arxiv: 2604.14580 · v2 · submitted 2026-04-16 · 💻 cs.CV · cs.MM· cs.SD

Recognition: unknown

TurboTalk: Progressive Distillation for One-Step Audio-Driven Talking Avatar Generation

Xiangyu Liu , Feng Gao , Xiaomei Zhang , Yong Zhang , Xiaoming Wei , Zhen Lei , Xiangyu Zhu

Authors on Pith no claims yet

Pith reviewed 2026-05-10 11:09 UTC · model grok-4.3

classification 💻 cs.CV cs.MMcs.SD

keywords talking avataraudio-driven video generationdiffusion model distillationone-step generationprogressive distillationadversarial distillationvideo synthesisreal-time inference

0 comments

The pith

A two-stage progressive distillation method reduces audio-driven talking avatar generation from many denoising steps to one while preserving quality.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that multi-step video diffusion models for audio-driven talking avatars can be compressed into a single-step generator without major quality loss. It does this by first using distribution matching distillation to create a stable four-step student model, then applying adversarial distillation with a progressive timestep sampling strategy and a self-compare adversarial objective to reach one step. A sympathetic reader would care because current models incur high computational costs that prevent real-time use in applications such as virtual communication or digital media production. The central mechanism allows the speed gains of one-step sampling while using intermediate references to keep training stable during the aggressive reduction in steps.

Core claim

TurboTalk is a two-stage progressive distillation framework that first applies Distribution Matching Distillation to obtain a strong and stable 4-step student model from a multi-step audio-driven video diffusion model, then progressively reduces the denoising steps from 4 to 1 through adversarial distillation; to stabilize this extreme reduction, it introduces progressive timestep sampling and a self-compare adversarial objective that supplies an intermediate adversarial reference, ultimately enabling single-step generation of video talking avatars.

What carries the argument

The self-compare adversarial objective paired with progressive timestep sampling, which supplies an intermediate reference to stabilize adversarial training as the number of denoising steps is reduced from four to one.

If this is right

Single-step inference achieves a 120-fold increase in speed over the original multi-step denoising process.
Generation quality for audio-driven talking avatars remains high despite the reduction to a single denoising step.
The same two-stage distillation procedure can be used to accelerate other multi-step diffusion models for video synthesis tasks.
Real-time deployment of talking avatar systems becomes feasible in settings that previously could not support multi-step sampling.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same stabilization techniques could be tested on diffusion models for other modalities such as image or 3D generation to see whether they also allow extreme step reduction.
If the one-step model runs on edge devices, it could support live video applications like mobile virtual meetings that currently rely on slower or cloud-based methods.
Combining this distillation approach with other acceleration methods such as quantization might yield even larger speed gains while still using the same progressive framework.

Load-bearing premise

The progressive timestep sampling strategy together with the self-compare adversarial objective can keep training stable when the denoising steps are cut from four to one without causing large drops in generation quality.

What would settle it

Training the final one-step model from the four-step student without the progressive timestep sampling or self-compare objective and then measuring whether output quality falls substantially below the original multi-step model or whether training diverges.

Figures

Figures reproduced from arXiv: 2604.14580 by Feng Gao, Xiangyu Liu, Xiangyu Zhu, Xiaomei Zhang, Xiaoming Wei, Yong Zhang, Zhen Lei.

**Figure 1.** Figure 1: We propose TurboTalk, a progressive distillation framework for audio-driven talking avatar generation. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Figure 2: Overview of TurboTalk. We first distill a multi-step diffusion teacher into a 4-step student via distribution matching, and then [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Qualitative comparison with other audio-driven talking avatar generation methods under different NFE settings. Under the 4-NFE [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

**Figure 4.** Figure 4: Qualitative ablation of progressive distillation under the 1-NFE setting. Step reduction enhances motion continuity, and self [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

read the original abstract

Existing audio-driven video digital human generation models rely on multi-step denoising, resulting in substantial computational overhead that severely limits their deployment in real-world settings. While one-step distillation approaches can significantly accelerate inference, they often suffer from training instability. To address this challenge, we propose TurboTalk, a two-stage progressive distillation framework that effectively compresses a multi-step audio-driven video diffusion model into a single-step generator. We first adopt Distribution Matching Distillation to obtain a strong and stable 4-step student, and then progressively reduce the denoising steps from 4 to 1 through adversarial distillation. To ensure stable training under extreme step reduction, we introduce a progressive timestep sampling strategy and a self-compare adversarial objective that provides an intermediate adversarial reference that stabilizes progressive distillation. Our method achieve single-step generation of video talking avatar, boosting inference speed by 120 times while maintaining high generation quality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

