🍥｜anime

[oak spoke]

[nova remnant]

you'll have to use something

other than prompt

[junior quest]

ehh its pain

[junior quest]

https://tenor.com/view/cat-smiling-smile-cat-orange-smile-cat-gif-23133359

[nova remnant]

@junior quest Best tag i could find is like (dark room)

[junior quest]

it doesnt work either

[nova remnant]

it doesnt do a whole lot, too few images on that tag

600 on danbooru

[junior quest]

lets add bunch of ((((())))))

[nova remnant]

and pony isnt very good at backgrounds

[junior quest]

EZZZZ

[nova remnant]

just do black background:1.4

Easy

[junior quest]

its long

[oak spoke]

prompt weight give my pony heart attack

🤣

[oak spoke]

dark enuff?

[nova remnant]

some stuff breaks it, specially if you go below 1.0 and on negatives

[oak spoke]

maybe it's also affected by resolution

[junior quest]

yesh but I need with a room

[nova remnant]

basically a bunch of math going on and it can break

[junior quest]

dark room

[nova remnant]

inpaint it from a black image

[oak spoke]

I'm getting good result with pure t2img rn

gib 1 min

[nova remnant]

Good result = tomato

[oak spoke]

nah that's GREAT result

[oak spoke]

dark room

[junior quest]

seee there is a lightt

there are some loras for dat, but smting smthing arent right for me

[oak spoke]

if there's no light then it's just pitch black background

[junior quest]

these loras work tho

but dun feel like use them

[junior quest]

https://tenor.com/view/cat-hello-cat-peek-cat-door-cat-peek-door-gif-17361470248875227911

[nova remnant]

Cozy in bed

[oak spoke]

bed? you mean butchering table?

[nova remnant]

go away

[junior quest]

purr, scratching

[gusty palm]

sd forge reset my user.bat ><

does anyone have the example with the vae and upscaler directories added

this is as far as i got but not sure how to add the esrgan ones

[lyric zealot]

  --vae-dir %A1111_HOME%/models/VAE^
  --realesrgan-models-path %A1111_HOME%/models/RealESRGAN^
  --esrgan-models-path %A1111_HOME%/models/ESRGAN^```

[gusty palm]

tyty

couldnt find that info listed anywhere after searching all over

and its weird how it uses 'model-path' instead of dir

[lyric zealot]

but it work right?

[gusty palm]

i believe so, thats looks like the one i was using before it got reset, im still reloading server

ill be making abackup copy of this bat now that i know its prone to resetting

[lyric zealot]

[gusty palm]

yep working thank yy

[lyric zealot]

[gusty palm]

is there a line for controlnet models

i think ve already redownloaded most of the controlet things into forge though

oh nvm thats in the web ui settings u can se the dir

[lyric zealot]

yup yup

[gusty palm]

okk

[pure oyster]

I need to expand my horizons with different tools and prompts, but I can't stop making waifus with midriffs

someone help

[nova remnant]

actually i can't, those images I was talking about are also nsfw

its in human nature to appreciate beauty uncensored or rather not putting limitations on expressions

btw one thing i wanna correct, autismmix / pony models are great for anime, not realism, like you might have already known, but some ppl are using loras to make images that are like realistic magazine covers with female characters, while those look very realistic in some form, but the model itself is meant for anime

these are all autismmix anime

[nova remnant]

too much water

[nova remnant]

squish

[steady grail]

there are ai to remove the background

i usually just do it, gen the bg i like and than copy-paste the character on it and than img2img the result

yeah, to fix the little mistactes and get a better end result

with low denoise, the base concept don't get altered

at the bottom you can set the denoise. the higher, bigger are the alteration

you can blend them with any painting program

but programs like sketchbook, krita or corel painter are better because they have a transparency settings and can set layers

to remove the bg, i use lama cleaner

forge have an extension like that too

[nova remnant]

just inpaint it

[steady grail]

i honestly hate inpaint

very unpredictable and gives weird mutations

just good for some small fast fixes

[nova remnant]

I'd say the opposite lmao

very bad for small fixes, takes so long to get something right when I try it

[steady grail]

luck is paramount when using inpaint XD

what's the deal with iopaint? i only find results in asiatic

not sure witch part of asia

https://www.youtube.com/watch?v=G_67tX2uTWM

[nova remnant]

I have no idea what it is

[nova remnant]

It should be changing a bit

hmmmm

I don't really know what you want to do

if you're generating a character, generate one on a white background and put it on top of the background you want lmao

very low denoise img2img after should fix some lighting issues if you have

[steady grail]

cute

[nova remnant]

glitch prompt really is glitchy

[steady grail]

but the effect is out if this world

[nova remnant]

[nova remnant]

https://tenor.com/view/ohnepixel-ohne-pixel-grab-grabby-hands-gif-26126978

[supple raptor]

You can try this. The tutorial looks promising, I haven't used it yet though. https://stable-diffusion-art.com/soft-inpainting/

[distant remnant]

cute foxes

[pure oyster]

your denoising strength is too high. anything below 0.3 is not noticeable and anything above 0.55 and you'll start generating entire images inside your images.

[nova remnant]

big eyes

[steady grail]

the upper part of her body is too long in a funny way XD

the legs too XD

[vital raptor]

anyone know if there's a better wildcard extension for a1111? dynamic prompts isn't... dynamic enough

something that supports conditionals and % chance would be amazing

[steady grail]

it's hard to use wildcards on auto 😦

i had the same problem on forge

[sage sluice]

I've heard good things about unprompted.

