218 xl

[floral mulch]

I'd need to do some xyz

[glass tulip]

it's loads better then either one of them though, in term of line tracing

[floral mulch]

I need to find out what styles work

or maybe the model can be finetuned idk

[glass tulip]

booting up the rng style generator

it spits these out :

[floral mulch]

can you

tell me how the rng style works?

do you have a file for thgat?

[glass tulip]

it's just a wildcard with a list of the a3.1 artists and for values its a list of 0.5 to 1.2 by 0.02 increments

it's not that fancy

should probably start by doing them one by one instead of mixing them

but maybe a lucky hit

i ll leave it runing for a while, i ll dump the results in a dataset in hf

from the few im seeing there is an effect but it's really minor

have you tried the pony's cropped artists tags on it ?

[floral mulch]

can u send me the value files

[glass tulip]

if the style changes a lot i know one of those 3 artist tag have an effect, most of them doesn't seem to have one atm

i ll try with the pony ones, they all look the same expect a few of them atm with the 3.1 tags

[floral mulch]

good idea

thanks

[glass tulip]

mmh it's worst with the pony ones, im not getting any change at all

but it seems to work with both sdxl and pony loras

this lora i made for pony works fine with it, just need to bumb the weight a bit

[floral mulch]

looks nice

I have to redo the weights

webui is . not ,

you can also do a step wildcard

for :: and :

how many steps after to add the tag

and how many to remove

[glass tulip]

we probably need the list of tags that actually works before that though

[glass tulip]

it's all the same

[floral mulch]

I dont feel like checking 1 by 1 tbh

but it might be the best way

I use models other than 3.1

maybe low lr finetune with like 50-100 of the best artists

would be a good thing

I've done a 5k step lora on 18k images before, altho it took like 6 hours

but with better and newer algos maynbe faster

[glass tulip]

did the lora turned out good ?

this might have potential, just need to find a way to not always make a child

([tianliang duohe fangdongye:0.9:14]),[(by quasarcake:0.01):5],[as109::15],

[floral mulch]

wtf does :0.9:14 do

also tianliang will always make a loli

[glass tulip]

adds "tianliang duohe fangdongye:0.9" at step 14 ?

[floral mulch]

not sure

especially with it being in ( )

would probably have to be like

([tianliang duohe fangdongye:14]:0.9)

or the other way around

[glass tulip]

i would do it the other way around yeah

[floral mulch]

[(tianliand duohe fangdongye:0.9):14]

[glass tulip]

[(hyde \(tabakko\):([to \(tototo tk\):natto soup:10]:0.5)::10],[(quasarcake:0.01,(tianliang duohe fangdongye:0.01)):23],[(tianliand duohe fangdongye:0.9):14],[(by quasarcake:0.01):5],[as109::15]
don't search for a logic, it just works

anything far gets nuked though

thats without any tags

the ones with tags looks a lot better

thats with bbwm-rotatetd11kosaani3ntdtpn4-comradeshipXL_v9-umbra_mecha.fp16-best-16bit.safetensors

doing the weight between steps mumbo jumbo change it more then if we just use a big weight on a single one

how did you get these eyes ? #1179290821409910844 message

[floral mulch]

hm, some of these are really nice

that's umbra mecha

I'll look into what algo is good for small scale finetune

[glass tulip]

it's massively inpainted

[floral mulch]

looks good at least

but tbh it looks like 99% quasarcake

[glass tulip]

yeah i inpainted with quasarcake

[floral mulch]

feel like the only styles that still work are the ones that were already strong

[floral mulch]

also lost char knowledge

nice pic

[glass tulip]

thats the original

they are just missing a little final facedetailer and maybe hand detailer too

do get a face with more style

[floral mulch]

damn

bg in 2nd one looks like umbra

so much shit going on

I think the fingers are deecent in this merge, at the cost of looking less stylized

[glass tulip]

i think its easier to get a style then decent fingers

cant get hands with a lora

at least not as good as a style

[floral mulch]

for me it's usually at this level

[glass tulip]

thats bbwm-rotatetd11kosaani3ntdtpn4-comradeshipXL_v9-umbra_mecha.fp16-best-16bit.safetensors ?

