PaperlayerView exact revision on arXiv

Source-linked AI summary

DINOv3

Oriane Siméoni, Huy V. Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yi, Michaël Ramamonjisoa, Francisco Massa, Daniel Haziza, Luca Wehrstedt, Jianyuan Wang, Timothée Darcet, Théo Moutakanni, Leonel Sentana, Claire Roberts, Andrea Vedaldi, Jamie Tolan, John Brandt, Camille Couprie, Julien Mairal, Hervé Jégou, Patrick Labatut, Piotr Bojanowski

arXiv:2508.10104v1cs.CVcs.LG

TL;DR

Long self-supervised training can improve global benchmarks while degrading dense features through patch-level inconsistencies. DINOv3 addresses this with large-scale training and Gram anchoring, achieving state-of-the-art dense-task performance with a frozen backbone across diverse domains.

  • Problem

    Prolonged self-supervised training improves global benchmarks but degrades dense-task performance as patch-level feature inconsistencies emerge.

    Source: §4 · p. 11

  • Method

    DINOv3 scales self-supervised training across data and model size, using Gram anchoring to preserve dense features and distillation to produce deployable model variants.

    Source: §1 · pp. 3–4

  • Results

    With a frozen backbone, DINOv3 achieves state-of-the-art across dense tasks, including 66.1 mAP on COCO detection, and surpasses prior approaches on satellite imagery.

    Source: §1 · p. 5

  • Takeaways & Limitations

    DINOv3 provides a single frozen visual encoder and scalable model family for computer-vision tasks spanning diverse domains and resource constraints.

    Source: §1 · pp. 2–3

    Continued training can produce additional feature outliers, whose different types are discussed separately in the appendix.

    Source: §4.1 · p. 11

Abstract

from arXiv · show

Self-supervised learning holds the promise of eliminating the need for manual data annotation, enabling models to scale effortlessly to massive datasets and larger architectures. By not being tailored to specific tasks or domains, this training paradigm has the potential to learn visual representations from diverse sources, ranging from natural to aerial images -- using a single algorithm. This technical report introduces DINOv3, a major milestone toward realizing this vision by leveraging simple yet effective strategies. First, we leverage the benefit of scaling both dataset and model size by careful data preparation, design, and optimization. Second, we introduce a new method called Gram anchoring, which effectively addresses the known yet unsolved issue of dense feature maps degrading during long training schedules. Finally, we apply post-hoc strategies that further enhance our models' flexibility with respect to resolution, model size, and alignment with text. As a result, we present a versatile vision foundation model that outperforms the specialized state of the art across a broad range of settings, without fine-tuning. DINOv3 produces high-quality dense features that achieve outstanding performance on various vision tasks, significantly surpassing previous self- and weakly-supervised foundation models. We also share the DINOv3 suite of vision models, designed to advance the state of the art on a wide spectrum of tasks and data by providing scalable solutions for diverse resource constraints and deployment scenarios.

1 Introduction

DINOv3 advances self-supervised learning at scale by combining data and model scaling, Gram anchoring for dense features, and post-training strategies into versatile frozen vision models. It achieves state-of-the-art dense-task performance while matching competing models on global tasks across diverse domains and deployment constraints.

Superior Feature Maps Through Gram Anchoring

  • Gram anchoring mitigates dense-feature collapse in large models and long training schedules, producing cleaner high-resolution maps and stronger geometric-task features than DINOv2.The strategy improves both parametric and non-parametric dense tasks while preserving high-level semantic performance.Source: §1 · pp. 3–4

The DINOv3 Family of Models

  • DINOv3 scales self-supervised learning to a 7B-parameter model and distills it into ViT Small, Base, Large, and ConvNeXt variants for diverse resource constraints.The efficient ViT-L variant performs close to the 7B teacher across a variety of tasks, while the family matches or exceeds competing models on global benchmarks and significantly outperforms them on dense prediction.Source: §1 · p. 3, Figure 2 · p. 3

Strong & Versatile Foundational Models

  • Carefully curated background data mixed with ImageNet-1k enables scaling from unconstrained image collections while supporting generalization across tasks and domains.The approach targets observational domains lacking metadata, including histopathology, biology, medical imaging, remote sensing, astronomy, and high-energy particle physics.Source: §1 · pp. 2–4

Model Architecture and Training

  • A custom 7B ViT uses axial RoPE, positional-artifact regularization, and constant hyperparameter schedules for 1M iterations instead of multiple cosine schedules.These architecture and training changes are presented as enabling stronger performance at scale.Source: §1 · pp. 2, 4

Post-training and Distillation

  • High-resolution post-training and single-teacher multiple-students distillation transfer the frontier model’s capabilities into practical smaller models.The pipeline includes a high-resolution phase and an efficient distillation procedure designed for common usage.Source: §1 · pp. 4–5

Benchmarking

  • mAP 66.1 on COCO detection and mIoU 63.0 on ADE20k establish state-of-the-art performance with a frozen backbone, outperforming specialized fine-tuned pipelines.DINOv3 also surpasses prior approaches on satellite imagery and performs comparably to CLIP derivatives on global tasks.Source: §1 · p. 5

2 Related Work

This section situates DINOv3 within self-supervised learning, scalable vision foundation models, and methods for improving dense transformer features. It highlights Gram-based regularization and native high-resolution feature stability as distinctions from prior work.

