Controlled retraining experiments isolate global-view alignment, local crops, patch masking, and crop scale in DINO-style vision transformers. Aligning two geometrically displaced global crops from the same image is sufficient to produce strong semantic structure.
Local crops substantially improve classification but add little semantic correspondence at fixed compute. The iBOT masking objective improves correspondence by 10.5 PCK points when combined with global alignment, but cannot create that structure alone. Semantic correspondence also predicts performance across dense 2D and 3D tasks more reliably than classification accuracy does.