A two‐stage active learning pipeline that uses diffusion‐based feature sampling and entropy‐augmented disagreement to pick the most informative pixels under extreme labeling constraints.