I found this older photo of myself and a friend, 25 years old now, in some newspaper scan.

The photo was of poor quality, but one could certainly see all the features - so I figured, why not let ChatGPT try to play around with it? I got three different versions where it simply tried to upscale it, "enhance" it. But not dice.

So I just wrote the prompt "render this photo as a hyper realistic photo" - and it really did change us - the people in the photo - it also took the liberty to remove some things, alter some other background stuff.

It made me think - I wonder what all those types of photos will be like 20 years from now, after they've surely been fed through some AI models. Imagine being some historian 100 years from now, trying to wade through all the altered media.

This is similar to my experience trying to get Stable Diffusion to denoise a photo for me. (AIUI, under the hood they're trained to turn noise into an image that matches the prompt.) It would either do nothing (with settings turned way down) or take massive creative liberties (such as replacing my friend's face with a cartoon caricature while leaving the rest of the photo looking realistic).

I've had much better luck with models specifically trained for denoising. For denoising, the SCUNet model run via chaiNNer works well for me most of the time. (Occasionally SCUNet likes to leave noise alone in areas that are full of background blur, which I assume has to do with the way the image gets processed as tiles. It would make sense for the model to get confused with a tile that only has background blur, like maybe it assumes that the input image should contain nonzero high-frequency data.)

For your use case, you might want to use something like Real-ESRGAN or another superresolution / image restoration model, but I haven't played much in that space so I can't make concrete recommendations.

>hyper realistic photo

Never use the words "hyper realistic" when you want a photo. It makes no sense and misleads the generator. No one would describe a simple photograph as "hyper realistic," not a single real photo in the dataset will be tagged as "(hyper) realistic."

Hyperrealism is an art style and only ever used in the context of explicitely non-photographic artworks.

I think that upon closer inspections the (current) technology cannot make 'perfect' fake photos, so for the time being, the historian of the future will have no issue to ask his/her AI: "is that picture of Henry Bemis, with Bruce Willis, Einstein, and Ayrton Senna having a beer real?" And the AI will say "mos-def-nope!"