jaakkoc

Check the 2nd note of the blog post (best viewed with desktop)

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.

Thanks for the feedback, I explicitly added that now to the blog post. The wording on this comment was off. That model is listed in the "Acknowledgements"-tab of the application.

The end goal on working on this program has been that the user shouldn't care much about what model/architecture/whatever tech is used, and should be abstracted away as much as possible. Though in a blog post like this it should be mentioned, it was hard to break away from that self-imposed mental jail :)

It is what it is with the title length

Thanks. Yeah, I used to do the tug-and-pull of scaling, then adding noise and sharpening, scaling a little more, adding more noise, adding more sharpening, etc. to ad nauseam. And photoshop sharpening was too much always for my taste.

At scale, this tool should speed that process up a "bit". Let's say 100s of photos.

Full disclaimer: the Enlarger app is something I did.

Fuller disclaimer: this kind of a "postmark"-sized pic blowup test case is borderline unfair to every model in the comparison, mine included.

Crazy times we live in that a president can do this.

Cool. Now do an accessible version.

(/s)

Seems like it was $25k worth of tokens