[sage sluice]

Personally, if you want true dynamic language with conditionals, you might be better off writing a standalone script up in python, php or something else and just use the "import from textbox" script. We're definitely due for a better wildcard extension.

[steady grail]

that was a fast experiment to test the workflow

[vital raptor]

yeah, if i can't find a way to work with existing tools i might have to make my own extension

[sage sluice]

I remember trying that jinja2 support in dynamic prompts, couldn't get anywhere with that.

[vital raptor]

I want to be able to do things like "insert a wildcard with 50% chance" and if it happens, another 30% chance of a second entry, 10% of a third, etc while making sure they don't duplicate

[steady grail]

that's waild

[vital raptor]

and also if X wildcard is chosen, don't allow certain other wildcards

[sage sluice]

I think you can do that first half in dynamics. But the 2nd half, no.

[vital raptor]

like the prompt might end up choosing "from behind" and "looking away" but then try to set the eye color, which will either not matter or confuse the image

[steady grail]

or gen god eyes in the sky that see everything, transforming a romantic image in a pure nightmare

[vital raptor]

or just gen a second person facing a direction where you can see their eyes

[sage sluice]

Yeah...just checked, unprompted can do that. https://github.com/ThereforeGames/unprompted/blob/main/docs/GUIDE.md

[vital raptor]

sometimes the weird mistakes are cool, but also sometimes you get a prompt that just has "hair clip" 3 times i a row

[sage sluice]

Check the section on conditionals.

[vital raptor]

yeah, looking at that, i'll have to play with it after work

that one looks like it might work, you can set all the conditionals and everything per-line in the 'wildcard' file so it can be fine tuned for individual entries

definitely have to give it a try

[sage sluice]

Kinda gives me Alex(mc) vibes.

Lately I rediscovered the old bioluminescent clothes lora. It gives such a nice magical effect.

[cerulean bane]

10/10

[slate narwhal]

thx

[cerulean bane]

https://tenor.com/view/pat-pat-pat-pat-head-head-pat-headpat-gif-2825247253179444134

[clever shore]

hey yall i recently got a gpu capable of running sd; used to use colab notebooks. are there any ui, setting, extension recommendations you guys have specific to anime type genning

lmk if there's a better place to ask that

[nova remnant]

You using Nvidia GPU?

[clever shore]

yea

[nova remnant]

Try forge

[clever shore]

what is that?

[nova remnant]

https://github.com/lllyasviel/stable-diffusion-webui-forge

[clever shore]

like a a ui fork?

[nova remnant]

Ye

[clever shore]

cool thanks

[nova remnant]

What is your gpu

[clever shore]

a 3060

[nova remnant]

Extension idk much, adetailer and auto complete would be the first ones I download

Forge comes with some like controlnet by default

[clever shore]

I see

[cerulean bane]

would.

[pure oyster]

She's 8ft in my story's canon.

[dire bridge]

woah buddy

[native halo]

so shes a meltrandi then

[cerulean bane]

that makes it even better.

[opaque bane]

Сат

[pure oyster]

Robo harpies? I approve

[pure oyster]

This is not midjourny, you have to do a local install.

[slate narwhal]

It was supposed to be robo angels ^.^

[pure oyster]

Try inpainting an eye on the chest emblem for a more biblically accurate style.

[junior quest]

a1111 tagcomplete and sd-webui-aspect-ratio-helper?

[junior quest]

prompt and model?

https://tenor.com/view/onegai-gif-4874664

[slate narwhal]

it's in the image,the prompt is

(street), (slender), (pale skin), (lady), //deep-red_orange hair, absurdly_long_wavy_hair//BREAK, //shiny-green eyes, eyelash//BREAK, one eye closed, nose, //pastel-red thin lip//BREAK, smile, (tongue out:1.2), (huge bre...sts), clea..g, bell-choker, //white frilled_bandeau, underbust, belt//BREAK, midriff, navel, //pink frilled_miniskirt, kitty-headphones//BREAK, looking at viewer,

The first one I didn't take note, the second and third are by puuzaki puuna, the fourth I 'didn't take note and the last one (angel) is by tachimi basue

[junior quest]

@wooden coral

https://tenor.com/view/cat-hello-cat-peek-cat-door-cat-peek-door-gif-17361470248875227911

[junior quest]

puuzaki puuna is a lora?

[slate narwhal]

No, all of these are direct prompts in the model. The model is 4th Tail; 4th Tail is IMHO (and after a lot of testing) the best Pony variant there is. 750+ artists, directly in the prompt with all the Pony flexibility and it just works (and works better than others like JS2prony, TponyNAI, Confetti, or the Autism line); no need to download Loras, only if you want though (and they will work with the model). Some artists work so nice that you don't even need to upscale.

Puuzaki puuna if you do portraits you don't need to upscale

and the results are really good

[junior quest]

woow thats a lot of detes, thanks

[copper star]

can we post NSFW stuff?

