/vt/ - Virtual Youtubers

Request anchor

>>36847468
IS THAT A WHIP

>>36848366

A cat is fine too.

>>36844207
>>36844085
>>36843956
That's actually a really cool combination.

>>36848479

>>36848479
holy sex BIGGER

>>36848479
non architect ver. please.

>>36848312
I would like to see all proompters, grug and {{{{{{le ebin meme tags}}}}}} alike, try to get a spoken or inset uoh. This was the closest I got and it was an accident.

>>36848575
Also this combo was hilarious. Changing cfg makes it cycle through just about everything it can throw into "spoken" - I only have the spoken tags you see in the filename.

>>36848540
>>36848562
The duality of man.

>>36848540

>>36848480
I wish Sora would put a collar on me.

>>36838592
Here, these are the three you responded to. Take what you need.
https://files.catbox.moe/35we6l.png
https://files.catbox.moe/dk0h0c.png
https://files.catbox.moe/305hub.png

>>36848562
Ireland...

>>36848653
Wiwa got larger bread than that anon

>>36848575
As in painful or?

>>36848562
Like this?

>>36848809
Oh god yes!

>>36848734
Based. Also puffy vulva.

>>36848886
Have one more

>>36848774
Oh yeah, I was adapting an older IRyS prompt, but forgot to change small breasts out.

>>36848774
No, not painful. "Perfect" would be an inset chibi likeness reacting with textbook "uoh". But I'd take written "uoh", a sobbing emoji, etc.
Tried this one all the way from cfg 2-22, but no decimal increments. Some of them got close, some are quite sad, and even a "q_q" looking kaomoji, but none are "sobbing". I'm also surprised the image kept its general structure throughout. No major pose change, just the level of detail in her outfit, the shit in and the amount of "spokens", and her expression. I guess that's what a bloated prompt + exclude list can do.

what is the advantage of embeddings compared to hypernetwork or dreambooth?
I've been wondering whether I should switch to use embeddings to match with everyone else or keep using dreambooth
especially with dreambooth being recently added to automatic1111'svmsygd

>>36849007
Belly grab is the most sex tag.

>>36849173
I actually didn't use it there, the AI just gave her that pose.
That image was a bit fucked up, she had 4 arms like an Indian god, so I fixed it up in paint.net quickly and then img2img'd it.

>>36848840
And one more for you as well.

>You don't mind if I take off my jacket, right anon?

>>36848575

Do you guys ever run old prompts through the AI when you're bored?

>>36849414
Of course, grandma, go ahead. I'll turn the fan on, I know you get pretty tired after our walks. Would you like some lemonade or anything?

>>36849414
bigger cuz why not

Took the Yopi embed for a short spin
https://files.catbox.moe/1g2h4y.png

>>36849159
I have no idea how to dreambooth anon. Embeddings are good for characters or objects, anything you can add as a promt or keyword.
Hypernetworks are "styles", these usually mimic other artists styles and take a significant investment to train,

>>36849465
I just roll Kroniis with huge tits and wide hips, I save none, but my dopamine gets going

>>36849465
Yeah. I have a bunch of flawed images I might get around to too.
Like this one. One day I'll fix her fangs.

>>36849331
One last sexxo catto for today.

>>36848414

>>36849518
Try bodypaint tags. Showed up a bunch in the dataset files.

>>36849643
sankq for the sex cats

>>36849567
I think you can train styles using dreambooth too but I haven't tried it before
I'm feeling a little bit left out when everyone talk about embeddings

Trained a hypernetwork for footjobs overnight but it still sucks lol

https://files.catbox.moe/zzi7wv.png

THIGHS
ʷᶦᵗʰ ᵗᶦnʸ ᶠᵉᵉᵗ

>>36849806
I made a v0.1 guide in previous threads on embeddings. https://files.catbox.moe/vv8gwa.png

>>36849690

>>36849499
Surprise me, sweety, you know what's best for this old body.

>>36849830
>Trained a hypernetwork for footjobs overnight but it still sucks lol
It's about 1000x better than anything I have ever managed to generate

>>36849992

>>36849830
You need a big toe embed

got a circled "T" on ddim with a different seed. it knows. give me a "ToT" or something pls you cold machine.

>>36849450
not front and center uoh - inset or in a speech bubble. how'd you get it bang on though? inpainting?
>>36849465
Yes, I've gone back and fixed a few up, or tried to run them through with different characters without altering them.
>>36849711
Can it put paint on her face? When I look at Iofi her model at least three things stand out - face paint - the weird shape of her eyes - the weird shape of her face - I see none of the above. Granted I have no idea what the training set looked like, nor have I ever really looked at yopi art up close. If it deepdanboru'd bodypainting, did it catch facepaint at all? Meant constructively, be mad if you want.
Went back to look through the 20 or so pages on danbooru, mainly to find this one: https://danbooru.donmai.us/posts/4861741?q=airani_iofifteen which also isn't particularly on model. This thumbnail practically jumped out at me: https://danbooru.donmai.us/posts/5569414?q=airani_iofifteen as a terrible example of what she looks like and this one as a good example: https://danbooru.donmai.us/posts/5548938?q=airani_iofifteen

>>36850017
Here you go, grandma.

