Real-Time Interactive Human Avatars with Streaming Diffusion Models

This paper addresses the challenge of creating real-time, interactive human avatars, a crucial area in digital human research. It tackles the limitations of existing diffusion-based methods, which are computationally expensive and unsuitable for streaming, and the restricted scope of current interactive approaches. The proposed two-stage framework, incorporating autoregressive adaptation and acceleration, along with novel components like Reference Sink and Consistency-Aware Discriminator, aims to generate high-fidelity avatars with natural gestures and behaviors in real-time. The paper's significance lies in its potential to enable more engaging and realistic digital human interactions.