README.md

---
license: other
license_name: per-file
license_link: https://huggingface.co/voxide/voxide-models#license-and-attribution
library_name: onnx
pipeline_tag: mask-generation
tags:
  - onnx
  - segment-anything
  - sam2
  - sam-med3d
  - microscopy
  - medical-imaging
  - interactive-segmentation
  - feature-extraction
  - pixel-classification
  - dinov2
  - vgg16
  - cellpose
  - micro-sam
base_model:
  - facebook/sam2.1-hiera-tiny
  - facebook/sam2.1-hiera-small
  - facebook/sam2.1-hiera-base-plus
  - facebook/sam2.1-hiera-large
  - facebook/dinov2-with-registers-small
  - facebook/dinov2-with-registers-base
---

# Voxide segmentation models (ONNX)

ONNX exports of **SAM 2.1** and **SAM-Med3D (turbo)** used by
[Voxide](https://huggingface.co/voxide), a GPU volume viewer for microscopy, for interactive
segmentation. Each model is an encoder/decoder pair run on the CPU with ONNX Runtime.

Voxide ships the SAM 2.1 Tiny pair. It downloads the others on request from this repository at a
pinned revision and checks every file against the SHA-256 in [`SHA256SUMS`](SHA256SUMS), so a file
that does not match is never loaded.

It also downloads five **network-feature** models for its pixel classifier (a DINOv2 ViT-S/14,
a DINOv2 ViT-B/14, the first two VGG16 blocks, the Cellpose cyto3 backbone and micro-sam's ViT-T
light-microscopy encoder), pinned and checked the same way: see
[Network-feature models](#network-feature-models-pixel-classifier) for their files, provenance and
licences.

## Files

| Model | Encoder | Decoder | Use in Voxide |
|---|---|---|---|
| SAM 2.1 Hiera Tiny | `sam2.1_hiera_tiny_encoder.onnx` (110 MB) | `sam2.1_hiera_tiny_decoder.onnx` (17 MB) | 2D slices (default, fastest) |
| SAM 2.1 Hiera Small | `sam2.1_hiera_small_encoder.onnx` (139 MB) | `sam2.1_hiera_small_decoder.onnx` (17 MB) | 2D slices |
| SAM 2.1 Hiera Base+ | `sam2.1_hiera_base_plus_encoder.onnx` (278 MB) | `sam2.1_hiera_base_plus_decoder.onnx` (17 MB) | 2D slices |
| SAM 2.1 Hiera Large | `sam2.1_hiera_large_encoder.onnx` (853 MB) | `sam2.1_hiera_large_decoder.onnx` (17 MB) | 2D slices (best masks, slowest) |
| SAM-Med3D turbo | `sammed3d_turbo_encoder.onnx` (373 MB) | `sammed3d_turbo_decoder.onnx` (31 MB) | 3D volumes |

## Using them without the in-app download

Download the files you need and either use **Models → Import model…** in Voxide or copy them
into Voxide's model folder. The file names must stay as they are: Voxide finds each model by its
`<name>_encoder.onnx` / `<name>_decoder.onnx` pair.

```bash
pip install -U huggingface_hub
hf download voxide/voxide-models --include "sam2.1_hiera_small_*" --local-dir voxide-models
```

## How they were made

Unmodified inference exports of the upstream checkpoints, produced with PyTorch 2.9.0 (ONNX opset
18, IR version 8). No fine-tuning, quantization or other change to the weights.

## License and attribution

Both upstream model families are released under the **Apache License 2.0** (see [`LICENSE`](LICENSE)),
and so are these exports.

- **SAM 2.1** © Meta Platforms, Inc. — <https://github.com/facebookresearch/sam2>
- **SAM-Med3D** © the SAM-Med3D authors — <https://github.com/uni-medical/SAM-Med3D>

The network-feature models carry their own licences, so this card's licence is per file
(`license: other`): Apache-2.0, BSD-3-Clause or CC-BY-4.0 (micro-sam's ViT-T encoder), as listed for
each file in [their section](#network-feature-models-pixel-classifier), with the attribution each
licence requires under [Licences and attribution](#licences-and-attribution).

