You can also enter the seed value in the advanced options. I haven’t played around enough yet to know whether you can recover the seed value used for an already-existing image, though.
If you click on one of your creations, it tells you the seed value that was used.
Maybe someone with lots of extra credits can try recreating a past image with the same text prompts, settings, and manually entered seed value?
The claustrophobic nature of 1:1 framing is part of the reason I’m trying to combine images.
The site gave me some bonus credits tonight (for getting “likes” so I decided to experiment a little with getting a widescreen version of my “Yokai Parade” image. The “duplicate” button seemed like a possibility for recreating the same image with tweaks, so I tried that, just changing it to widescreen. I got something completely different from the original. So I then made a brand new image with the same prompt but the widescreen option. I now expected something yet again different, and got it. It wasn’t wasted credits, though, because both new images are great.
I then did a duplication of “Ghosts Rising From a Graveyard”, no longer expecting a widescreen version of the original. I like it, too.
So the lessons I learned is if I want new images similar to a previous one I just have to enter the same prompt again instead of coming up with a new one.
Prompt 1: Cyberpunk 2077 - coherent
Prompt 2: Cyberpunk bar - artistic
Huh. I entered Psychobilly Cadillac – as in Johnny Cash’s "One Piece at a Time – for Steampunk style and it gave me something along the lines of an Art Deco railroad engine.
First link is broken.
There is no link, just the image.
The failures are sometimes amusing at least.
Here is what ALL the following prompts got me (with accuracy boost)
“Information Superhighway” - weight: 1
“Detailed | vast network| pinpoint lights | 8K 3D 8k resolution VRay volumetric lighting maximalist hyperrealism flickering light” - weight: 1
Yet another experiment in reframing an existing image. I used the same prompt and settings as the image I’ve been working with but this time uploaded the original as a reference (something I hadn’t tried before) but with the 16:9 switch. Unfortunately it came out 4:3 anyway, but holy crap, the result.
Okay, I thought “prompt 1” and “prompt 2” were supposed to be two different images.
Seems the “Art Deco railroad engine” turned into a Sci-Fi Tribute to Ed “Big Daddy” Roth.
I did screenshots of the various iterations for this one. (Linked because it is an animated gif.) I missed the initial image of a static of pixels, but this is 25%, 50%, 75%, and 100%. (Which is, I think, 50, 100, 150, and 200 iterations.)
The prompt was “And suddenly, he heard a familiar voice whispering to him. It was Godzilla.” That was taken from a passage by that AI text generator:
Summary
Godzilla ate back his mother’s ribs with her. And it was good enough to last for a good one. He had been there, not for the first time, then-and-today. He saw a picture of a lion in a blue cloak, standing off a cliff in the sky, at his side. He’d thought it might be a sign for “I’m all right, Man,” then he heard someone say something about how little he looked like that and that, when they met, he would never be one of him. “It was a big day,” he said. When the day rolled around, it wasn’t the first time Godzilla had eaten something, but the first time he’d been to his father’s grave. When King Gungans broke in through the door and knocked—in the night by an overgrown wild beast—he realized the woman was nowhere to be found. He looked around frantically. There was still nothing to do but walk. He heard his mother’s voice. And suddenly, he heard a familiar voice whispering to him. It was Godzilla. This is, of course, all very strange. Now, though, it all was a good day. In the evening, he had to call one of Tokyo’s famous kurabata shops and ask to buy a small amount of his own blood to kill his mother. But to his horror, what he saw didn’t surprise anyone; he was not
(There are more lines from that that I’m thinking of trying.)
Did much better with creating a Fallout 4 settlement.
Looks like it’s being guarded by a sentry bot.
“Halloween children trick or treat” in oil painting style should produce a nice wholesome Rockwellesque image, right?
Nope.
I wanted to add that upscaling widescreen images is a real bargin. With square images you get only 1024x1024 (1 megapixel) for one credit, but with widescreen, you get 2560x1536 (nearly 4 megapixels) for one credit.
Hm. I thought maybe the blank-canvas thing was a fluke, and so I re-tried my “Flag of Ukraine” prompt. Again it started by blanking the whole image, and again it eventually added a clothlike texture and (bizarrely) a bit of some sort of foreign (Arabic?) script in the corner, but this time, it didn’t even bother to put in the tiny highlights of blue and yellow near the edge. Apparently something about how it treats that prompt (and a few others) is just plain broken.
Try a different flag. Bhutan would be a nice choice.
(It also occasionally tries to “sign” mine in the corner.)
What settings are you using? for the default setting you get when you go to the page for the first time (I think it’s matte, artistic, I get this:
For “ukrainian flag” as oil painting, artistic I get this:
And, if it hasn’t been mentioned, or if it’s already obvious, you can play around as much as you want in Incognito mode if you don’t mind losing your creations.
“Halloween Children Trick or Treat | Style of Norman Rockwell” gives you a more Rockwellesque painting but also drives home the limitations of the AI to draw people.
I get the neighborhood street and pumpkins, etc and guess that’s a white sheet ghost in there but the rest is fairly mangled.
(Edit: I know you weren’t really TRYING to get a Rockwell painting but thought it was an interesting experiment)