Download
DINOv3 is the third-generation iteration of Meta’s self-supervised visual representation learning framework, building upon the ideas from DINO and DINOv2. It continues the paradigm of learning strong image representations without labels using teacher–student distillation, but introduces a simplified and more scalable training recipe that performs well across datasets and architectures. DINOv3 removes the need for complex augmentations or momentum encoders, streamlining the pipeline while maintaining or improving feature quality. The model supports multiple backbone architectures, including Vision Transformers (ViT), and can handle larger image resolutions with improved stability during training. The learned embeddings generalize robustly across tasks like classification, retrieval, and segmentation without fine-tuning, showing state-of-the-art transfer performance among self-supervised models.
Features
- Simplified self-supervised learning framework with improved scalability
- Teacher–student distillation without labeled data or heavy augmentation
- Support for multiple backbones including Vision Transformers
- Stable high-resolution training and distributed multi-GPU setup
- High transferability to classification, retrieval, and segmentation tasks
- Ready-to-use scripts for training, feature extraction, and benchmarking
Project Samples
