Perhaps the method as you describe is more accurate the generative deepfakes, but it's still adding information that doesn't exist in the original image, no? Like once an image is encoded to video, compressed, then recompressed whatever number of times before it becomes a blurry mess, there a percentage of information lost each time. I would assume the upscaler is only as good as whatever image is as close to the source resolution is as possible, otherwise, again, it's just filling in the gaps with assumptions.
