Copilot and ChatGPT both use the same image generator, but have different guidelines. Copilot refuses to address a niece in Japanese, but ChatGPT had no complaints.
(Neither did Gemini, which of course has a completely different AI.)
Summary
A Japanese man is worried about his niece (8 yo) becoming lost on an upcoming trip, so he is writing her home address in Japanese on her lower arm with a sharpie. Iphone 15 photo. 9:16
ChatGPT and Google Gemini have very different guidelines. When trying to generate the “build-a-boob” image, ChatGPT refused to let me do any refinement to make it more like I wanted it to look. Something about “violating guidelines” or whatever. Gemini had no problems at all.
Funny story about gemini. I had a silly idea in flow (the main google animations tool) where a professional wrestler dropped out of the sky and did an elbow drop on a dolphin who was swimming in his water tank. Google clearly would not allow the dolphin to take a hit, to be the target of violence. But what was funny is that it didn’t just say “this prompt is inappropriate, I’m not going to try” – it instead rendered the video, but in a way where the dolphin had a magic force field of protection, and honestly the results are pretty hilarious. Which is way better than if it had just refused to write the prompt. So google clearly has something between total prompt rejection but still not letting certain things happen, and I wonder if the image generator accepts a prompt but steers away from certain parts of what you asked for.
Was the first one a rare example of an anti-Freudian slip?
or in this case, a Freudian bra, I suppose?
That’s definitely a psychokinetic dolphin. Not only does it deflect the wrestler, but it also splashes the water while it’s in the air, and then wormholes from entering the water directly to exiting the water.
Yeah, it started with the name popping into my head and was prompted from there.
Those of you who lived in my alternate-reality version of the 80s surely remember the popular Goldwell Estate series, featuring spunky young and popular debutante Jennifer Goldwell, a girl was always ready to act.
I made them each separately but put them together into two images to not spam the thread. Sample prompt:
Art by Clyde Caldwell: Book cover of a prim debutante teen girl in a cardigan and skirt firing her shotgun in a bowling alley. Teens are running away in a panic. “Goldwell Estate Stories”, “When Samantha is caught cheating at bowling, Jennifer had to act!”
Well, dang it. Actually it seems like I was copying the preview image rather than the full one. To avoid being indulgent…
If they have coherent text and tell some sort of story / textual idea going on you can tell they’re not midjourney, midjourney doesn’t work like that. almost certainly one of the autoregressive models (nano banana, gptimage, a few minor others)
Gonna nitpick on otherwise pretty excellent images. I’ve noticed this problem; the generators really want to have characters facing the camera, no matter how inappropriate that might be, even when you explicitly describe them doing something where they should not be facing the camera. You need to work a lot harder on the scene composition and stage directions to get facings right. When I’m trying to get that to happen, I often overload it and other important details start getting dropped. Grr.
ETA: If I hadn’t seen the prompt, I wouldn’t have noticed anything wrong about the directions the characters were running in the scene. But once I saw the prompt I knew it had not obeyed it properly.
No argument. Really, the whole thing was just me joking around with friends on Discord. I mentioned that Caldwell was a good artist for getting 80’s vibe images and made the Commie cover as a demonstration off the top of my head. Then I ran with it in increasingly insane situations (as a good series does). None of them were more than two or three gens and picking out the better looking ones and I attribute any errors to a lack of quality control in 1980s YA Scholastic Book Fair level publishing
If it mattered, I could have gone back and tweaked everyone’s heads, etc but it didn’t matter in context of me making them.
A friend of mine was making “Galactic Tax Collector” images at the time, which led to the crossover episode.
The combination of artist with the simple description of prim blonde debutante teen led to some pretty predictable images and any that didn’t fit weren’t picked.
