DINOv3 ImageNet-1k Linear Classification Probes

Upon its release in August 2025, DINOv3 marked a milestone in self-supervised representation learning for image processing. The 7-billion-parameter flagship model was distilled into a family of smaller ViT and ConvNeXT checkpoints, whose sizes make them much more suitable for most CV tasks.

Sadly, only one ImageNet-1k (IN1k) linear classification probe was released: the one for the 7B model.

Here, we release pretrained linear probes for the smaller DINOv3 ViT models. They can be used directly with Meta's official checkpoints.

As in the original DINOv3 paper, we used 512x512 inputs (1024 input tokens), and trained the probes on the IN1k training set with Inception-crop augmentation.

All of our probes match or exceed the best IN1k-ReAL top-1 validation accuracy reported by the DINOv3 authors, as seen in Table 14 of the original paper.

We note that the raw IN1k top-1 validation accuracy was not reported by the DINOv3 authors, only the ReAL top-1 accuracy. Here, we report both.

Released Probes

• ViT-S/16 @ 512×512

  • Base: facebook/dinov3-vits16-pretrain-lvd1689m
  • Probe: canvit/dinov3-vits16-lvd1689m-in1k-512x512-linear-clf-probe

• ViT-S+/16 @ 512×512

  • Base: facebook/dinov3-vits16plus-pretrain-lvd1689m
  • Probe: canvit/dinov3-vits16plus-lvd1689m-in1k-512x512-linear-clf-probe

• ViT-B/16 @ 512×512

  • Base: facebook/dinov3-vitb16-pretrain-lvd1689m
  • Probe: canvit/dinov3-vitb16-lvd1689m-in1k-512x512-linear-clf-probe

• ViT-L/16 @ 512×512

  • Base: facebook/dinov3-vitl16-pretrain-lvd1689m
  • Probe: canvit/dinov3-vitl16-lvd1689m-in1k-512x512-linear-clf-probe

• ViT-H+/16 @ 512×512

  • Base: facebook/dinov3-vith16plus-pretrain-lvd1689m
  • Probe: canvit/dinov3-vith16plus-lvd1689m-in1k-512x512-linear-clf-probe

See the corresponding HuggingFace Collection.

Performance

Probe	IN-ReAL val top-1 (official / ours)	IN1k val top-1 (ours)
ViT-S/16	87.0% / 87.08%	81.40%
ViT-S+/16	88.0% / 88.08%	82.89%
ViT-B/16	89.3% / 89.54%	85.00%
ViT-L/16	90.2% / 90.42%	87.44%
ViT-H+/16	90.3% / 90.31%	87.65%

Accuracy and full Optuna hyperparameters can be queried using uv run print_metrics.py.

Usage

We recommend using uv.

Quick demo

Run the demo directly (no clone needed):

uv run https://raw.githubusercontent.com/m2b3/dinov3-in1k-probes/main/demo.py

Using from_pretrained

from dinov3_in1k_probes import DINOv3LinearClassificationHead

probe = DINOv3LinearClassificationHead.from_pretrained(
    "canvit/dinov3-vitb16-lvd1689m-in1k-512x512-linear-clf-probe"
)
# DINOv3LinearClassificationHead(in_features=768, out_features=1000, bias=True)

See demo.py for a complete example including DINOv3 backbone loading and preprocessing.

To get an interactive shell with the package:

uvx --with 'git+https://github.com/m2b3/dinov3-in1k-probes.git' ipython

Evaluation

Verify published numbers on your own IN1k val set:

uv run python eval_in1k.py --imagenet-val /path/to/ILSVRC2012/val
uv run python eval_in1k.py --imagenet-val /path/to/val --variant vitb16

Image size and backbone are read from the probe's HuggingFace config — no manual configuration needed.

Training

The training code is included in this repo under dinov3_in1k_probes/training/. Install the training dependencies with uv sync --group training.

Two-phase pipeline

1. Extract CLS tokens from a frozen DINOv3 backbone into .pt files on disk (GPU, DALI — Linux/CUDA only).
2. Train a linear probe via Optuna HP search over the cached features (all in GPU memory, fast).

This decouples the expensive backbone forward pass from the probe optimization, making 100+ Optuna trials practical over the full ImageNet training set.

Environment variables

Variable	Required by	Description
IMAGENET_ROOT	extract	Path to ILSVRC2012/ (must contain train/ and val/ subdirs)
FEATURES_DIR	extract, train	Where to read/write cached .pt feature files
CHECKPOINTS_DIR	train	Where to save probe checkpoints (default: checkpoints)
COMET_API_KEY	train	Comet ML API key for live experiment tracking. If unset, training writes Comet offline logs under $CHECKPOINTS_DIR/comet_offline/.
COMET_WORKSPACE	train	Comet workspace name for live logging (pass via CLI --comet-workspace or env var)

Extracted features are written to $FEATURES_DIR/<model_name>/<image_size>/<split>/ (e.g. $FEATURES_DIR/dinov3_vitb16/512/val/). Each extraction run writes *_cls_tokens.pt, *_labels.pt, and *_filenames.pt. Multiple train extractions (for augmentation epochs) are concatenated automatically by the training script.

Running locally

# Phase 1: extract CLS tokens
uv run --group training python -m dinov3_in1k_probes.training extract --split val
uv run --group training python -m dinov3_in1k_probes.training extract --split train  # run N times for N augmentation epochs

# Phase 2: Optuna HP search
uv run --group training python -m dinov3_in1k_probes.training train

Running on SLURM

sbatch --account=my_project_name slurm/extract.sbatch                        # val
sbatch --account=my_project_name --array=1-2 slurm/extract.sbatch --split train  # 2 train epochs
sbatch --account=my_project_name slurm/train.sbatch

Use --help on either subcommand for all options (model repo, image size, batch size, etc.).

Development

Push probes to HuggingFace Hub:

uv run push_to_hub.py --checkpoint dinov3-vitb16-lvd1689m-in1k-512x512-linear-clf-probe.pt --owner YOUR_HF_USERNAME

Citation

@article{berreby2026canvit,
  title={CanViT: Toward Active-Vision Foundation Models},
  author={Berreby, Yoha{\"i}-Eliel and Du, Sabrina and Durand, Audrey and Krishna, B. Suresh},
  year={2026},
  eprint={2603.22570},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2603.22570}
}

License

MIT. See LICENSE for details.