[ home / bans / all ] [ qa / jp ] [ spg ] [ f / ec ] [ b / poll ] [ tv / bann ] [ toggle-new / tab ]

/ec/ - エッチ/Cute

Cute Posts

New Reply

Whitelist Token
Password (For file deletion.)
Markup tags exist for bold, itallics, header, spoiler etc. as listed in " [options] > View Formatting "

[Return] [Bottom] [Catalog]

 File:00916-2435206029.png (1.4 MB,864x1152)


As I've blogged about in various threads, I've been doing AI stable diffusion model merging and LORA creation and stuff for a while now. I like to think that I'm pretty good at it.
Well, I think I'm almost fully satisfied with the general model, but it still needs some testing. Alas, I have only one brain so I'm limited to the stuff that I can come up with. Since this is something other people will eventually be using, I'd like people to give me prompts to test out. Nudity and other stuff is fine, but I'll probably post the more hardcore stuff on /megu/.
NOTE: This is the general model that is a poor at recognizing character tags. I'll figure out a way around it, but for now this is more general. Although I guess we could test which characters are recognized, but it's not very many of them. Reimu and other super popular ones work. (Kuon works because I trained a LORA for her half a year ago)

Please give me prompts to test out, using various poses and tags. It's a hybrid model that uses booru and probably e621 (furry) tags, so if you're familiar with either of those it'd be good. It is a bit of a "jack of all trades, master of none" situation though apart from how beautiful it is. It can make use of standard real life Stable Diffusion tags, too, like "dinosaur on a motorcycle" or something, but I don't usually use those so I'm not how good they'd be.

As an example, mermaid Kuon here is:
1girl, seashell bikini, rainbow hair, gradient hair, (mermaid:1.2), gold armor, very long hair, jewelry, crown, arms out, fish tail, lying, floating, scales, marine, coral reef, (plant humanoid), petals, underwater, smile, :d

Useful booru tag info:
https://danbooru.donmai.us/wiki_pages/tag_group%3Ahands (don't count on these being good)

E621 has its own set of tags and I'm not very familiar with them, but here's their wiki:


 File:00009-1475836378.png (1.3 MB,864x1152)

Hmm, but the Reimu recognition isn't 100% there it would seem


Hmm, I don't have a particular scene in mind, but how does the Ume LORA look?


 File:grid-0005.png (4.93 MB,1728x2304)

Alright then, with a simple "1girl, smile, upper body" prompt with the Ume LORA active.
Hmm, I think style LORAs might be something I need to address, so I'll have to keep that in mind. It looks a bit more grainy and noisy here than it should, but the LORA block weight thing can probably fix that up.


 File:grid-0006.png (4.5 MB,1728x2304)

from IRC:

masterpiece, award-winning, absurdres, 8k wallpaper, immaculate quality, superbly detailed, peerless rendering, stunningly beautiful, divinely inspired, big butt


 File:grid-0012.png (4.9 MB,1728x2304)

from IRC
loli, plump, flat chest, thick thighs, wide hips, belly, reverse bunnysuit, dark-skinned female, red eyes, white hair, huge ass


 File:grid-0015.png (4.71 MB,1728x2304)

With emphasis on loli and flat chest and removed 'huge ass'


 File:grid-0017.png (4.84 MB,1728x2304)

(loli:1.2), (chibi), plump, (flat chest:1.3), thick thighs, wide hips, belly, reverse bunnysuit, dark-skinned female, red eyes, white hair, (huge ass:0.8)


 File:grid-0020.png (4.76 MB,1728x2304)

with (dark skin:1.3)


loli, chair, pink, no legs, white gloves, eyes, frown, sad, summer, 1girl


 File:grid-0037.png (4.9 MB,1728x2304)

Mmm, "no legs" really isn't going to work because it will read "legs" and put legs in there. Such is the problem with AI.
Also maybe I should start asking people how smooth or "painted" they want stuff to look, from a scale of 1 to 3. Anyway, here is 2.


 File:grid-0038.png (4.61 MB,1728x2304)

I think you were probably after some sad amputation thing, but I doubt that's possible. If you want just an upper body look, then "upper body" works.
Here is the "smoothest" look with the least amount of paint strokes, but really it will need more booru tags to begin to look flatter, color-wise.


i was thinking something more like an amputee but thought maybe it would do an upper body instead
still the result is pretty darn close to what i had in my head, only thing missing from my imagination of it, is a gently breeze with some petals but i didn't specify that so not surprised there


1girl, loli, pink hair, dark hair, very long hair, camisole, bare shoulders, bare feet, pout


 File:grid-0189.png (5.03 MB,1728x2304)

Bleh, I was messing with checkpoints and made the server crash or someting. Well, running locally...
I'm going to assume anyone telling me to prompt loli will want me to add 'chibi' and 'flat chest' so it produces a loli body, so I added it.
Pout is something that never works. There's a LORA for it, but I haven't messed with it on the new model yet and I don't want to tinker with it as I'm about to fall asleep.


