Visual Representations inside the Language Model