TurboTalk's two-stage progressive distillation aims for fast one-step talking avatars, but verification of the quality and stability claims is still needed.

read the letter

The main point for you is that TurboTalk claims to deliver single-step generation for audio-driven talking avatars by progressively distilling a diffusion model from many steps down to one, with a reported 120x inference boost while keeping quality high. What is new here is the two-stage setup: first using Distribution Matching Distillation to create a stable 4-step student, then applying adversarial distillation with a progressive timestep sampling strategy and a self-compare adversarial objective. The self-compare part gives an intermediate reference to help with stability when cutting steps so aggressively. This combination looks like a targeted extension for the talking avatar task rather than a generic distillation trick. The paper handles the motivation well by pointing out how multi-step denoising limits real-world use in digital humans and virtual communication. It also identifies the instability problem in one-step distillation and proposes specific fixes for it. On the downside, the provided abstract stays at a high level with no metrics, baselines, or ablation results. This makes it hard to judge whether the stabilization techniques actually prevent quality drops or temporal issues in the video output. The concern that the self-compare objective might not fully bound distribution shift or avoid collapse during the 4-to-1 reduction is worth checking against the full experiments and method details. This work would interest researchers focused on efficient video generation and avatar systems. A reader looking for practical speed improvements in diffusion models for this domain could get something out of it, assuming the results support the claims. I think it deserves serious peer review because the idea is timely and the method seems thoughtfully adapted from existing distillation literature, even if more evidence is needed to confirm the central claims.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces TurboTalk, a two-stage progressive distillation framework that first applies Distribution Matching Distillation to obtain a stable 4-step student from a multi-step audio-driven video diffusion model, then uses adversarial distillation with progressive timestep sampling and a self-compare adversarial objective to compress it to a single-step generator for talking avatars. The central claim is that this yields single-step video generation with a 120x inference speedup while preserving high visual quality and lip-sync accuracy.

Significance. If the empirical claims hold under rigorous validation, the work would be significant for real-time audio-driven avatar synthesis, as it directly tackles the inference latency barrier of multi-step diffusion models in video generation. The progressive stabilization techniques could influence efficient distillation methods for other temporal diffusion tasks, particularly if the self-compare objective proves generalizable beyond this setting.

major comments (1)

[Progressive adversarial distillation stage (method description)] The central claim of quality preservation after 4-to-1 step reduction rests on the self-compare adversarial objective and progressive timestep sampling mitigating instability and distribution shift in video diffusion. However, the manuscript provides insufficient detail on the exact formulation of the intermediate adversarial reference, its generation process, and quantitative interaction with video-specific losses (temporal consistency and lip-sync), leaving the stabilization mechanism unverified against common artifacts like flickering or detail loss.

minor comments (1)

[Abstract] Abstract contains a grammatical error: 'Our method achieve' should be 'Our method achieves'.

Simulated Author's Rebuttal

1 responses · 0 unresolved

Thank you for reviewing our manuscript and providing valuable feedback. We address the referee's major comment below and will update the manuscript to provide greater detail on the proposed stabilization techniques.

read point-by-point responses

Referee: [Progressive adversarial distillation stage (method description)] The central claim of quality preservation after 4-to-1 step reduction rests on the self-compare adversarial objective and progressive timestep sampling mitigating instability and distribution shift in video diffusion. However, the manuscript provides insufficient detail on the exact formulation of the intermediate adversarial reference, its generation process, and quantitative interaction with video-specific losses (temporal consistency and lip-sync), leaving the stabilization mechanism unverified against common artifacts like flickering or detail loss.

