Ensuring Consistency for In-Image Translation