DINOv3
1. Paper overview
In one sentence: DINOv3 preserves spatial patch relationships during large-scale self-supervised training, producing versatile frozen features at the cost of substantial pretraining and task-dependent downstream adaptation. e05e06e07e08e10e12e14
| At a glance | What to know |
|---|---|
| Research problem | Source description A global image descriptor can improve while spatial patch representations become less useful. During extended ViT training, classification rises while VOC segmentation declines after roughly 200k iterations; patch-similarity maps become noisy despite register tokens controlling patch norms. DINOv3 asks how to scale data, model capacity and training duration without sacrificing local consistency. This matters whenever downstream systems need both object identity and spatial correspondence from the same encoder. e05e03 |
| Core mechanism | |
| A key reported result | COCO object detection with a frozen backbone: 65.6 simple; 66.1 TTA. mAP at IoU thresholds 0.50:0.05:0.95. Objects365-pretrained Plain-DETR adapter, then COCO training; COCO-VAL2017 evaluation. Simple inference uses a 2048-pixel short side; TTA combines scales and flips. PEspatial + DETA: 65.3 / 66.0; EVA-02 + Co-DETR: 65.4 / 65.9. The strongest listed COCO result trains a 100M-parameter detector over a frozen 7B encoder. Different detectors and training recipes prevent attributing the entire margin to the backbone. e14e15 |
| Reading caution | Reader analysis Self-supervision does not eliminate dataset bias: OCR-heavy tasks still expose weaknesses against PEcore, and the fairness evaluation retains substantial income and regional gaps. These are measured limits to the paper's broad versatility claim. e22 |
Core contributions
Figure 10. Gram refinement makes the illustrated patch similarities more coherent and localized. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each column vertically. The top row is the original image, the middle row is the model without high-resolution Gram refinement, and the bottom row is the refined model. The small red marker locates the reference patch whose cosine similarities are displayed across the image. Compare the scattered bright responses in the middle row with the more coherent regions in the bottom row, including the flower, geese and sheep examples. These are feature-similarity maps, not labeled segmentation predictions. Equation (2) constrains all pairwise normalized patch similarities during training; the displayed maps expose selected relationships from that much larger matrix. e06e07e08e12
What it supports. The examples make the numerical ablation easier to interpret: reduced background noise and more localized similarity can improve the features consumed by a dense predictor. The source caption states that these inputs are 1024 by 1024 pixels. The accompanying quantitative evidence comes from the Gram ablation and dense probing, rather than from interpreting a color as a predicted class.
Where the evidence stops. The original visual has no numerical colorbar or ground-truth masks. It supports a qualitative comparison for selected images, not a calibrated similarity threshold, average boundary accuracy, or a guarantee that every object is correctly separated.
2. Motivation
2.1 The problem and the proposed response
A global image descriptor can improve while spatial patch representations become less useful. During extended ViT training, classification rises while VOC segmentation declines after roughly 200k iterations; patch-similarity maps become noisy despite register tokens controlling patch norms. DINOv3 asks how to scale data, model capacity and training duration without sacrificing local consistency. This matters whenever downstream systems need both object identity and spatial correspondence from the same encoder. e05e03
2.2 What this reading follows
An image encoder can become better at recognizing a scene while becoming worse at locating its parts. DINOv3 diagnoses this divergence during long self-supervised training and introduces Gram anchoring to repair it. An earlier teacher supplies the relationships between patches, allowing the evolving model to retain useful spatial structure. Read the figures as a chain of evidence: training curves expose the failure and recovery, the teacher ablation tests the target construction, and dense probes measure the resulting features. The distillation diagram explains how this expensive teacher serves smaller models. Finally, the detection table shows what a frozen encoder can support when combined with a trained detector. e05e06e07e08e10e12e14
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Foundational work |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
This table preserves the labels recorded at reading time. The current major category is Components of WAMs. View the current classification.
