I can only speak for Nightcafe, but other sites are similar.

You simply go to https://creator.nightcafe.studio/?ru=Dark_AiArt (not sure if you have to create an account)
Then you can use a creation method. I always use Stable, it is the newest and most realistic algorithm so far.
On the next site you have to type in your prompt, for example: "Woman with dark hair and black wings"
After that you can choose a style and create the image.
Instead of using a style preset, you could also choose the advanced mode where you can add modifiers to your prompt. There are also others options there, like using a starter image, higher resolution, and so on.

Yeah, I've been using Dream, and the process is pretty much the same as DaRkStAr described. You put in your prompt, then select a style. Dream has a bunch of different styles, though I have had the most luck with "HDR," "Realistic," and "Comic." A lot of the others tend to turn out pretty surreal, which is fine if you're into that kind of thing, but in trying to produce character art, as I have been, the more realistic styles are my go-to. Then you just have it create it and see what you get. If you don't like it, there's a handy "Generate Again" button that lets you use all the same settings, or you can go back and start from scratch. I usually copy my prompt so I can just quickly paste it back in to try a different style.

As DaRkStAr mentioned with Nightcafe, you can also upload a base image and let the system base their image off of that. I have had mixed results with this, but occasionally it works pretty well. There's even a slider that lets you choose how much the original image affects the result. You can also download the picture they produce. They almost always have a border around them, but in terms of what I've posted here, I just used an editor to crop those out.

It looks like there have been some improvements in the AI Art program, so I thought I would give it another go.

Daneera

Kahr-ret-Taris

Orida Vise

I was never able to get any good Orida pictures the first time around. This time, things went a bit better for the foul-mouth Kithkin.

I like to think that Kahr-Ret-Taris just shows up to high school looking that way. Shirtless with a backpack. Clearly in his mid 20's, but we just roll with it.

It sort of fits with canon in a weird way, with Kahr spending nearly a century off-plane after his planeswalker ascension. Maybe he spend a few years trapsing around Europe (or, I don't know, Foraine?) between his junior and senior years.

Scar

The AI had a lot of problems creating centaurs (although it had far more trouble trying to create a drider, which is why there are no pictures of Lady Nasina, unfortunately). It also had issues making scar properly scarred, so I did my best. Here are the results.

The Tower of Tears

I've wanted to do a depiction of the infamous Tower of Tears, from Thorneau, for a long time now. None of these really capture it for me, and the AI had all kinds of trouble understanding that I wanted water cascading down the walls from the top, but these at least got to an approximation of what I was looking for.

I'm curious what service you're using for these generations (I'm assuming a service as you need a decent PC and some knowledge to set up a local model)? I've ended up being gifted a NovelAI subscription (which I am friends with a few of the people working on that) and have some small understanding of how to tune them in now, not that I could really help with fine-tuning these things of yours.

Well, I do know that scars are difficult if not impossible -- at least without some editing skills and feeding images back into the AI to get the perfect look -- because there's so little character art out there with scars already. One model I was messing with (a furry model) visually confused scars with musculature, so it was nigh impossible to get a character that was broken instead of ripped.

I do have some small desire to get a proper version of On-W, though, but due to my color choices on the original image I made like 8 years ago I'll need to do some editing before it can read the image properly. Maybe I'll share those here when I manage a good output.

I'm still using Dream (by WOMBO, whatever that is). It is free to use, although several of their features are behind a paywall. You can get their "Premium" subscription, which I think is something like 90 bucks or something for a year, but I don't have that. I've just been using the free version. I then edit them with a (really) old photo editing program. My photo editing skills are not particularly advanced, but to demonstrate, here is the oriignal of the picture with Daneera and her apparent daughter:




Yeah, and obviously I'm posting the best of the ones I've managed to create. Some of them are downright crazy-bad. And some of my characters are easier to make than others. Humans have gotten fairly good, which is why I have a lot of Daneera and Kahr. Races like Foxfolk (Antine), Houndfolk (Donagut), Rhox (Dorn) and others have proven much more difficult to get the AI to figure out. I cannot get it to give me ANYTHING approximating a good Lady Nasina at all. But the AI is getting better, it seems, and I will keep trying, at least occasionally.


I certainly wish you luck! As I mentioned, the more abnormal the description, the less likely it is the AI will give you what you are looking for, but somethings you might just get lucky and get something that either works, or is close enough that you can edit it.

So, since asking questions here I managed to fire up my own natively-run stable diffusion, using the Anything v.3.0 model card (I really should update to 4 or 5, but I have a project that started in 3 so I need to start it over or keep 3). Normally I'm quite fine with the anime style, but with what I was trying here I'm not sure it worked as well as I;d like

Very cool, Tevesh! I feel like you have a lot more continuity between the designs than my scatter-shot approach tends to produce. I particularly like the maniacal expression and pose on the "Ellia the Endbringer" one.

Thanks for posting!

Yeah, one of the nice things about the Stable Diffusion UI I work with is that I can set a random seed, or if I let it roll, it will tell me which seed was used for which image, so if I like one I can make variants on the same base canvass. The ability to do that was a big part of what I was looking for at first as I have ambitions to do a visual novel, and I need to be able to have the same character with multiple expressions, so having someone still look like a given person when I add "Smiling" or then change it to "Crying" or "Angry" is important. It's not been perfect, there are usually subtle changes in the details of their costume, but it is fairly good at that.

Yeah, the Dream AI is by my understanding way behind in terms of image generation, though from what I recall I was more concerned by their privacy policies the like 2 years ago or so when it was spreading around (like getting images of Hitler to sing to Queen songs). Of course, I can't speak for their UI but I know a majority are doing like Tevish and setting up Stable Diffusion versions to run on their GPUs. There's been a lot of developments in the last year that means the AIs are a lot less demanding (and a leak of the latest model so a decent chunk of people are using that through morally- and legally-grey avenues), though I'm not one to really talk to about requirements. I've even heard of this new thing called LORAs which are, like, really compressed training data.

I could maybe gift you a month of NovelAI for image generation if you're really interested in it.


Yeah, my method with most of my experiments has been to try dialing in some details by adding one tag at a time, then when I get a pretty good composition, use that seed to start making smaller changes, like increasing or decreasing the influence of certain tags (as an example, "muscular" is a *very* strong tag and usually has to be reigned in). Also, if you're not aware, there was a gigantic project someone put together about how a large collection of nounds, adjectives, adverbs, etc. affect the image, so you might be able to find better tags with greater or lesser influence if it suits you.

Annalee V'ray

Raiker Venn

The Clock Tower on Cartrevard

Considering how much trouble I had getting a good Tower of Tears, I was pleasantly surprised how good some of the Clock Tower pieces came out. I suppose a Clock Tower is much more common a thing than a tower in an ocean with torrents of water flowing down its walls.

Would you mind sharing your prompt for Rainer 2? I love that style.

The prompt I wrote was something like "A handsome man with black hair and a goatee wearing a nice, black suit and holding an ornate cane," but more importantly, I used the Art Style "Ink v2" on Dream, which I imagine is what you are looking for.

Using that same prompt and style, it came up with this one just now:



EDIT: I've been testing it, and it seems like "Ink v2" will pretty much always give you a black-and-white image unless you use a color "Input Image," so if you want some flesh tones in there, you may want to use a different style first, save the image, and then use it as the Input Image, although I could be wrong.

EDIT 2: Nope, I'm wrong again. I think it was because I mentioned the color black a couple of times that it kept making Raiker's pictures black-and-white. I tried a prompt that mentioned specific colors and it included them, so you can get color in the "Ink v2" style without using an Input Image.