Self-Supervised Learning

  • Self-supervised vision develops supervisory signals from image structure, including patch prediction, reordering, inpainting, recoloring, and transformation prediction.Inpainting later connected masked image modeling to denoising auto-encoding and latent-space prediction, while JEPA extended the paradigm to video.Source: §2 · p. 5
  • Discriminative self-supervised methods use instance classification, contrastive or information-theoretic objectives, clustering, and masked reconstruction to learn strong benchmark features, but often struggle to scale to larger models.This line of work includes iBOT, which combines discriminative losses with masked reconstruction.Source: §2 · p. 5

Vision Foundation Models

  • Vision foundation models progressed from supervised ImageNet representations and architectural advances to larger datasets, ViTs, weak supervision, and multimodal contrastive training.Scaling efforts include JFT, pseudo-labeling, ViT-22B, CLIP-style methods, and datasets exceeding 40B image-text pairs.Source: §2 · pp. 5–6
  • Unsupervised image pretraining has comparatively little scaling work, although DINOv2 showed that careful algorithms, larger architectures, and extensive data could match or surpass open-source CLIP variants across tasks.Subsequent work further explored larger models, open datasets, and improved training recipes.Source: §2 · p. 6

Dense Transformer Features

  • Dense transformer features support multimodal, generative, 3D, video, robotic, detection, segmentation, and depth applications, motivating local consistency losses, register tokens, distillation, and post-hoc enhancement.DINOv3 instead uses a Gram objective with earlier SSL iterations as teachers to regularize student-teacher patch similarity, while delivering stable, high-quality dense maps across resolutions.Source: §2 · pp. 6–7, Figure 4 · p. 7

3 Training at Scale Without Supervision

DINOv3 scales self-supervised learning through carefully curated data, a combined global-local training objective, and a 7B-parameter architecture optimized for robust visual representations. The section covers complementary data curation and sampling strategies, plus design and optimization choices intended to improve global and dense features.

3.1 Data Preparation

  • DINOv3 combines balanced clustering, task-relevant retrieval, and raw vision datasets from an approximately 17-billion-image web pool.The clustering branch produces the 1,689-million-image LVD-1689M subset, while retrieval targets concepts represented in selected seed datasets and raw data includes ImageNet1k, ImageNet22k, and Mapillary Street-level Sequences.Source: §3.1 · p. 8
  • The full data-curation pipeline performs best overall, because clustering and retrieval alone each fail to dominate across all downstream benchmarks.The ablation compares curated and raw datasets using a shorter 200k-iteration schedule; the full mixture combines their complementary strengths.Source: Table 1 · p. 8, §3.1 · p. 9

3.2 Large-Scale Training with Self-Supervision

  • The training objective mixes image-level DINO, patch-level iBOT reconstruction, Sinkhorn-Knopp normalization, and Koleo regularization across global and local features.Dedicated heads and crop-specific layer normalization support feature specialization; the normalization change improves ImageNet kNN accuracy by 0.2, ADE20k segmentation by 1 mIoU, and NYUv2 depth by reducing RMSE by 0.02.Source: §3.2 · p. 9
  • DINOv3 scales to 7B parameters with RoPE-box jittering and uses constant learning rate, weight decay, and teacher momentum during large-batch multi-crop training.The setup uses AdamW, 4096 images across 256 GPUs, two global crops, eight local crops, and a 3.7M-token sequence per batch, with warmup retained for learning rate and teacher temperature.Source: §3.2 · p. 10

4 Gram Anchoring: A Regularization for Dense Features

Extended training improves global metrics but degrades dense-task performance through worsening patch-level consistency. Gram anchoring regularizes these relationships, restoring dense performance and gaining further improvements by distilling smoothed high-resolution features.

