import mlx.core as mx
from canvit_mlx import load_from_hf_hub, load_and_preprocess, Viewpoint, extract_glimpse_at_viewpoint

model = load_from_hf_hub("canvit/canvitb16-add-vpe-pretrain-g128px-s512px-in21k-dv3b16-2026-02-02-mlx")
image = load_and_preprocess("path/to/image.jpg", target_size=512)

state = model.init_state(batch_size=1, canvas_grid_size=32)
vp = Viewpoint.full_scene(batch_size=1)
glimpse = extract_glimpse_at_viewpoint(image, vp, glimpse_size_px=128)
out = model(glimpse, state, vp)
mx.eval(out.state.canvas, out.state.recurrent_cls, out.local_patches)

canvas_spatial = model.get_spatial(out.state.canvas)  # [1, G*G, canvas_dim]

Source: CanViT-MLX

Citation

bibtex

@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}
}