>>36849518
Mooner embed as well
https://files.catbox.moe/y30w9r.png

>>36849465
>>36849946
I'm liking these way more than I should..

so. close. pls. gib. ToT

>>36850138
>mainly to find this one: https://danbooru.donmai.us/posts/4861741?q=airani_iofifteen which also isn't particularly on model
nor does it have better facepaint than what NAI slaps onto things when you ask for it, kek. or better tit...gravity. still a nice image though

>>36849974
thanks, I'll try it later

>>36849568
I feel that, I usually do something similar with Okayu.
>>36849592
Gotcha. Those are some almost really nice fangs...
>>36850138
>Yes, I've gone back and fixed a few up, or tried to run them through with different characters without altering them.
I'm almost embarrassed with how unga bunga my old prompts used to be, but when I "fix" them sometimes it breaks something else in the output for some reason. Character swaps are pretty fun too.

I MISS FAUNA

>>36850227
It's an old prompt but it stills works pretty well.

>>36850138
See picrel, a bunch of stuff works with it.
Examples: paint splatter, body writing, paint, facepaint, bodypaint, facial mark, facial tattoo

Unrelated but heres some loods
https://files.catbox.moe/nkhx2r.png
https://files.catbox.moe/bfd1sg.png
https://files.catbox.moe/hxhobp.png
https://files.catbox.moe/s1bjqn.png
https://files.catbox.moe/67t6xj.png
https://files.catbox.moe/3pyn0n.png
https://files.catbox.moe/cuys3w.png

>>36850180
Thanks, little one

I love these cute sisters

"We have Reine at home", please cooperate with me. Are you trying to speak in Bibi or something?

>>36850220
Mooner is spot on. Her model's got sleepy eyes too but you hardly see it in her art anyways.
>>36850397
Oh, for "fixes" I usually end up playing with excludes, scale, and weight more than anything else. Or just trying different samplers. The only proompt I've "fixed" and felt like I actually got a better image out of was that bloody loli Kiara incoming kiss. "Fixed it" to get outstretched arms when anon asked for a kiss from Kanata, re-rolled it with Marine and Pekora right then, got bored the other night and did Ayame, Chloe, Iroha...and a Kiara with outstretched arms and bigger tits
>>36850656
Picrel has that "cfg too high on euler_a" look though.
>https://files.catbox.moe/hxhobp.png
sexxo and looks the most like her
>https://files.catbox.moe/3pyn0n.png
>https://files.catbox.moe/cuys3w.png
These two are close as well. Maybe because she's looking down? I don't know how else to describe her dumb little alien face, always looks like her face is pointed down, kek.

anyone have any tips for two girls in one image? i always either end up with clones of one of the girls or two copies of a mix of the two girls features

>>36850832
yeah my bad, forgot to post a highresfixd one.

I usually run cfg at 9. You mean the star in her eyes right? Seems like a rouch promt to get right.

>>36850962
2girls
no, we have the same problem, someone said to use inpaint to force features on the girls

>>36850832
>>36850321
>>36850138
Would you accept this?

...and now to teeter on these tags to try and make it right

>>36850967
Mainly the shape. There's probably a Japanese word for it that means "eyes without a lot of white showing" https://static.wikia.nocookie.net/virtualyoutuber/images/c/cb/Airani_Iofifteen_-_Portrait.png but yeah the star and color too. That highres fix does look considerably more like her, down to the chin and all.
>>36851033
Even with inpainting it can still weasel its way back into swapping traits.
>>36851162
Excellent timing. You win by default.

>>36851345
nice

ayy BayoMio!

Did a quick comparison run between
>>36833626 https://files.catbox.moe/tuh4nj.png
>>36843523 https://files.catbox.moe/yh8ora.png

No offence to slower anon but the other one is significantly better. Is is the same dataset?

I know which one I would use.

fuck I played Also sprach Zarathustra too early this time, lmao

it made some neat reine at home outfits along the way though

>>36850404

>>36850962
You can always generate two images, stick them together, and then run it through img2img.

Supposedly "(subject1 detail1 detail2) AND (subject2 detail1 detail2)" syntax will more often generate images with both subjects but I haven't experimented with it much. It definitely combines them anyway sometimes.

>>36851933

>>36851937
2girls, (gawr gura from hololive with white hair and blue eyes and small booba) AND (calliope mori from hololive with pink hair and red eyes and a tiara and big booba) having a fist fight
retard attempt 1/3

>>36851937
2/3
2girls, (gawr_gura hololive white_hair blue_eyes small_breasts) AND (calliope_mori hololive pink_hair red_eyes tiara large_breasts) having a fist fight
>>36850404
>>36851970
fucking kek

>>36851937
I'm a neanderthal and I forgot to lock the seed.
here's 3/3 on the same seed as >>36852141
2girls, (gawr gura, hololive, white hair, blue eyes, small breasts) AND (calliope mori, hololive, pink hair, red eyes, tiara, large breasts), having a fist fight

>>36851937
>>36852243
>>36852141
2girls, (gawr_gura hololive white_hair blue_eyes small_breasts) AND (calliope_mori hololive pink_hair red_eyes tiara large_breasts) having a fist fight

>>36851970
I got you anon.

since using the "_" that NAI isn't supposed to need with the AND seemed to work best...
2girls, (gawr_gura hololive white_hair blue_eyes small_breasts short_hair shark_hair_ornament shark_hoodie) AND (calliope_mori hololive pink_hair red_eyes tiara large_breasts long_hair leather)

