Google AI Choreographer

Google AI Choreographer is a research project developed by Google in collaboration with choreographer Wayne McGregor, focused on music-conditioned 3D dance generation using advanced AI. It leverages the AIST++ dataset, a large-scale 3D human dance motion database containing over 1,400 sequences, 30 subjects, 10 dance genres, and more than 18,000 seconds of motion data paired with music. 

The core model, FACT (Full-body Action and Choreography Transformer), uses a crossmodal transformer architecture to generate realistic, smooth 3D dance motions that follow music rhythm.  Given a short seed motion (2 seconds) and a piece of music, the model can produce long-range, non-freezing dance sequences suitable for applications like automatic motion retargeting to virtual avatars (e.g., Mixamo). 

This project is part of broader AI research at Google, including the AI Choreographer: Music Conditioned 3D Dance Generation with AIST++ paper presented at ICCV 2021.  The code and training infrastructure are publicly available via GitHub under the google-research/mint repository.