CanViT-B fine-tuned on ImageNet-1k, 8 × 8 canvas (JAX / Flax NNX)

canvit/canvitb16-add-vpe-pretrain-g128px-s512px-in21k-dv3b16-2026-02-02 fine-tuned end to end for ImageNet-1k classification by linear probing then fine-tuning (LP-FT) with its canvas at 8 × 8, in the checkpoint format of canvit-nnx.

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

Install

uv add canvit-nnx

Usage

import jax.numpy as jnp
from PIL import Image

from canvit_nnx import CanViTForImageClassification, Viewpoint, sample_at_viewpoint
from canvit_nnx.preprocess import preprocess

scene = jnp.asarray(preprocess(512)(Image.open("scene.jpg").convert("RGB")))[None]
model = CanViTForImageClassification.from_pretrained("canvit/canvitb16-add-vpe-finetune-g128px-s512px-c8-in1k-2026-08-01-nnx")
model.eval()
state = model.init_state(batch_size=1, canvas_grid_size=8)
viewpoint = Viewpoint.full_scene(batch_size=1)
glimpse = sample_at_viewpoint(spatial=scene, viewpoint=viewpoint, glimpse_size_px=128)
logits, state = model(glimpse=glimpse, state=state, viewpoint=viewpoint)

Results

Top-1 accuracy (%) on the ImageNet-1k validation set (C2F viewpoints, 512 px scenes, 8 × 8 canvas): the mean over 11 runs, with a 95% bootstrap interval where runs differ.

Details

License

MIT. CanViT is pretrained by distillation from DINOv3, 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}
}