>>36850962
https://files.catbox.moe/6mx83l.png
You can take the meta data from this, this syntax is what you should try to use. Do note however that some tags are just flat out stronger than others so success in this field really depends on who the two characters are. For example, Okayu has an ahoge where as Korone doesn't, however there is no tag to indicate that Korone shouldn't have one so it gave an ahoge to both. If one of the girls has animal ears and the other doesn't it will try to give the animal ears to both, etc.

got a cute outfit swap
2girls, (gawr_gura hololive white_hair blue_eyes small_breasts short_hair shark_hair_ornament shark_hoodie) AND (calliope_mori hololive pink_hair red_eyes tiara large_breasts long_hair leather bustier)
why can I hear Q playing?
>>36852420
Adorable

Anya misses petwa
trying to make her hugs a stuffed penguin but it always generate a fucked up looking penguin instead

>>36852420
Some girl's names are very powerful, such as Botan's, to the point where including her at all can completely over power an image and lead to funny things like this.

That's some interesting basketball hoop design on the right.

>>36852380
now chibi

>>36851482
BIGGER

>>36852725
Eh, not a fan.

>>36852950
Bu... but that's the meme D:

>>36852841

>>36852950
Dose she have a garden of eyes inside her blouse?

>>36853076
And just to make sure picrel
Kson is that you?

>>36851937
the best advice I can give you is "exclude 'heterochromia' - always" and like the other anons said, some tags and characters are just too strong. try putting "character name, hololive, tentacle hair" in sometime. you will get squishies more often than anything else. so some tag combinations are also just super overwhelming.
here's the culmination of my retard prompt process
I just went for broke with really stylized tags and turned cfg way down. 2girls indeed. touching it at all makes the eyes change colors. trying to upscale it gives Mori shark teeth. kek. and this is why I won't try for two characters. it's not worth the headache and you have even less say as to what it looks like
2girls, (gawr_gura hololive white_hair blue_eyes small_breasts short_hair shark_hair_ornament hoodie) AND (calliope_mori hololive pink_hair red_eyes tiara large_breasts long_hair leather bustier), motion lines, monochrome, speed lines, opposing sides, sketch, flat color, colorful, action, album cover, absurdres, abstract background, serious
>>36852460
>why can I hear Q playing?
Now I can really hear it
https://www.youtube.com/watch?v=aetXqd9B8WE

>>36853155
...

>>36848312
Mika Pikazzo artstyle but hololive

>>36851937
>>36853225 cont
and with proompts this fucky, cfg changes of 0.1 on k_euler_a can mean the difference between the eye colors, garments, or clothing being correct, or being wrong.

>>36853001
okayitried

>>36853345
WOAH MAMA

>>36853380

>>36853578
UUUUUUUU

>>36853578
uuuu, thx

Felt like testing out that rat embed by making a bubble bath pic, but the bathtubs all looked real dumb so I swapped it out for an onsen instead but forgot to remove the bubbles part of the prompt.

If you actually stop and think about it it's sort of nonsense but it looks kinda nice somehow anyway

>>36853703
This needs to be on the miss fauna page

>>36853600
kek

>>36853816
Try jacuzzi

>>36848312
Gawr Gura, but {{{bigger}}}

>>36848312
Requesting Subaru, or whoever is easiest for you, in a Katamari Damacy level.

>>36853945
As in fat or "thicc"? I'll do either I guess.

>>36853945
>>36853990
Given no reply, the man prompted anyway.

>>36854230

>>36854230
These proportions are more a long the lines of what I was hoping for.

>>36854300

>>36853951
???

>>36853600
I kneel.

>>36853943
Good idea, I don't think it really looks that much like a jacuzzi but it's way better than what I had before

>>36853951
>>36854423
I should have included examples.
https://files.catbox.moe/f30w5y.jpg
https://files.catbox.moe/70ssja.jpg
https://files.catbox.moe/pxmpem.jpg

>>36853380

>>36853951
As the Katamari? As the Prince? Player? Bystander?
>>36854693
I don't think they didn't know what it was, mainly showing that the AI doesn't inherently know what it is. If you're too young to remember Katamari Damacy aren't you too young to be on 4chinz?

>>36854423
I love that she looks confused too.

>>36855361
I was thinking as the prince. Not necessarily trying to roll anything, though you could if you want, more so just little Subaru in a big environment, be it a house or wherever. Katamari Damacy came out 18 years ago.

>>36853951
Here she is as King of all Cosmos, kek.

>>36855361
I tried adding the artist name and its definitely doing something.

>>36855518
I can hear this image.

>>36855754
How did it know?

>>36855509
>more so just little Subaru in a big environment, be it a house or wherever
Oh. Turning her into the Katamari was easier, lmao. I'll try for that later.

>>36855920
Holy shit my sides.

Kronii, excuse you, nobody invited you.

>>36856010
Hmm?

>>36856010
HMMM?! I swear to God this wasn't intentional but it rolled two WRXs in a row. lmao. you heard it here first folks, there's a 3 door hatch WRX coming.

>>36856678
That's a lot of damage!

>>36856010
HMMM?!?!
ear rape warning https://www.youtube.com/watch?v=TYxifmvDJ_c

>>36856715
>>36856789
I SAW'D THIS BOAT IN HALF!

