Ehh maybe, the images will be too perfect. I don't think they can model every source of bias, that would be a huge training problem.