Overview

Flow3r augments visual geometry learning with dense 2D correspondences (`flow') as supervision, enabling scalable training from unlabeled monocular videos. Flow3r achieves state-of-the-art results across eight benchmarks spanning static and dynamic scenes, with its largest gains on in-the-wild dynamic videos where labeled data is most scarce.

Quick Start

1. Create the environment

conda create -n flow3r python=3.11
conda activate flow3r

pip install -r requirements.txt

2. Download and place checkpoint

• flow3r.bin: Flow3r trained on ~834k video sequences.

Please fetch the checkpoint manually from Google Drive and drop the file into checkpoints/.

3. Launch the Gradio app

python gradio_app.py 

Acknowledgements

• Our work builds upon several fantastic open-source projects. We would like to acknowledge and thank the authors of:
  • Pi3
  • VGGT
• We also thank the members of the Physical Perception Lab at CMU for their valuable discussions.

Citation

If you find our work useful, please cite:

@inproceedings{cong2026flow3r,
    title={Flow3r: Factored Flow Prediction for Scalable Visual Geometry Learning},
    author={Cong, Zhongxiao and Zhao, Qitao and Jeon, Minsik and Tulsiani, Shubham},
    booktitle={CVPR},
    year={2026}
}