DINOv2

Released April 2023, DINOv2 is Meta's self-supervised method for training computer vision models that learn from unlabeled images alone — no image-text pairs, no manual labels. Trained on 142 million curated images with a 1B-parameter ViT distilled into smaller variants, it produces all-purpose visual features that work across image distributions and tasks without fine-tuning.

DINOv2 surpassed OpenCLIP on most image- and pixel-level benchmarks and became a go-to backbone for depth estimation, semantic segmentation, and instance retrieval. Being purely visual (no language supervision), it captures pixel-level understanding that text-guided models miss. Code and models are openly released by Meta Research.