[slate narwhal]

No it's against the rules, read the pinned messages

[wooden coral]

[near elbow]

i find it nice how the AI insists on belly buttons in tight clothing 🙂

fun with fish eye lense

and miku, of course

[oak spoke]

omg the amount of artifact lmao

[gusty palm]

ya i decided to stop using that model lol

not totally blaming the model i got some more decent results but

[nova remnant]

Grabby

[keen hollow]

been away a few days

hope all is well

[oak spoke]

black and white is nice sometime

[junior quest]

skill issue

[vital raptor]

🤨

[stable crane]

damn this channel once again making me hungry

😄

[supple raptor]

[near elbow]

maybe use another sampler? and another VAE?

miku dolls, anyone?

or classical chibi?

[junior quest]

https://tenor.com/view/luffy-luffy-entrance-wano-luffy-entrance-badass-wano-gif-3541204707576213310

its W

respect

[steady grail]

what's that?

[grizzled moat]

what she peeking at ?

[gloomy scroll]

Checking if there is any shreks in sight so she can run away in time

[acoustic locust]

Siege gameplay:

[ivory pebble]

I think it's a Pony based model called 4thTail, something like that.

[acoustic locust]

Does anybody have an idea of when sd3 is coming out?

[ivory pebble]

I have no clue honestly, matter fact I don't think anyone knows

Some people have access to the beta version and they've been playing with that but as far as the official release goes nobody knows

[nova remnant]

sd3 when @lyric zealot

[lyric zealot]

when it's time

[steady grail]

let's let the baby grow up

[raven rivet]

Hello!

Do y'all know if NovelAI support LORA?

[lyric zealot]

as far as i know, NovelAi is closed source, and as such, you will not be able to include your own LORA

correct me if im wrong tho ^ (I don't actually use NovelAI so am just answering based on my limited knowledge on it)

[vital raptor]

so, after a few hours of figuring it out and getting it all setup, Unprompted is the way, thanks for suggesting it

so much better than wildcards and dynamic prompt

[cerulean bane]

@lyric zealot more bunny girls!!!

[lyric zealot]

not bunny but

[nova remnant]

Why

@cerulean bane

pout kinda cute

[supple raptor]

No, they have txt2img, img2img, inpainting, upscaling, and vibe transfer (which is basically a better version of ipadapter). Depending on what exactly you want to do with a lora, vibe transfer may be good enough to cover the need.

[clever shore]

how many sketches would I have to draw of an OC before I can train a Lora or sumn of it?

And what kinda variations would I need for good results

[nova viper]

a gril

[feral grail]

怎么用啊

[oak spoke]

lora training size varied from like 20 up to few hundreds

it depend really how single-minded you want the lora to be or how versatile

also for variation keep 1 thing in it exactly the same and make everything else varied so you get 1 specific aspect of it for lora

let's say for example you tryna make hat lora all image must be wearing hat but everything else can be varied

simple really

[supple raptor]

for variations, you would want a varied set of angles and ideally, diferent distantances so that you can generate the character from any angle and everything frrom full body to `upper body

[wooden sail]

Pony Diffusion XL is a different beast. Have to learn prompting again but it worth it

[fleet basin]

The Elemental Room, a sacred meeting place for the Guardians of Elementia City, - Marina Water Waters, a beautiful girl with blue eyes and long blue hair, wearing a sky dress representing the power of the water stars, standing in front of the window, looking at the city being destroyed in complete chaos and the buildings breaking, with a sad face and disturbed by the chaos in the outside. Pixar 3D

[gusty palm]

is contrast mult dependant on tonemap multiplier

[gusty palm]

which of these is best

nvm one has like 4x the stars

[floral wasp]

@floral wasp

hii

[gusty palm]

hiya

[vital raptor]

because i can set it up to use certain prompts only if certain other things get randomly picked

like, not having 'midriff' with a dress

it also supports recursive references, so one prompt tag can call one file that randomly calls multiple other files that each have their own things and possibly call additional files

it's like wildcards plus programming logic

[terse ore]

Any good X/Twitter accounts to follow that do unusual anime AI art? Like multiple characters interacting vs the usual single character portrait stuff for example

[nova remnant]

Sounds good

[fringe drum]

can you describe how you managed to get this?

[unkempt cape]

lmao I just wanna say whenever I do sd i always do interaction lol. like i did some hug pics a few billion years ago

posted that one here a while ago but i guess the view is still sorta generic

[terse ore]

Yeah no hate against single character stuff but it's super common. With interaction and consistency it's possible to make your own manga so I was interested in that aspect of content creation

[unkempt cape]

ye internet filled with mid stuff no wonder people hate ai art

[unkempt cape]

the view is still just front but i did like it very much

[terse ore]

I've been using blender with 3d anime characters and then use the screenshots to put in SD make it 2d style

[unkempt cape]

back then i just use prompt lol

it works it works

[terse ore]

Yeah. Now with Lora it's possible to make characters consistent. I just use 3d for that. Backgrounds are more inconsistent with just AI I think. But 3d to 2d trees don't look too good :/

[unkempt cape]

oh yeah. cant relate too much since the only thing i cared is pose i gave no shit to characters

[terse ore]

I see I see. You use just prompt to pose or controlnet also?