3.1 Evidence-based assessment
Supports the recorded classification
The recorded foundational-work / visual-encoder classification is supported. DINOv3 learns reusable image descriptors with self-distillation and masked-patch reconstruction. Its student, EMA teacher and Gram teacher are training roles, not a world-model/policy decomposition. There is no action output, inverse-dynamics policy or joint future/action predictor in the proposed encoder. The architecture, prediction-paradigm and quadrant fields are therefore appropriately not applicable to the WAM taxonomy. e03e04e06e17
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Why an older teacher can improve a later model
The early checkpoint is valuable for a specific property, not because it is the best model overall. During initial training, the global DINO objective continues to improve image recognition, while the dense features lose useful locality. The earlier model retains cleaner patch relationships. Gram anchoring transfers those relationships through the matrix of pairwise cosine similarities, while the other objectives keep training the representation. Reader interpretation: this is a constraint on how patches relate to one another, leaving more freedom than requiring every student vector to equal an old feature vector. The Gram teacher is also distinct from the main EMA teacher: it starts from an early checkpoint and is refreshed periodically from that EMA network. The paper's curves support late repair, but do not prove that every source of degradation has been isolated. e05e06e08e04
Figure 12. One teacher computation supplies several separately trained student models. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The horizontal arrow indexes GPUs; iteration time runs downward. Each device loads a portion of batch B, and the green band performs fixed-teacher inference across all N_T GPUs. The all-gather line distributes samples and teacher outputs before the blue student-training blocks. Student S_i uses its own N_S_i-device group, so each group can process the same teacher-labeled batch with a different model. The red synchronization lines belong to training coordination, not an image-processing network. Wait blocks reveal unequal group runtimes; the paper proposes allocating GPU counts to balance them. Section 5.2 confirms that this fixed teacher differs from the moving EMA teacher used during initial pretraining. e10e04
What it supports. Sharing the expensive teacher pass makes producing several student sizes more economical in the paper's compute accounting. Each extra student adds its own training work without requiring another full teacher inference over the batch. The resulting students can be used independently at deployment; this diagram is a training arrangement, not a multi-model inference ensemble.
Where the evidence stops. This is a procedural diagram, not a measured throughput benchmark. Communication, synchronization and group imbalance can affect realized speed. The students use fixed-teacher distillation without Gram anchoring, so the diagram does not depict the Gram-refinement architecture.
5.2 How a larger teacher image becomes a same-size target
Doubling the Gram teacher's input resolution yields a denser feature map. Directly comparing that map with the student's lower-resolution features would leave incompatible spatial grids. DINOv3 first downsamples the teacher features with bicubic interpolation, then uses their patch relationships as the target in the Gram objective. Section 4.3 motivates this combination as obtaining fine detail and smoothing inconsistent patch responses. Figure 9(b) lets the reader test the idea without confusing it with simply using a different teacher checkpoint: the two 200k rows differ in teacher resolution and report improved dense scores at doubled resolution. This refinement also precedes the separate high-resolution adaptation phase, where the flagship sees mixed crop sizes while retaining Gram anchoring. Smaller fixed-teacher students are trained without that additional Gram objective. e06e07e09e10
5.3 From an encoder to a useful downstream system
The evaluation has two levels. Dense linear probing measures how much task-relevant information a lightweight trained readout can extract from frozen final-layer features. Detection then asks whether those representations can support a competitive complete system. The detector uses intermediate-layer features, a combination of global and windowed image views, and an independently trained Plain-DETR adapter. Reader interpretation: the strong system result makes the encoder practically interesting, while the simpler probe gives cleaner evidence about representation quality. Neither experiment demonstrates a control policy or action-conditioned environment predictor. The distillation pipeline addresses a different deployment concern by transferring the large teacher into smaller encoders. Its shared computation occurs during training; using a distilled model later does not require retaining the teacher or the other students. e12e14e10e18e03
5.4 Training and inference
During training
