Live Avatar: Streaming Real-Time Audio-Driven Avatar Generation, Infinite Length (liveavatar.github.io)

The newly unveiled Live Avatar is a cutting-edge framework that facilitates the real-time streaming of interactive avatar video generation. Driven by a powerful 14-billion-parameter diffusion model, it can achieve an impressive 20 frames per second using five H800 GPUs with just four sampling steps. This technology is noteworthy for its ability to support streaming video that lasts over 10,000 seconds, allowing users to engage in seamless natural conversations with avatars via their microphone and camera, with instant visual responses. The significance of Live Avatar lies in its ability to overcome challenges faced by traditional talking-avatar systems, particularly identity drift and color shifts during extended interactions. By implementing Block-wise Autoregressive processing, the system can continuously generate video streams faster than playback speeds, enabling unbounded streaming expansion based on previous frames. Furthermore, integration with Qwen3-Omni opens the door for fully interactive dialogue agents, enhancing user experiences in virtual environments, gaming, and telecommunication. This development heralds a new era of realistic, engaging, and prolonged interactions with AI-driven avatars.