function during pre-training to achieve more thorough decoupling. Furthermore, we integrate a Coarse-to-Fine module to refine the rendered images, alleviating blurriness caused by head movements and enhancing overall image quality. Extensive experiments demonstrate that D^3-Talker outperforms state-of-the-art methods in both high-fidelity rendering and accurate audio-lip synchronization with limited training data. Our code will be provided upon acceptance.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 39 canonical work pages

  1. [1]

    L. Chen, Z. Li, R. K. Maddox, Z. Duan, and C. Xu. Lip movements generation at a glance. In Proceedings of the European Conference on Computer Vision (ECCV), pages 520–535, 2018

    work page 2018

  2. [2]

    L. Chen, R. K. Maddox, Z. Duan, and C. Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, pages 7832–7841, 2019

    work page 2019

  3. [3]

    K. Cho, J. Lee, H. Yoon, Y . Hong, J. Ko, S. Ahn, and S. Kim. Gaus- siantalker: Real-time talking head synthesis with 3d gaussian splatting. In Proceedings of the 32nd ACM International Conference on Multime- dia, pages 10985–10994, 2024

    work page 2024

  4. [4]

    Chu and T

    X. Chu and T. Harada. Generalizable and animatable gaussian head avatar. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    work page 2024

  5. [5]

    X. Chu, Y . Li, A. Zeng, T. Yang, L. Lin, Y . Liu, and T. Harada. GPA- vatar: Generalizable and precise head avatar from image(s). In The Twelfth International Conference on Learning Representations, 2024

    work page 2024

  6. [6]

    J. S. Chung and A. Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 Inter- national Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251–263, 2017

    work page 2016

  7. [7]

    D. Das, S. Biswas, S. Sinha, and B. Bhowmick. Speech-driven facial animation using cascaded gans for learning of motion and texture. In European Conference on Computer Vision, pages 408–424, 2020

    work page 2020

  8. [8]

    K. Deng, D. Zheng, J. Xie, J. Wang, W. Xie, L. Shen, and S. Song. Degstalk: Decomposed per-embedding gaussian fields for hair- preserving talking face synthesis. In ICASSP 2025, 2025

    work page 2025

Show all 44 references
  1. [9]

    Goodfellow, J

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020

    2020

  2. [10]

    Y . Guo, K. Chen, S. Liang, Y .-J. Liu, H. Bao, and J. Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Pro- ceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 5784–5794, 2021

    2021

  3. [11]

    Hannun, C

    A. Hannun, C. Case, J. Casper, B. Catanzaro, G. Diamos, E. Elsen, R. Prenger, S. Satheesh, S. Sengupta, A. Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014

    2014 arXiv

  4. [12]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022

    2022

  5. [13]

    X. Ji, H. Zhou, K. Wang, W. Wu, C. C. Loy, X. Cao, and F. Xu. Audio- driven emotional video portraits. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition , pages 14080– 14089, 2021

    2021

  6. [14]

    Johnson, A

    J. Johnson, A. Alahi, and L. 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, pages 694–711, 2016

    2016

  7. [15]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4):139–1, 2023

    2023

  8. [16]

    H. Kim, P. Garrido, A. Tewari, W. Xu, J. Thies, M. Niessner, P. Pérez, C. Richardt, M. Zollhöfer, and C. Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1–14, 2018

    2018

  9. [17]

    D. P. Kingma, M. Welling, et al. Auto-encoding variational bayes, 2013

    2013

  10. [18]

    Krizhevsky, I

    A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. Advances in Neural Informa- tion Processing Systems, 25, 2012

    2012

  11. [19]

    D. Li, K. Zhao, W. Wang, B. Peng, Y . Zhang, J. Dong, and T. Tan. Ae- nerf: Audio enhanced neural radiance field for few shot talking head synthesis. In Proceedings of the AAAI Conference on Artificial Intelli- gence, volume 38, pages 3037–3045, 2024

    2024

  12. [20]

    J. Li, J. Zhang, X. Bai, J. Zhou, and L. Gu. Efficient region-aware neural radiance fields for high-fidelity talking portrait synthesis. In Proceed- ings of the IEEE/CVF International Conference on Computer Vision , pages 7568–7578, 2023

    2023

  13. [21]

    J. Li, J. Zhang, X. Bai, J. Zheng, X. Ning, J. Zhou, and L. Gu. Talk- inggaussian: Structure-persistent 3d talking head synthesis via gaussian splatting. In European Conference on Computer Vision, pages 127–145, 2024

    2024

  14. [22]

    J. Li, J. Zhang, X. Bai, J. Zheng, J. Zhou, and L. Gu. Instag: Learning personalized 3d talking head from few-second video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion, 2025

    2025

  15. [23]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2021

    2021

  16. [24]

    Prajwal, R

    K. Prajwal, R. Mukhopadhyay, V . P. Namboodiri, and C. Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international Conference on Multimedia, pages 484–492, 2020

    2020

  17. [25]