[floral mulch]

yes

[glass tulip]

what tag did you use to get these ?

oh i found it, it has metadata 👍🏻

[nixeu:quasarcake:0.5], [omone hokoma agm:fkey:0.3], [ciloranko|ningen mame|potg], tianliang duohe fangdongye, solo, long hair, alternate costume, best quality, masterpiece, absurdres ,

[floral mulch]

nixeu works really well I think

it's a realistic artist

[glass tulip]

can see the inpainted bits but already spent so long doing it while not even liking the scene

[floral mulch]

[glass tulip]

it's doing the thing again

[floral mulch]

💀

[glass tulip]

that fuking goat ruining it

[glass tulip]

it's such a blast inpainting with it

[floral mulch]

cute af tho

anyways if u want any particular artists in the future u can tell me here

[glass tulip]

im not too knowledgeable about artists, i like parts, like feets from this one, eyes from an other, head shape, etc.

would that be possible with a lora ?

im guessing each image from the dataset would need to be inpainted to match that

can't really zoom in and tag it without the rest

[floral mulch]

Ehh

I mean if you have a dataset of the images you like you can literally turn it into a lora

if they share common themes

the most important thing is to share common things

lucky I uploaded that last merge

cuz I just accidentally overwrote it

[wide beacon]

https://danbooru.donmai.us/posts?tags=ssambatea
https://danbooru.donmai.us/posts?tags=konya_karasue nice bloom & bg
https://danbooru.donmai.us/posts?tags=dino_(dinoartforame) aid has dino but clunky, animagine is undertrained
https://danbooru.donmai.us/posts?tags=kaamin_(mariarose753) abstract paint oil style
https://danbooru.donmai.us/posts?tags=renzhi00334233 cake
https://danbooru.donmai.us/posts?tags=akakura very detailed slightly desaturated
https://danbooru.donmai.us/posts?tags=dowman_sayman dowman
https://danbooru.donmai.us/posts?tags=matchazi sashimi hair (danbooru missing a lot)
https://danbooru.donmai.us/posts?tags=kyouda_suzuka dhruv?
https://danbooru.donmai.us/posts?tags=poireau_13 rice shower (danbooru missing a lot from pixiv)
https://danbooru.donmai.us/posts?tags=mitsubachi_(roundbee) low contrast (danbooru missing a lot from pixiv)
https://danbooru.donmai.us/posts?tags=aamond dope desaturated style
https://danbooru.donmai.us/posts?tags=kame_(kamepan44231) ᗜˬᗜ (dis a joke btw)
https://danbooru.donmai.us/posts?tags=shinagire_(sinanohaka) another cool one
https://danbooru.donmai.us/posts?tags=takarai_shunsuke cute chibi like
@floral mulch #1148401296597057546 message

[floral mulch]

channel u linked is called unknown

I don't know if I saved the thing lyn used for downloading stuff from danbooru

but I guess that dataset helper will be enough

[glass tulip]

this one ? https://github.com/Bionus/imgbrd-grabber

[floral mulch]

I used this to download but nah

bruhg where the fuck is it

I literally asked someone to link it

and then I didnt save it?

übraindead

https://github.com/Particle1904/DatasetHelpers

maybe its this

yea

[wide beacon]

yeah it's the dm

I wondered if it would work on your side

maybe #1148401296597057546 message

or maybe #1148401296597057546 message

[floral mulch]

these work

[glass tulip]

i love that model so much @floral mulch

the inpainting is just flawless

blends so well with anything, adds details without fapping about with samplers and works with pretty much all the lora

at least the ones i tried yet

[floral mulch]

I only tried 1 style so far

[glass tulip]

i could make these all day long, it can make some crazy stuff like sd1.5 with colors and shapes

stuff that pony can't

and the lines are just so clean

like there is now way i can make anything that even looks like this on pony

[floral mulch]

wtf

cute af

[glass tulip]

it even works with my shitty spring rider lora that barely works on pony

and knows characters like animagine

[floral mulch]

interesting

wait you tried pony loras and they worked?

or

[glass tulip]

yeah

i made these with pony base

https://civitai.com/models/461926?modelVersionId=514049

the effect is less, but it's there and i can crank the weight a lot higher without it going nuts unlike pony

i didnt bother uploading the spring rider lora because it's really finicky to use on pony