4 Gram Anchoring: A Regularization for Dense Features

  • Gram anchoring targets the patch-level inconsistencies that emerge during prolonged training, addressing dense-feature degradation while preserving high global performance.The method is motivated by the unstable balance between global DINO and local iBOT objectives as training progresses.Source: §4 · p. 11, §4.1 · p. 12, §4.2 · p. 12

4.1 Loss of Patch-Level Consistency Over Training

  • During extended training, classification accuracy keeps improving, but segmentation declines after about 200k iterations and falls below early levels for ViT-7B.The degradation is associated with increasingly irrelevant patches showing high cosine similarity to a reference patch.Source: §4.1 · p. 11

4.2 Gram Anchoring Objective

  • The Gram loss preserves patch-similarity structure by matching the student’s Gram matrix to an early teacher, allowing local features to change without losing consistency.The loss uses pairwise dot products of L2-normalized patch features and is updated through a refinement step.Source: §4.2 · p. 12
  • Gram anchoring significantly improves dense tasks within 10k iterations, yields notable ADE20k gains after Gram-teacher updates, and mildly affects global benchmarks.Longer training further benefits ObjectNet, while the method operates on global crops and begins after 1M iterations for efficiency.Source: §4.2 · pp. 12–13, Figure 8 · p. 13

4.3 Leveraging Higher-Resolution Features

  • High-resolution Gram refinement distills improved patch-level consistency of smoothed high-resolution features into the student, yielding +2 mIoU on ADE20k beyond LRef; a 1M-iteration Gram teacher is detrimental because its patch-level consistency is inferior.Source: §4.3 · p. 13

5 Post-Training

Post-training extends DINOv3 to variable resolutions, efficient smaller models, and text-aligned zero-shot capabilities. These stages improve high-resolution dense features, preserve strong performance under reduced compute, and align global and local visual features with text.

5.1 Resolution Scaling

  • Resolution adaptation improves high-resolution dense prediction while maintaining stable classification across resolutions and producing stable feature maps above 4k resolution.The 7B model uses mixed-resolution crops for 10k iterations, with Gram anchoring essential to prevent dense-task degradation.Source: Figure 11 · p. 15, §5.1 · p. 15

5.2 Model Distillation

  • Distillation transfers the fixed ViT-7B teacher’s representations to ViT-S, ViT-B, and ViT-L students without applying Gram anchoring.The students use the first-phase objective, while the 7B model directly replaces the EMA teacher.Source: A Family of Models for Multiple… · p. 15

A Family of Models for Multiple Use-Cases

  • Distilled models deliver frontier-level performance across compute budgets, using 1M training iterations, 250k cooldown iterations, and subsequent high-resolution adaptation.The model family includes standard ViT-S, B, and L variants plus custom ViT-S+ and ViT-H+ models.Source: A Family of Models for Multiple… · pp. 15–16
  • Parallel multi-student distillation shares teacher inference across GPU groups, reducing per-GPU compute while adding only each new student’s training cost.Student groups are sized to balance iteration times and minimize synchronization idle time.Source: A Family of Models for Multiple… · p. 16, Figure 12 · p. 16

5.3 Aligning DINOv3 with Text

  • Text alignment freezes DINOv3, adds two transformer layers, and matches captions using concatenated mean-pooled patch embeddings and CLS features.This LiT-style contrastive training aligns both global and local visual features and improves dense prediction without additional heuristics.Source: §5.3 · p. 17

6 Results

This section evaluates frozen or minimally tuned DINOv3 representations across a broad range of dense, global, and geometric vision tasks. The results cover dense prediction, correspondence, object discovery, tracking, classification, retrieval, detection, semantic segmentation, monocular depth estimation, and 3D understanding.

6 Results

  • Frozen DINOv3 representations achieve strong performance across dense, global, and downstream vision tasks without fine-tuning the backbone.The evaluation covers lightweight probes and more complex systems for detection, segmentation, depth, and 3D understanding.Source: §6 · p. 17

6.1.1 Qualitative Analysis

  • DINOv3’s dense features are sharper, less noisy, semantically coherent, and linearly separable for segmentation and depth estimation.They exceed competing baselines on ADE20k and Cityscapes, including 81.1 mIoU on Cityscapes and a 0.278 RMSE improvement over DINOv2 on KITTI.Source: §6.1.1 · p. 18, §6.1.2 · p. 19

6.1.3 3D Correspondence Estimation

  • DINOv3 achieves the best geometric and semantic 3D correspondence recall, outperforming DINOv2 by 4.3% and 2.6%, respectively.The results indicate strong multi-view consistency and support downstream 3D applications.Source: §6.1.3 · p. 20

6.1.4 Unsupervised Object Discovery

  • DINOv3’s localized dense features improve unsupervised object discovery and video tracking, including 5.9 higher CorLoc on VOC 2007 and 83.3 J &F on DAVIS-L.It outperforms competing backbones across tracking resolutions and uses higher input resolution effectively.Source: §6.1.4 · p. 20, Figure 14 · p. 21, §6.1.5 · p. 21

