preprint · arXiv · Sep 28, 2026
EvolvingAvatar: Interactive 3D Head Generation That Adapts as Conversations Unfold
Junjie Chen, Fei Wang, Kun Li, Yiqi Nie, Xun Yang, Yanbin Hao, Linfeng Zhang, Meng Wang
In plain English
Most interactive avatar generators keep their weights frozen at run time, so they never learn anything from the conversation actually happening. This system keeps adapting during the interaction, using the user's own face video and the two-way audio as a self-supervised signal, with persistent weights that accumulate across the conversation and a faster jaw adjustment that tracks the current moment. The practical payoff is that output gets better the longer a session runs: on the hardest out-of-distribution split the authors report expression-statistic mismatch falling by up to 11.1% from the first interval, which is the opposite of the drift most long-session avatars show.