ImageNet-1k linear probe on DINOv3 ViT-S/16 at 512 px

A linear ImageNet-1k classifier on the CLS token of frozen DINOv3 ViT-S/16 at 512 × 512 px, released with CanViT.

CanViT, the Canvas Vision Transformer, is an active-vision foundation model: it sees a scene through a sequence of glimpses and remembers it on a scene-wide canvas.

Paper (NeurIPS 2026) · Code · Project page · All checkpoints

Usage

uv add "dinov3-in1k-probes @ git+https://github.com/m2b3/dinov3-in1k-probes.git" transformers pillow

import torch
from PIL import Image
from transformers import AutoModel
from dinov3_in1k_probes import DINOv3LinearClassificationHead
from dinov3_in1k_probes.data import make_val_transform

dinov3 = AutoModel.from_pretrained("facebook/dinov3-vits16-pretrain-lvd1689m").eval()
probe = DINOv3LinearClassificationHead.from_pretrained("canvit/dinov3-vits16-lvd1689m-in1k-512x512-linear-clf-probe").eval()
image = make_val_transform(512)(Image.open("scene.jpg").convert("RGB")).unsqueeze(0)  # [1, 3, 512, 512]
with torch.inference_mode():
    logits = probe(dinov3(pixel_values=image).last_hidden_state[:, 0])  # [1, 1000], from the CLS token

Results

Details

License

MIT. The probe reads features of DINOv3 ViT-S/16, which is released under the DINOv3 License.

Citation

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