LY Corporation · Text-to-Motion Research

MotionSpaceFlow

A flow-matching method from LY Corporation for text-driven human motion generation that works on full-resolution motion instead of a learned, downsampled latent space, with no motion encoder or decoder. It adds representation-aware noise scaling and a multimodal diffusion transformer that uses causal attention for frame-to-frame motion features and bidirectional attention for absolute joint coordinates. The authors report state-of-the-art text-to-motion results on HumanML3D and other datasets, self-reported, and the XYZ variant supports zero-shot control of any joint at any frame without control-specific training. Paper posted 28 September; code and pretrained models followed on 8 October under CC0, in a repo the authors call a temporary release.