Cellpose 3 cyto3 — ONNX export
This repository hosts the Cellpose 3 cyto3 segmentation model converted from PyTorch to ONNX (FP16 internal weights, FP32 I/O) for use in the open-source CellSeg Android app (source).
Provenance
This is a reproduction of the upstream cyto3 weights, converted via
torch.onnx.export (opset 17) and FP16-quantised internally with
onnxconverter-common, with FP32 input/output retained for compatibility
with the Android pipeline.
- File:
cyto3-fp16.onnx, 13.3 MB - Single output tensor
Attribution
This redistribution is published with the explicit permission of the upstream authors at the Pachitariu and Stringer labs, HHMI Janelia, granted by Marius Pachitariu (May 2026) under conditions of attribution and licence propagation.
If you use cyto3 (in any form, including this ONNX export), please cite:
Stringer, C. & Pachitariu, M. (2025). Cellpose3: one-click image restoration for improved segmentation. Nature Methods.
Stringer, C., Wang, T., Michaelos, M., & Pachitariu, M. (2021). Cellpose: a generalist algorithm for cellular segmentation. Nature Methods, 18(1), 100–106.
If you use the related Cellpose-SAM model, please also cite:
Pachitariu, M., Rariden, M., & Stringer, C. (2025). Cellpose-SAM: superhuman generalization for cellular segmentation. bioRxiv 10.1101/2025.04.28.651001v1.
Original project: github.com/MouseLand/cellpose
Licence and propagation
The Cellpose codebase is released under BSD 3-Clause.
The cyto3 weights are redistributed in this ONNX form with the explicit permission of the upstream authors at HHMI Janelia (granted May 2026), under the following terms:
- Attribution. Any use or further redistribution of these weights must cite the Cellpose papers above and credit the Pachitariu and Stringer labs at HHMI Janelia.
- Licence propagation. Any fork, mirror, or further redistribution of this repository must carry this notice forward — including this licence section verbatim, and the attribution block above.
For completeness on the upstream chain: Cellpose-SAM (used by the related cloud service the CellSeg app calls into) sits on top of upstream Segment Anything (SAM) weights. Per Michael Perham, Director of Janelia Innovations & Open Science at HHMI Janelia (May 2026), these upstream SAM weights carry a fully permissive licence that allows commercial applications.
This redistribution is published in good faith and in accordance with the explicit terms granted by the upstream authors. If you are an upstream maintainer and would like the wording adjusted or the redistribution withdrawn, please contact the maintainer at the email below.
How to reproduce
The conversion script is committed at: github.com/lynchaos/CellSeg/blob/main/tools/convert_cyto3_to_onnx.py
Run that script with cellpose v3 installed; it will produce the same cyto3-fp16.onnx
and validate equivalence before upload.
Independence
This is an independent open-source project. It is not affiliated with, sponsored by, or endorsed by the authors of Cellpose, Hugging Face, or any other party.
Contact
Maintainer: Kemal Yaylali — https://kemal.yaylali.uk — support@yaylali.uk