[floral mulch]

interesting

[glass tulip]

i trained it with the ponyv6 base

and it has the knowledge of supports like watercolor, hand-drawn and stuff, those got lost with pony

"watercolor"

pony needs lora to get that now

like you got a beast of a model here, at least its what it feels like to me

[floral mulch]

need to test it more

I'M checking the one I overwrote it with

used same models in it

[glass tulip]

it even does rollerblades with accurate wheels

i had to cheat with editing to get that reflection

and still did it wrong, legs are the wrong color

[floral mulch]

no pantyshot

damn thats sick tho

[glass tulip]

i fixed it

but those rollerblades

https://tenor.com/view/wario-nintendo-perfect-justright-meme-gif-5013133

[floral mulch]

:pag

send it in ai art

[jaunty vortex]

this is the new rotated model?

[floral mulch]

the rotate part was like 2 months ago

its just random merges

[floral mulch]

just realized cutesexyrobutts is kinda like the base pony style

[floral mulch]

https://huggingface.co/Enferlain/218xl/blob/main/test/bbwm-rotatetd11kosaani3ntdtpn4-comradeshipXL_v9-umbra_mecha.fp16-best-16bit2.safetensors alternate ver

maybe better maybe worse

@glass tulip actually not bad with latent nearest exact 2x .6 denoise

[glass tulip]

oof x2 thats big, i havent tried that, what about model upscale to x2 and then karras at 0.45 ?

[floral mulch]

whats that

u can only have different stuff for base and hires fix

this will just add more noise

not much difference

prob needs latent upscale

[glass tulip]

i injected noise into the latent, this was not the way

weird effect though

[floral mulch]

there's already a lot of noise

not as much as umbra but still

ur schizo prompt is weird

yep, lot of noise

[glass tulip]

how many steps ?

[floral mulch]

this is step 1

[glass tulip]

thats your highfix res no ?

[floral mulch]

no

its just the default gen at step 1

[glass tulip]

oh, wait when are you upsccaling the latent ?

from the start ?

[floral mulch]

no there is no upscaling

this is just a gen at step 1

to show how noisy it is already

[glass tulip]

oh, i tried injecting noise into the latent between the gen and highfix res

[floral mulch]

umbra at step 1

[floral mulch]

guess masterpiece is causing these art gallery type gens

[glass tulip]

i dont think masterpiece is a quality tag for sdxl

was for sd1.5

a3.1

[floral mulch]

yeah

just using it cuz used to it

did they use aesthetic for 3.0?

[glass tulip]

this merge is from 3.0 ?

[floral mulch]

ye

[glass tulip]

how does it know so much character ? it also has some 3.1 and pony inside ?

[floral mulch]

I think 3.1 added some anime ones but also caused a lot of game chars to be forgotten

the 3.0 finetunes I used had a few more characters each

[glass tulip]

oh i didnt know, i never tried 3.0 i was stuck with sd1.5 back then

hands became a mess

[floral mulch]

saj

3.1 was like an answer about ppl complaining about lack of anime chars

and nai artstyles

[floral mulch]

[glass tulip]

euler a for gen and highres then dpm++2m karra for inpainting

added a 0.05 film grain between the gen and highres

but i don't think it's that great, i normaly add the noise with blender, i think it looks better

[floral mulch]

blender?

[glass tulip]

and adding noise before inpainting probably not the best idea

[floral mulch]

isnt it for 3d shit

[glass tulip]

also for scenes and lighting

i got the workflow from AntlersAnon, it's the one he uses for his gen

[glass tulip]

with the blender noise

it's normaly for higher res images so it looks a bit too much on this but it gives it that artbook scan look

and it helps blend the inpainting marks

[floral mulch]

ah

that effect

[glass tulip]

it also makes the image takes like 10% more space

[floral mulch]

not a huge fan of it after everyone started using it tbh

[glass tulip]

i liked it to hide the inpainting marks back when i was using sd1.5

but the node on comfyui makes it add a bit more shadow and texture when placed between the gen and the highfix res

adding gaussian adds a depth blur effect

[floral mulch]

this artist sequence actually looks kinda like vils nai gens

[plucky dove]

long ass legs

or torso

who knows

[glass tulip]

normal / euler a normal with 1px noise
karras / dpm++2m karras with 1px noise

[floral mulch]

karras is more scuffed

[glass tulip]

i like the last one the most

[floral mulch]

