Avatar Forcing: Real-Time Interactive Head Avatar Generation for Natural Conversation

This paper presents Avatar Forcing, a new diffusion-driven framework that enables real-time interactive head avatar generation for natural conversation. It addresses the challenges of real-time motion generation under causal constraints and learning expressive reactions without extra labeled data. The framework processes multimodal inputs with low latency (approx. 500ms) and generates reactive, expressive avatar motions, outperforming baseline systems in user evaluations with over 80% preference. This has significant potential for virtual communication, entertainment, and human-computer interaction.