For pretraining we use multi-resolution MAE: we mask tokens across scales and reconstruct each level from the others (cross-scale inpainting). This yields strong, scale-consistent representations and speeds up downstream training, and lets us leverage large unlabeled microscopy datasets.