1st one for me

[glass tulip]

it has less contrast though

[floral mulch]

vs

euler is more smoother for details

[glass tulip]

mmh true

i usally inpaint the face though

normal looks better here too

has that little stiches

but it also makes hairs out of thin lines like eyelash but not on this model

karras does better on messy models

and messy loras

[floral mulch]

cuz karras is messy

I use it on aid

[glass tulip]

but the details on this

each strands of string

[floral mulch]

[nixeu:quasarcake:0.5], [omone hokoma agm:fkey:0.3], [(hyde \(tabakko\):([to \(tototo tk\):natto soup:10]:0.5)::10], [(tianliand duohe fangdongye:0.9):14],[(by quasarcake:0.01):5], [as109::15]

this is so schizo

[glass tulip]

it's 2 together, forgot to remove one at some point and liked how it looked

im lazy to type them out, so i put them on little nodes and connects the ones i want

at some point 2 stayed connected and that was the result

how much tokens does that even take ? i dont see it on comfyui

[floral mulch]

1girl, [nixeu:quasarcake:0.5], [omone hokoma agm:fkey:0.3], [(hyde \(tabakko\):([to \(tototo tk\):natto soup:10]:0.5)::10], [(tianliand duohe fangdongye:0.9):14],[(by quasarcake:0.01):5], [as109::15], solo, very long hair, full body, best quality, very aesthetic, absurdres

60/75

[glass tulip]

[floral mulch]

I like the heads in this tags

[glass tulip]

almost a full one, that explains why i need to place my triggers before it

[floral mulch]

arknighs lookin ass

[glass tulip]

because of the skizo prompt i can add details in without it changing too much other stuff

[floral mulch]

late af

gotta sleep

[glass tulip]

same, gn 👍

[floral mulch]

spent all day genning coomer shit

no more gens

[floral mulch]

continunig the merge but keeping this model

not satisfied with the loss of artstyles

doing 1 more pony merge then aid

then probably back to umbra then finetune

[glass tulip]

itss insane with lora

sdxl/pony doesnt matter everything works

pony lora adds more details because of extra noise

that last frieren one is 3 loras, 2 for pony and 1 for sdxl

[floral mulch]

I'll try some loras too

btw

I uploaded an alternate version to this

can u take a look whenever, just better seeing someone else try stuff

cuz I dont know how to compare

[jaunty vortex]

I downloaded vers 1

by yuming li by blackbox (blackbox9158) by quasarcake by dino miku hatsune 1girl solo close up profile from below dynamic pose realistic dark aqua background bare shoulders rainbow blue hair expressionless faux traditional media floating clothes gold earrings very long hair nose painterly beautiful color detailed aesthetic amazing quality

prompt was for artiwaifu originally

artiwaifu with that prompt

it's a bit off bc of the 2x hires probably

[floral mulch]

arti lot more aesthetic

feel like a lot of arststyles just dont work in the merge

[jaunty vortex]

yeah

this isn't bad though

[floral mulch]

it's not bad but it kinda has that base modelish vibe to it

like bland

happens when u merge different models

I'm planning to do a lokr finetune

[jaunty vortex]

I think it tends to add stability but average things out

You get a model that's really good at making a few pictures

Artiwaifu can't do dark teal background for some reason

[floral mulch]

the strongest styles stayed but the weaker ones are completely gone

wish there was some way to check what styles actually work at normal weight

as in alter the overall style

not just change parts

[jaunty vortex]

just prompt them all ig

[floral mulch]

[jaunty vortex]

200 hours later

[floral mulch]

and then u still gotta check 1 by 1 manually

[glass tulip]

there was an extension that made heatmaps of words from the prompt

[floral mulch]

ohh

I'd like that

one day I'll make a merge method based on token efficiency

[glass tulip]

but it was for sd1.5 and it was a while ago, idk if it works for sdxl

[floral mulch]

I'll try this now

[glass tulip]

https://github.com/toriato/stable-diffusion-webui-daam

[floral mulch]

ah

[glass tulip]

[floral mulch]

this is similar to cross attention visualizer or something

[jaunty vortex]

https://danbooru.donmai.us/posts?page=1&tags=yuming_li here's a style I'd like to make a lora for the colors

