DINOv2: self-supervised visual features

A family of vision transformer backbones trained without labels. Their features are good enough that a simple linear head often works for classification, depth or segmentation, with no fine-tuning.

• Smaller models are distilled from the largest ViT-g
• "With registers" variants give cleaner attention maps
• Some add-on models in the same repo (Cell-DINO, XRay-DINO) are non-commercial; check before use

Summary written by SolxLab. Weights and code belong to their authors and are distributed under the license shown; always check the official page before use.