canvas-vision-transformer
CanViT: Toward Active-Vision Foundation Models (NeurIPS 2026). Model, pretraining, task specialization and evaluation code.