Digital art creator algorithm website

Renders using a custom Lora I made based on Playboy cartoonist Doug Sneyd*. First one is using ElldrethsRetro model, the other two using… umm… Deliberate? I think? …with slightly different strengths. As you can see, all three ignored her brown pants.

Dude is going to be let down when it sees that AI got her nose and eyes totally wrong.

*None of the AIs or models I’ve used have Sneyd in them so I went and made my own style for it.

This thread is nearly a year old. It is pretty funny looking back at what impressed us a year ago.

This shit is moving fast.

Something I have noticed when generating images of Charles Darwin with the new prompt settings–it seems like SD may have conflated images of Darwin with images of Neanderthals or similar (a kind of image that does seems fairly likely to be sometimes be linked to a mention of Darwin).

The first image is especially apey.

I watched a video on YouTube on syntax of Stable Diffusion. I’m not sure how according to Hoyle it is, but I’ve been utilizing it lately and I’ve been coming up with some interesting results.

As it was explained in the video, the prompts should come in this order:

  • medium, (Subject), descriptor nouns, descriptor adjectives, location, rendering cues, artists

The subject should be simple (A Beautiful Woman, A Crumbling City, A Cute Puppy).
Descriptor nouns to indicate what you want to see, so if you want a full body render put in descriptors like shoulders AND feet, so it will try to fill in everything in between.
Descriptor adjectives should apply to the entire render; if you try “blue dress” you might get a lot of other things that are blue.
Location is indicated by, “in.” This should try to fill out the background.
Rendering cues, for lack of a better term, are things like photographic techniques or rendering techniques (telephoto or Unreal Engine).
Artists are indicate by, “by” and can be linked with “and,” so the prompt would be, “by This Guy and That Guy” for a mix of styles.

Commas delineate separate concepts and parentheses indicate higher importance. Also, according to the video, the algorithm assigns more importance to prompts at the beginning and the end of the line; things in the middle can be overlooked if there’s a lot going on.

The negative prompts are kind of the same story, but in reverse. I usually just snag something that keeps anatomy from going crazy.

A full on sample (no negative prompts):

Oil on canvas renaissance painting, (Charles Dawin with a Monkey), beard, bolo tie, sepia, soft lighting, in an ancient library, Octane Render, ray tracing, by Jack Kirby and Leonardo DiVinci

… which I just made up on the fly, so I ported it over to Playground AI, and this is what I got:

Imgur

You should tag that crotch demon as NSFW :rofl:

I re-ran the red-haired woman prompt with this syntax and I did get some brown pants but also 50% of the six-image grid had brown hair as well. Aside from that, they were largely similar to my previous attempts allowing for general seed randomness.

Weirdly, I saw that I forgot the “mid 30s” part and added it in. Four redheads and two brunettes but the pants now range from brown to khaki to blue jeans to purple to olive. Probably not meaningful (again, random seeds) but strange anyway.

https://creator.nightcafe.studio/creation/C7byKOjrFXas9gRXkPyD

So I made a series of images at Playground AI with a prompt modified from that example to fit with my artist mods, and got somebody that looked nothing like Charles Darwin. I stripped out everything superfluous from the prompt and got Darwins. I then a few minutes later went back to make a few more to balance the number of weird ones, and suddenly the streamlined prompt was making weird (well, weirder) images too. Bafflement followed by a close inspection showed that I had mostly been making images of Charles Dawin (no “r”), including in the above Night Cafe image. (And, eta, I see I copy/pasted the typo from Jack_Batty, who also unbeknownst to him made Charles Dawin.) :rofl:

Oh shit. I didn’t even notice.

Okay, modifying Jack’s prompt, I got

(Charles Darwin with a Monkey), beard, bolo tie, sepia, soft lighting, by Dan Witz and Mark Ryden and Margaret Keane and Pino Daeni

And the stripped down version is

(Charles Darwin with a Monkey), by Dan Witz and Mark Ryden and Margaret Keane and Pino Daeni

This gallery has 12 shots of the complex version, 12 of the simple version, and 12 each of the “Charles Dawin” images I was making before catching on:

(The prompts use a negative setting for “frame” because
Mark Ryden art often includes them.)

Is it just me, or does SD 1.5 seem to be getting better with hands? Not perfect, and not without some hideous monstrosities, but it seems to be producing what are merely mildly defeformed five-fingered hands much more often than it used to. Look at my Darwin and Dawin images, there are quite a few not-all-that-bad hands in there. Think there is refinement going on?

Here’s a meandering tale about meandering renders.

One of my usual test subjects when working on new settings is “Matilda and Leon from The Professional”. Stable Diffusion actually has a pretty clear idea about the basics of the characters, tending to draw a bearded man in glasses and a hat and a young dark-haired girl–sometimes they really do resemble Jean Reno and Natalie Portman. But of course there are also plenty of times SD goes wildly different. So I’m playing around with the characters using Jack_Betty’s recent prompt settings, then as always wander through changes, some of which yield results worth further editing.

One of several renders of “(Matilda and Leon from The Professional in a post-apocalyptic landscape), medium shot | by Dan Witz, Mark Ryden, Margaret Keane, Pino Daeni” was, very weirdly, a young woman standing near a severed head with some gore attached (which came closer to looking like a krasue than any of my multiple deliberate attempts.)