6.2 DINOv3 has Robust and Versatile Global Image Descriptors

  • DINOv3 provides robust global descriptors, surpassing prior self-supervised methods on out-of-distribution and fine-grained classification and leading instance retrieval benchmarks.It reaches 89.8% on iNaturalist21 and improves over DINOv2 by 10.8 points on Met and 7.6 points on AmsterTime.Source: §6.2.1 · pp. 23, 25, Table 7 · p. 24, Table 9 · p. 24

6.3.2 Semantic Segmentation

  • DINOv3’s foundation-model evaluation proceeds from object detection to semantic segmentation using decoders trained on top of the frozen backbone.This part introduces semantic segmentation as another dense per-pixel prediction task requiring localized representations.Source: §6.3 · p. 25, §6.3.2 · p. 26

6.3.1 Object Detection

  • A 100M-parameter detector trained on a frozen DINOv3 backbone reaches state-of-the-art object detection and outperforms prior systems with fewer trained parameters.The detector uses Plain-DETR and evaluates COCO and shifted-distribution COCO-O benchmarks.Source: §6.3.1 · p. 26, Table 10 · p. 26

6.3.2 Semantic Segmentation

  • DINOv3 semantic segmentation reaches state-of-the-art performance on ADE20k, matching ONE-PEACE at 63.0 mIoU and outperforming prior models on COCO-Stuff and VOC 2012.The model uses a frozen DINOv3 backbone with a ViT-Adapter and Mask2Former decoder; high-resolution features may eventually reduce reliance on such heavy decoders.Source: Table 11 · p. 27, §6.3.2 · p. 27

6.3.3 Monocular Depth Estimation

  • DINOv3 combined with Depth Anything V2 achieves a new state of the art for relative monocular depth estimation on all evaluated datasets except DIODE ARel, where DPT remains better.This result uses a frozen backbone and validates DINOv3’s inherited sim-to-real capability for synthetic-data-trained downstream tasks.Source: Table 12 · p. 28, §6.3.3 · p. 28

6.3.4 Visual Geometry Grounded Transformer with DINOv3

  • Replacing DINOv2 with DINOv3 in VGGT produces clear, consistent gains over VGGT’s prior state of the art across camera pose estimation, multi-view estimation, and two-view matching.The swap required minimal tuning and supports DINOv3’s suitability for tasks spanning scene-level abstraction, dense geometric prediction, and pixel-level correspondence.Source: §6.3.4 · pp. 28–29, Table 13 · p. 29

7 Evaluating the Full Family of DINOv3 Models

Across compute budgets, DINOv3 models significantly outperform competing collections on dense prediction while remaining competitive on classification. ViT, ConvNeXt, and text-aligned variants extend this result across architectures, resource constraints, and zero-shot vision-language tasks.

7 Evaluating the Full Family of DINOv3 Models

  • DINOv3 significantly outperforms other model collections on dense prediction and achieves similar classification results across compute budgets.The family spans ViT and ConvNeXt variants designed for diverse computational budgets and deployment scenarios.Source: §7 · p. 29

7.1 A Vision Transformer for Every Use Case

  • DINOv3 models consistently outperform counterparts on dense prediction tasks, exceeding the best competitor by over 6 mIoU points for ViT-L and approximately 3 points for ViT-B on ADE20k.They also achieve competitive results on global recognition benchmarks including ObjectNet and ImageNet1k.Source: §7.1 · pp. 29–30
  • The ViT-H+ student achieves performance on par with the 8-times-larger ViT-7B teacher, validating effective distillation with little or no quality loss.Figure 16b states that ViT-H+ has almost 10× fewer parameters than the 7B teacher while remaining close in performance.Source: Figure 16 · p. 30, §7.1 · pp. 30–31

7.2 Efficient ConvNeXts for Resource-Constrained Environments

  • DINOv3 ConvNeXt models deliver large dense-task gains over supervised ConvNeXt baselines, including +17.9 mIoU for CNX-T and +14.5 mIoU for CNX-L.They slightly lag at 256-resolution in-distribution classification but scale with increased resolution, show significant out-of-distribution classification gaps, and combine performance with computational efficiency.Source: §7.2 · pp. 31–32, Table 15 · p. 32

7.3 Zero-shot Inference with DINOv3-based dino.txt

8 DINOv3 on Geospatial Data

DINOv3’s generic self-supervised recipe transfers to satellite imagery through a 7B model trained on a large RGB dataset, achieving strong results across canopy-height and semantic geospatial tasks. Domain-specific satellite pretraining is especially beneficial for metric tasks, while the web model generalizes strongly to semantic tasks.