>>36856793
>>36856715
>>36856678
AI car commercials are the future

...this is so difficult without turning her into a chibi/loli

>>36853380
Here. Complete fucking accident.

there goes that idea

>>36857460
Oh no no no

Shuba no!

>>36855509
If you can find a proompt from someone who's successfully used "minigirl" on curated, lmk. It's half busted there, and on Full it makes kinda fugly stuff but works sometimes.

>>36858070
I didn't know minigirl was even a tag.

anyone know about an "anything" model? seen it mentioned in other threads as some better version of the nai model or something but can't find any more details on it

>>36858070
It kinda works sometimes
{minigirl}, oversized object, houshou marine, 1girl, red hair, chess pieces

>>36858322
cute

>>36858244
https://rentry.org/sdmodels#anything-v30-38c1ebe3-1a7df6b8-6569e224
It's better at some things and worst at others, you need to test it out

>>36858426
Aww, now I want more little chess chuubas.

>>36858322
>oversized object
that's what I needed.
...
Nope. kek

>>36858627
very cute though

>>36858482
any specifics on what it sucks at?

>>36855509
Best I can do is tree fitty. Holy oof. Prince Subaru on an upturned kitchen dish. Katamari in the background.

>>36858520
Towa did not want to be mini

>>36858627
yeah, with more stuff it gets harder to get it working

>>36859003
"Amiibo Subaru gets stolen by a Klonoa monster" 2022, Colorized.

>>36859044
Still adorable though.

I just put in Nene talking.

>>36859731

>>36859783
english power charge, huh?, what?, oh my god, go around, no no no no back, yes go around, yes very yes, future all okay chicken okay, believe in my heart, nice, yes my honey, okay husband all okay future in believe to you, i no understand english, i don't know, no believe, no look future, no believe in my heart, i don't in my heart, together twelve morning, nice ice cream future, in the morning, {{masterpiece}}, {{momosuzu nene}}, hololive, bangs, blonde hair, double bun, flower, {gradient hair}, {green eyes}, hair bun, long hair, multicolored hair, {pink hair}, two side up, 1girl, indoors,

>>36853380

>>36848774
>>36849711
>>36850574
My primal desire to fucking grab this fat ass and destroy it.

I wonder how many people in these threads prefer cute loli stuff vs people who want them lewd instead.

>>36861348
Why not both?

nice helmet

>>36861388
Like at the same time?

There was a function on imgtoimg where you could just pick colors and paint in details, what happened to it and how do I get it back

>>36861486
Nice! Here's one I made weeks ago, back when I was still trying to work out webui

>>36848575
no emotes yet, hypernet may be fucking with it, but she did start speaking the language of the gods

>>36848283
Anon who trained kobo embed yesterday, her fountain hair keeps on failing. Any suggestion?

>>36863036
that's still decent, even embeds aren't perfect so you'll have a lot of duds

Is there a way to get the png info out of people's posts here? Whenever I download images and put them in the PNG Info tab, it says there's no data. Is that being scrubbed? Also how do I get my filename to correctly show the prompt? I have the checkbox in settings turned on for "Add Extended Info"

>>36863158
Yeah 4chan strips all meta data, catbox posts should still work though.

>>36853703
Your fau fau is so good. I got an embed a few threads back for here which I'm using here >>36863158
but its not fau-fau-NAI-t8, its fau-fau-t8-NAI-1000. Is that the same embed, or is there a different one thats better?

>>36863036
Idk try maybe water fountain ahoge, water ahoge, something like that.

>>36863158
Use the catbox links for png info , 4chan scrubs metadata.

>>36863254
Do you know why my filename gets cut off after masterpiece though? I see other people have part of their prompts in the filename at least.

>>36863486
Thanks, it's not the same embed, the t8 ver doesn't produce as much noise as the t16 one. t16 bad, t8 good. Find it in past threads until it gets put on sdupdates2.

Try it with antlers for better ones. Inpaint the mole if you want to.

>>36863573
not sure what controls filename length, are you using colab?

>>36863158
You can take anything from catbox and open it with notepad++, that's what I do anyway.

>>36863486
Sorry misread. the final embed I'm using as author is the same one as the 10000 step one. They are the same embed.

So I decided to try training on my gpu again, and it's working. Not sure if it's training right as it seems to think Wiwa's hair is silver currently. I think I see which dataset images are to blame, not too fussed about it at this speed if it goes bad.

>>36861348
is this cute or lewd?

pantylines didn't work how I wanted, now I have a collection of Mori stank
I was expecting something like this >>36860933
anyways, good night.

>>36864450
She doesn't have any defined sexual features so I don't think it's really lewd at all. Also she appears to be melting so you might want to take her to a hospital.

>>36864450
prompt?

>>36863895
Nah I'm using webui.

>>36864123

HOLY CUTE!

