Meta recently announced the launch of DINOv3, a general-purpose image processing AI model that requires no labeled data. The model was trained using self-supervised learning on 1.7 billion images and has built up 7 billion parameters, enabling it to handle various image tasks and domains with minimal adjustments.
This feature makes DINOv3 particularly valuable in specialized fields where labeled data is limited, such as satellite image processing. Meta stated that DINOv3 performs well in challenging benchmark tests that previously required dedicated systems, and its performance surpasses the previous generation model DINOv2, although the improvement is not as significant as the jump from v1 to v2.
To promote the adoption and application of this technology, Meta has released multiple pre-trained model variants on GitHub, along with adapters and training and evaluation code. These resources are available under the DINOv3 License, allowing commercial use.
Address: https://github.com/facebookresearch/dinov3
