Skip to main content
Back

Meta's DINOv3 Scales Vision AI to 7B Parameters Without Task Labels

Meta releases DINOv3, a self-supervised vision backbone family scaling to 7B parameters and 1.7B images, with frozen features that beat specialized baselines.

  • Meta released DINOv3, a family of self-supervised vision backbones from ViT-S up to a 7B teacher.
  • Trained on 1.689B curated web images (LVD-1689M) plus a 493M satellite variant, with no human labels.
  • Introduces Gram anchoring, a regularizer that stops dense features from degrading during long training.
  • Frozen features beat specialized baselines on classification, segmentation, depth, retrieval, and 3D keypoints.
  • ViT-L/16 hits 90.2 on ImageNet-ReaL and 54.9 mIoU on ADE20k with a linear head.
  • Full code and 12 checkpoints at facebookresearch/dinov3, gated under a custom Meta license.

Meta has released DINOv3, a family of self-supervised vision backbones trained to produce global image embeddings and dense patch features. Its largest model is a 6.7B-parameter Vision Transformer trained on 1.689 billion curated images, then used to distill smaller models for practical deployment.

Meta reports that frozen DINOv3 features outperform previous self-supervised, weakly supervised, and supervised backbones across classification, retrieval, segmentation, depth estimation, and correspondence benchmarks. The 300M-parameter ViT-L/16 offers the clearest starting point for teams balancing quality, memory, and inference cost.

Frozen features simplify downstream training

DINOv3 pretraining uses images without task-specific human annotations. Developers can keep the backbone weights frozen and train a lightweight head for the target task, reducing the data, compute, and engineering required for full-model fine-tuning.

The released representations support several evaluation and deployment patterns:

  • k-nearest neighbors or logistic regression on the class token for image classification
  • Nearest-neighbor image retrieval
  • Linear probes for semantic segmentation and depth estimation
  • Geometric and semantic 3D keypoint correspondence
  • Unsupervised object discovery and video segmentation tracking
  • Video classification with a four-layer attentive probe

Frozen does not mean training-free for every task. Classification, segmentation, and depth systems generally need a fitted head, labeled examples, or a reference index. The base checkpoints are visual encoders without a native text encoder, so they do not directly replace CLIP or SigLIP in text-to-image retrieval and prompt-based classification.

Gram anchoring protects dense features

Extended DINO-style training can improve global classification features while degrading the spatial structure of patch features. That trade-off weakens the representations needed for segmentation, depth prediction, and point correspondence.

DINOv3 addresses the failure mode with Gram anchoring. A Gram matrix records pairwise relationships among feature vectors. The added objective anchors those relationships during training, limiting drift in the dense feature map while the global representation continues to improve. Meta combines the method with post-training adaptations for resolution, model size, and text alignment.