[floral mulch]

guess I'll add it to the list

https://danbooru.donmai.us/posts?tags=ssambatea
https://danbooru.donmai.us/posts?tags=konya_karasue nice bloom & bg
https://danbooru.donmai.us/posts?tags=dino_(dinoartforame) aid has dino but clunky, animagine is undertrained
https://danbooru.donmai.us/posts?tags=kaamin_(mariarose753) abstract paint oil style
https://danbooru.donmai.us/posts?tags=renzhi00334233 cake
https://danbooru.donmai.us/posts?tags=akakura very detailed slightly desaturated
https://danbooru.donmai.us/posts?tags=dowman_sayman dowman
https://danbooru.donmai.us/posts?tags=matchazi sashimi hair (danbooru missing a lot)
https://danbooru.donmai.us/posts?tags=kyouda_suzuka dhruv?
https://danbooru.donmai.us/posts?tags=poireau_13 rice shower (danbooru missing a lot from pixiv)
https://danbooru.donmai.us/posts?tags=mitsubachi_(roundbee) low contrast (danbooru missing a lot from pixiv)
https://danbooru.donmai.us/posts?tags=aamond dope desaturated style
https://danbooru.donmai.us/posts?tags=kame_(kamepan44231) ᗜˬᗜ (dis a joke btw)
https://danbooru.donmai.us/posts?tags=shinagire_(sinanohaka) another cool one
https://danbooru.donmai.us/posts?tags=takarai_shunsuke cute chibi like
https://danbooru.donmai.us/posts?page=1&tags=yuming_li sab - here's a style I'd like to make a lora for the colors

[glass tulip]

these colors are trippy to look at

[jaunty vortex]

it's similar to nathan fowkes colors I think

[jaunty vortex]

I'd add him but I need to find hires pics

[floral mulch]

going for around 50-100 artists I think

first try

[jaunty vortex]

https://danbooru.donmai.us/posts?tags=pino_ko smaller artist but also aesthetic

https://x.com/yeejchan/status/1730748615143096768

Yeej (@yeejchan) on X

侧颜
#初音ミク #HatsuneMiku

Twitter · Dec 2, 2023, 12:40 AM

I have a lot of artists saved

https://danbooru.donmai.us/posts?tags=mura_karuki

[floral mulch]

very cool

[jaunty vortex]

https://danbooru.donmai.us/posts?tags=siuroi&z=2 / https://www.pixiv.net/en/users/35271351

[floral mulch]

same

[floral mulch]

this image is awesome

[jaunty vortex]

https://x.com/iotaectoplasm https://danbooru.donmai.us/posts?tags=iotaectoplasm

白川柏川 (@iotaectoplasm) on X

工作邮箱[email protected]
ins: https://t.co/UktaGnS73G

Twitter

https://www.pixiv.net/en/users/529590 https://danbooru.donmai.us/posts?tags=starshadowmagician

https://x.com/pix_bun https://danbooru.donmai.us/posts?tags=pixiescout

Airi @ Fanime 1513 (@pix_bun) on X

I draw a bunch!!!! Principal Artist @ Riot | Illustrator イラストレーター| ✉️[email protected]

Twitter

https://x.com/ponn_mame https://danbooru.donmai.us/posts?tags=ponn_mame

ぽん豆° (@ponn_mame) on X

✰ illustration 🇯🇵￤色と少年と.￤✎オンライン講座(https://t.co/cBI2HNHMsH )￤All rights reserved.

Twitter

[floral mulch]

cool stuff so far

ty

[jaunty vortex]

https://x.com/BUspuzPMrropHwh

唠兽 (@BUspuzPMrropHwh) on X

🚫禁止AI

Twitter

these are all pretty much my personal taste

https://x.com/Kana_bgyjs

(@) on X

#illustration #原神 #八重神子 #GenshinImpact #genshinimpactfanart

Twitter

https://www.pixiv.net/en/users/31383108

https://x.com/nat_nim_

NAT (@nat_nim_) on X

나트/ 콜로소 coloso : https://t.co/XANO5xTHBQ / 문의 [email protected]