>>36864604
we lost her
I failed to get her to the hospital fast enough
>>36864859
masterpiece, best quality, intricate, ultra-detailed, 1girl, (((loli))), (((younger))), (petite), ((skiny)), flat chest, (anya melfissa), purple eyes, long hair, two side up, (streaked hair|gradient hair|blonde hair:0.2|brown hair), (detached sleeves|sleeves past fingers|puffy long sleeves|black sleeves), wrist cuffs, jewelry, thigh ribbon, ribbon, white bowtie, (backless dress|black dress|frilled dress|frills), bare shoulders, kris hair ornament, ((((((anymlfss)))))), happy, open mouth, smile, (ass), (pantyshot), ((((wet clothes)))), ((see-through)), (river), (riverbank), (wariza), (splashing), bokeh, (soft lighting:1.2)
Negative prompt: lowres, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, jpeg artifacts, signature, watermark, out of frame, extra fingers, mutated hands, (poorly drawn hands), (poorly drawn face), (mutation), (deformed breasts), (ugly), blurry, (bad anatomy), (bad proportions), (extra limbs), cloned face, censored, (black hair), (white hair), (blue hair), (pink hair), (purple hair), (orange hair), (red hair)
Steps: 28, Sampler: Euler, CFG scale: 12, Seed: 367528720, Size: 512x768, Model hash: f70bb064, Clip skip: 2, ENSD: 31337

>>36865464
>(backless dress|black dress|frilled dress|frills)
never seen this syntax, is this new, or is it just for easy reading?

>>36865785
not sure if that does anything
I'm just lazy and copy paste several prompts and only edit the necessary
changing it to normal syntax only generate a slightly different image tho

>>36866704
gotcha, thanks

>>36865246
here's my save settings

>>36858849
Feels like it's better than NAI at making good characters and pretty backgrounds, but it has a tougher time making lewds. I should lurk the /h/ thread to figure out a good lewd model to merge into it, or just use different models for different content I guess.

>>36868471
try finding your images in the output or log folder instead of using the save and download button on the webui

>>36858482
whats the difference between these fp16, fp32 and full ema models?

>>36869491
I was replying to him but if he's saving manually then that could be it

>>36865464
>((((((anymlfss))))))
is this an embed, or is it that dreambooth thing?

I'm a little curious about dreambooth after seeing it brought up a few times but to me it just seems worse? Since it's writing it into the model itself you can't share as easily as you can with embeds? Maybe you get higher quality images in exchange, I'm not really sure since I don't know if anyone is trying both

>>36848312
Drunk Kronii

>>36870053
It's dreambooth
other than things you've already mentioned creating multiple dreambooth models instead of embeds also takes too much space
I'm trying to see how far I can mix-trained a model with multiple characters before it starts generating ungodly chimera

>>36863036
+ water hair
- don't fuck up water hair
max strength

>>36870416

>>36870416

>>36871723

>>36871767
I tried to give her a glazed over look in her eyes.

>>36863561
fountain ahoge just puts her in a fountain, kek

>>36871831
>>36871767
>>36871723
>>36871319
Ok dude

>>36871833
Kobo REALLY loves to show off her midriff unprompted

>>36871831

>>36871993
??

I never realized how much Kiara suits the neon 80's aesthetic until now.
Especially when looking like a vhs fitness trainer.

>>36872133
doesn't her outfit is like that too?

>>36872349
Yeah, unless you de-emphasize the embed a little, it'll bring in some of her lewd see-through raincoat outfit to most pictures

china dress shion uohh

>>36862546
kek my fucking sides
>>36869794
nta fp16 and fp32 refer to "full precision" 16bit and 32bit. as opposed to half precision. I don't know if that's what they used while making it, or the environment it's meant to be run on.
>full ema
no idea
>>36872449
embed kobo so good

>>36873154

I tried training kobo embed on my own, why does my training result looks so awfully unlike kobo on 1500 steps?
I saw anon embed training yesterday even on 500 steps already looks like pseudo kobo

>>36873702
Having never tried to make an embed in my life, my guess from what I've read is that you didn't quite tag every image you trained it on properly.

>>36870053
I've tested a lot of the model modifications and here are my thoughts on them:
embeds: these are tiny files which find the best representation of whatever you're training them on in the base model. By far the most flexible option and will have very good results if the goal is to group or emphasize things the model already understands
hypernetworks: there are like instructions that slightly modify the result of the base model after each sampling step. They are quite powerful and work decently for everything I've tried (subjects, styles, compositions). The cons are they can't be easily combined like embeds. They are also harder to train because good parameters seem to vary wildly so a lot of experimentation is needed each time
dreambooth: modifies part of the model itself and is the only method which actually teaches it something new. Fast and accurate results but the weights for generating adjacent stuff will get trashed. These are gigantic and have the same cons as embeds

>>36872148
That'll do proompt that'll do...

>>36873826
This is my tag and I already tagged every single picture on my training folder

>>36873890
nta but don't be a faggot. we have few enough request fillers as it is without people like you

>>36873922
Well I got nothin' then. Like I said, I've never tried it myself.

>>36873922
How many images did you use to train kobo?
I wonder if it's better to use more images or use less but with more consistent artstyles

>>36873854
Thanks for the explanation, definitely seems like embeds are the way to go for most of the generating that goes on here then

>>36848312
Robot Sana pole dancing at a burning city surrounded by people while wearing a competition swimsuit

mmmuuuuuuu

>>36875265
Bigger!

