TurboTalk: Progressive Distillation for One-Step Audio-Driven Talking Avatar Generation
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866909015924015104 |
|---|---|
| author | Liu, Xiangyu Gao, Feng Zhang, Xiaomei Zhang, Yong Wei, Xiaoming Lei, Zhen Zhu, Xiangyu |
| author_facet | Liu, Xiangyu Gao, Feng Zhang, Xiaomei Zhang, Yong Wei, Xiaoming Lei, Zhen Zhu, Xiangyu |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2604_14580 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | TurboTalk: Progressive Distillation for One-Step Audio-Driven Talking Avatar Generation Liu, Xiangyu Gao, Feng Zhang, Xiaomei Zhang, Yong Wei, Xiaoming Lei, Zhen Zhu, Xiangyu Computer Vision and Pattern Recognition Multimedia Sound 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. |
| title | TurboTalk: Progressive Distillation for One-Step Audio-Driven Talking Avatar Generation |
| topic | Computer Vision and Pattern Recognition Multimedia Sound |
| url | https://arxiv.org/abs/2604.14580 |