Including some of the AI sites themselves.
Early in the thread there is a short back-and forth where I was trying to generate a Great Pumpkinesque pumpkin creature rising above a pumpkin patch. I developed a complex, two-lined prompt in the process, and generated some interesting images, but never quite got anything close to what I was imagining. When Stable Diffusion became available I condensed the prompt into one line and tried it there. The results were very different from “coherent”/Disco Diffusion. The images in the upper right approachs the kind of scene I was originally aiming for.
And now that I can access Dall-E 2, I have tried it there. Quite different yet again. Even though it isn’t exactly what I was imagining (no pumpkin patch, for instance) I really, really like the fat pumpkin guy on the upper right.
Then I realized that Dall-E 2 might be sophisticated enough to just directly ask for what I want instead of building an arcane formula like in “coherent”, so I simply asked for “pumpkin monster rising above a pumpkin patch”. Here are example results from that. The upper right image there approaches perfection. I could make some small complaints, such as I wanted a night sky, and the pumpkins should still be on the vines, but otherwise it is just about as good as I could hope for.
It looks like Dall-E 2 uses small blobs of color in building its images, most visible in the red devil-looking pumpkin image (which looks like it needed more iterations). “Artistic” in Night Cafe uses obvious horizontal-ish ribbons, “coherent” is more subtle but there does seem to be often fine hairlike elements in the images. And Stable Diffusion is doing something complex-parts of some images are obviously mirrored at some stage of construction, and elements often copied several times across the image, so there are some kinds of processing going on beyond extracting image from noise.
That site has content filters even more strict than Night Cafe. Such as the word “skull”, or “ugly”. And apparently it has a dictionary that it uses to block prompts with words that it doesn’t know
I’m having a lot of fun over there. At least it’s free. You can run all sort of experiments without burning through your coffee budget for the week. It seems to allow you to add many more “words” than Night Cafe, but I don’t think it reads them all. I’m pretty sure it just ignores everything after a point. I’ve been porting stuff from Night Cafe over with a full compliment of prompts and adding more prompts at Playground AI doesn’t do much - I just get variations on an original theme.
Another example of (almost certainly) false positive censorship at Net Cafe. 5 out of 16 images censored.
3 out of 7 credits were refunded, which meant 11 images for 4 instead of the normal 9 for 4, so I was ahead at least, yesterday. But then today they dropped their Stable Diffusion “cost”. It is now 1 credit for 4 images/4 credits for 16.
I can’t figure out what is going on with the censorship at the Playground. Sometimes it rejects a prompt for content and other times runs it fine (at the beginning, some times it refuses to show images after they have finished, too.) It appears to have a problem with the word “pile”? The times I tried making a pile of skulls it refused to run it, but a mound of skulls? That’s no problem. Weird.
On the other hand, they keep making improvements on the interface. For instance, now you car run up to 4 images at once. It is no wonder Night Cafe is having to cut prices, competing with “free”. I suspect the main Dall-E 2 site will have to significantly adjust their pricing before the site goes out of beta, too.
Night Cafe has added Dall-E 2 to their options. Expensive, though. One image per credit as opposed to four per credit for Stable Diffusion.
Playground AI has gone from free (“here, kid, no charge … this time”) to money-grubbers. First they started charging if you wanted more than 300 images a day at high resolution and after 1000 images you had to wait 30 seconds to create again. Fine, I paid. I suddenly had the ability to post 2000 images a day, but only at 512x512 and only one image at a time. They tweaked it a little and fixed the 512x512 issue, but the only other thing they addressed is now I can create two images at a time. Today, they went one better; if you want to run DALL-E at all you have to pony up another subscription, basically doubling what you already pay.
I like messing around with they way they have things set up over there, but … come on, man.
That’s too bad. You can still get 5 free Dall-E images a day at Night Cafe, though.
Meanwhile, some of my favorite Halloweenish creations with SD and DE2:
Suitably spooky:
This was my entry for the Halloween contest:
Stable seems to lean quite a bit into the Fallout universe:
DeviantArt has produced their own AI image generator. DeviantArt members can run five prompts (of three images each) for free with paid membership required after that.
It isn’t especially good, from my five tries.
Here is “a cat riding a shark” from their AI:
For comparison, samples from Stable Diffusion
And Dall-E 2
Plus just out of curiosity one try with the old Night Cafe “coherent” renderer
Perhaps of interest: here’s a series of three free comic books featuring AI-generated illustrations (although I suspect some additional Post-Processing) with a broadly Lovecraftian/Kaiju bent. See also this article.
If anyone hasn’t visited in a while, NightCafe is now doing daily challenges. Voting on the challenges gets you an additional 2 credits.
I haven’t cracked the top 20% yet,but its kinda fun.
My Ai image generator of choice is Midjourney. I’ve been very active on it lately. My vibe is Afrofuturism, decopunk, and rendering portraits in the styles/cameras/lenses of different photographers.
Link to some of my recent renders:
I’m having a lot of fun with that too. I’ve cracked the top 20% a couple of times and top 5% once, but I’ve been on a dry streak lately. There is some really impressive work winning top honors and I’m not even really sure how they do it. When I come up with something impressive it’s usually sheer luck.
I like to use these things in a way that is slightly outside of the recommendations. I asked Stable Diffusion for ‘A short poem about cheese’ - I got this (actual image further below):
Ched
wwed ie bed ounrice. Ll
these too, lec thte ichese,
tnia be-he.
lou luee tar tnie liul se,
in tu, s cled woina ii.
cu ia lestointe in. uo prese
to wieun aj cheed berx.
When I asked for a haiku about autumn, I got:
Lutem folo fim
is fake fooun tom fatus fa
gofil meases fats
Stable Diffusion 2.0 is out:
Haven’t tried it yet.