Authors: We agree that additional explicit details are warranted to fully substantiate the stabilization claims. The manuscript introduces the progressive timestep sampling and self-compare adversarial objective in Section 3.3 as a means to generate an intermediate reference by comparing student outputs against a dynamically updated reference at sampled timesteps, thereby reducing distribution shift. To address the concern directly, we will revise the paper with: (1) precise mathematical formulations of the self-compare loss and its integration with temporal consistency and lip-sync objectives; (2) pseudocode detailing the reference generation process; and (3) new quantitative ablations and visual results measuring artifact mitigation (e.g., optical-flow-based flickering scores and LSE-D lip-sync metrics) against direct 1-step baselines. These additions will verify the mechanism without altering the core claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity; method extends prior distillation with independent components

full rationale

The abstract and described framework present a two-stage process: first applying Distribution Matching Distillation to obtain a 4-step student model, then using progressive adversarial distillation with a new progressive timestep sampling strategy and self-compare adversarial objective to reach one-step generation. No equations, claims, or steps in the provided text reduce the final generator or its performance claims to a fitted parameter renamed as prediction, a self-definition, or a load-bearing self-citation chain. The stability mechanisms are introduced as novel additions to address known instability in extreme step reduction, and the 120x speedup with quality maintenance is framed as an empirical result rather than a mathematical necessity derived from the inputs by construction. The derivation chain remains self-contained against external benchmarks and prior techniques without circular reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the effectiveness of an existing multi-step diffusion model as teacher and on empirical stability of the new distillation objectives; no explicit free parameters or invented entities are detailed in the abstract.

free parameters (1)

distillation hyperparameters and timestep schedule
Progressive reduction and adversarial objectives typically require tuned weights and sampling strategies chosen to achieve stability.

axioms (1)

domain assumption A strong multi-step audio-driven video diffusion model exists as the starting teacher.
The framework begins by distilling from such a model, assuming it produces high-quality outputs.

pith-pipeline@v0.9.0 · 5465 in / 1218 out tokens · 47511 ms · 2026-05-10T11:09:33.061446+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

38 extracted references · 14 canonical work pages · 4 internal anchors

[1]

Diffusion forcing: Next-token prediction meets full-sequence diffu- sion.Advances in Neural Information Processing Systems, 37:24081–24125, 2024

Boyuan Chen, Diego Mart ´ı Mons´o, Yilun Du, Max Sim- chowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffu- sion.Advances in Neural Information Processing Systems, 37:24081–24125, 2024. 1

2024
[2]

Pose: Phased one-step adversarial equilibrium for video diffusion models.arXiv e-prints, pages arXiv– 2508, 2025

Jiaxiang Cheng, Bing Ma, Xuhua Ren, Hongyi Jin, Kai Yu, Peng Zhang, Wenyue Li, Yuan Zhou, Tianxiang Zheng, and Qinglin Lu. Pose: Phased one-step adversarial equilibrium for video diffusion models.arXiv e-prints, pages arXiv– 2508, 2025. 2, 3

2025
[3]

Lightx2v: Light video genera- tion inference framework.https : / / github

LightX2V Contributors. Lightx2v: Light video genera- tion inference framework.https : / / github . com / ModelTC/lightx2v, 2025. 6

2025
[4]

arXiv preprint arXiv:2510.02283 (2025)

Justin Cui, Jie Wu, Ming Li, Tao Yang, Xiaojie Li, Rui Wang, Andrew Bai, Yuanhao Ban, and Cho-Jui Hsieh. Self- forcing++: Towards minute-scale high-quality video genera- tion.arXiv preprint arXiv:2510.02283, 2025. 1

work page arXiv 2025
[5]

Wan-s2v: Audio-driven cinematic video generation,

Xin Gao, Li Hu, Siqi Hu, Mingyang Huang, Chaonan Ji, Dechao Meng, Jinwei Qi, Penchong Qiao, Zhen Shen, Yafei Song, Ke Sun, Linrui Tian, Guangyuan Wang, Qi Wang, Zhongjian Wang, Jiayu Xiao, Sheng Xu, Bang Zhang, Peng Zhang, Xindi Zhang, Zhe Zhang, Jingren Zhou, and Lian Zhuo. Wan-s2v: Audio-driven cinematic video generation,
[6]

Video dif- 9 fusion models.Advances in neural information processing systems, 35:8633–8646, 2022

Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video dif- 9 fusion models.Advances in neural information processing systems, 35:8633–8646, 2022. 1

2022
[7]

The gan is dead; long live the gan! a mod- ern gan baseline.Advances in Neural Information Process- ing Systems, 37:44177–44215, 2024

Nick Huang, Aaron Gokaslan, V olodymyr Kuleshov, and James Tompkin. The gan is dead; long live the gan! a mod- ern gan baseline.Advances in Neural Information Process- ing Systems, 37:44177–44215, 2024. 4

