CoTalker: Audio-Driven Interactive Conversational Head Generation

Despite advances in audio-driven avatar generation, existing methods often model one side of conversation or rely on explicit role switching, failing to capture truly interactive conversational dynamics. To this end, we introduce CoTalker, a two-stage audio-driven framework that synthesizes interactive conversational head avatars from a reference portrait image and multimodal conversational signals. The framework integrates a diffusion-based motion generator with a GAN-based avatar renderer, enabling expressive conversational motion prediction and high-fidelity video rendering. A novel state guider is introduced to fuse audio–visual cues and model latent interaction patterns in real conversational scenarios. Furthermore, a progressive training strategy is employed to leverage single-role motion priors before dyadic finetuning, enhancing the model's ability to generate semantically aligned speaking motions, diverse listener responses, and seamless role transitions without explicit role assignment. Extensive experiments demonstrate that CoTalker outperforms state-of-the-art methods in producing photo-realistic and context-aware interactive conversational avatars.