flowfusion

Generative modelling and density estimation using diffusion models and flow-matching.

The code in this repository was developed as part of Alsing et al. (2024), and Thorp et al. (2024, 2025). In lieu of a more specific reference, please cite those papers if you make use of the code included here. Please also cite the papers associated with any dependencies of the code, particularly Chen et al. (2018), which describes torchdiffeq.

Installation

To install the code, please clone this repo:

  git clone https://github.com/Cosmo-Pop/flowfusion

Then move into the top level directory and run:

  pip install .

To install flowfusion without updating the dependencies:

pip install poetry
poetry install --no-update

This will obtain any dependencies and will install the code, which can then be imported in Python by doing:

import flowfusion

References

The code in this repository was developed and applied in the following papers:

• J. Alsing et al. (2024). ApJS 274, 12. arXiv:2402.00935
• S. Thorp et al. (2024). ApJ 975, 145. arXiv:2406.19437
• S. Thorp et al. (2025). ApJ, 993, 240. arXiv:2506.12122

For the mathematical underpinnings of the different modules within the code, please see (and consider citing) the following references, which our implementations largely follow:

flowfusion.diffusion

• R.T.Q. Chen et al. (2018). NeurIPS 2018. arXiv:1806.07366
• Y. Song et al. (2021a). ICLR 2021. arXiv:2011.13456
• Y. Song et al. (2021b). NeurIPS 2021. arXiv:2101.09258

flowfusion.flow

• Y. Lipman et al. (2023). ICLR 2023. arXiv:2210.02747

flowfusion.symplectic

• P. Toth et al. (2020). ICLR 2020. arXiv:1909.13789