Initial training uses AdamW, 4096 images over 256 GPUs, two 256-pixel global crops and eight 112-pixel local crops. Appendix C specifies learning rate 0.0004, 100k-step warmup, weight decay 0.04 and teacher EMA factor 0.999. The student uses both crop types; the EMA teacher sees only global crops. Gram loss weight is 2. e04
The separate satellite model uses SAT-493M Maxar RGB imagery, adapted normalization and a shorter schedule. This demonstrates reuse of the training recipe with domain-specific data, rather than a single unchanged web checkpoint covering every experiment. e19
During inference
The encoder produces features in a forward pass. Classification uses CLS features; dense probes use normalized patch outputs, with trained batch normalization and a linear predictor. Gram anchoring is a training objective and requires no extra teacher at deployment. Detection and segmentation add trained decoders; VGGT explicitly fine-tunes its image backbone. e06e12e13e14e16e17
Final-layer normalization suppresses feature-dimension outliers. Applying that same learned normalization to intermediate layers can hurt some tasks; Appendix A.2 recommends appropriate feature scaling for those layers. e21
5.5 Implementation flow
- Curate and sample
From approximately 17 billion Instagram images, DINOv2 embeddings and hierarchical clustering produce the 1,689-million-image background subset. Retrieval-curated images and public vision datasets complement it. ImageNet1k-only homogeneous batches occupy 10% of training; other iterations mix the remaining components. e02
- Learn global and local representations
The 6.7B-parameter backbone has 40 blocks, 4096-dimensional embeddings, 16-pixel patches, four registers and RoPE with coordinate-box jittering. Separate heads implement DINO class-token matching and iBOT masked-patch latent reconstruction. Sinkhorn–Knopp target normalization and distributed KoLeo regularization help stabilize the representation. e03e04
- Anchor spatial relationships
After 1M initial iterations, introduce a separate Gram teacher initialized from an early checkpoint. Match its within-image patch-similarity matrix on global crops while continuing the other objectives. Refresh the Gram teacher from the main EMA teacher every 10k steps, for at most three updates. The higher-resolution variant forwards twice-resolution images through this teacher and bicubically downsamples its features to the student grid. e06e07e04
- Adapt resolution and distribute capacity
The flagship undergoes 10k additional mixed-resolution iterations with Gram anchoring. Smaller students instead learn from a fixed 7B teacher, without Gram anchoring, for 1M iterations followed by a 250k cosine cooldown and high-resolution adaptation. Multiple GPU groups share teacher outputs, then train their respective students separately. e09e10
- Optional text alignment
Train a text encoder contrastively against captions while freezing the vision backbone. Two added visual transformer layers provide flexibility; concatenate mean-pooled patch embeddings with the CLS token for matching to text. This is an additional paired-data stage, separate from image-only backbone pretraining. e11
6. Experiments & results
DINOv3 scales self-supervised image encoding while repairing the dense features that deteriorate during long training. Its Gram anchoring loss transfers patch-similarity structure from an earlier teacher, followed by resolution adaptation and distillation. The resulting frozen features support strong dense probes and trained downstream systems. The contribution is reusable visual representation learning; action selection and predictive environment dynamics are outside the demonstrated system.
6.1 Read the original evidence
Table 3. Lightweight trained probes expose the quality of the frozen patch features. Original paper, p. 19 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across a row to distinguish segmentation from depth. ADE20k, Cityscapes and VOC report mIoU, where larger is better; NYUv2 and KITTI report RMSE, where smaller is better. Model families are grouped by pretraining approach, and the ViT column preserves the different sizes and patch widths. The caption specifies 1,024 patch tokens for segmentation: 512-pixel inputs for patch size 16 and 448 for patch size 14. Appendices D.1–D.2 add a key detail: these probes include final-layer normalization and trained batch normalization before the linear predictor. The DINOv3 row therefore evaluates a frozen encoder with supervised readouts, not an entirely annotation-free downstream system. e12
What it supports. DINOv3 reports 55.9 ADE20k mIoU versus 49.5 for DINOv2 and 53.0 for AM-RADIOv2.5. Its depth RMSE is 0.309 on NYUv2 and 2.346 on KITTI. These results support useful dense representations across several tasks, beyond the appearance of clean similarity maps.
Where the evidence stops. Model sizes and pretraining data differ, and no repeated-run uncertainty is supplied. Cityscapes values 81.1 and 78.4 imply a 2.7-point gain over AM-RADIO; the accompanying source prose says 2.5. The edition preserves the table values and flags the discrepancy.