I thought the 384x576 image merited higher resolution so I switched from Night Cafe to Playground and started modifying it in 512x768, modifying the prompt to “girl standing beside severed head and guts”. SD understood what I wanted, and good results in that made me want higher resolution. With their highest free resolution being 1024x1024, I needed to infill the sides with something.

I don’t know if any of you have worked at 1024x1024, but that resolution is far enough outside SD 1.5’s training that the results are weird when it is having to create novel material (as opposed to editing existing image areas, which it does fine). If you ask for a person, for example, it tends to produce multiple intertwined variations of the person. It is worth trying with free credits, but too often random gibberish to be worth spending anything on. In this case I wanted simply extensions of the background. I got images with weirdness, including one with creepy gory weirdness. So I went on a tangent there and infilled the “good” center with more of the weirdness, then expanded outward into a weird gory panorama. (I had to choose between more than one good option for each segment, and left out the ones with nudity.)

So after the creepy gory weird tangent, back to the merely morbid one. I got suitable sides by infilling at 512x512, enlarging that to 1024x1024, and patching in the higher resolution central figure. I did a number of variations, including various facial expressions. The somber is pretty good. Some had open-mouthed smiles, but none of those worked for me. But one had a subtle, very Mona Lisa grin that I thought was terrific. She’s here, both with the bright, smiling head that was original to that render and with a more subtle one pasted in.

(I always feel a little self-conscious about dominating this thread with examples, but hopefully there are useful for some. I’d like to see other people post more, though.)

Months back in the thread I posted an image of a family picnic by Wane Barlowe where one of the images ended up with people seeming to be feasting on human heads sitting on the table. A day or two ago, I attempted to recreate that intentionally at Night Cafe, asking for heads on a table in different ways (as an aside, NC blocks “decapitated” but not “severed”). One of the new images was three (and a half) heads that immediately brought to mind the severed heads in jars from Futurama, and also maybe three stern judges of some type.

So today I trimmed the sides and reframed the image to be near the top of a 512x768 frame and moved over to Playground. I was wondering how much effort it was going to take to get it to attach bodies to the heads. It turned out to be incredibly easy, all I had to do was infill mask everything but the faces and describe it as simply as “people standing” plus my artist prompts. (The hardest part turned out to be making them not be naked or in skin-tight clothes, and it was never clear on the genders of each person. So I added various clothing description mods, including Victorian, Samurai armor, and loose casual.)

I then went on to try turning them into heads in jars on shelves, with some success. (Thanks to the aspect ratio, SD adds a second row of newly-invented heads.) I still want to work with those more and add more pre-made heads, but I set that aside for later.

I then set to work on one of the grisly gristly heads featured in an earlier post, and was wildly successful in making it look good on a shelf and in a jar. The jar in particular took quite a few rounds of editing to get it as good as it is. The sort of snowglobe/paperweight thing popped up mostly on its own during the variant generation process.

So then I headed back to Night Cafe to try some groups of “severed heads in mason jars sitting on shelves” images without a start image. A couple of them produced something like heads on shelves (with or without jars) but most of them produced small, head-free jars packed together. Some of which looked useful for later editing, though. One in particular had a red-headed person facing away from the “camera” in front of shelves.

I took that one over to Playground and described it as “woman with short red hair facing away looking at wooden shelves filled with colorful glass jars” and later as “woman with short red hair standing in front of shelves filled with colorful glass jars” (both with “Dan Witz, Mark Ryden, Margaret Keane, Pino Daeni” artist modifiers). These samples include more than one generation of evolution and noise levels ranging from little to lots. The one most similar to the original image is the third one in the middle row. I really like some of these and may later go to some effort to tweak the jars in at least one of the images so they all look good even on closer scrutiny.

(The images today range from funny to hideous to beautiful. Let’s just say my tastes are eclectic.)

They need to rename this thing “ReallyDamnFarInTheJourney”.

Adobe sticks its toe in the pool:

https://www.fastcompany.com/90868402/adobe-firefly-generative-ai-photoshop-express-illustrator

While recently playing with a couple of new artist names (Mab Graves, Carne Griffiths) and rolling them into an earlier prompt I discovered that I could make some nice alien (but still usually more or less humanoid) aliens, and body horror images including a look of flayed skin. I’ve only recently started posting on reddit, and this is my first use of the fancy-schmancy slideshow feature, where I posted 72 sample images from the new prompt family, so rather than reinvent the wheel I’m posting the link to that.

One point not made in the reddit post is that combination of artists tends to toss out nudes unprompted. And they tend to slip through the content filters at NC and Playground but, frustratingly, whenever I try to edit one then they refuse to work with it. Here is a (NSFW) sample of those. The two leftmost images are especially worth further editing, but nope. (Playground was willing to work with the top right one, though, and I ended up with a slightly expanded framed version that I like.)

Bing Chat now has ‘Image Creator’. I don’t know which model it’s running right now.

I had been thinking about book covers soon being all AI generated, so I tried: “An image of an attractive female astronaut in a battered space suit after a long day’s work on an asteroid.”

Here’s what I got:

Dall-E. I hear that it is an unreleased new version.

Here is a guide, from Tom. (I haven’t tried it yet).

Okay, I’m jealous of someone.