8.1 Pre-Training Data and Benchmarks

  • The satellite DINOv3 7B model is pretrained on SAT-493M, comprising 493 million 512 × 512 Maxar RGB ortho-rectified images at 0.6-meter resolution.It uses the web model’s hyperparameters except for satellite-adapted RGB normalization and training length, then distills the 7B model into ViT-Large for lower-budget use.Source: §8.1 · p. 33
  • The recipe applies across diverse geospatial benchmarks, including GEO-Bench, LoveDA, iSAID, DIOR, SatLidar1M, and Open-Canopy, spanning classification, segmentation, detection, and canopy-height estimation.GEO-Bench covers six classification and six segmentation tasks across spatial resolutions and optical bands, while Open-Canopy requires adapting the backbone for a fourth infrared channel.Source: §8.1 · p. 33, §8 · p. 33

8.2 Canopy Height Estimation

  • Canopy-height estimation tests recovery of continuous spatial structure under variation in slope, viewing geometry, illumination, atmospheric scattering, and quantization artifacts.The evaluation trains a DPT head on frozen DINOv3 features using SatLidar1M and also evaluates training and testing on Open-Canopy.Source: §8.2 · pp. 33–34
  • The 7B satellite model sets new state-of-the-art canopy-height results, reducing MAE from 2.4 to 2.2, 3.4 to 3.2, and 2.42 to 2.02 on SatLidar validation, test, and Open-Canopy.The distilled ViT-L satellite model performs comparably overall and achieves the lowest Neon-test MAE of 2.4, versus 2.6 for the 7B satellite model and 2.9 for Tolan et al. (2024).Source: Table 17 · p. 34, §8.2 · p. 34

8.3 Comparison to the Earth Observation State of the Art

  • DINOv3 satellite and web models achieve state-of-the-art results on 12 of 15 Earth-observation classification, segmentation, and detection tasks, despite using frozen RGB-only backbones.The satellite model outperforms prior methods on three unsaturated classification tasks and five of six segmentation tasks, while the web model leads on GEO-Bench, LoveDA, and DIOR.Source: §8.3 · p. 34, Table 18 · p. 35
  • DINOv3’s results indicate task-dependent pretraining benefits: the satellite model excels at physically grounded metric tasks, whereas the web model achieves state-of-the-art semantic geospatial performance.The complementary strengths suggest that satellite-specific priors help canopy-height estimation, while diverse web pretraining supports universal semantic representations.Source: §8.2 · p. 34, §8.3 · p. 35

9 Environmental Impact

DINOv3’s environmental impact is estimated using fixed grid and hardware assumptions, with one ViT-7B pre-training run consuming 47 MWh and the full project estimated at roughly 2600 tCO2eq. The project estimate excludes emissions beyond GPU electricity, such as cooling, manufacturing, and disposal.

Estimation methodology

  • The carbon estimate assumes a PUE of 1.1 and US-average carbon intensity of 0.385 kg CO2eq/KWh, following prior methodologies.GPU power is modeled at 400W for A100s and 700W for H100s.Source: §9 · pp. 35–36

Model training footprint

  • 47 MWh is required to train one DINOv3 model, roughly equivalent to 240,000 km of driving with an average electric vehicle.The report provides detailed ViT-7B pre-training computations and analogous DINOv2 and MetaCLIP data in Table 20.Source: §9 · p. 36, Table 20 · p. 36

Whole-project footprint

  • Based on roughly 9M GPU hours, the whole project is estimated at roughly 2600 tCO2eq, equivalent to half of all 12 daily Paris–New York Boeing 777 return flights.This estimate considers only electricity for powering the GPUs and ignores emissions from cooling, manufacturing, and disposal.Source: §9 · p. 36

10 Conclusion … A.1 High-Norm Patch Outliers

DINOv3 advances scalable self-supervised vision through larger datasets and models, Gram anchoring, and post-hoc polishing, achieving strong task performance without fine-tuning. The appendix examines high-norm patch outliers and adopts register tokens as the most effective mitigation among tested strategies.