Table 10. A trained Plain-DETR adapter over frozen DINOv3 reaches 66.1 COCO mAP with TTA. Original paper, p. 26 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with FT: fire icons denote fine-tuned encoders and the snowflake marks DINOv3's frozen backbone. Read Encoder, Decoder and Trainable separately; freezing a 7B encoder leaves 100M trainable detector parameters, not a 100M total system. Compare COCO Simple with Simple and TTA with TTA. The latter combines image scales and flips, with SoftNMS merging described in Appendix D.9. COCO-O reports a separate shifted-distribution evaluation, with ER denoting effective robustness. Dashes are missing results rather than zeros; the caption explains that InternImage-G could not be evaluated on COCO-O because its detection model was unavailable to the authors. e14e15
What it supports. The DINOv3 system reports 65.6 simple and 66.1 TTA COCO mAP, compared with PEspatial's 65.3 and 66.0. This establishes that a frozen encoder can support a strong detector after supervised adapter training. It does not make the downstream pipeline training-free or remove the cost of forwarding a large backbone.
Where the evidence stops. Different detectors and training recipes limit causal attribution. An exact reproduction also faces a source conflict: Section 6.3.1 specifies one epoch for the second Objects365 stage, whereas Appendix D.9 specifies four. The supplied v1 does not resolve it.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Dense linear probing Frozen final-layer features with trained normalization and linear heads on ADE20k, Cityscapes, VOC, NYUv2 and KITTI. Segmentation inputs have 1,024 patch tokens; D.1–D.2 do not explicitly name evaluation splits. | 55.9 / 81.1 / 86.6 mIoU; 0.309 / 2.346 RMSE. Segmentation mIoU ↑; depth RMSE ↓ | DINOv2: 49.5 / 75.6 / 83.1; 0.372 / 2.624. AM-RADIOv2.5: 53.0 / 78.4 / 85.4; 0.340 / 2.918. Table 3 supports improved dense features under lightweight supervised probes, with no reported repeated-run uncertainty. Its Cityscapes values imply a 2.7-point gain over AM-RADIO; nearby prose says 2.5. e12 |
| Gram teacher and resolution ablation Figure 9(b), Section 4 refinement comparison; the panel does not state evaluation split names or repeated-run uncertainty. | 200k teacher at twice resolution: 55.7 / 0.281 / 88.0. ADE20k mIoU ↑; NYUv2 RMSE ↓; ImageNet linear accuracy ↑ | Baseline: 50.3 / 0.307 / 88.2; same teacher at native resolution: 53.6 / 0.285 / 88.0. Higher teacher resolution adds 2.1 mIoU beyond native-resolution anchoring. The dense gain accompanies a small classification reduction; the baseline-to-refinement comparison does not isolate every changed loss weight. e07e06 |
| ImageNet classification and ObjectNet transfer Linear CLS probe trained on ImageNet training data; ImageNet validation selects hyperparameters; transfer to ObjectNet; 1,024 patch tokens. | ImageNet validation 88.4%; ObjectNet 79.0%. Top-1 accuracy | DINOv2: 87.3% / 66.4%; PEcore: 89.3% / 80.2%. Substantial transfer gains over DINOv2 coexist with a remaining gap to PEcore. Separately tabulated supervised models use a different protocol. e13 |
| COCO object detection with a frozen backbone Objects365-pretrained Plain-DETR adapter, then COCO training; COCO-VAL2017 evaluation. Simple inference uses a 2048-pixel short side; TTA combines scales and flips. | 65.6 simple; 66.1 TTA. mAP at IoU thresholds 0.50:0.05:0.95 | PEspatial + DETA: 65.3 / 66.0; EVA-02 + Co-DETR: 65.4 / 65.9. The strongest listed COCO result trains a 100M-parameter detector over a frozen 7B encoder. Different detectors and training recipes prevent attributing the entire margin to the backbone. e14e15 |
| ADE20k semantic segmentation with a trained decoder Frozen 7B encoder with modified ViT-Adapter and Mask2Former; COCO-Stuff and Hypersim pretraining, then ADE20k training; validation evaluation at base resolution 896. | 62.6 single-scale; 63.0 TTA. mIoU | ONE-PEACE: 62.0 / 63.0 with backbone fine-tuning. Matches the listed TTA leader, but requires 927M trainable decoder/adapter parameters. This is a different protocol from the 55.9 linear-probe result. e16 |
| Open-Canopy canopy-height estimation DPT heads trained and evaluated on Open-Canopy with frozen backbones; the source adapts the patch embedding for RGB plus infrared inputs. | DINOv3 satellite ViT-7B: 2.02. MAE ↓ | Tolan et al. backbone with retrained head: 2.42; DINOv3 web ViT-7B: 2.17. Supports domain-specific pretraining for this metric task. Table 17 does not print an MAE unit; no unit or uncertainty is added here. e19 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 8. Dense prediction recovers when Gram refinement begins after prolonged pretraining. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the horizontal axis as training iterations, with the shaded region marking the late refinement phase. The blue trajectory is the original training; the green and orange legends identify ordinary Gram refinement and its higher-resolution teacher variant. VOC and ADE20k use mIoU, while ObjectNet uses classification accuracy, so compare trends within each panel. The segmentation curves deteriorate before refinement even as the ObjectNet curve rises. In the shaded region, the segmentation scores recover, with the higher-resolution variant above the ordinary refinement. Equation (3) on page 12 defines the refinement objective; Section 4.3 explains how its higher-resolution target is constructed. e05e06e07e08
What it supports. The curves support the paper's central diagnosis: global improvement does not guarantee useful spatial features. They also show that late refinement can repair dense performance after degradation has already occurred. ObjectNet changes much less abruptly, consistent with the intended separation between retaining local relationships and continuing to learn global representations.
