LiveAvatar

LiveAvatar is a research AI system for real-time, streaming avatar video generation from a single portrait image and an audio or motion driving signal. Built on a 14-billion-parameter diffusion model, it produces high-quality talking-head and full-body avatar videos at interactive frame rates using multiple GPUs. The system supports continuous, theoretically unlimited-length output, making it suitable for live broadcasting scenarios such as virtual presenters, AI streamers, digital humans, and interactive agents.

LiveAvatar focuses on maintaining visual identity, lip-sync accuracy, and natural head and body motion, while reducing artifacts common in traditional video synthesis. The project provides technical details, benchmark comparisons, and demonstration videos, primarily targeting researchers, engineers, and advanced developers interested in generative video, diffusion models, and real-time graphics. While not a plug-and-play SaaS product, LiveAvatar showcases the underlying technology that could power future tools for virtual events, online education, gaming characters, and personalized marketing avatars, highlighting the state-of-the-art in real-time portrait video generation.