10 Conclusion

  • DINOv3 scales self-supervised learning through careful data preparation, design, and optimization, while Gram anchoring prevents dense-feature degradation during extended training.These strategies support robust visual representations without dependence on manual annotations.Source: §10 · p. 37
  • Post-hoc high-resolution training and distillation deliver state-of-the-art performance across visual tasks without fine-tuning the image encoder.The DINOv3 model suite also targets varied resource constraints, deployment scenarios, and application uses.Source: §10 · p. 37

A Artifacts and Outliers in Large-Scale Training

  • Large-scale DINOv3 training reveals patch- and feature-dimension-level outliers whose impacts and mitigation strategies are analyzed.The discussion connects these artifacts to observations in both language and vision models.Source: A Artifacts and Outliers in… · p. 53

A.1 High-Norm Patch Outliers

  • In 7B models, high-norm patch outliers emerge mainly in low-information backgrounds such as skies, echoing findings in DINOv2 and CLIP.The reported initial signals come from models trained for 150k iterations, with patch norms measured before layer normalization.Source: §A.1 · p. 53
  • Four register tokens redirect internal communication between patch tokens and CLS, eliminating visible high-norm outliers and improving ImageNet-1k performance.The strategy follows Darcet et al. (2024) and was not further ablated because of the high experimental cost.Source: §A.1 · p. 53
  • Value gating adds a learnable value bias to attention, while attention bias introduces learnable key and value biases as alternative outlier mitigations.These approaches were selected from An et al. (2025) because they appeared relevant and required minimal attention changes.Source: §A.1 · p. 53
  • Register tokens achieve the best observed performance, whereas attention-based mitigation reduces but does not fully eliminate high-norm patches; registers are therefore used throughout the paper.Value gating removes clear outliers but generally produces higher patch norms.Source: §A.1 · p. 53, Figure 20 · p. 54

A.2 Feature Dimension Outliers · B Additional Results · B.1 Evolution Over Years

DINOv3 identifies feature-dimension outliers in 7B models, shows that final-layer normalization or alternative scaling suppresses them for downstream use, and documents their implications for intermediate features. Additional results also provide precise references and ImageNet performances underlying the reported evolution of state-of-the-art performance over time.

A.2 Feature Dimension Outliers

  • Feature-dimension outliers emerge in 7B models as a small set of channels with exceptionally large, depth- and training-amplified values across patches and images.Unlike patch outliers, these channel outliers remain consistent across patches and images, persist through model layers, and peak at the output layer.Source: §A.2 · p. 54
  • Applying L2 regularization to suppress feature-dimension outliers reduces performance, whereas removing them at inference causes no significant change, indicating mostly trivial downstream signals.The final layer normalization is trained to substantially scale down these dimensions.Source: §A.2 · p. 54
  • For downstream use, apply the final layer norm to final-layer features; batch normalization can also suppress outliers because their elevated values are consistent across patches and images.Batch normalization is specifically used for semantic segmentation and depth estimation experiments based on intermediate-layer features.Source: §A.2 · pp. 54–55
  • Earlier-layer features require caution because feature-dimension outliers can make them ill-conditioned, and applying the final layer norm may reduce performance on some tasks.Standard scaling methods such as batch normalization or principal component analysis can be effective instead.Source: §A.2 · p. 54

B.1 Evolution Over Years

  • Table 21 standardizes the evolution comparison by using first arXiv appearance for dates and ImageNet top-1 accuracy from each algorithm’s largest model.For weakly supervised and self-supervised models, the reported performance is based on linear probing.Source: Table 21 · p. 55
  • The paper’s additional results report precise references and performances supporting the rough year-by-year evolution of state-of-the-art performance shown in Fig. 1.Table 21 lists publication years, ImageNet top-1 accuracy for each algorithm’s largest model, linear-probing results for weakly and self-supervised models, and each number’s reference.Source: §B.1 · p. 55, Table 21 · p. 55

B.2 Per-Layer Analysis

The DINOv3 7B features are evaluated across classification, segmentation, depth, tracking, and 3D correspondence tasks at every backbone layer. Performance generally rises toward later layers, but geometry-heavy tasks peak around layer 32, while the final layer remains a strong default because intermediate layers offer only slight gains.

B.2 Per-Layer Analysis

  • Across five representative vision tasks, DINOv3 7B features are assessed layer by layer using linear probes for classification, segmentation, and depth, and non-parametric methods for tracking and 3D correspondence.The tasks include ImageNet and ObjectNet classification, ADE20k segmentation, NYU depth estimation, DAVIS tracking, and NAVI 3D correspondence estimation.Source: §B.2 · p. 55
  • Performance increases smoothly across layers for classification and dense prediction, whereas depth estimation, tracking, and 3D correspondence peak around layer 32.The earlier-layer peak indicates that geometry-dependent tasks can benefit from features before the final backbone layer.Source: §B.2 · p. 55
  • Intermediate layers provide only slight improvements over the final layer, making the last layer a strong default choice.This conclusion applies despite the earlier-layer advantages observed for geometry-heavy tasks.Source: §B.2 · p. 55

B.3 Additional Results to Main Results Section

This section complements the main results with additional evaluations spanning fine-grained classification, instance recognition, intermediate-layer features, and semantic segmentation across several datasets and metrics.

