Research map / Label-efficient and robust learning / Self-supervised and contrastive
DINOv3
Oriane Siméoni, Huy Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seung Eun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza and 14 more
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Technical report on DINOv3, a self-supervised visual foundation model aimed at learning general-purpose representations from diverse image sources (natural to aerial) with a single algorithm.
Problem
Self-supervised learning should remove the need for manual annotation and scale to massive datasets and larger architectures, learning visual representations across diverse domains with one algorithm.
Approach
Scaling a self-supervised vision training recipe (DINO lineage) to learn general visual representations from diverse sources using a single algorithm; details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use combined curated data mixture (clustering + retrieval + public datasets) instead of raw data (ablation) | DINOv3 (200k iterations) · iNaturalist 2021 | accuracy: 70.1 -> 87.0 | Improved |
| use data mixture instead of retrieval-only curation (ablation) | DINOv3 (200k iterations) · Paris retrieval | accuracy: 82.7 -> 85.9 | Improved |
| use clustering-only curation instead of raw data (ablation) | DINOv3 (200k iterations) · IN1k k-NN | accuracy: 80.1 -> 79.4 | Hurt |
| add Gram anchoring refinement (ablation) | DINOv3 ViT-7B · ADE20k | mIoU | Improved |
| use high-resolution Gram teacher features (L_HRef vs L_Ref) (ablation) | DINOv3 ViT-7B · ADE20k | mIoU: +2 mIoU | Improved |
| use a much later Gram teacher (1M iterations) instead of 100k/200k (ablation) | DINOv3 · dense tasks | mIoU | Hurt |
| train longer without Gram anchoring (ablation) | DINOv3 ViT-7B · Pascal VOC segmentation | mIoU | Hurt |
| add dedicated layer norm for local and global crop outputs (ablation) | DINOv3 · ImageNet kNN / ADE20k / NYUv2 | kNN acc / mIoU / RMSE: +0.2 acc, +1 mIoU, -0.02 RMSE | Improved |
| apply high-resolution adaptation (ablation) | DINOv3 ViT-7B · ObjectNet at low resolution | accuracy | Mixed |
| omit Gram anchoring in high-resolution adaptation (ablation) | DINOv3 ViT-7B · dense prediction | mIoU | Hurt |
| scale to ViT-7B with SSL and Gram anchoring vs DINOv2 | DINOv3 · KITTI depth | RMSE: 2.624 -> 2.346 | Improved |
| use SSL-only DINOv3 vs weakly-supervised PE-core on OCR tasks | DINOv3 · GTSRB | accuracy: 94.8 -> 87.5 | Hurt |
Limitations
- Only the abstract was available; experimental details are unknown
Experimental setup
Builds on: DINO DINOv2
Closest papers by meaning
- DINOv2 Meets Text: A Unified Framework for Image- and Pixel-Level Vision-Language Alignment (CVPR 2025 · similarity 0.87)
- Simplifying DINO via Coding Rate Regularization (ICML 2025 · similarity 0.86)
- Beyond Random Augmentations: Pretraining with Hard Views (ICLR 2025 · similarity 0.86)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.86)
- DistillNeRF: Perceiving 3D Scenes from Single-Glance Images by Distilling Neural Fields and Foundation Model Features (NeurIPS 2024 · similarity 0.86)
- VESSA: Video-based objEct-centric Self-Supervised Adaptation for Visual Foundation Models (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