2024
[8]

Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion

Xun Huang, Zhengqi Li, Guande He, Mingyuan Zhou, and Eli Shechtman. Self forcing: Bridging the train- test gap in autoregressive video diffusion.arXiv preprint arXiv:2506.08009, 2025. 1, 3

work page internal anchor Pith review arXiv 2025
[9]

Live avatar: Streaming real-time audio-driven avatar generation with infinite length, 2025

Yubo Huang, Hailong Guo, Fangtai Wu, Shifeng Zhang, Shi- jie Huang, Qijun Gan, Lin Liu, Sirui Zhao, Enhong Chen, Ji- aming Liu, and Steven Hoi. Live avatar: Streaming real-time audio-driven avatar generation with infinite length, 2025. 2, 6

2025
[10]

HunyuanVideo: A Systematic Framework For Large Video Generative Models

Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024. 1

work page internal anchor Pith review Pith/arXiv arXiv 2024
[11]

Let them talk: Audio-driven multi-person conversational video generation.arXiv preprint arXiv:2505.22647, 2025

Zhe Kong, Feng Gao, Yong Zhang, Zhuoliang Kang, Xi- aoming Wei, Xunliang Cai, Guanying Chen, and Wenhan Luo. Let them talk: Audio-driven multi-person conversa- tional video generation.arXiv preprint arXiv:2505.22647,

work page arXiv
[12]

arXiv preprint arXiv:2402.13929 (2024) 5

Shanchuan Lin, Anran Wang, and Xiao Yang. Sdxl- lightning: Progressive adversarial diffusion distillation. arXiv preprint arXiv:2402.13929, 2024. 3

work page arXiv 2024
[13]

arXiv preprint arXiv:2501.08316 (2025) 2, 3, 4

Shanchuan Lin, Xin Xia, Yuxi Ren, Ceyuan Yang, Xuefeng Xiao, and Lu Jiang. Diffusion adversarial post-training for one-step video generation.arXiv preprint arXiv:2501.08316,

work page arXiv
[14]

arXiv preprint arXiv:2506.09350 (2025) 2, 4

Shanchuan Lin, Ceyuan Yang, Hao He, Jianwen Jiang, Yuxi Ren, Xin Xia, Yang Zhao, Xuefeng Xiao, and Lu Jiang. Autoregressive adversarial post-training for real-time inter- active video generation.arXiv preprint arXiv:2506.09350,

work page arXiv
[15]

Sora: A review on background, technology, limitations, and opportunities of large vision models, 2024

Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jian- feng Gao, Lifang He, and Lichao Sun. Sora: A review on background, technology, limitations, and opportunities of large vision models, 2024. 1

2024
[16]

arXiv preprint arXiv:2503.06674 (2025) 2, 4, 5, 10, 23

Yihong Luo, Tianyang Hu, Jiacheng Sun, Yujun Cai, and Jing Tang. Learning few-step diffusion models by trajectory distribution matching.arXiv preprint arXiv:2503.06674,

work page arXiv
[17]

Echomimicv2: Towards striking, simplified, and semi- body human animation

Rang Meng, Xingyu Zhang, Yuming Li, and Chenguang Ma. Echomimicv2: Towards striking, simplified, and semi- body human animation. InProceedings of the Computer Vi- sion and Pattern Recognition Conference, pages 5489–5498,
[18]

A lip sync expert is all you need for speech to lip generation in the wild

KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Nambood- iri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. InProceedings of the 28th ACM international conference on multimedia, pages 484–492, 2020. 2

2020
[19]

Fast high- resolution image synthesis with latent adversarial diffusion distillation

Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high- resolution image synthesis with latent adversarial diffusion distillation. InSIGGRAPH Asia 2024 Conference Papers, pages 1–11, 2024. 3

2024
[20]

Adversarial diffusion distillation

Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. InEuropean Conference on Computer Vision, pages 87–103. Springer,
[21]

Soulx-flashtalk: Real-time infinite streaming of audio- driven avatars via self-correcting bidirectional distillation,

Le Shen, Qian Qiao, Tan Yu, Ke Zhou, Tianhang Yu, Yu Zhan, Zhenjie Wang, Ming Tao, Shunshun Yin, and Siyuan Liu. Soulx-flashtalk: Real-time infinite streaming of audio- driven avatars via self-correcting bidirectional distillation,
[22]