[unkempt cape]

so i have this cool style model i really liked but the prompt coherence is 0 so i use this more generic model and then controlnet

It also changes the pose just a slightly bit makes it fun to look at

[terse ore]

Ah yeah it's hard to get everything you want I guess. It's like steering a car with a stick

[unkempt cape]

its defo possible. I mean for what I did back then its basically 100% of my requirement done

theres this vroid depth method like you use vroid model or something to pose

and then depth controlnet but i havent tried

because eh

[unkempt cape]

oh yeah ai manhwa exist

[terse ore]

Yeah vroid is good. Ive found that using anime lineart preprocessor and anime lineart alone is pretty good, no need for depth or other stuff. I have a few test images, let me check

[unkempt cape]

theres this 1 manhwa about artists or smth. Its mid asf tho, because the art is generic and mugshot portrait everywhere

its boring ASF

unrelated ~~ wtf my tv is showing some chinese drama and theres this guy with garfield mask wtf ~~

[terse ore]

So, this was the input image screenshot:

Turned into this:

[unkempt cape]

Heres mine its still boring i guess but i love it lul

[terse ore]

Kinda messed up. Weird aura thingy around them. Maybe could be fixed by making an actual background. The white disembodied finders are because I didn't delete the extra faces around the fingers. Also it looks at the viewer because I used SDXL. And clipping because I didn't rig the skirt

[unkempt cape]

Maybe just maybe you could try explaining the pose to dall e

[terse ore]

I like that style you made. And that it has multiple characters. Different than the usual kind

[unkempt cape]

and then controlnet it

[unkempt cape]

yeah i fucking dig that style for some reason

[unkempt cape]

so that its less stiff

[terse ore]

I think 3d is faster because I can directly make the pose exactly as needed. Dalle would be faster if it's just 1 image but for manga/comic panels this way is faster. But yeah the stiffness needs to be adjusted, but this was just quick posing

[unkempt cape]

have you make images that werent stiff using this method? cuz its pretty hard i think

[terse ore]

Yeah they're not on my phone though. It just takes longer to make to get a more relaxed pose correct

[unkempt cape]

ok then this method is pretty viable then

also if ur serious on making a manga, try making the style not look like it was generated by ai

this weird mid point between realism and cartoon is frustrating to look at to say the least

[terse ore]

There is the other style I make with 1.5

This is the messed up version, but you can understand the style. The better version is on my PC

The 1.5 can transfer even the mouth and eyebrows, for the facial expressions.

[nova remnant]

look in dm's

[cerulean cosmos]

is there any alternative ways to download checkpoints from civitai? i cant log in and i rly wanna try the autism mix sdxl

[nova remnant]

I can download and upload for you on another site if you want

im not doing anything atm

[cerulean cosmos]

id love that

im with civitai support right now, for some reason i cant log into their website

[nova remnant]

which version do you want

[cerulean cosmos]

let me check

lol i cant even check them since they are nsfw

[nova remnant]

I think confetti is the most popular out there

I'll try that one cx

[cerulean cosmos]

i wanna use what hes using here = /

https://discordapp.com/channels/1002292111942635562/1091193032273043516/1222780561291087902

no wait, thats not right

[nova remnant]

what you mean

[cerulean cosmos]

anyway, yes plese do upload whatever version u see fitting

i appriciate it

[nova remnant]

5 minutes or something to upload

[cerulean cosmos]

thanks a bunch!

[nova remnant]

[near elbow]

needs more horns 🙂

[cerulean cosmos]

im trying to svd her right now with, not very good results lol

not sure what is happening

turns out im using a sdxl image for svd thats for 1.5

[cerulean cosmos]

@nova remnant hiya man, can u confirm what model it is ur using in the second image here please?

https://discordapp.com/channels/1002292111942635562/1091193032273043516/1221140359539851335

[nova remnant]

they are both using LightningBloomXL

v20

[cerulean cosmos]

i cant believe those are lightning, thanks !

[nova remnant]

np, i should add, i was testing some prompting cross mixing pony styles

[cerulean cosmos]

i just got access to sdxl so im trying some autism and stuff

[nova remnant]

autism is great

well autismmix confetti

[cerulean cosmos]

all the stuff i posted earlier is confetti, i love it

[nova remnant]

do you have a list of artist styles?

[cerulean cosmos]

no i dont

[nova remnant]

here is a curated list based on aesthetics, but there are bunch more that i still have to go through

[cerulean cosmos]

oh, nice

[nova remnant]

i wish there was a pony variant for realistic images

i love its prompt coherence and neat outline

[cerulean cosmos]

im trying some xy plots now, its intresting

unlike 1.5, its really stable

[nova remnant]

[nova remnant]

gimme this prompt

[cerulean cosmos]

im not sure i can

[nova remnant]

why not

[cerulean cosmos]

i think it was this

score_9, score_8_up, score_7_up, score_6_up, score_5_up, score_4_up, source_anime, BREAK,wednesday addams, horns, gentle smile, black hair, blush, yellow eyes, shiny, sparkes, from above

[nova remnant]

@cerulean cosmos another example, not sure i can explain it except for the prompt i used

[cerulean cosmos]

nice

lol i have so many things i wanna try out

theres not enough hours

[nova remnant]

good sword

[cerulean cosmos]

its pretty darn good considering its almost no effort into prompting

[nova remnant]

Ye

The thing I have with the confetti one is that it looks ugly

[cerulean cosmos]

oh u dont like it?

