ADE20K probe on CanViT's 64 × 64 canvas
A linear ADE20K semantic segmentation probe on CanViT's 64 × 64 canvas, trained on frozen features with the paper's probing protocol.
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 "canvit-pytorch>=0.2"
import torch
from PIL import Image
from canvit_pytorch import CanViTForSemanticSegmentation, Viewpoint, sample_at_viewpoint
from canvit_pytorch.benchmarks.ade20k import CLASS_NAMES
from canvit_pytorch.preprocess import preprocess
model = CanViTForSemanticSegmentation.from_pretrained_with_probe(
pretrained_repo="canvit/canvitb16-abl-no-vpe-2026-03-03",
probe_repo="canvit/probe-ade20k-40k-s1024-c64-abl-no-vpe",
).eval()
scene = preprocess(1024)(Image.open("scene.jpg").convert("RGB")).unsqueeze(0) # [1, 3, 1024, 1024]
state = model.init_state(batch_size=1, canvas_grid_size=64)
with torch.inference_mode():
viewpoint = Viewpoint.full_scene(batch_size=1, device=scene.device)
glimpse = sample_at_viewpoint(spatial=scene, viewpoint=viewpoint, glimpse_size_px=128)
logits, state = model(glimpse=glimpse, state=state, viewpoint=viewpoint) # [1, 150, 64, 64]
labels = logits.argmax(dim=1) # ADE20K classes, named in CLASS_NAMES
Details
| Features | the 64 × 64 canvas of canvit/canvitb16-abl-no-vpe-2026-03-03 |
| Probe | LayerNorm, dropout, BatchNorm, 1 × 1 convolution |
| Training rollouts | 10 glimpses of 128 px, R-IID viewpoints, 1024 px scenes |
| Training steps | 40,000, batch size 16 |
| Released weights | step 35,500, selected on ADE20K validation mIoU |
| Optimizer | AdamW, peak learning rate 0.0003, weight decay 0.001 |
| Schedule | 1,500-step linear warmup, then cosine decay |
| Augmentation | DINOv3's segmentation training transforms: scale jitter by a factor of 0.5 to 2, random 1024 px crops with its class-ratio crop rejection, horizontal flips with probability 0.5, photometric distortion |
| Dropout | 0.1 |
| Precision | bfloat16 autocast |
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}
}