>>36873854
Its not chuuba related, but since you have some experience with these, I'm interested in training the model to be able to prompt for Au'ra, a playable race from Final Fantasy. I don't know if an embed or hypernetwork would be more appropriate in this case, what do you think? I gathered ~100 images to use (see below), do you think it is too much?

https://mega.nz/folder/ZWcXCYpB#Zo-dHbp_u30iIz-LxLUGyA

>>36875599
ogey

>>36875857
belly too?

you ever wake up in the middle of the night just to take a massive shit?

>>36875798
NTA, no experience, gut says hypernetwork because it looks like a different artstyle to what the model usually makes. and because they have "unique horns" like IRyS and that one was an HN.
>>36849465
>>36850397
this is one I just decided to go back to and make "look cooler", the fact that it's of mori is a coincidence

>>36875951
looking pretty squishy

>>36875970
What did it look like before?

>>36876125
Cute tum.

>>36876126
Not nearly as cool, not picrel either, still wanted to give it something to really make you FEEL like ohhhhhhhhhh I have to shit. Before: https://litter.catbox.moe/evx78u.png

>>36876156
All the instant noodles and delivery food...

>>36876222
Now we're talkin. Even more?

>>36876195
I love that she just looks offended. She knew the prompt.

>>36876276
I feel bad for fattening the hag, sorry

>>36870416

>>36876393
That's fair. Could you tell me Reine embed you have or link me to it please? And also the base you use with it to make Reine, I don't really know how this embed thing works but I figure I might as well learn to use it myself instead of trying to beg.

>>36876393
Based

>>36876487
https://files.catbox.moe/b9cn4z.png
white hair should help get the hair color right, add long hair, side ponytail,french braid for the classic look, short hair for the casual look.

kronerino, where did you find that blue wine? are you sure it's safe to drink???

>>36876629
>no results for "how to seduce chicken"

>>36876629
Mine didn't come out looking anything like her so I must be doing something wrong. Thank you though.

>>36876941
Upload some results to catbox so I can peep the png info and see what's the matter.
I also have this older Reine embed that was baked for much longer and uses up way more tokens, it'll force her and her outfit pretty hard if you want to try it out
https://files.catbox.moe/m0he7i.png

>>36877021
Holy shit 35 vectors and almost 40k steps?
Take 3 of Elira is coming along nicely it seems, should be done by the time I wake up. I did a 4 vectors run up to 10k earlier but wasn't really satisfied with it.
Good night thread.

>>36848312
post hilarious fuck ups and happy little accidents. me on the bottom
>>36877322
gn

>>36848312
Shishiro Botan in Princess Peach cosplay

>>36876804
(inflatable pool toy) actually returns what I wanted, technology sure is crazy

>>36877021
Sorry, I had to step away.
Just as a prempt that I'm not a complete retard, I did set it up properly: https://files.catbox.moe/219g6h.png
But Reine just looks like nothing:
https://files.catbox.moe/n6zjbe.png
https://files.catbox.moe/0ltcom.png

>>36877890
Try using "weine-10000" in the prompt instead? Or whatever it says in the cmd window when listing your loaded embeds.

>>36878083
Oh that's probably not what I want to see. https://files.catbox.moe/n1wd73.png
But then it says it loaded 39 textural inversions anyway?

>>36878227
when did you last git pull? that could help if it's a software problem

>>36878373
I set this up on Oct. 16th and haven't done anything to it since then. But I will git pull right now if you tell me how.

>>36878459
Put git pull below your command args. Like this

set COMMANDLINE_ARGS= --precision full --no-half --always-batch-cond-uncond --opt-split-attention --deepdanbooru --xformers
git pull
call webui.bat

>>36878459
In your stable-diffusion-webui folder, right click in the explorer window, select "git bash here", write "git pull" in the window that opens up and press enter
>>36878540
and this does it automatically on startup

>>36878540
>>36878572
I did the git pull and now it does list the embeds and using the "weine-10000" from the list gave me this. https://files.catbox.moe/y6k5i0.png
Could someone recreate this for me so I know that its working properly please?

masterpiece, best quality, weine-10000, white hair, short hair, side ponytail, french braid, standing, outdoors, beach, facing viewer, bikini,
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name
Steps: 50, Sampler: Euler, CFG scale: 12, Seed: 698620752, Size: 512x768, Model hash: 925997e9, Clip skip: 2, ENSD: 31337

Used embeddings: weine-10000 [04f7]

>>36878838
stuck it into png info and ran the generation, here's my output. do you have the VAE loaded in settings / stable diffusion / SD VAE dropdown?

>>36874193
If you have a problem you can leave, not a single person in this thread is special (You) primarily are not. In fact we should be looking out for people like you in these threads as you think you are or have something of worth.

>>36878955
It was set to auto but I changed it to load the nai.vae.pt, the one the guide told me to rename when I was initially getting everything set up, and it gave me the same image but now it looks more clear. https://files.catbox.moe/bfpz7j.png

>>36879171
I mean it looks like the embed is working fine and your setup doesn't have anything disastrously wrong with it, it prints out reines like intended. The rest is messing around with model mixes, tags, sampling methods, steps, highres fix and so on to get outputs you like

>>36879262
I'm just wondering why they still have those few differences. Thank you for all the help though.

>>36879490
Do you have --xformers in your webui-user.bat? it can cause some difference in outputs. I also don't use medvram, precision full or no half in those arguments, that might cause some deviation too.

>>36879564
I added xformers to my .bat but the image still looks the same. The other things you said probably are the difference though, thanks again! Out of curiosity could you put that image in a catbox link so I could generate it and see what it looks like on my end?