[nova remnant]

Not much

The default style at least

[cerulean cosmos]

i see

[nova remnant]

I like using the original pony and autism pony sometimes

[cerulean cosmos]

im trying to make a comic, so i would really need something a bit more basic

but its really clean

heres a autism pony

base pony with anime lora, meh

[nova remnant]

[nova remnant]

im not using any negative prompt

and very specific to positive prompts between realism and anime

same basic prompt but with minor changes to style specific prompt, in this case realism

[nova remnant]

metadata in img

[clever shore]

ive been p out of the loop for almost a year in ai imaging so i've been tryna find models/embeddings n stuff by looking around at ppls gens here. i love this artstyle do you mind sharing the img info if you still have it 😭

[nova remnant]

You can download that image and put it in your stable diffusion PNG info tab to see the prompt etc

[clever shore]

i tried that and it showed empty

[nova remnant]

Open in your browser and right click to save first

It does have info on it

[clever shore]

Oh I see I didn't know you had to open in browser

The last time I was here discord was still stripping it

[nova remnant]

It's SDXL

[clever shore]

Wait i tried that and it's still showing nothing

are u putting it in png info

[nova remnant]

I'm just in my phone, there are discord bots that can see PNG info

Well

Look in your dm

[lyric zealot]

well, because the bot is using stealthpng info

[lyric zealot]

do you already have the metadata?

[clever shore]

He sent me it

though i couldnt get it to work for myself

[nova remnant]

Sounds stupid

[lyric zealot]

well, because im using a certain extension called stealth pnginfo, and only those with that extension can view it

[nova remnant]

[lyric zealot]

unless i copy and paste from my folder, which i'm too lazy to do

[clever shore]

i see

[lyric zealot]

you stoopid

[clever shore]

i leave a couple months and i come back feelin like a boomer

[lyric zealot]

well, ai tech is moving really fast right now

and maybe if you wait another month, SD3 would be out

[junior quest]

meanie

[unkempt cape]

Half of the sd3 announcement is about safety so bleh

oh also human anatomy understanding is uh not good to say the least

[slate narwhal]

A few notes about this one:
Pony getting better and better. Prompting that transformed the 2d style by puuzaki puuna into a 2.5 plus with a very different look from the typical Pony, while keeping the original eye style perfect as it is in lower quality initial versions trained in the model. Got freckles right in her nordic viking face. Happier and happier with these results. Lots of fun.

[bitter latch]

hmm ok then

[nova remnant]

hey tora

[nova remnant]

lol i made an image some days ago about a cute cartoon character sheltering under a mushroom that got lots of likes and got featured on main page

[bitter latch]

nice

[nova remnant]

:p

[bitter latch]

now thats detailed!

I really should be in bed

[nova remnant]

lol are you using comfyui?

[bitter latch]

nope

[nova remnant]

how did you make that image?

girl with mushroom on her head... prompting?

[bitter latch]

Forge the branch of a111

lora:more_details:1> Chibi Miku girl sheltering under a Large mushroom in a spoor forest in rainy weather, large raindrops, out of focus background, focused sharpness up close, moss on wood, far view

[nova remnant]

heh

[bitter latch]

yeah it's a messy prompt

tired

[nova remnant]

naww. its supposed to be that way

miku is too big for mushroom lol

[bitter latch]

very far wide view, miku got lost

[nova remnant]

haha

[bitter latch]

Fail horror

[nova remnant]

texture and colors look pretty nice

[bitter latch]

ummmm what da fuk?

[nova remnant]

trippy !!

[bitter latch]

watches the currernt genarion and gooooes oooh

waits

sigh, hi-rez and upscaling ruined it. gah

there we go

1 more and i'm going to bed and pass out

[nova remnant]

over cooked

[bitter latch]

well I did use the extra detail lora x 2 😄

turns it down to .5

gotta go gotta go!

switches focus's

grins

hmm didn't change like I thought

hehe

there last one more anime and bed.. thud

[nova remnant]

thud for me too

ive been generating images for hours lol

[gusty palm]

is there anything special needed to run pony models or do they work similar to a normal sdxl model

ew why is it 6gb

[zinc flax]

Just run it like any anime models

Except now it's larger in filesize

[gusty palm]

im too cheap to dedicate that much hdd space

[unkempt cape]

first sd pic since fucking forever. had to learn stuff from basically scratch (no webui, going code mode 😀). Cant say im 100% happy but its ok

still using tech from dino age tho thats 4 sure
still swimming in artifacts. 🏊

discord pls for fuck sake

[gusty palm]

nice

[unkempt cape]

https://twitter.com/sagawa_gawa/status/1770767312964215039

さがわ/sagawa (@sagawa_gawa) on X

「今日もあの人見てるアル」
「ほらまた見たアル」

Twitter · Mar 21, 2024, 11:00 AM

HOW

HOW

THIS WAS AI

HOW THE FUCK

Legitimately asking if any one can recreate this kind of quality

[vital raptor]

define quality

there's nothing special about that image

[unkempt cape]

This guy is (assuming it was actually ai) the best ai img creator out there holy crap

[unkempt cape]

can u teach me how to recreate it

😭

[vital raptor]

like i said, there's nothing special about it, i don't know what you want to be taught

prompting for 2girls instead of 1girl?

