Oh, yeah. I didn’t notice that.
I tried your variation, but told it to split it into 16 images, and it pretty much blew ChatGPT’s mind.
16 x 3 x (average of at least 2 traits per item in the scene) requests you wanted it to get right in the same image, not surprised it failed trying.
I knew I was pushing it when I wanted what amounted to 16 x (average of 3) (I think) critical requests plus a few minor arrangement hints in that complicated group scene I posted yesterday. It took a lot of tries before it got all but one of them close enough and I called it quits and posted. Nobody tried to guess at the connection those had to one another. 
I only recognized Indy and John Wick. With the stylized purple objects it seemed to be some gaming reference that I’m not really familiar with.
Celebrities are definitely hit or miss. You often get their stunt doubles.
I’ve been doing a rewatch of the 1960s Batman and last night watched the episode with the Novosaurus (3x09). I tried to recreate that, but I just can’t get it to look as bad as the original (and I wanted accurate awfulness).
My prompt (also using the reference image):
Summary
A realistic photo of a large green, grey and yellow humanoid reptile. It has large 3-fingered webbed hands that end in claws and large webbed feet. It has an elongated snout that contains a large red tongue and several large, fang-like teeth. The snout ends in prominent nostrils and a tall spine/spike. Starting on the crest of the head above its large black eyes is a row of the tall, thin spines that run all the way down its back and short tail. It has a frilled collar around the neck and is covered with fine scales. It is at the same time menacing, terrifying, and somewhat comical. It lives in a swampy jungle with liana vines and dappled sunlight. Full-length Candid photorealistic kodachrome dslr nature wildlife photo with shallow dof and forced perspective. 9:16
Yeah, I tried redoing the same set of 16 connected ideas with a slightly different framing suggested by ChatGPT. The joystick refused to come out purple, and the characters in the photos came out looking even less like the actors than in my other try:
BTW, Nobody apparently guessed why Chell from Portal was connected to the other 3 in that earlier image, or maybe it didn’t need to be said.
Roger Moore’s not bad, but Keanu, George, and Harrison keep getting progressively worse, like the stunt double budget kept getting slashed.
I made sure that each generation was in a separate branch without any earlier pics to contaminate things in the context. I was actually going for Sean Connery in that last one. And Pierce Brosnan and/or Daniel Craig in others. I never called out Roger Moore specifically in any of the 3 I posted from this round. Several other attempts weren’t good enough to bother posting.
ETA: I tried one of them using o3, the reasoning model, rather than 4o. Before it hid its reasoning, I managed to catch bits where it was questioning whether to show recognizable people was acceptable within its guidelines. In that one, it might have deliberately obfuscated the recognizable actors. 4o might be doing that too.
Here are the cropped thumbnails from my ChatGPT library, they show in reverse chronological order.
#2 was just giving it the 16 things and go straight to making an image giving it no hints how to unify them at all. #3 was one where I gave it the 16 things, asked it to first write out a good prompt unifying everything, and then I told it to use it as written. It didn’t follow its own prompt well.
And here’s a quick screengrab of all the cropped thumbnails from yesterday. I had already started on using ChatGPT’s suggestion for how to frame it using the celebrity photos on the wall, but after the first try with multiple screwups I said to myself that last one from the previous set was good enough, post and done.
It took me a few tries before I figured out I’d accidentally deleted the word “purple” from my prompt text. Arrgh, at least one of those was pretty good otherwise. And, why, no, I don’t want to show you the other 3 I was working on before these.
From my “Mexican standoff series.” It’s notable it does not appear to fully understand how a Mexican standoff is supposed to work, even if you explicitly state who should be pointing what at whom.
This was supposed to be Django, Seth Gecko (Clooney from From Dusk Til Dawn), and Hugo Stiglitz from Inglourious Basterds. I got Django, Ewan McGregor, and some random guy who doesn’t know how a Mexican standoff works.
This was supposed to be Mr. Pink, Jules Winnfield, and Clarence Worley from True Romance. Got the outfits right, at least.
This should be self-explanatory. Not sure where that random extra crossbow and disembodied hand came from. The office itself is fairly accurate. Good job on Rainn Wilson again, and John Krasinski came out pretty good. Steve Carell is recognizable but a miss.
Again, bafflingly can’t figure out how a Mexican standoff is supposed to work. That’s supposed to be John McClane on the right.
I don’t think that’s a scenario where you want a weapon with spin-up time.
It’s fine when Ahnold can just impale the others with it after they’ve exhausted their magazines.
ChatGPT came close with Burton, but nailed Davis and Taylor.
The Petrified Forest meets Who’s Afraid of Virginia Woolf?
Three Clint Eastwood characters playing poker. Dirty Harry Callahan is still a bit confused about the rules.
Speaking of that, when I was trying for the “Rainn on your wedding day” image I just gave the names and it gave modern versions. An older Alanis was acceptable but a bearded Rainn was not, so I re-ran specifying Alanis Morissette from Jagged Little Pill and Rainn Wilson from The Office. ChatGTP asked me if I wanted to upload photos or for it to make its on versions of them “from those eras”. (I let it pick on its own.)
In that thread on randomness I asked for random words including non-English ones. I chose a few of those to make a prompt
Schatten Shizen Geist Peregrino. Yūgen Hikari Fulgur Tomodachi Amare Umi. Candid 9:16 photo taken with an iphone 12.
Here’s the result.
ChatGPT also chose to relist the meanings
Your phrase blends poetic and philosophical words from different languages. Here’s a breakdown of what each part could mean:
- Schatten (German): shadow
- Shizen (自然, Japanese): nature
- Geist (German): spirit
- Peregrino (Spanish): pilgrim
- Yūgen (幽玄, Japanese): a profound, mysterious sense of the beauty of the universe
- Hikari (光, Japanese): light
- Fulgur (Latin): lightning
- Tomodachi (友達, Japanese): friend
- Amare (Latin/Italian): to love
- Umi (海, Japanese): sea
Together, this could be interpreted as a poetic description or title, possibly for an image or a story—suggesting themes of a journey, light and shadow, nature, friendship, and wonder.
It just started doing that in the last few days and I’m really annoyed. I’m like, if I thought you needed a reference image, I’d already have uploaded it. Just draw what I asked for!
And then it might say that it can’t do images of real people.
A few days back (inspired by food images generated by my “random word” prompts) I tried making glamorous photos of terrible food combinations. I then decided to try adding Mikey from the old cereal commercials being horrified by it. It didn’t understand Mikey, but refused to do a redraw with an uploaded photo. (In ChatGPT.)
Just yesterday I tried images involving Go Go and Kiddo from Kill Bill. It did a bad job with Chiaki Kuriyama. The more permissive Sora refused to work with an uploaded photo of her, either the full character or just her face.
(It also seems to have a problem with the idea of the tardigrade rearing like a horse.)