DINOv3
Paper • 2508.10104 • Published • 319
You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
Access is approved automatically. By requesting access, you share your Hugging Face username and email with HistoForge. We may get in touch to ask for your feedback.
Log in or Sign Up to review the conditions and access this model content.
H&E-12 HoVer Starter
A starting model for training HistoForge to separate individual tissue structures. Consider HoVer for large objects: it uses watershed separation instead of Cellpose's repeated flow-following steps. It can also be trained to segment nuclei.
It reuses H&E-12 nucleus-pretrained features with newly initialized segmentation layers that need training.
Compare starter models
H&E-12 starters include histology pretraining. DINOv3 starters use a pretrained image encoder with newly initialized segmentation layers.
| Method | Start here for | H&E-12 weights | DINOv3 weights |
|---|---|---|---|
| Cellpose | Nuclei, cells and irregular or elongated objects | H&E-12 Cellpose | DINOv3 Cellpose |
| HoVer | Large individual tissue structures | H&E-12 HoVer | DINOv3 HoVer |
| Dense semantic | Tissue compartments, rather than separate cells | H&E-12 Semantic | DINOv3 Semantic |
Cellpose and HoVer can separate touching nuclei. These are starting recommendations, not an accuracy or speed ranking; compare trained results on held-out images.
Using the module
Install histoforge/starter-he12-hovernet and select it as the starting model when training from your QuPath annotations. Define your target classes and image resolution, train, then validate the resulting model on held-out images before using its predictions.
| Output class | What it marks |
|---|---|
| Object | Individual segmented objects |
This starter has no instance-classification head. Classification can be added when defining your training task.
Measurements
No task-specific measurements are included.
Limitations
- The HoVer head is untrained; its predictions are not meaningful before fine-tuning.
- This starter has no task-specific accuracy benchmark or prediction gallery.
- Input images must be RGB with valid pixel-size calibration. The initial recipe uses 0.25 microns per pixel; choose the resolution appropriate for your training data.
Models and training
| Component | Initialization |
|---|---|
| DINOv3 ViT-L/16 backbone | H&E nucleus-pretrained |
| Spatial adapter and feature readout | H&E nucleus-pretrained |
| Task head | Newly initialized horizontal/vertical and foreground head |
The model uses the shared 256-channel spatial adapter and four CNN-transformer interaction blocks. Newly initialized components use seed 42. Neural tiles are 256 pixels with 10% overlap; the complete settings are in recipe.json.
H&E-12 used 5,022 eligible training images from twelve nucleus datasets. See training data for sources and licences. These datasets have already influenced the initializer and cannot automatically serve as independent downstream holdouts.
Licences and citations
- Weights: DINOv3 licence.
- Foundation model: Simeoni et al., DINOv3 (2025).
Model tree for histoforge/starter-he12-hovernet
Base model
timm/vit_large_patch16_dinov3.lvd1689m