AI Algorithm Path

Aug 16, 2025 · Artificial Intelligence

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

Meta's DINOv3 is a 70‑billion‑parameter self‑supervised visual foundation model trained on 17 billion Instagram images without any labels, introducing dense feature extraction, Gram‑Anchoring to prevent feature collapse, high‑resolution adaptation, and multi‑student distillation that together enable out‑of‑the‑box performance on segmentation, depth estimation, 3D matching, and tracking while surpassing prior models such as DINOv2, CLIP, and SAM.