Where the evidence stops. These trajectories are not repeated-run confidence intervals. Refinement changes the objective mixture and continues training; the plot alone cannot attribute every gain specifically to Gram loss or establish equal-compute superiority.
Figure 9(b). An early teacher and doubled input resolution give the strongest dense results in this ablation. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First compare the baseline with the 200k-teacher rows. IN1k Linear measures global classification, ADE mIoU measures segmentation, and NYU RMSE measures depth error; only RMSE improves when it decreases. The two 200k rows hold teacher age fixed while changing its input-resolution multiplier from one to two. Section 4.3 specifies that the doubled-resolution features are downsampled before matching the student grid. Next compare the doubled-resolution rows across teacher ages. The early 100k and 200k teachers reach the same ADE score, whereas the 1M teacher has already experienced degraded patch consistency. The table reports refinement-stage results, distinct from the final backbone's Table 3 scores. e06e07
What it supports. With the 200k teacher, doubled resolution changes ADE20k from 53.6 to 55.7 mIoU and NYU RMSE from 0.285 to 0.281 while IN1k remains 88.0. Relative to the baseline, dense performance improves but IN1k falls from 88.2 to 88.0. The 1M teacher still helps, though less than the early teachers on ADE.
Where the evidence stops. The panel supplies no repeated-run uncertainty or explicit evaluation splits. Native-versus-doubled resolution is a cleaner contrast than baseline-versus-refinement, which also changes the weighted loss mixture. Teacher age does not isolate target quality from every other checkpoint property.
7. Analysis & limitations
7.1 What the evidence leaves open
Self-supervision does not eliminate dataset bias: OCR-heavy tasks still expose weaknesses against PEcore, and the fairness evaluation retains substantial income and regional gaps. These are measured limits to the paper's broad versatility claim. e22
Strong visual features and VGGT integration establish perception capabilities, not action-conditioned dynamics or robot execution. VGGT also changes input resolution, learning rate and layer aggregation, so it is not a pure backbone-only comparison. e03e17
The reported tables largely omit repeated-run uncertainty. Full-system comparisons vary architecture and supervision, while Gram refinement changes an objective mixture; the cleanest resolution contrast is the same 200k teacher at native versus doubled resolution. e07e14e16
7.2 Questions for discussion
- How much of late dense recovery survives when every non-Gram refinement setting is held fixed?
- Can smaller distilled encoders retain the spatial relationships most useful for downstream control, when tested in an actual action-learning system?