Twitter

https://x.com/mumu53_

黄晓牧mumu (@mumu53_) on X

Xiamen University （2025毕业

Twitter

https://x.com/SuteinuA

尾粥 (@SuteinuA) on X

成人済み松尾です、絵を描いたり描いてなかったり | ねことハムスターがすき | commission closed！ 絵の無断転載・使用（AI含め）は禁止です| うちの子 #ナースロボ_タイプＴ

Twitter

https://x.com/washanapple

摸鱼斋 (@washanapple) on X

中国人/prohibit reproduce without permission
伊吹鸡腿子代理搬运号。
请不要盗图。
不接推广，不接广告。
中文主页地址：https://t.co/1pObtDh8Tg
https://t.co/yRifqNqMFI

Twitter

[floral mulch]

[floral mulch]

oh animagine was supposed to have this

[jaunty vortex]

I wonder what artist 3.0 was finetuned on bc the poses are the best (at least in terms of foreshortening, dynamic)

https://x.com/se_5eeeee

anyway as you can see I like girly styles with lots of color variation

https://x.com/rei_17?lang=en might as well add rei 17 https://www.pixiv.net/en/users/74184
https://danbooru.donmai.us/posts?tags=rei_(sanbonzakura)

零 (@rei_17) on X

中文/English/簡単な日本語おk 基本的にPF専用アカウントです。 魂音泉/ジョジョ/FGO/Mahabharata R18垢： https://t.co/BUcrquPD2N

Twitter

https://danbooru.donmai.us/posts/3851948?q=parent%3A3851948 I remember seeing this years ago

[wide beacon]

Post has 5k messages

[floral mulch]

[floral mulch]

guess u were right

pony loras work maybe?

@glass tulip

[floral mulch]

if u can send me some pony loras to test that'd be nice

[jaunty vortex]

I tried model 1 a bit more, the base model vibes really are strong

maybe I'm just prompting wrong

by quasarcake by dino at the beginning of the prompt did something

model 2 same prompt same seed

I think the quasarcake influence is weaker in that 2nd pic but the top of the hat makes more sense

it doesn't know zumizumi anymore (or mizumi zumi)

1st model keeps doing this to background flowers

[glass tulip]

they work for me yes but the effect is not as strong, and i just tried training one using bbwm-rotatetd11kosaani3ntdtpn4-comradeshipXL_v9-umbra_mecha.fp16-best-16bit as a base

and it's really good

i need to do more testing

[jaunty vortex]

counterfeit lora for pony on 1st

some doujin pony lora

[glass tulip]

they feel usable to me

[jaunty vortex]

yeah seems to work

[jaunty vortex]

I managed to mix loras trained on sdxl and pony and get something decent on 1st model

[floral mulch]

Does the first one seem better

[jaunty vortex]

Didn't try second as much

https://www.pixiv.net/en/users/62834074 also another good artist https://danbooru.donmai.us/posts?tags=potg_(piotegu)

[nocturne cosmos]

if you guys want some good art style, could checkout who I am following https://www.pixiv.net/users/13633275/following

[jaunty vortex]

Lol I see ctd on there

Good list

[wide frost]

hm, isnt it better to do "by niwa_uxx" instead of "by niwa uxx"

[floral mulch]

doesn't really matter

nothing uses underscores though

so I won't either

[wide beacon]

extra token for no reasons

[wide frost]

i see

[floral mulch]

need to find lyns config for booru scraping

[hidden matrix]

So what is bbwm-rotatetd11kosaani3ntdtpn4-comradeshipXL_v9-umbra_mecha.fp16-best-16bit2.safetensors supposed to be?

[floral mulch]

same thing as bbwm-rotatetd11kosaani3ntdtpn4-comradeshipXL_v9-umbra_mecha.fp16-best-16bit1.safetensors just different alpha/beta

[dawn pecan]

This model has a problem with the output image in Fooocus, what is causing this, the style ticker has been unchecked in Fooocus, and the output in stable-diffusion-webui-forge is fine!

