FlashMotion proposes ultra-fast long-horizon 3D motion generation framework
What happened
Kaixing Yang and colleagues proposed and released the paper and code for FlashMotion, a framework for long-horizon 3D motion generation that covers three tasks: text-to-motion, music-to-dance, and speech-to-gesture. The authors state that the goal is to generate high-quality long-horizon motion even under a strict inference budget. Method-wise, FlashMotion implements few-step generation in 3D motion space using a MeanFlow formulation and adds kinematic and dynamic consistency constraints; for sequence modeling it uses a lightweight, non-autoregressive BiMamba architecture that fuses conditioning information via AdaIN, reducing the compute and memory cost of single-step inference over long sequences. Its "ultra-fast" positioning therefore comes from these two design choices, few-step generation and a lightweight architecture, rather than from measured results afterward.
AI-written from the coverage · updated 9 d ago
Coverage timeline
Follow the coverage to see the story from different sides.
- arXiv Game AIPickFlashMotion proposes ultra-fast long-term 3D motion generation framework
FlashMotion is a framework for generating high-quality long-term 3D motion under a strict inference budget, covering three tasks: text-to-motion, music-to-dance, and speech-to-gesture. It achieves few-step generation using a MeanFlow formulation in 3D motion space together with kinematic and dynamic consistency constraints, and reduces per-step computation and memory overhead on long sequences with a lightweight non-autoregressive BiMamba architecture and AdaIN-based conditional fusion.