My point being that ‘image’ models are predominantly trained on ‘images’, not on UIs.

They may be ‘underused’ as you say, but that addresses a different point than the one I’m making.