AVAPrintDB: A Public Multi-Generator Photorealistic Talking-Head Database and Benchmark for Avatar Fingerprinting

Researchers released AVAPrintDB, a public multi-generator photorealistic talking-head database and benchmark for avatar fingerprinting, a task used to determine whether two avatar videos are driven by the same real person, to address the risk of identity impersonation in AI-mediated communication. The database is built from two audiovisual corpora and three advanced avatar generators (GAGAvatar, LivePortrait, HunyuanPortrait), and contains two types of videos, self-reenactment and cross-reenactment, simulating normal use and impersonation scenarios respectively. On this basis, the authors defined a standardized, reproducible benchmark, evaluated publicly available advanced avatar fingerprinting systems, and explored new methods based on the foundation models DINOv2 and CLIP, while also analyzing under generator and dataset shift. Results show that identity-related motion cues are still preserved across synthetic avatars, but existing avatar fingerprinting systems are highly sensitive to changes in the synthesis pipeline and source domain. The database, benchmark protocol and related systems have all been made public.