[unkempt cape]

I want to know what I need to do to achieve a similar style

like as exact as possible. like i can put next to his posts and no difference

[vital raptor]

well, if it's style you want, then you have to find the model and prompt that is closest to the style

but that has nothing to do with quality or some secret technique

[unkempt cape]

yeah im searching for that on civit

but these images for sure are leagues above the stuff i usually see

[vital raptor]

i doubt you'll find it, people that take the time to make their own art don't use generic styles

[unkempt cape]

yeah but it looks so unachievable with anything we got rn

[vital raptor]

that person is either using a custom training, or a prompt combination that only they know, but you can use a model like autismmix and test different artists mixtures

[unkempt cape]

i will do some digging but heavily doubt its possible

[vital raptor]

i mean, they're using AI, so obviously it's possible

[unkempt cape]

by that i meant i find a way

[vital raptor]

the easiest way is to use something like autismmix or something else with a lot of styles trained and run images for every known artist name, then find the ones that match the closest

maybe mix a couple together to get what you want, it takes time, but it's not difficult

[unkempt cape]

is it that versatile

[vital raptor]

this #🍥｜anime message is the most complete list of artist names I know of

yes, that list has over 650 artist names

ranging from pencil art to pixar style

[gleaming flame]

this talk got me inspired... I really need to start doing something more serious with my life xD

[unkempt cape]

guys ur not creating good pics just by prompting 2 girls

[gleaming flame]

yeah, I guess... need 3 xD

jk

[unkempt cape]

am i really the only one who see how that stuff is miles above anything in this channel? (havent been around for long but eh)

no offense btw

[gleaming flame]

cool

checking that twitter account but for me it's just different style

not exactly the one I'd choose, but I guess it has it's charm

probably need to find out who is the main influence for this style, and use the model or lora that works well with that...

thin precise lines plus 3d glossy models

[nova remnant]

The qualities are very big

https://tenor.com/view/kok-cat-gif-25438805

It's not even that high quality

I mean literally

Zoom in it's super blurry

[junior quest]

NAI?

[junior quest]

symmetrical docking?

[nova remnant]

Bnuy

[lyric zealot]

it's easter sunday for most people right?

[nova remnant]

Doubt it's NAI

But it could be

Lots of artists in there

It all depends on the artists you choose for NAI anyway

[nova remnant]

I dunno

[zinc flax]

It's NAI

[nova remnant]

If sleepy says it

Then it is

[nova remnant]

Would

Eggs

[lyric zealot]

[zinc flax]

she's not amused

[ivory pebble]

Pony/AutismMix would be able to do the same if the model was trained on specifically anime images, but since the model was trained on a mix of stuff (Anime/Cartoons.. etc) it doesn't have the same base style. You can tell it's NovelAI since it has those soft colors/thin lines that a lot of other Japanese artists use.

Plus whoever that is they didn't even bother fixing the eyes. They didn't want to waste their anlas on upscaling either

[ivory pebble]

I didn't even know it knew that many omg, thank you for sharing this 🙏

[near elbow]

random anime

[acoustic locust]

What model?

[zinc flax]

pony diffusion anime

[acoustic locust]

thanks

[acoustic locust]

https://civitai.com/models/257749/pony-diffusion-v6-xl Is it different to this or is it this?

[zinc flax]

It's this one
https://civitai.com/models/315596/pony-diffusion-for-anime?modelVersionId=354017

[acoustic locust]

thanks

It's like steam lmao

Are there any special tags meant to make the image look better, or is it fine out of box?

[zinc flax]

I had to use multiple style loras to make it look good unfortunately

[acoustic locust]

For example?

[zinc flax]

Unreleased

[acoustic locust]

👍

[zinc flax]

Baked the loras just for myself lol

[acoustic locust]

are there any good ones that are released

[acoustic locust]

nice

[zinc flax]

They have tons of style lora you can choose
https://civitai.com/user/luuna/models

[acoustic locust]

i can't see anything on their page

nvm my filters were beefed

fixed it

[nova remnant]

@magic wave

[magic wave]

Catto :O

[magic wave]

Pog

[magic wave]

Hope you like em :o

[supple raptor]

They're most likely using something close to the super common artist list: artist:wlop,[artist:ATDAN],artist:ciloranko,{{chiaroscuro}} on NAI, with a maybe 1 or 2 more added.

[bitter latch]

well was suppose to be her in the easter bunnies lap but ok....

😄

[magic wave]

Karin from Fear & Hunger 2

[nova remnant]

what should she do with all those eggs?

[unkempt cape]

no bruh dont compare this to that

[native halo]

oh you mean something detailed and very coherent like a nai output ?

[unkempt cape]

about what makes those images good to me im still trying to figure that out myself

[native halo]

try to draw anything with your own hands and compare it to any model output,that should give u an idea

[keen hollow]

what's the question? that image style or something?

[unkempt cape]

ye

[keen hollow]

hmm

simple background, thin lines

doesn't really stand out other than that

says they're retouched, so maybe this person adds extra white reflection?

the two characters together is a lora ive definitely seen before

[native halo]

i think the model is animageXL

could also be pony with a style lora

[keen hollow]

probably, that's a popular model

trying to generate an example that i can actually post 😐

[native halo]

yea,not 1.5 tho because of the lines

[keen hollow]

looks like this person does a lot of tavern ai prompting

[unkempt cape]

the way eyes are drawn too

colors arent deep fried(??? idk english)

[native halo]

balls

[keen hollow]

style is harder than colors I think

[native halo]

yea colors can be edited after,style cant

[keen hollow]

lemme see

He is using NAI Diffusion Anime V3. (Stable Diffusion XL C1E1DE52)

mystery solved

[native halo]

https://tenor.com/view/chiyo-scream-chiyo-chan-azumanga-daioh-gif-26165611

now back to stealin pics

[keen hollow]

I have a prompt example for his work also. But I do not have a NAI subscription so I cannot provide examples. If you do, @unkempt cape , then you can make some. It's their new V3 model.