Linsen Song, Wayne Wu, Chaoyou Fu, Chen Change Loy, and Ran He. Audio-driven dubbing for user generated con- tents via style-aware semi-parametric synthesis.IEEE Trans- actions on Circuits and Systems for Video Technology, 33(3): 1247–1261, 2023. 2

2023
[23]

Emo: Emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions

Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive generating expressive portrait videos with audio2video diffusion model under weak conditions. In European Conference on Computer Vision, pages 244–260. Springer, 2024. 2

2024
[24]

Nonlinear 3d face morphable model

Luan Tran and Xiaoming Liu. Nonlinear 3d face morphable model. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7346–7355, 2018. 2

2018
[25]

Wan: Open and Advanced Large-Scale Video Generative Models

Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianx- iao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jin- gren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianghua Huang, Mengyang Feng, Ningyi Zhang, Pandeng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fan...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[26]

Fanta- sytalking: Realistic talking portrait generation via coherent motion synthesis

Mengchao Wang, Qiang Wang, Fan Jiang, Yaqi Fan, Yun- peng Zhang, Yonggang Qi, Kun Zhao, and Mu Xu. Fanta- sytalking: Realistic talking portrait generation via coherent motion synthesis. InProceedings of the 33rd ACM Interna- tional Conference on Multimedia, pages 9891–9900, 2025. 2

2025
[27]

Aniportrait: Audio-driven synthesis of photorealistic portrait animation

Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024

work page arXiv 2024
[28]

Hallo: Hierarchical audio-driven visual synthesis for portrait image animation.arXiv preprint arXiv:2406.08801, 2024

Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation.arXiv preprint arXiv:2406.08801, 2024. 2

work page arXiv 2024
[29]

Ufogen: You forward once large scale text-to-image gener- 10 ation via diffusion gans

Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image gener- 10 ation via diffusion gans. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196–8206, 2024. 3

2024
[30]

LongLive: Real-time Interactive Long Video Generation

Shuai Yang, Wei Huang, Ruihang Chu, Yicheng Xiao, Yuyang Zhao, Xianbang Wang, Muyang Li, Enze Xie, Ying- cong Chen, Yao Lu, et al. Longlive: Real-time interactive long video generation.arXiv preprint arXiv:2509.22622,

work page internal anchor Pith review arXiv
[31]

Infinitetalk: Audio-driven video generation for sparse-frame video dubbing,

Shaoshu Yang, Zhe Kong, Feng Gao, Meng Cheng, Xi- angyu Liu, Yong Zhang, Zhuoliang Kang, Wenhan Luo, Xunliang Cai, Ran He, et al. Infinitetalk: Audio-driven video generation for sparse-frame video dubbing.arXiv preprint arXiv:2508.14033, 2025. 2, 5, 6

work page arXiv 2025
[32]

To- wards one-step causal video generation via adversarial self- distillation.arXiv preprint arXiv:2511.01419, 2025

Yongqi Yang, Huayang Huang, Xu Peng, Xiaobin Hu, Dong- hao Luo, Jiangning Zhang, Chengjie Wang, and Yu Wu. To- wards one-step causal video generation via adversarial self- distillation.arXiv preprint arXiv:2511.01419, 2025. 3

work page arXiv 2025
[33]

Im- proved distribution matching distillation for fast image syn- thesis.Advances in neural information processing systems, 37:47455–47487, 2024

Tianwei Yin, Micha ¨el Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and Bill Freeman. Im- proved distribution matching distillation for fast image syn- thesis.Advances in neural information processing systems, 37:47455–47487, 2024. 3

2024
[34]

One-step diffusion with distribution matching distillation

Tianwei Yin, Micha ¨el Gharbi, Richard Zhang, Eli Shecht- man, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 6613–6623, 2024. 3

2024
[35]

From slow bidirectional to fast autoregressive video diffusion mod- els

Tianwei Yin, Qiang Zhang, Richard Zhang, William T Free- man, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast autoregressive video diffusion mod- els. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22963–22974, 2025. 1, 3

2025
[36]

Sadtalker: Learning realistic 3d motion coefficients for stylized audio- driven single image talking face animation

Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio- driven single image talking face animation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8652–8661, 2023. 2

2023
[37]

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. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3661–3670, 2021. 6

2021
[38]

Celebv- hq: A large-scale video facial 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 facial attributes dataset. InEuropean conference on computer vision, pages 650–667. Springer,