I tried a couple through their web site
Definitely experimentation is required to see the qualitative effect of not only engineering the prompt, but also messing with various parameters (ideally at the level of the actual Python code, because it’s not immediately clear how these are exposed in their interface)
Yeah, I don’t get the new method yet. Why are some of my 9 images blurry? And where’s the dancing?
If the head was better, this would be great!
ETA: Better without the names
ETA2: I published the 2nd one and it was immediately reported. I can’t imagine why.
MidJourney also (coincidentally? No?) dropped a new Beta model today. Naturally, I had to try my standard test: Eldritch Cheerleader Carwash.
I feel like the dog that caught the car.
There is some censorship filter but I do not have time right now to figure out what it thinks it is doing.
I’m not sure I want to know HOW she’s cleaning the car…
The new system doesn’t understand the meaning of prompts as well as some, you’ll need to go simpler. When it does understand the output can be pretty awesome, though:
Boring failures include giant crabs not attacking a NASCAR race:
And Godzilla not eating Donald Trump (I do love the crude drawing of Godzilla giving another kaiju a bouquet of greenery, though.)
Fun failures include cats holding signs that do not say “a cat holding a sign”
And “who dat is?” (Who could complain about “who drat?” and “whch dis”?
(“by Mark Ryden” is an excellent prompt suffix in the new system.)
See some eldritch cheerleaders here (along with a couple of Mark Ryden prompts)
https://creator.nightcafe.studio/u/IEAf6ufciDe5z9lrDkACHE6Q1PX2
Nothing shows up. Just says no creations yet.
Oops, never mind, I didn’t share them. The cheerleader carwash one did have cheerleders and cars, but nothing eldrich.
50 steps (probably worth increasing to 100), guidance scale = 12.7
If you are reading anything political into this one, it’s just in your head, I’m sure.
ETA: OMG, this one got reported to. People have too much time on their hands.
I have to admit, I’m really enjoying the results of this goofy algorithm sometimes, but I feel it kind of spoils us. It’s much more like someone on Twitter described as “Slot Machine Art”. You just pull the little lever and see what you get. The old method of Artistic and Coherent evolution seems like it came up with much more original type stuff.
What you could do is take the output of an “artistic evolution”, or your own original sketch, and evolve it using the diffuser. Generating from random noise is not the only option.
Same discussion/debate is going on at MidJourney. As one person put it, “The new beta creates stock photos, the old version created art”. I don’t 100% agree with that but I get what they were getting at. Previous works had more (virtual) creativity versus just plugging a couple concepts together. Other people noticed that the prompts worked differently and maybe a good deal of it was just relearning the system. Or running two systems concurrently and having the user decide which way they wanted it interpreted/rendered,
I’m enjoying the original avatars I’m making
Can’t believe I have never perused this thread. Really amazing. Maybe Cecil should revive the Threadspotting feature…. Obviously some Dopers have spent many hours on the site. How much time did it take to learn the basics?
Do you know how to type words? You’re up to speed.
There were already two vastly different systems with vastly different understandings of prompts and types of images generated—this is just a third. Having to learn it all over from scratch again is just a natural consequence of vastly complicated self-teaching AIs. Nobody knows the rules without trial and error, even the programmers. Each of the three systems have things that they do pretty well and things that they are mind-blowingly terrible at.
I’ve tried plugging a few of my past complex prompts for the “coherent” system into “stable” and got vastly different results, some of which I quite like. The new system has mostly solved facial symmetry (in some cases it is obviously mirroring select parts of images for faces and bodies, but far from always) but unfortunately is very, very bad at hands.
One prompt I’ve been playing with in the new system is “Matilda and Leon from the Professional by Mark Ryden”. First is the “artistic” AI (scaled to fit the other images), next is “coherent”. Both are, as you can see, total crap. Next are two from the new “stable” AI. Do they look anything like Natalie Portman and Jean Reno? No, they do not. But they are interesting. (I generated a couple of dozen, portrait and landscape, and they are all over the place.) Next up is the same prompt but with da Vinci and Rembrandt instead of Ryden. Those, I think, did a really good job—I could hardly ask for better.
