TIMotion: Temporal and Interactive Framework for Efficient Human-Human Motion Generation

Project page: https://aigc-explorer.github.io/TIMotion-page/

Human-human motion generation remains challenging due to inadequate temporal and interaction modeling in existing methods. We propose TIMotion, a unified framework with Causal Interactive Injection, Role-Evolving Scanning, and Localized Pattern Amplification for more effective and efficient generation. Results on InterHuman and InterX demonstrate clear improvements over prior approaches.

one demo

MARRS: Masked Autoregressive Unit-based Reaction Synthesis

Project page: https://aigc-explorer.github.io/MARRS/

Introducing MARRS: a new framework for human action-reaction synthesis that generates coordinated, fine-grained reactions conditioned on another person’s motion. By avoiding VQ and modeling body/hand units with UD-VAE + ACF + MUM, MARRS captures cross-unit perception more effectively and efficiently. It achieves state-of-the-art quantitative and qualitative results.

Overall framework

Demo