MoGen: Detailed Neuronal Morphology Generation via Point Cloud Flow Matching

ICLR 2026

Franz Rieger Jan-Matthis Lueckmann Viren Jain Michał Januszewski

Google Research

📄 OpenReview Paper 💻 Code

MoGen is a flow matching model that generates high-fidelity, controllable 3D point clouds of neuronal morphology, significantly reducing manual proofreading efforts in connectomics.

Generation Process. MoGen iteratively transforms noise into high-fidelity neuronal morphologies.

Controllable Interpolation. Smoothly interpolating between different neuronal shapes by varying the conditioning vectors.

Demo Notebooks

We provide a ready-to-run Jupyter notebook to explore MoGen interactively. You can download the demo notebook and its required supporting data below:

mogen_demo.ipynb (Interactive Notebook)
demo_archive.zip (Supporting Data)

Pretrained Models

Our pretrained model checkpoints are publicly hosted on Google Cloud Storage at gs://mogen-release/models. The repository contains the following subdirectories corresponding to different species and datasets:

drosophila_10um
mouse_mixed
mouse_negative
zebrafinch_10um
zebrafinch_50um

Note: You can easily download these using the gsutil command-line tool. For example: gsutil -m cp -r gs://mogen-release/models/drosophila_10um .

Citation

@inproceedings{rieger2026mogen,
    title={MoGen: Detailed Neuronal Morphology Generation via Point Cloud Flow Matching},
    author={Franz Rieger and Jan-Matthis Lueckmann and Viren Jain and Michal Januszewski},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026},
    url={https://openreview.net/forum?id=HpIxllcNtb}
}