I'm surprised there isn't a readily available water-mark remover at this point. A synthetic training set for such a model could be created trivially.