    S. Shen, W. Li, Z. Zhu, Y . Duan, J. Zhou, and J. Lu. Learning dynamic facial radiance fields for few-shot talking head synthesis. In European Conference on Computer Vision, 2022

    2022

  18. [26]

    J. Tang, K. Wang, H. Zhou, X. Chen, D. He, T. Hu, J. Liu, G. Zeng, and J. Wang. Real-time neural radiance talking portrait synthesis via audio-spatial decomposition. arXiv preprint arXiv:2211.12368, 2022

    2022 arXiv

  19. [27]

    Thies, M

    J. Thies, M. Elgharib, A. Tewari, C. Theobalt, and M. Nießner. Neural voice puppetry: Audio-driven facial reenactment.European Conference on Computer Vision, 2020

    2020

  20. [28]

    Trevithick and B

    A. Trevithick and B. Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 15182–15192, 2021

    2021

  21. [29]

    Q. Wang, Z. Wang, K. Genova, P. P. Srinivasan, H. Zhou, J. T. Barron, R. Martin-Brualla, N. Snavely, and T. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690– 4699, 2021

    2021

  22. [30]

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transac- tions on Image Processing, 13(4):600–612, 2004

    2004

  23. [31]

    Y . Xie, T. Feng, X. Zhang, X. Luo, Z. Guo, W. Yu, H. Chang, F. Ma, and F. R. Yu. Pointtalk: Audio-driven dynamic lip point cloud for 3d gaussian-based talking head synthesis. InProceedings of the AAAI Con- ference on Artificial Intelligence, volume 39, pages 8753–8761, 2025

    2025

  24. [32]

    Q. Xu, Z. Xu, J. Philip, S. Bi, Z. Shu, K. Sunkavalli, and U. Neumann. Point-nerf: Point-based neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 5438–5448, 2022

    2022

  25. [33]

    Z. Ye, J. He, Z. Jiang, R. Huang, J. Huang, J. Liu, Y . Ren, X. Yin, Z. Ma, and Z. Zhao. Geneface++: Generalized and stable real-time audio- driven 3d talking face generation. arXiv preprint arXiv:2305.00787 , 2023

    2023 arXiv

  26. [34]

    Z. Ye, Z. Jiang, Y . Ren, J. Liu, J. He, and Z. Zhao. Geneface: Gen- eralized and high-fidelity audio-driven 3d talking face synthesis. The Eleventh International Conference on Learning Representations, 2023

    2023

  27. [35]

    Z. Ye, T. Zhong, Y . Ren, Z. Jiang, J. Huang, R. Huang, J. Liu, J. He, C. Zhang, Z. Wang, et al. Mimictalk: Mimicking a personalized and expressive 3d talking face in minutes. Advances in Neural Information Processing Systems, 37:1829–1853, 2024

    2024

  28. [36]

    Z. Ye, T. Zhong, Y . Ren, J. Yang, W. Li, J. Huang, Z. Jiang, J. He, R. Huang, J. Liu, et al. Real3d-portrait: One-shot realistic 3d talking portrait synthesis. The Twelfth International Conference on Learning Representations, 2024

    2024

  29. [37]

    A. Yu, V . Ye, M. Tancik, and A. Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 4578–4587, 2021

    2021

  30. [38]

    H. Yu, Z. Qu, Q. Yu, J. Chen, Z. Jiang, Z. Chen, S. Zhang, J. Xu, F. Wu, C. Lv, and G. Yu. Gaussiantalker: Speaker-specific talking head synthe- sis via 3d gaussian splatting. In Proceedings of the 32nd ACM Interna- tional Conference on Multimedia, page 3548–3557, 2024

    2024

  31. [39]

    Zakharov, A

    E. Zakharov, A. Shysheya, E. Burkov, and V . Lempitsky. Few-shot ad- versarial learning of realistic neural talking head models. In Proceed- ings of the IEEE/CVF International Conference on Computer Vision , pages 9459–9468, 2019

    2019

  32. [40]

    Zhang, P

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The un- reasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 586–595, 2018

    2018

  33. [41]

    Zhang, L

    Z. Zhang, L. Li, Y . Ding, and C. 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 Recogni- tion, pages 3661–3670, 2021

    2021

  34. [42]

    H. Zhao, H. Wang, C. Yang, and W. Shen. Chase: 3d-consistent human avatars with sparse inputs via gaussian splatting and contrastive learn- ing. arXiv e-prints, pages arXiv–2408, 2024

    2024

  35. [43]

    Zhong, C

    W. Zhong, C. Fang, Y . Cai, P. Wei, G. Zhao, L. Lin, and G. Li. Identity- preserving talking face generation with landmark and appearance pri- ors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9729–9738, June 2023

    2023

  36. [44]

    X. Zhu, Z. Lei, X. Liu, H. Shi, and S. Z. Li. Face alignment across large poses: A 3d solution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 146–155, 2016

    2016

Pith tools

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