ARDY Combines Autoregressive Diffusion With Hybrid Representation for Real-Time Human Motion

arXiv·medium signal

Offline 3D human motion generators offer precise text and kinematic control but are too slow for interactive use, while online methods hit real-time budgets at the cost of controllability and long-horizon coherence due to limited context. ARDY (arXiv 2607.08741, July 9) uses autoregressive diffusion over a hybrid representation to keep text controllability and long-horizon goals while meeting interactive latency, targeting animation, simulation, and humanoid robotics. The autoregressive-over-diffusion framing is the transferable idea here for any real-time generative pipeline.