Maybe I’m expecting too much.
It’s mostly the censoring. The uncensored open source models are a lot more willing and able to generate exact lookalikes. But they also have a tendency to inject porn randomly because that’s what a lot of them were trained and fine tuned on.
In other words, probably not that different from our own celebrity crush daydreams 
Ah, that makes sense.
I mean, maybe? People get so used to really advanced things happening in science fiction that they take it for granted. But just because AI image generation isn’t 100% flawless yet doesn’t mean that it isn’t a really, really advanced thing. I think it (and LLMs) is the most futuristic, science-fictional technology in existence today. (The only real competitor for the title being Space-X’s Falcon 9 first stage.)
The height markers on the left have two 5 foot marks, the one on the right has either two 5 foot or two 4 foot marks (or something else altogether) and the top right number is 7 inches.
It’s interesting to me to look past their blueness and greenness (and head tentacles 
) to see what ethnicity(ies) their features seem to resemble.
IMO they look like an 80/20 mix of Caucasian and African features. What seems to be fully absent is the various Asian or Amerindian influences. Which outcome probably fairly accurately represents the ethnic mix of a US-centric random trawl of internet photos of ordinary people.
It’s noticeably worse if you try to get around the censorship, lol.
Mantis looks like a surprised bastard daughter of Sandra Bullock and Michael Jackson.
And if you ask for stereotypes, well, you get stereotypes:
More fun to mix and match stereotypes.
It seems to struggle with the distinction between “Asian” and “Native American”, which I suspect some human artists would too (in terms of facial features).
And in this case, the conversation has also gone out long enough that it’s forgetting some of the previous contexts. The guy on the left is half-assedly blue and gave up on drumming completely.
It’s starting to look like a MMO video game with paid skins and costumes.
Before the alien line-up (BTW, it refused to render it when described as a "police line-up) I tried for an alien support group. It made Harry Vanderspeigle bluish-grey for some reason. But still recognizably Alan Tudyk.
BTW, a different AI called Dall-E used in Microsoft’s Bing (among other places) attempts to solve the diversity problem by quietly adding terms to your prompts in the background and undisclosed. But they sometimes “leak” into the rendered image. You see phrases like “ethnically ambigious” and “east asian” and “hispanic” and “female” (or more typically a mangled version) mixed in with any text in the image.
Obvious AI. The people look like department store mannequins, their skin is too smooth and perfect, their expressions are dead (especially the eyes), and the images are too sharp to be real.
ChatGPT 4o has a clear understanding of what sleestacks are. BJ and the Bear, not so much.
I find it interesting that ChatGPT knows that Bear is some sort of primate but isn’t sure which species, so it throws in a capuchin.
The two prompts:
Summary
BJ and the Bear slowing their truck down to stare at three sleestacks hitchhiking on the side of the interstate. The point of view is from inside the cab of the truck, looking out through the side window. Kodachrome DSLR image with shallow dof and forced perspective
BJ and the Bear slowing their truck down to stare at three sleestacks hitchhiking on the side of the interstate. Kodachrome DSLR image with shallow dof.
Lately I’ve seen Facebook’s algorithm has been showing me pictures that look like various cartoon families as real people, which I am 99.99% sure are AI generated. Some examples:
Realistic looking versions of Hank, Peggy, and Bobby from King of the Hill standing next to a propane grill, with Dale drinking a beer in the background.
Realistic looking versions Homer, Marge, Lisa, Bart, and Maggie Simpson sitting on the couch in a realistic looking version of the Simpson’s living room.
Realistic looking versions of Peter, Lois, Chris, Meg, Brian, and Stewie Griffin from Family Guy. Brian looks like a real dog. Stewie looks like a baby with an adult’s head. The AI obviously had trouble with Stewie.
I’d post the images, but I’m at work and the sites that host the images seem to be blocked by the corporate firewall.
Those are made with ChatGPT version 4o, the one I started this thread about. That’s another insanely impressive aspect. All you have to do is upload a screenshot and ask for it to be converted into a realistic image.
It is also the source of all of the action figure images: you just upload a photo, ask for it to be converted into a carded action figure, and add the text for the card and any accessories that you want. You don’t even have to be specific. With this photo of my cat, I asked for the text (and got the exact type of yellow star bubble I wanted, without asking) but just told it to include two or three appropriate cat toys as accessories. (I don’t know what the chicken leg thing at the bottom is supposed to be.)
As stand-alone apps a media converter an an action figure maker of those skills would be major undertakings to write and highly popular. But both of those are just unplanned, emergent properties that happened to form in the complexity of the model. (Another emergent property that is just starting to get press, the ability to determine the geographic location where a photo was taken.)
I think it kind of missed with that one. Not only does the cat not look at all like an action figure, but the plastic indentation isn’t nearly deep enough to hold it.
That’s a photo of a cat under plastic. No dimensionality and no shadow (like what the toys have). Looks like an attempt to ripoff someone on ebay.
You are really going out of your way to be dismissive of what this is doing. I opened a computer program, told it in a few words in plain English what I wanted done with the photo, the program understood what the “cat” was enough to perfectly extract it from a complex background, create a new version of it, place it inside of the backing and bubble of of an action figure card, generate clear and appropriate lettering, and choose a mouse, a fish, and a ball as appropriate toys for a cat. None of which it was specifically taught to do, all learned by examining many pictures of cats, cat toys, and action figures and internalizing the concepts. I expect your “too cool for school” dismissiveness of all that is based on a personal philosophical distaste for modern AI and that there is no level of competence you wouldn’t dismiss.
Here are examples generated with more elaborate prompts:
So? Isn’t the argument for AI stealing work from others that it makes “realistic” images? But none of your examples are realistic in the slightest. So your plagiarism doesn’t even achieve your goal.
The images look fake, so you’ve stolen them for nothing.
Nailed it.