I disagree - the human face is the hardest target exactly because (most) AI upscalers make up face traits and alter expressions. The blog post you linked uses

I disagree - the human face is the hardest target exactly because (most) AI upscalers make up face traits and alter expressions. The blog post you linked uses ChatGPT, which uses generative cloud AI with billions of parameters, which does indeed make up stuff. The models that are diff'd on this blog post uses a whole different local architecture which has nothing to do with ChatGPT or Nano Banana 2, which the likes the blog post critiques. It is more rooted in older tech pre-generative era. I do agree with the blog post, and "fool most people" -effect is especially strong when it's an old photo from a time that is not freshly in the viewers memory, and the subject no longer looks like it.