Diffeomorphic demons registration algorithms work with vector image types?

I might be able to figure this out by plumbing the internals, or maybe trying to template a filter using a vector image, but I just wanted to see if anybody has the answer “at the ready”. In principle, registration by optimization based on intensity differences should work if the intensity differences are replaced by, for example, magnitude of difference vectors, but if the existing formulations aren’t set to handle such a difference calculation, things would break down. Any guidance would be greatly appreciated.