Motion Diffusion Models

Motion diffusion models learn to create or recover sequences of articulated movement, especially human motion, for applications such as text-to-motion, speech gestures, interaction, animation, and video-based reconstruction. Common approaches use latent or coordinate-space diffusion transformers, conditioned on language, audio, style references, skeletal structure, or past motion; autoregressive and streaming variants extend generation over time. Research also applies diffusion to motion editing and cleanup, while adding physics-aware losses, contact constraints, and explicit path controls to improve plausibility and user control. A central challenge is balancing semantic fidelity and diversity with physical consistency, long-horizon stability, and efficient inference across different bodies and settings.