DINOv2

DINOv2 is a family of self-supervised vision transformers from Meta AI, pretrained on 142 million curated images without any labels. The resulting features are general purpose: a frozen DINOv2 backbone with a simple linear head handles image classification, semantic segmentation, depth estimation, and retrieval without fine-tuning, reaching up to 87.1 percent linear-probe accuracy on ImageNet. Four distilled sizes are published, from ViT-S at 21M parameters through ViT-B and ViT-L up to ViT-g at 1.1B, each also available with register tokens that improve feature map quality. Pretrained classification, segmentation, and depth heads are provided for immediate use, and specialized variants exist for X-ray and cell microscopy imagery under non-commercial research licenses. The main code and weights are Apache 2.0, and the line has continued into DINOv3. Computer vision researchers and practitioners commonly adopt it as a frozen feature extractor when labeled data is scarce.