I am not sure what free models might also look like that one.

[unkempt cape]

wait what u actually found the model out? i checked and saw nothing

[keen hollow]

yes

[unkempt cape]

can i get the prompt pls

[keen hollow]

sure, but it wont help you without that model 🙂

[unkempt cape]

yeah just give me it pls thx

[keen hollow]

parameters: {{teranekosu, ciloranko, ariimaw}}, konya_karasue, ask_askzy, artist_ningen_mame, {{{"cute" + "Bunny girl" + "Black bunny ears" + "Dark black hair" + "Purple eyes" + "medium-large breasts" + "looking at viewer" + "Cowboy outfit" + "Holding revolver" + "Pointing revolver at viewer" + "Evil lines" + "Smiling" + "White background" + "full upperbody view" + "dutch angle"}}}, best quality, amazing quality, very aesthetic, absurdres Negative prompt: nsfw, lowres, jpeg artifacts, worst quality, watermark, blurry, very displeasing, Lowres, bad anatomy, bad hands, text, error, missing fingers, bad hands, bad fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, multiple views, (obese, skin folds, skin indentation, thin, skinny, ribs, small waist), anime style, flat color, crayon, cartoon, steam, fog, chibi, caricature, bloom lighting, (loli, children, toddler) Steps: 28, Sampler: Euler a, CFG scale: 5.0, Seed: 2371933763, Size: 832x1216, Clip skip: 2, ENSD: 31337

the artist tags appear to be at the beginning

maybe you can find a model that knows them

[unkempt cape]

thx bro

[keen hollow]

you bet

[vital raptor]

autismmix knows those artists for what it's worth

[keen hollow]

oh right on, maybe try that

[unkempt cape]

yeah thats what im planning too

gotta wait a few hrs tho run out of colab usage lul

[supple raptor]

animagine 3.1 is probably going to be a lot closer to the style

[keen hollow]

it's so different than the nai models though. so... i wouldn't get your hopes up

[unkempt cape]

i will hope lul

[keen hollow]

hmm ive got animagine v3... i can run a quick test

i'm terrible with this model though

[unkempt cape]

i mean from what i see of the nai model the output without artist tags are like the typical sdxl stuff

[keen hollow]

they add their 'secret sauce' so i dunno

[native halo]

that negative looks good specially on xl

[unkempt cape]

i search the model on internet and the example results look nothing crazy

so maybe its not that different but we'll see

few hrs later i will try that exact prompt on anything xl, animagine xl

autismmix not sure if on hugging face but if it is then that too

we'll see i have hopes

[keen hollow]

ignore the stupid image.. just trying style

[unkempt cape]

maybe crazy but i feel like he redrew the fucking lines

[native halo]

(obese, skin folds, skin indentation, thin, skinny, ribs, small waist)

[keen hollow]

running a 3060 here so... kinda slow 😄

[supple raptor]

probably does a few touchups here and there, or just plays the gacha until he gets a good pull, but the lines come out clean as-is for the most part with those artists on NAI.

[keen hollow]

there's gotta be a tag for that head-tilted-forward look

but yeah

that's as close as I can get

[supple raptor]

from side, head tilt, looking at viewer

[nova remnant]

Smelly feet

[keen hollow]

trying...

[supple raptor]

[keen hollow]

with full metadata

[keen hollow]

mm doesn't quite do it best I can tell

I need that deeper tilt... the almost.. menacing tilt

one more w/ meta

hope that helps @unkempt cape

[nova remnant]

(head tilt:1.8)

[keen hollow]

ok one more, thought it was cool

'til next time

[dire bridge]

til next time fellow friend

[supple raptor]

I'll just leave this here... https://huggingface.co/spaces/waifu-diffusion/wdxl-demo

[native halo]

no feet here

[dire bridge]

car crash

[nova remnant]

cute faces

[supple raptor]

bad image. no marisa toes.

[nova remnant]

correction, sweet faces

[native halo]

theres toes but hand merged so cropped

[nova remnant]

Image too detailed

Bad image

Had to crop this one hope no one mind it

[dire bridge]

I wonder why

[native halo]

https://tenor.com/view/i-saw-what-you-just-edited-i-saw-what-you-just-deleted-walter-white-muslim-walter-white-memes-gif-23994973

[nova remnant]

Hmmm because... The flower on the whole background was weird

So I decided to remove it

👍

This "Chinese dress, from behind" prompt doesn't look like it will generate something to post

[dire bridge]

perfect for you

agony and waiting and pain for us

[vital raptor]

unfortunately it doesn't adhere to prompts very well

[supple raptor]

wdym? it generates wd-xl every time?

[vital raptor]

but i specifically asked for short black hair

[native halo]

take your chinese dress

[supple raptor]

hmm.. probably skill issue

see, it can do short black hair

[nova remnant]

Needs more from behind

[dire bridge]

another car crash

[vital raptor]

that's not a car crash, that's a pillow fight

[dire bridge]

very funny

but you're right.

