Demo Paper Code Model
TL;DR

Flow matching + physics-based GRPO → physically plausible 3D motion

Explore in 3D

InputMonocular video
Loading 3D viewer…

Drag to orbit · Scroll to zoom · Right-drag to pan

Qualitative Comparisons

More Application

Data for Robotics

We retarget motions generated by FlowHMR to the Unitree G1 robot using GMR, then use SONIC to track the retargeted motions in physics simulation.

Animation Creation

We use GPT-6 Astra for scene reconstruction, motion retargeting, and alignment.

Text-to-Motion

We generate a video from a text prompt with MiniMax H3, then capture its 3D human motion with FlowHMR.

BibTeX

flowhmr.bib 
@misc{wang2026flowhmr,
  title={{FlowHMR}: Physically Plausible Motion Capture from Video},
  author={Zhanke Wang and Chengfeng Zhao and Qing Shuai and
          Jingzhong Lin and Heng Li and Zeyu Ling and Yuxin Wen and
          Jing Li and Di Kang and Chunchao Guo and Linchao Bao},
  year={2026},
  eprint={2610.03691},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.03691}
}