>>36879828
https://files.catbox.moe/0lh3xj.png
the model I'm using is a mix of the new anything model, f222, SD1.5 and r34, so you probably won't get anything close to identical with NAI, but see how it goes

>>36879977
Oh, being on a different stable diffusion checkpoint would definitely be why they are different then, that explains it. Could you point me to where I could get the model you are using?

>>36877422

>>36880040
I only did this because I'm bored but if you want to try it out, here goes
https://rentry.org/sdmodels
Download the following models:
SD 1.5
Anything-V3.0 - fp32 variant
Zeipher f222
r34_e4
put them all in your stable-diffusion-webui\models\Stable-diffusion folder and use webui's checkpoint merger tool, follow this guide
https://rentry.org/berrymix
but replace NovelAI with Anything, SD1.4 with 1.5 and f111 with f222. If your model hash ends up being 99a1a2f6 you have the exact same thing as I do. I'd just upload it somewhere but idk where to toss a 4gb ckpt

>>36880308
whatever happened here I like it

>>36880319
emaonly for SD 1.5 or just pruned?

>>36880376

>>36880460
emaonly

>>36880480
Okay, thanks. Downloadin all these gigabytes. I offer you this cute sheep in return.

>>36871833
>>36872133
>>36872449
Glasses Kobo halu...

>>36875798
Cute yopi

>>36875265
>>36875857
>>36876125
>>36876222
Sex

>>36876629
>>36876804
>>36879262
>>36879564
Rock hard more glasses please

>>36880593
good luck on making your own frankenstein's diffusion!
>>36880747
grown kobo is too powerful
(The Great Wave off Kanagawa) improved the gen, kek

>>36880919
I'm starting to think inpaint is best for fountain hair. Unless someanon finds the golden print for it

>>36881071
That or making yet another embed just focused on that. Or both, inpainting with a specific embed!

Good night anons

>>36880919
The anything.ckpt came with it's own vae.pt file. Do I need to include that in my stable diffusion folder as well or is the vae i already have fine?

>>36881172
I think I read on /h/ this morning that it's just the NAI vae, but you can put that in the VAE folder under models and try it out to be sure. I'm using the final merge with just the NAI vae

>>36881260
Okay, thank you.

>>36881142
I swear I have more lewds of yopi has been created in the past 24h than what exists on the internet previously

>>36851589
Yeah, mine's a piece of shit, i have no idea why. I just wanted to post it because i'd spent so long on it. Allegedly it's supposed to be the same images, same tags, same settings etc but i dont know. Probably something i fucked up on my end. I'll also be using his embed and give up on making my own. Also, thanks for actually showing me just how bad my thing is, i'd've spent another day just smashing my head against a wall and wondering why it didn't produce decent images
t. abstract moona maker

>>36879977
>>36879564
I followed the guide to make the mix with the changes you said. I dragged your image from catbox into png info and generated it using the mix and I got the exact same model hash as you but somehow it is still different. https://files.catbox.moe/lm79pl.png

>>36881880
Ok, last thing, do you have "Enable quantization in K samplers for sharper and cleaner results. This may change existing seeds. Requires restart to apply." checked under settings / stable diffusion? that's the only thing I can think of, outside of launch parameter settings or GPU as the source of the difference

>>36881835
Keep trying anon. You probably only need a few settings tweak. Double-check some thing on the guide i posted. https://files.catbox.moe/vv8gwa.png

>>36882026
It was not checked so I checked it and restarted. It gave me almost the exact same image again except one of her fingers is shorter. https://files.catbox.moe/qpbrro.png

>>36882236
Wild, what are your launch parameters and what GPU are you using? My webui-user.bat has the commandline args
set COMMANDLINE_ARGS= --xformers --no-half-vae --deepdanbooru
and I'm running a 3090

>>36882281
set COMMANDLINE_ARGS=--precision full --no-half --no-half-vae --medvram --xformers
and I have a GTX 1660 Super, 6GB VRAM

>>36882332
I'll give it a try with your commandline to see if that's where the difference is

>>36882072
Nah, i legit cant be fucked anymore. I'd rather leave it to the anons that know what the hell they're doing and then just reap the rewards. Also, it's pretty disheartening to start doing something, have someone finish it in a fraction of the time, produce amazing results, only for yours to come out 5 hours later looking like a volcano eruption that cant even get the hair length correct. Call me weak-minded, i dont care

>>36882388
Also i wont be pointlessly torturing my pc anymore

>>36882332
>>36882360
Tried it and still keep generating this
https://files.catbox.moe/ezyo0o.png
Can the GPU really make such a visible difference? wild

>>36882332
oh, one more thing, remember to check "Upscale latent space image when doing hires. fix" in upscaling settings, that has a big effect, picrel is the same generation without it checked

...how high is the possibility that i've downloaded a bunch of embeds but i'm somehow not actually using them to generate images and NAI is pulling them straight from the character tags as if said embeds dont exist?

>>36882613
LEWD! https://files.catbox.moe/feuxp1.png
That certainly made a difference but it's not the same image as yours. But yeah if I take "percision full" off I get just a black screen so I guess it can make a pretty big difference. I will try to take solace in the fact that at least Reine still looks cute, even if she is different from yours. And thanks for taking all this time to troubleshoot with me.

miko

>>36882757
check the info box under your generation in webui
>>36882783
np, good luck on your quest for cute weines

>>36882817
One last question I had. What is deepdanbooru? I saw it in your command line.

