I’ve been trying to get images similar to the first one I posted (Yokai Parade) by giving prompts that lead in that direction. Tonight I tried “Full Moon Over Tokyo Bakemono”. I didn’t get the nightmare creatures I was hoping for, but I did get one of my favorite images so far.
This city needs to be the setting for an SF book.
Has anybody taken a look at OpenAI’s improved text-to-image system, DALL-E 2? I don’t know how curated the images are that are presented on the site, but still, that’s a new level of impressive…
I’ve looked at pages showing off that and the landscape generator, but have never been able to actually use them.
That’s not Robin; that’s Violet Parr.
Speaking of invisible, has anyone else had cases where the AI starts by painting the whole canvas white, and then at the last minute realizes it needs to toss in a little bit of detail? It happened to me with “Flag of Ukraine”:
And then again with “Can we fix it? Yes we can!”
(but oddly, not with my earlier “Polar bear in a snowstorm”
)
Oh, and I think this one is recognizable, but then, I know what the prompt was:
What do we think? I pity the fool
Mr. T with a little more hair?
I mentioned Lens is pretty good at some of mine. In this case, it is suggesting just about every black guy on the planet.
It has got to be doing some sort of looking stuff up like Chronos suggested. Here is Fin Fang Foom and Doctor Doom Sitting in a Waiting Room.
It is obvious that the system has some idea of who Fin Fang Foom is. (Doctor Doom, on the other hand, looks more like a profile of Bender Bending Rodríguez.)
Here is The Great Wave off Kanagawa by Hokusai.
“At the mountains of madness with Pinocchio”. First take, no fine-tuning. I find it quite expressive.
Evolved it a bit by the cue “HP Lovecraft AND horror”, and this came out:
ETA: I like the little baby Cthulhu in the bottom left corner.
I pity the fool who subscribes to an arbitrary set of beauty standards centered around the worship of bilateral symmetry.
This is what I got for:
“Questionable Content” - weight: 1
“volumetric lighting Unreal Engine matte painting detailed painting 8K 3D” - weight: 1
And then I evolved it with accuracy boost and got this???
It may have gone unnoticed, but I spoilered it below the image: The prompt was, literally, “I pity the fool”.
I’m not sure why it gave him so much hair: Mr. T’s mohawk is probably his most distinctive feature (along with his heavy chain necklaces, which it also partially included). Maybe it has some sort of reversion to the mean going on, that it tries to make any picture of a person look more like “typical human”?
Experimentation continues. To my surprise, it was able to correctly deal with a prompt in Latin (albeit a very familiar bit of Latin):
‘In principio dixit Deus “Fiat lux!”, et facta est lux’
That’s definitely God creating light, if we accept that God looks like a character from a goofy Adam Sandler movie.
Emboldened by that, I tried Sindarin, which I think maybe worked?
“A Elbereth Gilthoniel silivren penna míriel o menel aglar elenath!”
No such luck with the Black Speech or Kuzdul, though:
“Ash nazg gimbatul”
“Baruk Khazâd! Khazâd ai-mênu!”
A giant djinn with an ice sword is cool and all, but it’s definitely not a dwarf with an axe.
Wow, I didn’t even notice the blurred text before. The image apparently kind of hides it from the mind.
I didn’t see it either and apparently guessed correctly.
In my ongoing quest to generate a whole series of similar creepy images (I want to make a series of “photographic slides”) I tried Gashadokuro at Fuji Shrine.
At first I couldn’t discern any sign that it knew what a Gashadokuro is. But then I realized, yeah, thats a skull and hunched back eclipsing the moon, and stunted arms reaching out. I concider it a failure at being a good image, but I can see the influence.
How did you get it to do widescreen?
It’s one of the selections you can make (16x9) if you toggle ‘advanced options.’
I wish I knew that before. I wonder if there is any way to generate a 16:9 (or even 4;3) recomposition of works I have already made? Would re-entering the same prompt generate the same image again? (I’m guessing not.)
Maybe if you enter the same seed value? There might be other randomness involved though.
I’ve never done anything like that. The titles that I give for my posts here are the entirety of what I enter. (And I’ve never evolved any of them.)