[floral mulch]

I have no clue, never used focus before

it's just a test model

[dawn pecan]

Maybe this model is special~~

[floral mulch]

surely

[glass tulip]

it works fine with a fooocus sampler in comfyui for me

[floral mulch]

Dunno what that means

[wide frost]

sharpness

[glass tulip]

https://civitai.com/models/463446/katorthoma
"pony" "score_9,score_8_up,score_7_up,score_6_up is not necessary"

[floral mulch]

already saw a few models claiming that

you dont need score tags if u merge in a shitload of loras or finetune for 1 artstyle i Guess

[wide beacon]

looks so bad tho

[floral mulch]

22 it not enough for this

ok lets test this shit

fresh out of the gemini schizo oven

and already cooked up another method

training an adversarial network for alignment of pony into animagine

or whatever

discriminator whateverthefuck

gemini schizo idea

ok this mergei s fucking dogshit

[floral mulch]

`by morikuraen by nardack by reiko, hina (blue archive), (beautiful color,detailed,best quality,amazing quality,very aesthetic:1.5)

blurry,lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract],[[ai-generated]],((sketch)),unfinished`

awa

[jaunty vortex]

pretty good

[floral mulch]

it stylistic

but same issue as aid

too stylistic = not consistent

weird fingers and details

[jaunty vortex]

yeah I guess it's somehow overfitted

[floral mulch]

https://civitai.com/models/391659/ponymagine-91 literally just animagine skull

[nocturne cosmos]

heart of apple v3 is out

[wide beacon]

[jaunty vortex]

Watch me release 0.98 animagine 0.01 pony 0.01 awa amazing merge

[plucky dove]

seems pretty good so far

[floral mulch]

yep

seems actually good

[quartz elk]

it really is and the guy is going to continue training it

[plucky dove]

what settings u guys run on it

ive just been doing 28 steps 7 cfg

[floral mulch]

I do 28 steps 8 cfg euler a

[wide beacon]

My settings are so weird

[plucky dove]

cute

[floral mulch]

is there a

wildcard for animagine for characters?

like character, copyright

in this format

@wide beacon do u know about any

[wide beacon]

There literally is one in the Animagine 3.1 hf repo

Sorted by how many images of them are in the dataset

[floral mulch]

huh

found it

its just characters

wait thats what i saked about

fuck you

thanks

[plucky dove]

Quasarcake ebifurya hard to get to work like it did in awa

Ebifurya is overbaked as hell

I did like 0.01 weighting and it still there

[floral mulch]

I dont ven know wha ebifurya looks like

gonna try

[plucky dove]

Uh

[lethal sierra]

I couldnt handle artist style mixing in heartofapple

[floral mulch]

thats concerning

[lethal sierra]

as if only 1 artist at time is ok

maybe my skill issue, but I prefer new umbra, so multiple artist tags is ok

[floral mulch]

I havent tried heart of apple yet, just looked at the civit images

eh

[plucky dove]

nice

looks nice

[floral mulch]

too many ways to combine artstyles

[plucky dove]

so mixing seems ok

guess i need to change weights around

[floral mulch]

its not heart of apple so dunno

but I'd think mixing works on any model with prompt editing

[plucky dove]

what does the colon do?

[floral mulch]

:: removes after x steps

: adds it

thing:thing:0-1 means at what part they switch in the steps

0.5 is half

thing|thing is swapping every step or thing|thing:2 swap every 10 steps etc

[plucky dove]

oh ok

that seems p useful actually

thank u

is 218xl a more trained umbra

[floral mulch]

no

its an older model of aid merges

[plucky dove]

is it good? i will try it later

[floral mulch]

its not bad, older aid versions are still some of the better ones for artstyles

but dont know a bunch of stuff

like almost no char knowledge

or nsfw

[plucky dove]

oh i see

[floral mulch]

uploading umbra on civit if u guys feel like u can add pics whenever if u have any lying around, would be cool

[floral mulch]

__user/charanimagine__, [__user/ai3artists__:__user/ai3artists__:0.3], [__user/ai3artists__:14], solo, long hair, alternate costume, best quality, very aesthetic, absurdres

[floral mulch]

https://civitai.com/models/477988 4 months old model

[olive onyx]

did you delete this :?

[floral mulch]

never uploaded it

just to hf

[floral mulch]

huh

got unpublished

autistic website

republished

[lethal sierra]

maybe, but imho h of apple is wrong. look at recommended neg prompt for it, this is ridiculous. wall of text. (worst quality:1.8) bruh bro are you serious? 💀 umbra managed to be better in general

[plucky dove]

yea that wall of negative prompt is crazy

works ok tho

[floral mulch]

They used (worst quality:1.5) in their previous recommendations so it's probably just messed up cuz language, but yeah either way model quality seems a lot worse than v2

[nocturne cosmos]

Any new good model? I broke my leg last week.