8. Reproducibility audit
8.1 Requirements and known gaps
Full pretraining requires the curated data mixture and distributed training infrastructure. Appendix C specifies fully sharded PyTorch training with bfloat16 and 8-bit matrix multiplications; Table 20 reports 61,440 H100-SXM5 GPU-hours for the 1M-step pretraining entry. This is not the compute budget for the entire research project. e02e04e20
The text leaves all non-ImageNet mixture ratios and some configuration settings unspecified. D.1–D.2 provide probe optimizer sweeps but omit explicit evaluation split names and training duration. Code references do not fill these gaps in the supplied evidence. e23
Resolve the detection schedule discrepancy before an exact reproduction: Section 6.3.1 specifies one epoch at Objects365 resolution 2048; Appendix D.9 specifies four. Both then describe 12 COCO epochs. Neither version can be silently preferred. e15
A smaller proposed mechanism check should branch one degraded checkpoint into matched continuations with and without Gram loss, using identical data and probe tuning. Test dense recovery together with global accuracy, then compare native- and doubled-resolution teachers. This would test the mechanism without claiming to reproduce the full flagship. e05e06e07e12
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Separate Gram anchoring from the rest of refinement
Reader-proposed check, not run: save early and degraded checkpoints from one training trajectory, then branch the degraded model into matched continuations. Keep data order, optimizer, duration and all non-Gram loss weights fixed. Compare no Gram loss, Gram with the early teacher at native resolution, Gram with the same teacher at doubled resolution, and Gram with the degraded teacher. Use the same held-out images and identical dense-probe tuning for every branch; measure ADE20k mIoU, NYUv2 RMSE, global ImageNet accuracy and patch-similarity structure across multiple seeds. The mechanism is challenged if reduced Gram error fails to recover dense performance, or if the same recovery occurs without Gram under matched settings. Treat any smaller-model experiment as a mechanism test, not a reproduction of the 7B scores. e05e06e07e12e23
Check 2: Test whether shared-teacher distillation preserves independent training
Reader-proposed check, not run: use one fixed teacher, identical augmented batches and the same student initialization to compare independent single-student distillation with the shared-teacher, multi-student procedure. Match each student's effective batch size, optimizer updates and training duration, allowing only numerical differences expected from distributed arithmetic. Verify that gathered targets remain paired with the correct images. Record teacher-inference work, communication time, synchronization waits, total GPU-hours and final probe quality; compare both fixed per-student resources and fixed total resources. The proposed efficiency mechanism is undermined if sharing changes learning targets or if communication and imbalance erase the saved teacher work. Report hardware and allocation explicitly, since Figure 12 provides a procedure rather than measured speedup. e10e18
8.3 Reading coverage
Visual audit: All 20 text chunks covering the complete 67-page v1 were read, including references and Appendices A–D. The declared PDF pages were rendered and visually inspected, including the title/author block, architecture table, losses, training details, supporting result tables and relevant appendix protocols. Six faithful crops were then individually viewed: Figures 8, 9(b), 10 and 12, and Tables 3 and 10. Figure 12's GPU/time arrow directions, group boundaries and synchronization markers were checked against Section 5.2; refinement labels were checked against Equations (2)–(3) and Section 4.3. No claim-relevant graphic/formulation conflict was found in those diagrams. The Cityscapes arithmetic and detection-stage duration discrepancies are explicitly retained. Other source visuals were not all inspected; no later venue edition, separate supplement, code or dataset was examined.
PDF pages inspected for this edition: 1, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 19, 23, 24, 25, 26, 27, 28, 30, 33, 34, 36, 54, 57, 58, 59, 62, 63, 64, 65, 66. Appendix coverage: reviewed.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- Title and abstract; Sections 1–2: introduction and related work
- Sections 3.1–3.2: data preparation, architecture and self-supervised pretraining
- Sections 4.1–4.3: dense-feature degradation, Gram anchoring and higher-resolution targets
- Sections 5.1–5.3: resolution adaptation, distillation and text alignment
- Sections 6.1.1–6.1.6: dense probing, correspondence, object discovery, tracking and video classification
- Sections 6.2.1–6.2.2: global classification and retrieval
- Sections 6.3.1–6.3.4: detection, segmentation, depth and VGGT systems
- Sections 7.1–7.3: distilled ViTs, ConvNeXts and dino.txt
- Sections 8.1–8.3: satellite pretraining and geospatial evaluations
- Sections 9–10: environmental impact and conclusion
- References, PDF pp. 38–52
- Appendices A.1–A.2, B.1–B.5, C and D.1–D.13, including Table 28 on p. 67
Outside the original text pass
- Identity/version note: the inspected title page identifies DINOv3, arXiv:2508.10104v1 [cs.CV], 13 August 2025, with all 26 catalog authors. The catalog records a 2026 TMLR publication; that venue edition and any intervening revisions were not supplied or compared. This report describes the supplied v1 only.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The preceding acquisition limitation was addressed for the retained claims by inspecting original PDF pages and all six final crops. Visual inspection was selective; the complete 67-page text, including references and appendices, was read in all 20 supplied chunks.
