You could also have a go at subclassing the existing image dataset to get this working