Additional evaluations

  • Additional results cover fine-grained classification, instance recognition, and semantic segmentation, complementing the main results in Sec. 6.The section provides per-dataset, full-metric, and cross-dataset evaluations in Tables 22–24.Source: §B.3 · pp. 55–56

Instance recognition

  • Table 23 provides full instance-recognition evaluation results with additional metrics.These results complement the instance-recognition evaluation described in Sec. 6.2.2.Source: §B.3 · p. 55

Intermediate-layer features

  • Figure 21 evaluates DINOv3 7B intermediate-layer features on five benchmarks using linear and non-parametric approaches.Evaluations (a–c) use a linear layer, whereas (d, e) use a non-parametric approach.Source: Figure 21 · p. 56

Fine-grained classification

  • Table 22 reports per-dataset fine-grained classification results on small datasets using linear probing (Fine-S).The evaluation follows Oquab et al. (2024).Source: §B.3 · p. 55, Table 22 · p. 56

Semantic segmentation

  • Table 24 reports complementary semantic-segmentation results on COCO-Stuff, PASCAL VOC 2012, and Cityscapes.These results concern the state-of-the-art semantic segmentation model from Sec. 6.3.2.Source: §B.3 · p. 56

B.4 Classification on OCR-Heavy Datasets

DINOv3 is evaluated on OCR-heavy street-sign, logo, and product classification, where it substantially improves over DINOv2 but still trails weakly supervised models. The remaining gap is attributed partly to the absence of pair image-text training data and left for future work.

B.4 Classification on OCR-Heavy Datasets

  • The evaluation covers street-sign, logo, and product classification, comparing DINOv3 with DINOv2 g and PE-core G under matched sequence-length settings.DINOv3 uses 512-resolution images, while other models are adjusted by patch size; results are reported in Table 25.Source: §B.4 · p. 56
  • DINOv3 drastically outperforms DINOv2 on OCR-heavy classification, but remains substantially behind weakly supervised models.The authors attribute the gap partly to DINOv3’s lack of pair image-text data, which makes learning glyph associations harder, and defer closing it to future work.Source: §B.4 · p. 57

B.5 Fairness Analysis

DINOv3 7B shows relatively consistent geographical fairness across income categories and regions, but substantial disparities remain for low-income data and between Europe and Africa.

B.5 Fairness Analysis

  • DINOv3 7B’s performance drops 23% in the low-income bucket versus the highest-income bucket, while medium- and high-income buckets perform comparably.The evaluation follows Goyal et al. (2022b) and compares DINOv3 with DINOv2 and SEERv2.Source: §B.5 · p. 57
  • The Europe–Africa relative performance difference exceeds 14% for DINOv3, improving on DINOv2’s difference of more than 17%.DINOv3 nevertheless achieves relatively good scores across regions overall.Source: §B.5 · p. 57

C Implementation Details

DINOv3 training uses multi-crop self-distillation with masked patch modeling and specialized losses, optimized for 1M iterations in a sharded mixed-precision setup. Gram anchoring and high-resolution adaptation use empirically selected update and crop-resolution schedules.

Core training setup

  • Multi-crop training uses two 256×256 global crops and eight 112×112 local crops, totaling 3.7M tokens; the teacher processes only global crops.LDINO operates on class tokens, LiBOT aligns masked student patches with visible teacher tokens, and LDKoleo uses small class-token batches.Source: C Implementation Details · pp. 57–58

Optimization

  • Training runs for 1M iterations with fully sharded data parallelism, bfloat16, 8-bit matrix multiplications, a 0.0004 learning rate, and 100k warmup iterations.Additional settings include 0.04 weight decay, 0.98 per-layer learning-rate decay, 0.4 stochastic depth, and a 0.999 teacher EMA factor.Source: C Implementation Details · p. 58

Gram anchoring and high-resolution adaptation

  • Gram anchoring uses loss weight wGram = 2 and updates its teacher every 10k steps for at most three updates, while high-resolution adaptation samples five crop-resolution configurations with probabilities from 0.05 to 0.3.The configurations are (512, 112, 768), (768, 112, 1152), (768, 168, 1152), (768, 224, 1152), and (768, 336, 1152), selected empirically.Source: C Implementation Details · p. 58

D Experimental Details … D.13 Geospatial

This section details the datasets, metrics, implementation choices, and evaluation protocols used to assess DINOv3 across a broad range of vision tasks, including dense prediction, correspondence, discovery, video, image, recognition, detection, depth, visual geometry, and geospatial evaluation.

