NovelAI and Stable Diffusion

I'm surprised no one has even thought to bring up these two AI Art things up in the forums at all. I mean, with these, literally any kind of supergirl, with the right prompts, can get art.

www.superwomenmania.com/index.php/forum?...1893&start=400#75578

I've played a bit, but I'm not very good at it. This from "a buxom cyberpunk superheroine with a narrow waist, flying high over Manhattan".
 

I've been playing with Stable diffusions free demo a lot, and the paid DreamStudio, attached are the best I've generated so far, and you should be able to guess which are from the free trial and which from the paid version (it's much better with faces in the paid version). After nearly 4,000 trials with both paid and free my opinion is this;

• The technology has HUGE potential for creating superheros and villains alike
• It's amazingly creative if you ask it to do an art drawing rather than a photo
• However it really struggles with faces, and hands, even in the paid version
• If you want a photo and not a drawing, including "latex costume" (assuming you're after a supergirl) seems to work 95% of the time 

Currently I think the failure rate with images is too high, even for the DreamStudio paid option (which is like 1200 basic pictures for 10 USD), unless you're explicitly after a single image type and know what you're asking it. This would be better marketed as an application for purchase, or an annual license.

But in summary, it's really cool.

Edit 2: please find below an example of the kind of text you can enter to get a reasonable art image;
korean supergirl, portrait, long hair, breathtaking beautiful amazing awesome awesome, punk supergirl, long hair, fantastic, beautiful, HD HQ art by, greg rutkowski, wlop, zdzislaw beksinski, volumetric light, ambient light, detailed, beautiful face, punk, 4k image, 8k drawing, ultra high details, HD wallpaper, concept art, intimidating

And this is an example to generate a photo;
beautiful Korean supergirl, black latex, face portrait, face photo, 4k, ultra high details, 8k photo, detailed photo, staring, stunning, breathtaking, awesome, beautiful face, black latex supergirl costume, 4k photo, face photo, 8k details, HD wallpaper

The order of the words (first to last) sets the priority order in image generation. You can add a numerical weighting as well but I've found it effective just to loop multiple phrases to ensure you get a good quality image. It's rather trial and error. 

Amazing new age of AI art.  The faces and eyes are a tell of AI.  
Here is Kweethul on DA.

 

For some reason, I really struggle to get quality images out of Stable Diffusion. However, it can do some pretty imaginative stuff. I have some Supergirl costume with Native American themes that came out very interesting. My platform of choice has been Midjourney. I can get a lot better quality images out of it.

I have a whole gallery of images from Stable Diffusion and Midjourney on my DA page below:

www.deviantart.com/alexapappas

Interestingly, I really struggle to get more fantastical action scenes out of both products. For example, I can't get my characters to bend steel pipes, lift cars, or shoot heat vision out of their eyes. Even getting their eyes to glow red or be in a flying pose is a struggle. So, I would say that AI is good to get a base image that can be further enhanced or corrected in Photoshop. Funny enough, if you want to create a profile picture/portrait of your character from a Dungeons and Dragons or other RPG game, AI can be pretty good. AI image generation is very much a roll of the dice. I have gotten near perfect images on the first try and other times created dozens on images before I get something close to what I am looking for.

So trying to have the AI generator do some kind of action picture is not a good option?

marcoasalazarm2 wrote:

So trying to have the AI generator do some kind of action picture is not a good option?
 

In my limited experience, yes. Check out the two websites below. Lexica allows you to search on Stable Diffusion images. Midjourney keeps a public gallery of all their images created. You can search by keywords. There are some amazing profile pictures and portraits, imaginative landscapes, and other amazing pieces of work. However, very very very limited action pictures.

lexica.art/

midjourney.com
 

Also, the AI uses existing images to interpret its own creation on. So, more popular the character, the closer the AI will come to imitating the costume. I have done experiments with that.

Superman and Supergirl are rock solid on costumes more often than not, and can have some interesting alterations. Maybe 80% of the time correct.

Superwoman is a bit iffy. Sometimes the AI gets it really good; other times it is completely off. Maybe more 50/50.

Power Girl is a complete mystery to the AI. The images that come out looking like a superhero outfit more often look like something closer to Supergirl than anything else.