Reproduce “MotionHiFlow: Text-to-motion via hierarchical flow matching” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Heng Li, Xiaotong Lin, Ling-An Zeng, Yulei Kang, Shuai Li, Jian-Fang Hu
MotionHiFlow introduces a hierarchical flow matching framework that generates realistic 3D human motions from text by progressively refining motions from coarse semantic structures to fine-grained details across multiple temporal scales. The framework achieves state-of-the-art performance on HumanML3D and KIT-ML datasets, notably recording the highest R-Precision (Top1: 0.563) and lowest FID (0.032) on HumanML3D.
Key insights
- Human motion conceptualization is inherently hierarchical, processing from coarse, high-level semantic structures to fine-grained temporal details, a process that can be effectively mirrored in generative models.
- A significant amount of semantic information about motions is preserved even at lower temporal scales, suggesting that training on coarser representations can foster stronger alignment with core textual semantics.
- Explicitly modeling structural dependencies among human joints through specialized architectural components and positional encodings within a hierarchical flow matching setup significantly improves both semantic alignment and motion detail.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.