Using AI during Scanning and Subsequent Editing

[blee1996]

Actually hallucination is not limited to AI, human do it all the time as well. Our perception and imagination are basically using the same brain structure (or computing units). That enables us to add in the missing features, fast predict the changes, as well as daydream. It is not exactly a bug, but a feature as well.

 

[_T_]

Well, that's just the term that has come into common use for now when an AI makes a prediction that is contrary to reality, but it's just a fanciful and evocative term to describe a common problem with the software. It's not really hallucinating in the way that a human would hallucinate due to a mental illness or while under the influence of a drug.

The ai is not capable of thought, it doesn't dream or see. It's just an algorithm that analyzes what usually comes next in a set of data that you feed it and uses that information to make a prediction of what you might want it to show you next based on the prompt that you give it.

 

[Alan Edward Klein]

No, it's just a hallucination. It's not a real person's face, but an amalgamation of faces from many many images upon which the software was trained.

Actually that's true for all the faces in the photo, none of the faces are the actual faces of the people originally in the scene.

Generative AI is not capable of making the kinds of decisions you're attributing to it. It doesn't have the capability to know who a person is, or what they look like. It's just predicting what it thinks you want the individual pixel values to be based on the pixel values it was trained on and the prompting you give it.

In his post, Les said: "I have other clearly focused images of these contestants and the guesses were reasonably close.'

 

[_T_]

And yet that's all it's capable of doing, guessing. It just happened to be "reasonably close."

 

[calebarchie]

My drum scanner from 1996 has AI...

 

[MattKing]

It is reasonably likely that contestants at a beauty pageant would have employed cosmetics and chosen hair styles that tend toward a certain set of expected, idealized looks. They may have also self-selected based on those sorts of idealized looks.
So AI may very well have a fair bit of reliable data to help reconstruct how those contestants were attempting to appear.

 

[koraks]

In his post, Les said: "I have other clearly focused images of these contestants and the guesses were reasonably close.'

And like I said, that may be the qualification attributed to the result by someone who doesn't know these women. If you'd show the reconstructed photo to a brother, parent etc. they'd go "Hahaha no you silly, that ain't our Rhonda!"

 

[Korbel]

I have current version of photoshop, so I can contribute a little from my part. Camera raw has an early access AI feature called "dust" under removal panel, it is supposed to find and clean it in one click, but it doesn't detect my dust at all. I have a suspicion it is trained on a dust from digital sensors, not scans.

Photoshop AI has one thing the others don't, it is the only AI tool I found, which fills only the selected portion of the image, so you can, for instance, fix a nasty light leak without affecting the rest of the image:

rb67 + portra 800


But take in mind the current versions of photoshop upload everything you put into it into their cloud to train AI, if you firewall it, then you loose the AI tools. Also, sometimes it fails, like here when I tried to remove the fence:

rb67 + portra 800

The only reason I see why to use the new photoshop nowadays is the AI denoise in LR, which let's me shoot on 25600 iso. Other than that it is quite useless, because you cannot rely on the AI retouching tools. If I didn't make a living from it, I doubt I would pay the expensive subscription.

 

[koraks]

like here when I tried to remove the fence:

It's kind of funny as you can tell that model training involved a lot of photos of horses.

Thanks for sharing those results; looks like we're not there yet!
Specifically for dust removal I personally would hold more stock in an AI-assisted routine that mimics a human user in (1) detecting dust and (2) healing/cloning it out using intelligent sampling within the image itself. This should be technically feasible, but there's evidently not a whole lot of money in it, so it's not something the major players (the ones that really have good access to LLM tech currently) would be investing in. Perhaps in the near future when a DIY/enthusiast developer makes something that you can connect your Claude sub to.