DINO v3

Intended use. Extract visual features from images without needing labeled training data. Ideal for insect classification, phenotyping, and behavior analysis when you have many images but few annotations. Works well as a starting point (backbone) for other tasks like segmentation or object tracking, but produces features—not final predictions—so you’ll need additional lightweight components for those outputs.

Architecture and training

DINOv3 uses Vision Transformer (ViT) and ConvNeXt architectures trained without human labels. Instead, it learns by having smaller models mimic a larger teacher model through self-distillation. The training uses 1.7 billion unlabeled images from the web, scaled up 6× in model size and 12× in data compared to DINOv2.

Key features include rotary positional embeddings (allowing arbitrary image resolutions), register tokens for better global context, and a novel Gram anchoring technique that preserves fine-grained spatial details.

Inputs and outputs

Inputs: RGB images at any resolution (256×256 to 4096×4096 tested). No text prompts required.

Outputs: Global image embeddings (for classification) and dense feature maps (for segmentation, depth estimation, detection). Normalization uses standard ImageNet statistics.

Performance

Frozen backbone evaluation (no fine-tuning required):

Domain-specific: Forest canopy height mapping achieved R²=0.86 (vs. 0.53 for DINOv2). Sources: Meta AI DINOv3 technical report (2025).

Limitations

Critical constraints for insect monitoring:

• Not end-to-end detection: Produces features only; requires separate heads for bounding boxes or masks
• Small insects: Dense features may degrade for very small specimens (<50px²); test higher-resolution variants
• Fine-grained taxonomy: Trained on general web data; specialized insect features may require fine-tuning
• Compute & memory requirements: ViT-7B needs 4-8 A100 GPUs; ViT-L (300M) minimum for research-grade accuracy. Smaller models are available for edge deployment.

DINOv3 differs from Grounding DINO—it lacks text-grounded detection. Use Grounding DINO for query-based detection, DINOv3 for unsupervised feature extraction.

How to obtain the weights

Hugging Face: All variants under facebook/dinov3-*.

Loading example (PyTorch Hub):

model = torch.hub.load('facebookresearch/dinov3', 'dinov3_vitb16')

Citation

@misc{simeoni2025dinov3,
  title={{DINOv3}},
  author={Sim{\'e}oni, Oriane and Vo, Huy V. and Seitzer, Maximilian and Baldassarre, Federico and Oquab, Maxime and Jose, Cijo and Khalidov, Vasil and Szafraniec, Marc and Yi, Seungeun and Ramamonjisoa, Micha{\"e}l and Massa, Francisco and Haziza, Daniel and Wehrstedt, Luca and Wang, Jianyuan and Darcet, Timoth{\'e}e and Moutakanni, Th{\'e}o and Sentana, Leonel and Roberts, Claire and Vedaldi, Andrea and Tolan, Jamie and Brandt, John and Couprie, Camille and Mairal, Julien and J{\'e}gou, Herv{\'e} and Labatut, Patrick and Bojanowski, Piotr},
  year={2025},
  eprint={2508.10104},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2508.10104},
}