- Separate supplemental material availability has not been fully verified.
- Code, linked datasets and configuration files were not inspected; no training, evaluation or reproduction experiments were run.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title, author block, affiliations and arXiv margin stamp
The observed title is DINOv3; 26 named authors and Meta AI Research, WRI and Inria appear. The artifact is arXiv:2508.10104v1, 13 August 2025.
Go to primary source ↓e02PDF pp. 8–9, Section 3.1 and Table 1
Data preparation starts from approximately 17 billion Instagram images; hierarchical DINOv2 clustering yields 1,689 million images, supplemented by retrieval and public datasets. ImageNet1k-only batches constitute 10%; Table 1 uses a 200k-step ablation.
Go to primary source ↓e03PDF p. 9, Table 2; p. 10, Updated Model Architecture
ViT-7B has 6.7B parameters, 40 blocks, patch size 16, four registers, dimension 4096 and RoPE. Coordinate-box jittering scales the normalized position box with a factor in [0.5, 2].
Go to primary source ↓e04PDF pp. 9–10, Section 3.2, Eq. (1), Optimization; pp. 57–58, Appendix C
DINO, iBOT and distributed KoLeo train the encoder; two global/eight local crops and an EMA teacher supply targets. Batch 4096 spans 256 GPUs. Appendix C specifies masking, optimizer settings, sharding, precision and Gram weight 2 with at most three teacher updates.
Go to primary source ↓e05PDF pp. 10–11, Figures 5–6 and Section 4.1
Global classification improves during extended training while dense segmentation declines after approximately 200k steps; patch similarities become noisy and less localized despite stable patch norms with registers.
Go to primary source ↓e06PDF p. 12, Section 4.2, Eqs. (2)–(3) and Figure 7
Gram anchoring matches the pairwise dot products of normalized patch features from the student and an early teacher. It is applied on global crops after 1M iterations; periodic replacement uses the main EMA teacher. The refinement objective reweights DINO and KoLeo and adds Gram loss.
Go to primary source ↓e07PDF pp. 13–14, Section 4.3 and Figure 9(a–b)
Twice-resolution Gram-teacher features are bicubically downsampled. Figure 9(b) reports baseline 88.2/50.3/0.307, 200k native teacher 88.0/53.6/0.285, 200k doubled teacher 88.0/55.7/0.281, 100k doubled teacher 87.9/55.7/0.284, and 1M doubled teacher 88.1/54.9/0.290 in IN1k/ADE/NYU order.
Go to primary source ↓e08PDF p. 13, Figure 8 and adjoining paragraph; p. 14, Figure 10 and caption
Refinement improves VOC and ADE20k shortly after it begins. Figure 10 compares cosine-similarity maps before and after high-resolution Gram refinement on 1024-by-1024 inputs; the refined maps are more localized.
Go to primary source ↓e09PDF p. 15, Section 5.1 and Figure 11; p. 58, Appendix C, crop-resolution tuples
Flagship high-resolution adaptation lasts 10k iterations, samples global sizes 512/768 and multiple local sizes, and retains Gram anchoring. Appendix C specifies the paired crop probabilities and Gram-teacher resolutions.
Go to primary source ↓e10PDF pp. 15–16, Section 5.2 and Figure 12
Fixed-teacher distillation omits Gram loss, trains students for 1M iterations plus 250k cooldown, then adapts resolution. Figure 12 shares teacher inference across all GPUs, all-gathers samples and outputs, and trains students in separate groups before synchronization.
Go to primary source ↓e11PDF p. 17, Section 5.3
Text alignment keeps the visual backbone frozen, adds two visual transformer layers, and trains a text encoder with contrastive caption matching using concatenated CLS and mean-pooled patch embeddings.
Go to primary source ↓e12PDF p. 19, Table 3 and results text; p. 58, Appendix D.1; p. 59, Appendix D.2
Table 3 gives DINOv3 segmentation 55.9/81.1/86.6 and depth RMSE 0.309/2.346, alongside DINOv2 and AM-RADIO values. Segmentation uses 1,024 patch tokens. Probe protocols specify final layer norm, learned batch norm and AdamW sweeps. Cityscapes table subtraction gives 2.7 rather than the prose's 2.5 points.