[native halo]

https://tenor.com/view/better-call-saul-meme-gif-468341328098124204

[nova remnant]

which model is it btw?

[native halo]

pony + 3 style loras

[nova remnant]

3 style loras is a name ?

or you mean its using 3 style loras

[native halo]

yea 3 different loras

[nova remnant]

ahh

cant guess which they are

the image has a good style, if some merges those loras with the checkpoint that would be cool

[bitter latch]

hhe

[supple raptor]

okok... the demo was not actually generating images, it was just retrieving them from a database. the actual model weights are out now though. https://huggingface.co/waifu-diffusion/wdXL

[native halo]

https://tenor.com/view/rdb-julo-barayan-pepiflux-martinz-gif-4151523561641641730

[supple raptor]

it's a real model

[near elbow]

that's not really SFW anymore 😉

shortish black hair and mischief 🙂

[tame dagger]

cat

[junior quest]

dress name?

[gleaming flame]

is this one nsfw? depends on the work probably... ;p

[slate narwhal]

I really like the contrast; very simple eyes vs the complexity of the rest. And the little colors it adds to the eyes in the makeup

[nova remnant]

I could make it safe if I worked with her

[nova remnant]

No way you can't guess it

[native halo]

whats this dress name? i wanna buy one

[nova remnant]

Tell her to take it off I'm buying right now

[gleaming flame]

any idea why I get those green artifacts? cfg is low

[nova remnant]

Which model you using

[vital raptor]

might be the sampler, or you need more steps

[gleaming flame]

tried a few samplers, sometimes it works ok, but sometimes can't get rid of them whatever I do

[nova remnant]

It's probably the model needing something more

[gleaming flame]

couterfeit v3 is the model

[nova remnant]

Oh yeah

[native halo]

prob a 1.5 vae issue,they sometimes show those corrupted dots on imgs

[nova remnant]

Go on the counterfeit page

I think it has a tutorial how to solve it

Or look at comments

[gleaming flame]

ok, thanks, it seems to be vae issue

[unkempt cape]

hey... any ideas on how to fix the faces? Like this kind of "<:D" face is the worst thing ive seen out of sd anime. NVM UR USING NAI IM STUPID

[nova remnant]

What does that mean

[native halo]

https://tenor.com/view/ez-noob-skill-issue-simply-a-difference-in-skill-gif-20780203

[unkempt cape]

idk how to english but eyebrows

its not //:D// but <:D which i hate

// for escaping chars

[nova remnant]

I summon @lyric zealot

Do you have any idea what that means

[lyric zealot]

[native halo]

skull issue

[nova remnant]

Frfr

[unkempt cape]

ok ill come back when i find words to describe it

[lyric zealot]

well, maybe show us an example of what you generated that you don't like?

[unkempt cape]

for example this pic (not mine): it doesnt look like a full smile it looks like a sad smile

[vital raptor]

try using happy in your prompt instead of sad

[lyric zealot]

that doesnt look sad to me. But maybe it's the open mouth smile that you dislike?

[unkempt cape]

pretty sure it doesnt have it but ok

[lyric zealot]

maybe can try "closed mouth, smile"

[unkempt cape]

is there a way to hide eyebrows all together?

like dont have it in the image at all

[nova remnant]

Maybe put it in negative

[lyric zealot]

might be possible if you put it in negative

but your quality tag might just make it so that it will always gen with eyebrows

[gleaming flame]

getting suggested vae didn't solve the problem, now I have spots in different colors xD but at least the green ones are gone

[nova remnant]

Grabby

[lyric zealot]

can you send one pic with pnginfo

[nova remnant]

Did you look in the model page

[gleaming flame]

yes, got the suggested vae from there

[native halo]

put what in negative :3

[nova remnant]

Clothes in negative

[gleaming flame]

huh, it seems I get those spots when I use this model for image2image

at least as I browse through recent files

[vital raptor]

just to check, when you say low CFG, we're talking under 7 right?

[nova remnant]

Use clip: openai/clip-vit-large-patch14-336

@gleaming flame

That's what it says in the model page

[gleaming flame]

I saw the patch, but not sure what to apply it to 😛

and what it does

[vital raptor]

it needs it's own clip and it's not merged in the model? might be better to just find a better model

[gleaming flame]

the model itself is probably the best for me

so I'll have to live with that, because as I understand the fix is for a1111

and I'm using invokeAI

[nova remnant]

I think if you throw that clip name on Google you'll find how

[junior quest]

cheongsam?

[junior quest]

that face, how did you get dat? any special expressions or smthign?

[nova remnant]

did you just curse me

its chinese dress

the one and only

stoopid

[nova remnant]

happy, smile

[junior quest]

the only chinese dress I know is cheongsam

[junior quest]

https://tenor.com/view/sad-cat-sunakook-tired-exhausted-gif-24687868

[native halo]

I prefer when theres no dress at all

[tidal phoenix]

One good hand, one wonky hand, still decent enough to share

[junior quest]

gimme dat dress name?

[tidal phoenix]

Wasn't prompted with a dress name XD

a beautiful swashbuckler rogue RPG character, detailed intricate clothes, layered clothing, (revealing:1.1) leather_armor, lowleg_pants, (japanese samurai:1.15)

Just kinda shotgunning character concepts right now for funsies XD

[native halo]

speaking of hands

[junior quest]

dress name? sundress?