FLAME: Free-Form Language-Based Motion Synthesis & Editing

Text-based motion generation models are drawing a surge of interest for their potential automating the motion-making process in game, animation, or robot industries. In this paper, we propose diffusion-based synthesis and editing model named FLAME. Inspired by recent successes diffusion models, integrate generative into domain. FLAME can generate high-fidelity motions well aligned with given text. Also, it edit parts motion, both frame-wise joint-wise, without any fine-tuning. involves new transformer-based architecture devise to better handle data, which is found be crucial manage variable-length attend free-form experiments, show that achieves state-of-the-art performances on three text-motion datasets: HumanML3D, BABEL, KIT. We also demonstrate FLAME’s capability extended other tasks such as prediction in-betweening, have been previously covered dedicated models.