preprint · arXiv
Avatar Forcing: Real-Time Interactive Head Avatar Generation for Natural Conversation
Taekyung Ki, Sangwon Jang, Jaehyeong Jo, Jaehong Yoon, Sung Ju Hwang
In plain English
A real-time head-avatar method that generates reactions to what the user says and does, trained without hand-labelled reaction data by treating versions of the clip with the user's input removed as negative examples. The authors report roughly 500 ms latency, a 6.8x speedup over their baseline, and human raters preferring its reactions over 80 percent of the time. The practical takeaway is that expressive listening and reacting no longer requires a labelled reaction dataset, which has been the main cost barrier to building one.
real-time avatars
talking head generation
inference speed
evaluation