Go to primary source ↓e13PDF pp. 23–24, Section 6.2.1 and Table 7; pp. 62–63, Appendix D.7
ImageNet-trained CLS probes use validation-based hyperparameter selection and transfer to ObjectNet. Table 7 reports DINOv3 88.4/79.0, DINOv2 87.3/66.4 and PEcore 89.3/80.2 for ImageNet validation/ObjectNet; supervised reference rows have a different protocol.
Go to primary source ↓e14PDF pp. 25–26, Section 6.3.1 and Table 10; pp. 63–64, Appendix D.9
COCO-VAL2017 detection uses a frozen 7B backbone and a 100M trainable Plain-DETR adapter, Objects365 then COCO training, intermediate layers and global/windowed features. DINOv3 scores 65.6 simple, 66.1 TTA and 66.4 COCO-O mAP. TTA uses flips, eight resolutions and SoftNMS; InternImage COCO-O results are absent.
Go to primary source ↓e15PDF p. 26, Section 6.3.1 Implementation; p. 64, Appendix D.9 Training, stages 2–3
The main text says one Objects365 epoch at resolution 2048; Appendix D.9 says four. Both specify a subsequent 12-epoch COCO stage.
Go to primary source ↓e16PDF pp. 26–27, Section 6.3.2 and Table 11; p. 65, Appendix D.10
A modified ViT-Adapter without its injector and Mask2Former consume frozen features. COCO-Stuff/Hypersim/ADE20k training yields 62.6 simple and 63.0 TTA validation mIoU; 927M decoder parameters are trained. ONE-PEACE reports 62.0/63.0 with fine-tuning.
Go to primary source ↓e17PDF p. 28, Section 6.3.4 Implementation; p. 66, Appendix D.12
VGGT uses the distilled ViT-L and fine-tunes the image backbone. Relative to original VGGT, the reported setup changes image size to 592, learning rate to 0.0001 and aggregation to four intermediate layers.
Go to primary source ↓e18PDF p. 30, Figure 16(a) and Table 14
The family contains ConvNeXt and ViT students. Figure 16 lists ViT-L at 300M parameters and 721 GFLOPs at resolution 512, versus 6716M and 14515 GFLOPs for ViT-7B. Table 14 gives ViT-L DINOv3 ADE20k 54.9 versus DINOv2 48.8.
Go to primary source ↓e19PDF pp. 33–34, Sections 8.1–8.2 and Table 17; p. 66, Appendix D.13
SAT-493M contains Maxar RGB imagery; satellite training changes normalization and duration. Open-Canopy uses an added infrared channel and a DPT decoder. Table 17 lists MAE 2.02 for satellite 7B, 2.17 for web 7B and 2.42 for the retrained Tolan baseline, without printing units.
Go to primary source ↓e20PDF p. 36, Table 20 and whole-project footprint paragraph
The 1M-step ViT-7B pretraining entry reports 61,440 H100-SXM5 GPU-hours. Energy/carbon figures are estimates under stated assumptions; the whole-project compute estimate is separately discussed.
Go to primary source ↓e21PDF p. 54, Appendix A.2
Channel-wise feature outliers can persist after register tokens resolve patch-norm outliers. Final layer norm suppresses them, but using its learned parameters on earlier layers can impair performance; batch normalization or PCA can help.
Go to primary source ↓e22PDF p. 57, Appendices B.4–B.5 discussion; p. 58, Tables 25–26
DINOv3 improves OCR-heavy classification over DINOv2 but trails PEcore on most listed datasets. The fairness table shows remaining differences between income groups and geographic regions.
Go to primary source ↓e23PDF pp. 8–9, Section 3.1; p. 58, Appendix C and D.1; p. 59, Appendix D.2
The text fixes the ImageNet1k batch fraction but does not enumerate all other mixture ratios. Appendix C refers remaining hyperparameters to release configuration files; dense-probe descriptions give optimizer sweeps without explicit split names or training duration.
Go to primary source ↓8.5 Primary sources
DINOv3 ↗
PDF · 34,553 extracted words
Source fingerprint
c68dc50b1f73e1641a592f777984fc360759e84b7fe65ac4f6e1377927f20b2e