But we can already generate these images via image editing models, so what's new here?