DINOv3

DINOv3 is Meta AI's family of vision foundation models for researchers and developers building image analysis systems on their own hardware. Its pretrained models extract features for individual image regions that can support multiple vision tasks without fine-tuning the backbone.

• Backbone choices: Choose Vision Transformer (ViT) or ConvNeXt models, including smaller distilled ViT variants and larger backbones.
• Satellite imagery: Separate ViT weights trained on satellite images support remote sensing work alongside the general-purpose models trained on web images.
• Task-specific models: Pretrained heads cover ImageNet classification, COCO2017 object detection, ADE20K semantic segmentation and single-image depth estimation trained on SYNTHMIX.
• Training and adaptation: Use code for training segmentation and depth prediction layers on pretrained features, plus distillation code for ConvNeXt backbones.

The models run locally through PyTorch and can load weights stored on your machine. CUDA support is strongly recommended. Backbones also work with Hugging Face Transformers and timm, and weights are available through Hugging Face Hub.

Access to model downloads requires approval, after which download links arrive by email. The related CHMv2 model uses DINOv3 to estimate canopy height from imagery.