Ultra Maiden

Version 2 is a huge update, and was made from an order of magnitude more data than version 1. I was actually planning to move on from SD, but it turns out I don't have enough compute to work comfortably with XL. So I'm stuck polishing SD, at least until GPU's are affordable again.

But when it comes down to it I think I'll actually be able to push good ole SD to a quality level where I don't even care about moving to XL!

...

Ultra Maiden 2.0 is a sexy realistic model that still needs some work before I'd call it perfect. It nails anatomy most of the time, but occasionally it spits out something that looks like a collab between Escher and Picasso. And it is currently way too nude biased. Hopefully next version will be better.

That said, it's already a very good model, which is why I'm posting it in this early state. It makes good images for very short prompts, the example images were made with just a single word in the prompt and nothing in the negative. It's actually so well tuned it produces near perfect images for most seeds without even having a prompt at all!

And it is pretty fast. I'm using 12 steps DPM++ 2M with AYS for the examples. Good enough for me. Feel free to use more steps if you have the compute or the patience.

The focus is primarily on sexy maidens, with or without clothes, and secondarily on environments to put them in, indoor and outdoor. I also stuff a lot of fantasy/scifi/horror in there, to make it possible to do cool stuff too, and not just sexy stuff.

But everything else is outside of the scope for the model and might perform poorly, due to lack of training. SD has a tiny neural net, so I think it's best to stay focused and not try to make it do everything.

When it comes to the fantasy/scifi/horror stuff, the model drifts away slightly from the near perfect photorealism it normally has. This is something I hope to be able to fix eventually, but for now the best you can hope to get from fantasy related prompts is something like this:

Prompt "fantasy":

Prompt "cleric":

Prompt "merchant":

It doesn't look bad, but it's drifting a bit towards semi-realistic, which isn't what I want.

And when it comes to things that aren't trained, that includes all kinds of things, like cars for instance. The model has old image data for cars in it, but the neural net has been modified a lot since it last saw an image of a car, so this old data is partly compromised. Which means proportions may be slightly off for cars, or image quality just suffers a bit in general.

Prompt "destroyed":

Wonky proportions if you ask me... So this model can't do everything equally well. It's really just meant to do sexy maidens. If that's what you want to do you might like it!

PS

I posted 20 example images (plus 4 on the model description page itself). Civitai only scanned 5 of them, the rest got stuck in limbo for ten hours. This site is buggy as hell. But we still love it!