If you use these models in published work, please cite the original papers:

```bibtex
@article{ravi2024sam2,
  title   = {SAM 2: Segment Anything in Images and Videos},
  author  = {Ravi, Nikhila and others},
  journal = {arXiv preprint arXiv:2408.00714},
  year    = {2024}
}

@article{wang2023sammed3d,
  title   = {SAM-Med3D: Towards General-purpose Segmentation Models for Volumetric Medical Images},
  author  = {Wang, Haoyu and others},
  journal = {arXiv preprint arXiv:2310.15161},
  year    = {2023}
}
```

## Network-feature models (pixel classifier)

Feature extractors for Voxide's pixel classifier: each turns an image tile into a per-pixel feature
map that the classifier trains on. Single files, run on the CPU with ONNX Runtime; Voxide finds each
by its exact file name, so keep the names as they are when importing them by hand.

| Model | File | Size | SHA-256 | Licence |
|---|---|---|---|---|
| DINOv2 ViT-S/14 with registers | `dinov2-with-registers-small.onnx` | 88 522 426 bytes (89 MB) | `815e440d222e60294bb5b165491c4760e7f08dc7ecd2488a5450e93d50887efe` | Apache-2.0 |
| VGG16 blocks 1–2 | `vgg16-blocks12.onnx` | 1 042 858 bytes (1 MB) | `230fb1431be9f9b6088fc68c24d4ff8c4894e7ef8ae3b16e14e0116b351f1b45` | BSD-3-Clause |
| Cellpose cyto3 backbone | `cellpose-cyto3-backbone.onnx` | 26 483 022 bytes (26 MB) | `e689706a9709ba7fcf70480019bdd1a6a4a94eff05277dfd64d024f2d015195d` | BSD-3-Clause |
| DINOv2 ViT-B/14 with registers | `dinov2-with-registers-base.onnx` | 346 622 747 bytes (347 MB) | `419f8937888b0d4cd607dccaed37b790b48d79cd1aad65a66bf96c7769e3b48b` | Apache-2.0 |
| micro-sam ViT-T (light microscopy) | `microsam-vit-t-lm-encoder.onnx` | 27 983 702 bytes (28 MB) | `f4ee03d6c4e5286ee2a4caf92212ce2cc90349f9f26b0d45c3490e989dc814b3` | CC-BY-4.0 |

```bash
hf download voxide/voxide-models --include "dinov2-*" "vgg16-*" "cellpose-*" --local-dir voxide-models
hf download voxide/voxide-models --include "microsam-*" --local-dir voxide-models
```

### Provenance

- **DINOv2 ViT-S/14 with registers** — Meta AI's `facebook/dinov2-with-registers-small`, in the
  fp32 ONNX export by onnx-community (`onnx-community/dinov2-with-registers-small`, file
  `onnx/model.onnx` at revision `ec7abea1a8757ec4f9f7b26399d31234ffaa6e6a`), copied here unchanged
  under a new name. Input `pixel_values` `[N, 3, H, W]` (H and W multiples of 14), output
  `last_hidden_state`; opset 14.
- **DINOv2 ViT-B/14 with registers** — Meta AI's `facebook/dinov2-with-registers-base`, in the
  fp32 ONNX export by onnx-community (`onnx-community/dinov2-with-registers-base`, file
  `onnx/model.onnx` at revision `fcd050203baf48e192055356f379b5949a111255`), copied here unchanged
  under a new name. Input `pixel_values` `[N, 3, H, W]` (H and W multiples of 14), output
  `last_hidden_state`; opset 14. Voxide reads 768 features per 14 × 14 patch (768 × 32 × 32 for a
  448 × 448 tile).
- **VGG16 blocks 1–2** — torchvision's ImageNet VGG16 (`VGG16_Weights.IMAGENET1K_V1`,
  `vgg16-397923af.pth`), cut after the second convolution block: the ReLU outputs of conv1_2 (64
  channels) and of conv2_2 (128 channels, upsampled ×2 inside the graph), concatenated to a
  192-channel stride-1 map. Exported by
  `scripts/export_vgg16_onnx.py` in the Voxide repository (pinned
  torch 2.9, torchvision 0.24, onnx 1.22; opset 17). The weights are unmodified.
- **Cellpose cyto3 backbone** — Cellpose 3's built-in `cyto3` model (cellpose 3.1.1), the U-Net's
  32-channel stride-1 decoder output before the final 1×1 output convolution (no flows or cell
  probability). Input `pixel_values` `[1, 2, H, W]` (H and W multiples of 16; channel 0 the
  1st–99th-percentile-normalised plane, channel 1 the nuclei channel or zeros). Exported by
  `scripts/export_cellpose_onnx.py` in the Voxide repository (pinned cellpose 3.1, onnx 1.22;
  opset 17). The weights are unmodified.
- **micro-sam ViT-T (light microscopy)** — the image encoder of micro-sam's SAM LM Generalist
  ViT-T (`vit_t_lm`), from the BioImage Model Zoo record `faithful-chicken` v1.1, file `vit_t.pt`
  (SHA-256 `32b92ab64959088878a1cc3e52b9b64aaa0ba4b24c98e9959746d7390bec7b7a`). Exported by
  `scripts/export_microsam_vit_t_onnx.py` in the Voxide repository (pinned torch 2.14, onnx 1.22;
  opset 17): the image encoder alone (no prompt encoder or mask decoder), with a bilinear resize of
  the 448 × 448 input to 1024 × 1024 baked in front of it and its attention biases set to zero, as
  micro-sam's runtime has them. Input `pixel_values` `[1, 3, 448, 448]` (fixed; ImageNet-normalised
  RGB), output `features` `[1, 256, 64, 64]`: one 256-vector per 7 × 7 input pixels.

