Post Generative/AI Art You've Made

Good point, at least Stable Diffusion is pretty good at clothes. Too bad the input boxes don't resize themselves.

I wonder what other Negatives are good for getting the whole person in frame. Trying to keep the whole prompt visible.

How in the hell your results are better than mine? Working from text only, I got a string of anatomical disasters like duplicated heads, multiple arms, absurdly long necks and the like. Basically, if the AI doesn't start from an image, it can't get the body proportions right, and even then it mangles the anatomy.

you can have it exclude tags like bad_anatomy, extra_hands, extra_fingers, etc etc, and including tags like masterpiece, best_quality &c can help too.

Suuuure. You wish the negative prompts worked like this.

You are using Stable Diffusion 2.1, right?

"pale woman with dark hair, black tanktop, pierced lower lip, tattooed arms, olive cargo pants, background is industrial street at night"

Edited by Malady on Dec 25th 2022 at 3:13:17 AM

For that specific shorttank image, it was version 1.5 inpaint hosted on mage.space - maybe they have their own dataset for it that just doesn't work that well unless "additionally motivated" with a reference pic. The one on the bottom right in your pic looks closest to my intentions.

Edited by NotSoBadassLongcoat on Dec 25th 2022 at 12:30:55 PM

Regarding NSFW content: the SD 2.X model (I think, rather than the software?) has NSFW blocked by default, which I find kind of hilarious. I think it's a model-level limit, 2.X with waifu diffusion just takes booru tags without question and spits out a relevant answer, for better or for worse.

As for the above, sometimes it can take multiple attepts to get a picture with decent anatomy. Negative prompts help, but the full list recommended by the UI version is very long for a reason.

Put together a bunch of foxgirls with largely the same prompt (second half added a pose and rain). (There were 20 generated, but one wound up missing some clothes, which was awkward). Mostly things that wouldn't usually all show up in a picture together; there's like... a dozen that have maybe a third of the tags, let alone all of them. I actually think it did pretty well for the most part, I think these◊ two came out particularly well.◊

And here's 99 more. I thought it'd do 100, but I guess not (was using a plugin to make sure the seeds were different each time). It's interesting to note how similar a lot of them are; whether that's because there's not THAT much to draw from for some particular inputs is one question of interest, or because the prompt is highly specific, I'm not sure. The guidance is a bit higher than the default of 7, but at 12 it still has room to play around, and there's seed variety. Main difference is that there were 25 with no pose-related tags, 25 with sitting, 25 with hands on hips, and 25 with hands on pockets.

One thing I do find to be an interesting conclusion is that the very specific prompt (and the negative prompt) might curtail variety somewhat, but it does do pretty good at ensuring relatively decent anatomy across the board. There's the occasional example where the fingers are clear but the wrong number◊, and the usual iffiness around hands, but it seems like a good approximation of amateur art here. And it DOES stumble a bit on how multiple tails are meant to look sometimes.

Edited by RainehDaze on Dec 26th 2022 at 12:53:35 PM

You didn't specify her belly size, did you? Because that's more variable than I expected from Bare Your Midriff stuff, since that's usually fanservice-y, a.k.a thinner.

Edited by Malady on Dec 26th 2022 at 6:37:12 AM

Actually, I did, the variety is pretty surprising. As I said, I was going for a tag combo that definitely wouldn't show up in a search, though you could get pieces of it. So, the prompt was:

1girl, original, white hair, very long hair, nature, red eyes, animal ears, fox tail, fox ears, fox girl, flat chest, dark skin, plump, multiple tails, jacket, jeans, navel, crop_top | {sitting, hands on hips, hands in pockets}

and the negative prompt (pulled from here) is

Deformed, blurry, bad anatomy, disfigured, poorly drawn face, mutation, mutated, extra limb, ugly, poorly drawn hands, missing limb, blurry, floating limbs, disconnected limbs, malformed hands, blur, out of focus, long neck, long body, ((((mutated hands and fingers)))), (((out of frame)))

So, it's actually the thinner ones that are a bit of an anomaly here. The guidance value was a bit lower for the longer second batch, which may also explain the tendency of the bust size to vary (presumably because some combinations are more common than others). First set forgot to specify crop top (and honestly I'm surprised it worked since I shouldn't have put underscore in, that's deprecated; but the results show it did) hence the removed image that only had the jacket.