Good images and good to know.


hmm well, I guess on closer inspection the bodies aren't very loli-like. Might need better tagging for this, or to use artist LORAs...


fingers, toes, hair, joints, patterns


 File:grid-0205.png (4.87 MB,1728x2304)

(without any blatant booru tags it tends to look more western)


 File:xyz_grid-0010-1170495876.png (17.72 MB,6912x3058)

An example of steering the art direction by choice of prompt.
The left is danbooru,
second one is the "natural language" of stable diffusion
third and fourth are e621 tagging with slightly different tags

Basically the more booru tags you use the closer to the expected 2D booru image you'll get and likewise for the furry direction


big penis, wrinkly penis, (((dick vein))), smelly ballsack, sounding


I was wondering how long it would be until someone did something disgusting...
Yeah, I think I'll post that one on /secret/ in the "wonders of AI" thread with a spoiler


poor forgotten /megu/


 File:00221-2554430168.png (1.2 MB,864x1152)

The existing /megu/ thread is more of a nsfw blog of mine so I'd need to make a new thread for fulfilling lewd AI requests. I'd want a better OP image than that.
Well, I guess I could make it later but so far people are not nearly as interested in prompting pornographic stuff as I am so ehhhh


 File:Fmw9DYpaEAI4WHP.jpg (197.3 KB,967x1598)

can you do CLIP interrogation or whatever it was called when you feed in an image and get a likely prompt out?


 File:grid-0056.png (4.73 MB,1728x2304)

If you're telling me to take that and turn it into a prompt, sure:

1girl, paw shoes, paw gloves, solo, gloves, animal hands, animal ears, tail, on back, cat ears, lying, ahoge, loli, navel, blush, brown eyes, cat tail, twintails, open mouth, bangs, shoes, long hair, bed sheet, hair between eyes, animal ear fluff, wavy mouth, light brown hair, @_@, flat chest, sweat, spread legs, swimsuit, bikini, panties, underwear, arms up, cameltoe, bare shoulders, fake animal ears

But you then need to weed out the stuff that is wrong or generic. I.E she has "paw shoes" so "shoes" is redundant and would take away from the generation, and it's not a swimsuit but "cat lingerie".

1girl, paw shoes, paw gloves, solo, on back, cat ears, lying, ahoge, loli, navel, blush, brown eyes, cat tail, twintails, open mouth, bangs, long hair, bed sheet, hair between eyes, animal ear fluff, wavy mouth, light brown hair, @_@, flat chest, sweat, spread legs, cat lingerie, arms up, cameltoe, bare shoulders, fake animal ears

Here's the result with the version of the model that is good at character recognition and booru tags


 File:grid-0057.png (4.79 MB,1728x2304)

And here's the 'regular' version. Although, hmm, it did a decent job with the pose. Just need to put more work into getting a loli result with more weights and added tags like 'chibi'


Interesting results. The set of tags is pretty accurate but the regenerated images do not capture the art style or the loli nature (look how big Mahiron's head, hands and feet are as a proportion of his body, that's missing). The @_@ tag doesn't have any discernable effect on the images either.


If you wouldn't mind doing one more CLIP interrogation, what does it say about this cat photo? >>>/qa/102927


 File:C-1688941510151.png (2.54 MB,920x1520)

Huh, those look relatively normal compared to what I get when I try to generate that image with that prompt list. I added in "Megumin" and used controlnet on the original image and each generation turns out a whole lot more painted than those.


 File:grid-0058.png (5.03 MB,2304x1728)

Well, I just plugged in what was there. I'd have to add tags like chibi or maybe big head to try and get the desired results. You expect too much from it if you expect it to replicate an art style with a list of tags. And yeah, good luck getting eye stuff to work. I think there are some LORAs for them, though.

The SD CLIP says:
a cat sitting in the grass next to a fence and bushes with green leaves on it and a fence behind it, Felix Octavius Carr Darley, regal, a photocopy, naturalism


 File:grid-0060.png (4.75 MB,2304x1728)

And here is the result with the booru-based tags:
no humans, cat, animal focus, plant, whiskers

Hmm, upload it to catbox so I can see the prompt settings


Here, it's what I tried for my prompts


 File:C-1688942542686.png (2.24 MB,920x1520)

Oh yeah, and this is with the Kissumono mix


my parents actually said once that in that pic our cat looked like a king


 File:xyz_grid-0016-1651180902.png (11.42 MB,3240x3597)

There's a few things:

The sampler is different.
Euler or Euler a for smooth appearance
DPM++ 2S a for, uhh.. "medium"
Karras stuff for more painty noise

The CFG is higher (higher CFG will introduce noise for some reason)

Also I started with a slightly larger image and upscale it less. I start at 576x768 and do 1.5 as my default. I haven't done much testing with this because my GPU can't really do images that are much larger.

And also of course denoise being lower will, uhh.. well, reduce the amount of noise that is being denoised when upscaled


could you use >>10596 as a basis to create a pout prompt and then mass produce pouts?


I don't know how good it would look, but yeah. Extremes like a super pout are more difficult to do than basic stuff, but I don't have much experience with the ControlNet stuff since I find it to be too much work and it eats up so much VRAM that I'd rather generate 20 random images than 2 "traced" ones

[Return] [Top] [Catalog] [Post a Reply]
Delete Post [ ]

[ home / bans / all ] [ qa / jp ] [ spg ] [ f / ec ] [ b / poll ] [ tv / bann ] [ toggle-new / tab ]