Manipulating Natural Images By Learning Relationships Between Visual Domains