preprint · arXiv · Apr 16, 2026
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
In plain English
Compresses a multi-step diffusion model for talking avatars into a single-step generator, reporting roughly a 120x inference speedup while holding visual quality, via distribution matching followed by adversarial distillation.
distillation
audio-driven animation
inference speed