DE3 does a pretty good job with “realistic photo of Ahsoka Tano dancing with x” (though it often tries to show them as action figures with seams at the joints).
Am I the only one offended by the snail shells with concentric rings rather than a spiral?
Sometimes you get spirals.
They finally get hands looking half decent (sometimes) and now it’s snail shells holding AI back.
Continuing to play around with Bing/Dall-E 3 when I can find a prompt that doesn’t get censored. Was doing that yesterday when a bumped Discworld thread inspired me to try “Tiffany Aching in Discworld”. As I mentioned in my first post on DE3, each set of images has a distinct style/content that isn’t in other sets. It is clear that ChatGTP must be adding additional descriptions to each prompt behind the scenes without informing you. (In the last set of images on the bottom right I tried “Tiffany Aching in Discworld not holding a frying pan”.)
Other Discworld tests:
Night Café evidently doesn’t know what an Assaultron is but it certainly got the setting right.
There have been a number of times I’ve asked a SD or DE version for a specific robot and got a random robot. So the systems know what class they are even if they don’t know specific details.
Well, the Librarian is pretty good, at least… but some of those are just completely clueless.
For the past couple of days I’ve been doing an insane number of images of characters doing the Say Anything pose whenever somebody occurs to me. Sometimes Bing is amazingly good at understanding the character, sometimes it is totally clueless.
Charlie Brown
Peppermint Patty
Snoopy
Woodstock
The nanny filter won’t let me do many modern real-life figures (and even trying much can result in suspension escalating to banning). Not even the generic “Pope” or “American President” without names are allowed. Dead ones seem fine, though. I’ve made Bob Ross, Elvis, Marilyn Monroe, and Stephen Hawking, for example.
Sometimes I try an abstract concept and am blown away by how well DE3 handles it. Here, for example, is anthropomorphic fingerprint:
Here is a brand new one. I love that Grumpy Cat is wearing a Grumpy Cat sweater.
Interesting to see what DALL·E 3 thinks the future holds for us. (Note: I commissioned these last week.)
Can’t see many hyoomans…
“What the world will look like 100 years from now”
Midjourney gave me a pretty accurate rendering of my 1950s era Lionel O-Gauge train set, in a hyperrealistic, cinematic/Tim Burton action scene with studio lighting, etc. Pretty much nailed the engine (that my brother still has: solid metal, very heavy). Close to how I imagined it as a kid, speeding around our Christmas tree.
The latest fad is using Bing/DE3 to make comics. Here are some of mine.
A four panel comic about Charlie Brown and Yoda learning to pluck a turkey.:
A four panel comic about a man and a woman saying “transubstantiation” to each other:
A four panel comic about Charlie Brown and Garfield saying “To hell with Mondays!” To each other.:
The Mad Scientists:
I tried Peppermint Patty in SDXL and got figures that showed a vague understanding of what Peppermint Patty looks like (reddish-brown bob, shorts) but completely unlike Patty. Amusingly, one had some Schultz-esque art on her shirt so it KNOWS, it just didn’t wanna do it. I explicitly didn’t try prompting for comics or Schultz.
I also made an unconvincing off-brand Don Draper and even less convincing Les Nessman. Once again, showing a vague understanding of the character but in a “heard a story from a friend of a friend” sort of way.
DE3 is good at “vaguely like the actors playing fictional characters but you’ll only recognize them if you are told”. For example, here are “Die Hard” and “The Fonz”:
Others are more recognizable. Asking for Sherlock Holmes I very clearly got the Benadict Cumberbatch version. (Asking for Sherlock Holmes and Doctor Strange weren’t quite as good, but still obvious).
BTW, here are a couple of galleries full of many of the tries I made with the prompt. All of the name celebrities in these should be recognizable without explanation.
I’ve mentioned before where Bing seems to be secretly adding details to prompts. These current comic generations are clearly revealing prompt details being added to increase diversity that are being turned into image text.