>>36848312
Mori but make her boobs big. Thanks.

>>36882903
Like comically huge or?

>>36882901
Lets you "interrogate" images in the img2img tab, putting the danbooru tags the machine spirit thinks fir the image into the prompt box. I use it for preparing embed materials in the image preprocessing tool.

>>36882903
gigantic enough?

>>36882388
>>36882428
Totally fair and understandable.
I didn't make the other moona embed though but you did upload the dataset at a request for someone else to compare. Can you take a picture of the entire step process logs in \textual inversion\date\moonmoon-nai-t8\images ?
Maybe that can give more clues and help me debug common training mistakes.

>>36882949
>>36883073
Yeah but in her normal outfit.

However this is pretty much perfect thank you anons.

>>36883212
There's nothing "wrong" with the progression imo other than the first ones having the wrong hair color, so i likely fucked up the settings without even knowing, like with the style_filewords thing

>>36883222
>normal outfit
You are going to have to settle for a breast size nerf then unless you are okay with the rest of her being fat as well.

Lui-mama here to give all the good boys and girls some milk

>>36883375

>>36883400

>>36848312
Requesting curvy Lize Helesta posing in blue latex

>>36883375
>>36883400
Perfect. And I'm fine with it. Thank you for your time anon it's appreciated.

>>36883424
Mori is a fat ass cow!

>>36883458
Happy to help.

>>36883375
>>36883400
>>36883424
>>36883538
Prompt for the clothes?

Anything is so good

>>36883805
Mori's complete outfit: anklet, armlet, armor, belt, black cape, black dress, black thighhighs, bracer, cape, cleavage, detached sleeves, dress, gold footwear, high heels, jewelry, long dress, pumps, see-through sleeves, shoulder spikes, side slit, single detached sleeve, single thighhigh, sleeveless, sleeveless dress, spikes, thighhighs, torn cape, veil, white veil, see-through veil,
For those images I took off the tags about her shoes and anklet so that she would be properly in frame.

Why the hell does NAI keep giving me multiple angles? I've tried putting them in the negative tags, using the de-emphasizing brackets, even tried straight up removing the prompt. For fuck's sake. Does this thing have a fetish towards that thing or something?

>>36884073
1girl, solo,
Put multiple views, in negatives.

I think I messed up with my embed learning rate
I set it at 2e-6 currently 2000 steps in and it doesn't looks like anything I want

>>36884367
I've been using the default 0.005 and the results have been good

>>36883437
I have no idea who that is, but with danbooru on the side i got this girl. Tell me if that's her and if you want ~5 more images like this but with subtle differences and ~5 more with an exposed navel

>>36884427
how far do you usually train using the default rate?

>>36884494
10k has been plenty so far

>>36884512
I'm getting 0.15 to 0.25 loss at this current rate, wouldn't I get even higher loss by increasing the rate?

>>36870027
>>36869491
I was doing that, I didn't realize that changed the filename. Okay I'll start checking the folder now.

>>36884676
loss rate really doesn't matter as long as the end product comes out alright

>>36883349
Honestly anon I got no clue. I would just double check settings if you ever feel like trying again.

>>36884890
Can you post the prompts please?

>>36885888
illustration,1girl,(weine:0.9),standing,spacesuit,skin tight,thighhighs,big ass,science fiction,space station interior
Negative prompt: lowres, bad hands, bad feet,bad anatomy,text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, (extra limb), (poorly drawn hands), (((mutated limbs, mutated feet, mutated hand))),
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 3647098280, Size: 1024x1536, Model hash: 99a1a2f6, Denoising strength: 0.64, Clip skip: 2, ENSD: 31337, First pass size: 0x0

good morning

>>36879056
>[Deleted]

>>36882236
>except one of her fingers is shorter.
>>36882332
>>36882360
I'm not going to try and read through the entire reply chain, but onegai, for the sake of troubleshooting, KISS not a chu, keep it simple, stupid. Also, a difference like "one of her fingers is shorter" can boil down to a different GPU, running with -medvram or different launch options...

>>36886081
good morning!

>>36886243
gonna roll for infinity until [Deleted] is holding a nice shotgun

page 10 moon

>>36887096
prompt?

>>36887186
masterpiece, best quality,high detail,1girl,illustration of mooner,purple hair,nude,straps,thigh strap,arm strap,beltbra,leather thong,harness,o-ring collar,bondage outfit,black leather,o-ring,skindentation,nsfw
Negative prompt: lowres, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, (extra limb), (poorly drawn hands), (((mutated limbs, mutated feet, mutated hand))),no nipples
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 10, Seed: 8864016, Size: 1024x1536, Model hash: 99a1a2f6, Denoising strength: 0.64, Clip skip: 2, ENSD: 31337, First pass size: 0x0

>>36887240
>99a1a2f6
What model are you using? That's not NAI or Anything v3.0.
Is it SD 1,5?

>>36887296
>>36880319
basically berrymix but with updated models

>>36887333
Did you mix it yourself? Is it 50:50 mix?

>>36887711
I followed the step-by-step text instructions linked in my post