DINOv2 models with registers lose the ability to semantically represent similar objects when performing PCA across images.

[aenoca]

It seems the DINOv2-S with registers does not display the same emergent behaviour as DINOv2 without registers for matching different semantically close elements across images. I've observed this consistently across different combinations of models.

[hiyyg]

Does the plain Diniv2-s have such ability? Did you also observe similar results for larger models?

[data-ant]

我已收到，谢谢。                       祝您诸事顺利!