### Licences and attribution

- **DINOv2** © Meta Platforms, Inc., Apache License 2.0 (see [`LICENSE`](LICENSE)) —
  <https://github.com/facebookresearch/dinov2>, <https://huggingface.co/facebook/dinov2-with-registers-small>;
  ONNX export by onnx-community — <https://huggingface.co/onnx-community/dinov2-with-registers-small>.
- **DINOv2 ViT-B/14** © Meta Platforms, Inc., Apache License 2.0 (see [`LICENSE`](LICENSE)) —
  <https://github.com/facebookresearch/dinov2>, <https://huggingface.co/facebook/dinov2-with-registers-base>;
  ONNX export by onnx-community — <https://huggingface.co/onnx-community/dinov2-with-registers-base>.
- **VGG16 weights** from torchvision — <https://github.com/pytorch/vision>, BSD-3-Clause
  (<https://github.com/pytorch/vision/blob/main/LICENSE>).
- **Cellpose** © Howard Hughes Medical Institute (HHMI) / MouseLand — <https://github.com/MouseLand/cellpose>,
  BSD-3-Clause (<https://github.com/MouseLand/cellpose/blob/main/LICENSE>). The same cyto3 weights
  are redistributed by the BioImage Model Zoo as `famous-fish` ("CellPose(cyto3)",
  <https://bioimage.io/#/?id=famous-fish>) under BSD-3-Clause. **Note:** per the Cellpose README,
  the datasets the built-in models were trained on include images under CC-BY-NC licences; check
  that your use fits those terms. Use of the built-in models in Voxide was confirmed with the
  Cellpose authors on 2026-10-24.
- **micro-sam ViT-T** (CC-BY-4.0) — SAM LM Generalist (ViT-T) image encoder from micro-sam, by
  Anwai Archit and Constantin Pape (University of Göttingen;
  https://github.com/computational-cell-analytics/micro-sam), BioImage Model Zoo `faithful-chicken`
  v1.1 (https://bioimage.io/#/?id=faithful-chicken), licensed under CC BY 4.0
  (https://creativecommons.org/licenses/by/4.0/). Changed: the image encoder alone, converted to
  ONNX, with a bilinear resize of a 448 × 448 input to 1024 × 1024 in front of it and its attention
  biases set to zero as micro-sam's runtime has them. Provided as is, without warranties. Please
  cite: Archit, A., Freckmann, L., Nair, S. et al. Segment Anything for Microscopy. Nat Methods 22,
  579–591 (2025). https://doi.org/10.1038/s41592-024-02580-4. Architecture: MobileSAM / TinyViT
  (Apache-2.0, https://github.com/ChaoningZhang/MobileSAM), built on Segment Anything (© Meta,
  Apache-2.0).

If you use these models in published work, please cite the original papers:

```bibtex
@article{oquab2023dinov2,
  title   = {DINOv2: Learning Robust Visual Features without Supervision},
  author  = {Oquab, Maxime and others},
  journal = {arXiv preprint arXiv:2304.07193},
  year    = {2023}
}

@article{darcet2023registers,
  title   = {Vision Transformers Need Registers},
  author  = {Darcet, Timoth{\'e}e and Oquab, Maxime and Mairal, Julien and Bojanowski, Piotr},
  journal = {arXiv preprint arXiv:2309.16588},
  year    = {2023}
}

@article{simonyan2014vgg,
  title   = {Very Deep Convolutional Networks for Large-Scale Image Recognition},
  author  = {Simonyan, Karen and Zisserman, Andrew},
  journal = {arXiv preprint arXiv:1409.1556},
  year    = {2014}
}

@article{stringer2021cellpose,
  title   = {Cellpose: a generalist algorithm for cellular segmentation},
  author  = {Stringer, Carsen and Wang, Tim and Michaelos, Michalis and Pachitariu, Marius},
  journal = {Nature Methods},
  volume  = {18},
  pages   = {100--106},
  year    = {2021}
}

@article{stringer2025cellpose3,
  title   = {Cellpose3: one-click image restoration for improved cellular segmentation},
  author  = {Stringer, Carsen and Pachitariu, Marius},
  journal = {Nature Methods},
  year    = {2025}
}

@article{archit2025microsam,
  title   = {Segment Anything for Microscopy},
  author  = {Archit, Anwai and Freckmann, Luca and Nair, Sushmita and others},
  journal = {Nature Methods},
  volume  = {22},
  pages   = {579--591},
  year    = {2025},
  doi     = {10.1038/s41592-024-02580-4}
}
```