DINOv2 — Self-Supervised Vision Foundation Model

DINOv2

DINOv2 is Meta AI's second-generation self-supervised vision transformer that learns powerful, general-purpose visual features without any labels. DINOv2 features transfer remarkably well to downstream tasks — often matching or beating supervised models — with no fine-tuning.

Key Features

• Universal features: linear probing achieves top-1 on ImageNet without fine-tuning
• Scales: ViT-S/B/L/g (21M–1.1B parameters)
• Tasks: image classification, semantic segmentation, monocular depth, retrieval, video understanding
• Curated 142M-image pre-training dataset (LVD-142M)
• Pre-trained weights released under Apache-2.0
• Torch Hub and Hugging Face Hub integration

Quick Start

import torch

# Load from Torch Hub (no clone needed)
dinov2 = torch.hub.load("facebookresearch/dinov2", "dinov2_vitl14")
dinov2.eval()

from torchvision import transforms
from PIL import Image

img = transforms.ToTensor()(Image.open("photo.jpg").resize((224, 224))).unsqueeze(0)
with torch.no_grad():
    features = dinov2(img)  # (1, 1024)
print(features.shape)

Install via ai-supply

npx ai-supply add dinov2-self-supervised-vision-features

Curated mirror of the open-source DINOv2 (Apache-2.0). Get it from the source.