Rogue, hallucinating A.I.?

[Fast Eddie B]

I recently discovered how powerful A.I. apps can be in modifying and improving photos. I may post a few prior recent examples that worked really well.

But one has me scratching my head, and a bit concerned. I had a scan of an old newspaper article from the late 1960’s. It’s a local paper and the photo had me and an old girlfriend looking at an artist at a sidewalk art show:

Seemed like a prime candidate to improve upon a bit, so I asked Gemini to simply crop the original to just the photo and to colorize it:

I find it bizarre that it would add people, and completely change the artist’s position. The poor little dog actually changed color as well. I can understand getting the dog’s color wrong, but the other mods seem bizarre, with the A.I. going far beyond what it was instructed to do. Maybe some ambiguity in my command, but still very, very odd to me.

 

[Myriad]

The first time i looked at the original, I thought the artist was a muppet teddy bear wearing a beret. So I'm not suprised the AI got the hair/hat distinction wrong for the artist.

You, behind the pole, must have evoked (that is to say, generated a strong signal of similarity to) the visual trope of a person reflected in the edge of a mirror. It added the "real you" in the foreground to create consistency with the reflection, and moved the artist's canvas to correspond better to where the "real you" is looking. (Since you know what you actually looked like, I would guess you don't see the similarity between the hallucinated foreground figure and your actual self in the picture, but to me the added figure in the left foreground looks like a plausible "un-reflection" of you.)

It couldn't make sense of the overexposed background to the right of the pole. so it filled in a generic street scene. It also failed to parse the overexposed background to the left of the pole, so it copied it with some arbitrary contrast increase. I'm not sure where the black jacketed walking-away man came from. At first I thought the white symbol "on the back of his jacket" had somehow evolved from the prominent black square in the original right background, but realized that's a functional icon rather than part of the generated picture. Nice to know I can hallucinate too, when trying to fill in a coherent picture based on insufficient information!

 

[Darat]

I think Myriad is spot on. I've tried using a few of the AIs to repair and colourise old photos, and what I've found works best is to describe the scene and then say something like "remove scratches and other blemishes, colourise the photo, keep everthing else the same". If I know the colours that should be used I would add those into the description stage. That seems to keep the hallucinations to a minimum BUT the resulting image will not be your photo only colourised, it's whether you are happy with the inaccuracies in the new "photo".

 

[Fast Eddie B]

I thought Gemini did an admirable job on this:

Before:

[

After:

I especially like how it got the red taillights and amber parking lights right, and did a credible job of guessing the color of my friend’s outfit.

 

[Fast Eddie B]

Another…

Before:

After:

I asked it to eliminate the beer can and to make the sky bluer. I especially like how it added a glow from the cigarette I was apparently lighting!

 

[Gawdzilla Sama]

Has Skynet cleared this?

 

[jadebox]

The first time i looked at the original, I thought the artist was a muppet teddy bear wearing a beret.

 

[Darat]

Another…

Before:

After:

I asked it to eliminate the beer can and to make the sky bluer. I especially like how it added a glow from the cigarette I was apparently lighting!

That's not bad.

Back in 2021 I wanted to repair and colourise a photo, it was in a very bad condition -and it was a teeny-weeny photo as you can see on the image below - I also "scanned" it by taking a photo with my phone. The middle image is my colourised version, this was done by hand and using photoshop tools such as "heal" and "remove". Thought I'd try it today with Gemini, that's the third version below. In some ways it has done a better job (less grain in image), the colours probably look better as well on a monitor as mine was tuned to get the best printed colours, which are not always the same as best monitor colours. However the new image Gemini created is subtly wrong, it is not quite the same woman, a sort of uncanny valley. I find it off-putting but if you didn't know her or couldn't compare it to the original I think it would look OK.


(Lots of downsizing and compression required to post it here but I think enough detail comes through.)

 

[TragicMonkey]

(Lots of downsizing and compression required to post it here but I think enough detail comes through.)

You still have one of those pens?! Those were the bomb when I was in junior high school!

 

[Thermal]

This is the whole creepshow thing with AI. It's trying to think for us, doing stuff we don't ask it to do. That's got a Skynet vibe that doesn't sit well.

 

[alfaniner]

I think Myriad is spot on. I've tried using a few of the AIs to repair and colourise old photos, and what I've found works best is to describe the scene and then say something like "remove scratches and other blemishes, colourise the photo, keep everthing else the same". If I know the colours that should be used I would add those into the description stage. That seems to keep the hallucinations to a minimum BUT the resulting image will not be your photo only colourised, it's whether you are happy with the inaccuracies in the new "photo".

Ah, I remember when colorization ("colourisation") alone was considered anathema by many people. (And still is...)

 

[Trebuchet]

Ah, I remember when colorization ("colourisation") alone was considered anathema by many people. (And still is...)

Depends on what you're colorizing!
Your mom? Fine. Casablanca? Not so much.

 

[EHocking]

Casablanca I could live with. Young Frankenstein would be missing the point.