Additional examples from Fig. 2 Right, containing first four PCA components for more images for supervised ViT and DINO-ViT. DINO-ViT produces "cleaner" components. Both ViTs demonstrate positional bias in early layers.