DINOv2: Learning Robust Visual Features without Supervision | VitaVision
PaperTMLR 2024 (arXiv 2023) · 2023

DINOv2: Learning Robust Visual Features without Supervision

Maxime OquabTimothée DarcetThéo MoutakanniHuy VoMarc SzafraniecVasil KhalidovPierre FernandezDaniel HazizaFrancisco MassaAlaaeldin El-NoubyMahmoud AssranNicolas BallasWojciech GalubaRussell HowesPo-Yao HuangShang-Wen LiIshan MisraMichael RabbatVasu SharmaGabriel SynnaeveHu XuHervé JégouJulien MairalPatrick LabatutArmand JoulinPiotr Bojanowski
arXiv 2304.07193DOIView in graph
1page built on it
7more pages cite it
4narratives
2registry papers cite it

In the Atlas

ModelPrimary source of
DINOv2

A self-supervised ViT trained on a curated 142M-image dataset that yields general-purpose visual features usable frozen — via kNN or linear probes — for classification, dense depth and segmentation without finetuning.

Cited as a reference by 7 pages, grouped by domain

Representation learning 3
CLIPDINOv3Self-Supervised Learning
Depth 2
Depth AnythingDepth Anything V2
Geometry 2
Depth Anything 3VGGT (Visual Geometry Grounded Transformer)

Appears in 4 narratives

NarrativeDepth Becomes General GeometryNarrativeFoundation Models for VisionNarrativeThe Geometry Pipeline CollapsesNarrativeWhere Did the Inductive Bias Go?

Lineage

Papers in the Atlas registry only · one dot per paper
Built upon by · 2
’23’24’25Vision Transformers Need Registers (2023)DINOv3 (2025)
Builds on · 0