Feasibility of changing foundation model (DINOv2 --> CroCov2)

[czanoli]

Good morning,

I was wondering about the feasibility of switching the foundation model from DINOv2 to CroCoV2, ideally without making heavy modifications to the existing codebase.

Additionally, I’m curious whether the clustering and PCA parameters (e.g., 2048 centroids and 256 dimensions) were specifically chosen for DINOv2, or if they would be equally valid for other models like CroCoV2.

By any chance, have you experimented with other foundation models besides DINOv2? I'd be really interested to hear about any insights or outcomes from those trials. If you were to perform such a substitution, do you have any important advice or considerations to keep in mind?

Thanks so much!