Edited by RainehDaze on Dec 26th 2022 at 2:39:52 PM

Felt like trying that blue-eyed white-hair girl again, by starting with the hair, and somehow got a centaur, apparently, in image 2.

And the image uploader checks file type based on extension, a bit.

Edited by Malady on Dec 26th 2022 at 8:00:50 AM

That thing haunts my nightmares

I think that thing is fucking with me, big time. I straight up copied the list of negative prompts from the wiki and still got something with two pairs of breasts and three arms.

I think someone should come up with a mechanism to generate ten sample images and teach the AI which ones failed so bad they should be rejected from the pool automatically.

What AI Art generator are you using? Different generators might wind up with extremely different results, even with the same prompt.

Could be that the guidance strength of the prompt is too low (don't know how online tools handle that), or something's confusing it in the active prompt.

Also, you've basically described the normal learning process—provided you don't put in bad pictures. The thing is, in operation, it can't do that. That sort of learning is enormously complicated and takes way more memory than simply using them does (and they already take a lot), and self-referential examination would be a radical departure in operation.

mage.space's implementation of SD 2.1 . Apparently "free" means "we just told it to browse Google Images and went out for a beer." It's good for nothing aside from inpainting fragments of already made images.

(yes, mage.space is the site's address)

Un-fucking-believable. Turns out that if I don't start dropping names in the prompt, I won't get SHIT.

Here's the third result of a new attempt on Lexperiment with mage.space, with enhanced text prompt "borrowed" from a different site.

Am I supposed to throw famous artist names at the goddamn thing to get anything reasonable?

I guess Centaur Girl was just random chance? And just figured out how to crop uploads

Anyway, I decided to take a crack at said RPG character. I'm getting that it doesn't know what piercings are, as it's been fairly obstinate about not including it.

Prompt only.

Drawing as input (image BG extended to be square so it wouldn't reshape anything). Didn't generate a background, though. This is the closest I could get◊; I reckon some sort of thing in vaguely the right colours would need to be sketched in (or pasted) behind it to get it to click.

Seems pretty okay, no specific artistic input here.

Edited by RainehDaze on Dec 26th 2022 at 4:38:59 PM

Yeah, I had to add a gradient to the background of the original drawing to even start getting backgrounds. Here's the version I recovered from the site, I can post the original edit when I get home.

What AI engine are you using? It seems to work better with tattoos.

Edited by NotSoBadassLongcoat on Dec 26th 2022 at 5:44:26 PM

Uh, it's just stable diffusion v2; I downloaded the checkpoint from their site. It is running locally, which does give a bit more control I guess.

Having pasted in something to use as a background, it does get a bit better. Terrible at integrating the two components into a scene or dealing with lighting, but eeeeh, the original input also has that sharp delineating and that's what it's working from. Reminded me of VN stuff, actually, so I got curious what happened if I told it to do anime style (quite messy, by and large). Could probably rework it into booru tags and then use WD instead if I wanted to explore that.

Ha, this is why I used a gradient and some feathering on the outline of the character, for less delineated and more abstract background.

But I do admit, I like the one with the SUV and "Cari the Pelstin" in the background.

And using Waifu Diffusion instead to generate anime-style pictures.

Are there any generators that don't require registering? I only know Craiyon that does that.

Stable Diffusion's online demo doesn't need registering.

If you have a good enough GPU† (has to be nvidia because of cuda), you can just run it locally and that also doesn't have a registration requirement.

† I'm not sure HOW good is required; the stable diffusion UI does have a low-memory mode that should work at least with a 1060, it's not really something I've thought about.

mage.space doesn't require registering, and you register only to keep your pictures saved. It doesn't require "credits" or any crap like that.