But wouldn't you need the original image first? At which point you are merely adding compute to downscale it after you already obtained it?
the original has to exist somewhere, but it doesn't have to be everywhere