That will be an incredibly questionable dataset used to train it. At that point, there would probably be no need to upscale these images. Just downscale then use this questionable DL to re-upscale any image.