AI-Generated

some

more

some

So, a bit of a revolution in Stable Diffusion has dropped recently, and I've been playing around with it a bit. It's called ControlNet, essentially what it is is multiple ways to keep composition in img2img. Which, before, you could only kind of do with denoising strength tweaking. ControlNet gives you options in regards to how the composition is maintained.

From creating an outline for the AI to sort of follow, to creating depth maps (Kind of like a template stencil.), to creating a rough idea of the pose of a composition that can then translate to the generation.

Here is a video on what it does and how to install it: hxxps://www.youtube.com/watch?v=OxFcIv8Gq8o The channel in general is a good way to get into using Stable Diffusion.

Here are some of my experiments. Initially I was fiddling with using Gmod and anime models for img2img, but with controlnet I'm actually getting outputs worth posting. Naturally this opens up a wide breadth of options for both artists and the layman. Doesn't hurt that it works with ANY standard Stable Diffusion models.

Girl in redish

some

davidboy:
more

Holy fuck! That drawing of the two cum covered lolis kissing is hot, too bad they're apparently sharing an arm...

(Near the bottom of page 8 by davidboy )

some

Usually I have some tidbits on AI generated goings on. This time it's just me messing around with bloodorangemix model. Good, but so one note as is the usual on NAI derived models.

johnsmithlahuser:
davidboy:
more

Holy fuck! That drawing of the two cum covered lolis kissing is hot, too bad they're apparently sharing an arm...

(Near the bottom of page 8 by davidboy )

*Cracks knuckles* Let's see what I can do.

Messing around a bit more with grapefruit hentai model. This is v31, and there are a few more versions that have been released since then... But I'm not made of HD space dagnabbit!

If I'm going to be honest, an AI generated board would be a nice addition. It would keep AI generated stuff segregated a bit from traditional art, people can make requests or create threads with themes. I dunno, I think it would be a nice addition rather than lumping it into LAH artists.

Anyway, pics.

mix

Red hair

more

Whats the best prompts from your experience are better for stable diff. I'm having mixed results with the prompts i'm using.

meric103:
Whats the best prompts from your experience are better for stable diff. I'm having mixed results with the prompts i'm using.

It depends a lot on the model, but there are a few things to keep in mind when making your prompts:

1). A lot of the anime styled models use Danbooru styled training, therefore you can use those for prompting.

e.g. 1girl, loli, nude, blonde hair, short hair

2). The earlier in the prompt something is, the more 'weight' it is going to have.

e.g. loli, 1girl, swimsuit, black hair, twintails
1girl, swimsuit, black hair, twintails, loli

The second prompt will have less emphasis on loli than the first unless you manually add more emphases. (I.E. (loli:1.25) )

3). Different models require different levels of negative prompting. Something like AnythingV3 requires very little to almost no prompting to get very good results. But something like WaifuDiffusion needs at least some negative prompting to get acceptable results.

In general, the more negative prompting you use, the less creativity you're giving to the AI. Hence why you can spot NovelAI and models trained on it a mile away since it has such a distinct and consistent style, for better or worse.

A simple negative prompt I use is:

bad anatomy, extra legs, extra arms, poorly drawn hands, poorly drawn feet, disfigured, out of frame, tiling, bad art, deformed, mutated

Which seems to work pretty well on both Booru tag styled prompting and 'Real language' prompting.

4). When downloading models, especially from places like Civitai, see if there's any documentation on it. For the most part, people who create models will give some semblance of what the models prompting should start with, or if there's a recommended negative prompting.

An example being the leaked NovelAi model benefited greatly from having high-res and masterpiece at the very start of prompts.

5.) Sometimes going off the beaten path can yield cool and interesting results. I've found that using words the model likely does not know or are complete nonsense can actually be used to make semi-consistent characters in AnythingV3. By that same vein, using 'real wording' prompts like you would use in the basic Stable Diffusion models in models trained on Booru tags will still work and can yield interesting results as well.

Despite what the crybabies on Twitter would have you believe, prompting is more of an art than a science. at least in my humble opinion having used it for a decent amount of time.

Anyway, learned that HentaiDiffusion is still getting new versions and as of this post is up to version 22. Haven't really deep dived into it, but here are some of my messing around.