D.1 Semantic Segmentation: Linear Probing

  • Linear probing evaluates semantic segmentation on ADE20k, VOC12, and Cityscapes with mIoU, and depth estimation on NYUv2 and KITTI with RMSE.Both protocols train a linear layer over normalized patch features from a frozen backbone and sweep AdamW learning rates and weight decay.Source: §D.1 · p. 58, §D.2 · p. 59

D.3 3D Keypoint Matching

  • Correspondence evaluation uses NAVI and SPair, reporting recall under specified distances after patch matching at task-specific image resolutions.NAVI uses 1cm, 2cm, and 5cm thresholds and averages recall across them; final-layer features are evaluated with and without the final layer norm, retaining the maximum.Source: §D.3 · p. 59

D.4 Unsupervised Object Discovery

  • Unsupervised object discovery follows TokenCut on VOC07, VOC12, and COCO20K, generating one box per image and measuring Correct Localization at IoU greater than 0.5.Patch similarities form a fully connected graph, Normalized Cut identifies salient patches, and the similarity threshold is swept from 0 to 0.4.Source: §D.4 · p. 59

D.5 Video Segmentation Tracking

  • Video segmentation tracking uses DAVIS 2017, YouTube-VOS, and MOSE with J &F-mean, propagating first-frame instance labels through cosine-similarity patch matching from a frozen DINOv3 backbone.Hyperparameters are selected on DAVIS training data and the best combination is applied to all dataset test splits.Source: §D.5 · p. 60

D.6 Video Classification

  • Video classification samples 16 frames, converts their patch features into a 4096-token sequence, and trains a shallow attention-based classifier with spatial and temporal positional information.The classifier uses four self-attention blocks, a cross-attention aggregation query, supervised training for 20 epochs, and dataset-specific validation-based hyperparameter selection.Source: §D.6 · pp. 60–61

D.7 Image Classification with Linear Probing

  • Image classification linear probing reports top-1 accuracy across ImageNet transfer, robustness, stylized, fine-grained, and larger datasets, while instance recognition uses cosine similarity and dataset-specific retrieval metrics.Instance recognition covers Oxford, Paris, Met, and AmsterTime, reporting mAP, GAP, and accuracy under their respective evaluation protocols.Source: §D.7 · pp. 62–63, §D.8 · p. 63

D.9 Object Detection

  • DINOv3 is evaluated for object detection on COCO and challenging COCO-O using mAP at IoU thresholds [0.5 : 0.05 : 0.95], with Objects365 also used for training.The detector keeps the DINOv3 backbone frozen, extracts layers [10, 20, 30, 40], and uses a separate six-block encoder and decoder with one-to-one and one-to-many queries.Source: §D.9 · pp. 63–64
  • 66.1 COCO mAP is achieved with multi-resolution test-time augmentation, compared with 65.4 at 1536 pixels and 65.6 at 2048 pixels.The augmentation processes each image at resolutions from 1536 through 2880 and merges predictions with SoftNMS.Source: §D.9 · p. 64

D.11 Monocular Depth Estimation

  • Monocular depth uses synthetic DAv2 training data, evaluates on five datasets, and reports zero-shot scale-invariant absolute relative error and δ1.Training images come from IRS, TartanAir, BlendedMVS, Hypersim, and VKITTI2; evaluation uses NYUv2, KITTI, ETH3D, ScanNet, and DIODE.Source: §D.11 · p. 66

D.12 Visual Geometry Grounded Transformer with DINOv3

  • The DINOv3 VGGT implementation uses 592-pixel inputs, a 0.0001 learning rate, and concatenated intermediate ViT-L layers, improving over original VGGT across tested benchmarks even when untuned.The four-layer input benefits DINOv3 but not DINOv2; the untuned comparison uses 512-pixel inputs, the original learning rate, and the final layer.Source: §D.12 · p. 66

D.13 Geospatial

  • Geospatial evaluations freeze the backbone and train lightweight task-specific heads, using linear classification for GEO-Bench and DPT decoders for segmentation with validation-selected learning rates.GEO-Bench training runs for 2400 iterations with batch size 32, while segmentation learning rates are searched over four values from 3e−5 to 1e−3.Source: §D.13 · p. 66
  • LoveDA and iSAID use 80k-iteration UPerNet training, while canopy-height and detection evaluations use DPT and Faster R-CNN heads, respectively.UPerNet uses batch size 8 and 1500 warm-up iterations; Faster R-CNN is trained for 12 epochs.Source: §D.13 · p. 66
  • Satlidar contains one million 512×512 Maxar images paired with dense lidar measurements from multiple locations and extracted from larger tiles.The numbers of tiles for each sub-dataset are specified in Table 28.Source: §D.13 · p. 66, Table 28 · p. 67

Related papers

Loading 2508.10104v1…