FlashMotion proposes ultra-fast long-term 3D motion generation framework
FlashMotion: Ultra-Fast Long-Term 3D Motion Generation
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.
The paper covers three tasks — text-to-motion, music-to-dance and speech-to-gesture — so you can compare its few-step generation approach against your own real-time pipeline.
Source: arXiv Game AI · arxiv.org