Meta Unveils DINOv3: A Universal Self‑Supervised Visual AI for All Image Tasks

Meta has released DINOv3, a visual foundation model that learns dense image features without any supervision. Unlike earlier DINO versions, DINOv3 extracts per‑patch semantic information, which benefits dense tasks such as segmentation, object tracking, depth estimation, and 3D matching, and it can be used directly without fine‑tuning.

The core model is a 70‑billion‑parameter Vision Transformer (ViT‑7B) trained from scratch on 17 billion raw Instagram images. No external datasets (e.g., JFT‑300M, LAION) or annotation labels are used. Data selection combines hierarchical k‑means clustering for visual diversity, retrieval‑based sampling for concept relevance, and a small amount of ImageNet data to maintain balance, avoiding the naïve "throw‑everything‑in" approach.

To address feature collapse during long training of large models, Meta introduced a novel loss called Gram Anchoring . This loss compares the similarity structure of block features at the current training step with that of an early checkpoint, allowing moderate drift while strictly preserving the relational structure of features, thereby stabilizing training for the 70‑billion‑parameter model.

DINOv3 also incorporates a high‑resolution adaptation stage. After the main training, the model is fine‑tuned on crops of 512, 768 or higher resolution images, combined with Gram Anchoring, which dramatically improves generalisation to varying input sizes, enabling robust performance on 4K satellite imagery, aerial maps, and complex street scenes.

Once trained, the backbone can be frozen and applied to multiple tasks without adding task‑specific heads. Simple linear classifiers, k‑nearest‑neighbors, or lightweight clustering layers suffice for semantic segmentation (ADE20K, Cityscapes, Pascal VOC), monocular depth estimation (NYUv2, KITTI), 3D correspondence, and video understanding, where DINOv3 consistently outperforms DINOv2, CLIP‑based models (e.g., SigLIP), and even composite models such as AM‑RADIO.

For users with limited compute, Meta distilled the 70‑billion‑parameter model into smaller ViT variants (ViT‑S ≈ 21M, ViT‑B ≈ 86M, ViT‑L ≈ 300M, ViT‑H+ ≈ 800M). A parallel multi‑student distillation framework shares teacher outputs across GPUs, preserving about 90 % of the original performance on dense tasks while delivering substantial inference speed gains.

Although DINOv3 is purely visual, a text encoder can be added using a CLIP‑style contrastive objective to align pooled visual features and patch features with textual embeddings, enabling zero‑shot classification and retrieval at both global and fine‑grained levels.

In summary, DINOv3 breaks the supervision bottleneck by learning solely from raw pixels, provides a universal dense feature extractor, scales reliably to 70 billion parameters, and demonstrates strong cross‑domain generalisation, making it a compelling foundation for researchers building their own visual AI pipelines.