AI-Generated

That's a bummer, went through all that work getting Python functioning for all of this and in the end it tells me my drivers are to old to do this...

mix

more

some

Man, what models are you guys using? Custom, a mix, trained? You've all completely eclipsed what I can do.

Personally I like the results from Waifu Diffusion 1.3. NovelAIs model is really good, it's the only one I've found thus far that can actually do tanlines from Txt2img, but it's VERY 'one-note' art style-wise.

Recently someone has released a further trained model based on the leaked NAI one called the 'Anything' model, haven't tried it myself. Admittedly I'm kind of reticent to jump into a model that came from China. There have been a number of other more lewd models released thus far, but I haven't really done much of a deep dive into them. A number of development in training optimizations have been quite big as well, Dreamstudio being able to be run natively on pretty low Vram, Aesthetic embeddings, I even hear tell of a theoretical (?) method that could allow training with just one image!

One of the more fun things though is an extension add-on for A1111 that lets you use Stable Diffusion from within Krita (Kind of a photoshop/GIMP thing) which makes it amazing for in-painting things.

*Ahem* Anyway, Waifu 1.3 and the leaked NovelAI model are really easy jumping off points, since you just use a bunch of Danbooru styled tags in your prompt it allows you to get really good results without more esoteric prompt crafting. Personally I've really been into doing Img2Img things as of late, but nothing really 'on theme' for here. Maybe I'll delve into some of the models I've acquired and post my findings here later if I get the chutzpah. Remember .ckpt's are not inherently safe, so exercise appropriate caution.

Out of curiosity, what model/method are you using?

mix

mix

more

mix

Found a nice artist that make really good AI generated lolis. By far the best looking i saw yet.

hxxps://www.pixiv.net/en/users/88171380

more

So... Yeah, can confirm that AnythingV3 model for stable diffusion is pretty top tier. Apparently it is a continued training of the leaked NovelAI. Very good stuff with pretty minimal prompting. None of the example pictures here had any negative prompting at all.

mix

Any spaces where a lot of this stuff is getting posted?