[floral mulch]

wtf?

how

and nothing really new, no

working on merge methods for like a week +

[floral mulch]

lost to a basketball?

seems like it can do touhou chars maybe

[floral mulch]

hm cant decide if 1 or 2 is better

1 quasarcake:cutesexyrobuttss:0.9
2 quasarcake:cutesexyrobuttss:0.9

1 2

1 quasarcake:cutesexyrobutts:0.4
2 quasarcake:cutesexyrobutts:0.4

I'm thinknig 2 might be better tbh

even though it has big boob bias sometimes

[floral mulch]

if u feel like it sometime u could run these 2 and see which one is better. both aren't really great but they could be good ingredients

just dunno which one to use

ugh

really don't like the ungabunga boob bias in 2

would be easy to fix tho if it really is better than 1

hmm

I guess these models are more arti

by aid artist tags seem to work

a lot better compared to non

gonna try some aid prompt

nope

by iwzry vs just iwzry

1

2

quite different huh

@nocturne cosmos so apaprently these 2 merges are more like arti than umbra/animagine

as far as artist tags and chars go

adding by iwzry instead of just iwzry

[nocturne cosmos]

missing the good old aidv6

[floral mulch]

same

dont see a reason why it couldnt be merged tbh

almost literally works the same exact way

no adetailer

doesnt look like yogisya otherwise, missing the schizo hair

ok yea this is literally just a weaker aid with characters

[plucky dove]

2 seems better, i do see the big boob bias

[floral mulch]

it's not that bad actually

just needs different prompting slightly than I'm used to

but gonna work on merging it back into what I'M used to

in a way that still works

[plucky dove]

looking forward to it 🫡

[wide beacon]

It’s so funny seeing imi evolution

(old hf repo readme)

[heady echo]

Beautiful

[olive onyx]

@floral mulch

https://civitai.com/models/270238/animat-background-v1

could you try to extract the difference between this and animagine as a LoRA and see if it would help in generating better backgrounds ?

[wide beacon]

You can do that on StableSwarm

There’s a tool

[olive onyx]

I know but I was lazy and didn't want to download the model :p

I'm almost out of space on my harddrive

[wide beacon]

A

I’ll try to do it later then

[olive onyx]

thanks

[floral mulch]

the pics in the link don't look that different from stuff u can do already tbh

[jaunty vortex]

Aesthetic

https://openmodeldb.info/models/1x-Manhwa-Null I wonder if filtering out some of jpeg compression could help with some styles

[wide beacon]

https://civitai.com/models/301208?modelVersionId=338284

other Chinese dude also calls it Ziyun lol

[wide beacon]

https://huggingface.co/AshtakaOOf/lora-extract/blob/main/sdxl/AnimatBG_04.safetensors

Lora extract of Animat 0.4 from SDXL 1.0

[olive onyx]

sank you

I will test it out

[frail zealot]

imi left server lol

[wide beacon]

buh

wat

[plucky dove]

Wuh

[olive onyx]

was he banned ?

[frail zealot]

he's on here on another alt

wait no he's not

what

[wide beacon]

rip post

[olive onyx]

Did they leave or get banned?

I still don't know

[wide beacon]

Left

[olive onyx]

Reason unknown?

[wide beacon]

Not unknown just vague

[olive onyx]

great

rip bozo

(I'm going to hell)

anyways you're probably gonna come back to this thread one day so I just wanna say the umbra model was pretty good, so good job on that 👍

[hidden matrix]

Which one are you using?

[olive onyx]

umbramecha

[hidden matrix]

And then there's umbra_mecha.safetensors itself.

[olive onyx]

I think it was this one

[wide beacon]

he's back

btw

[olive onyx]

https://tenor.com/view/xqc-gif-18046692

when umbramecha on SD3 ?

[wide beacon]

when anime finetune on SD3

[olive onyx]

i hope linaqruf can get the creator license

so we can get peak fiction

rn It's just a waiting game

[wide beacon]

as if yodayo cares about licenses

[olive onyx]

rn I'm just waiting for someone to finetune ANY kind of model so we can see if the finetuning results are good or not

#pleasefinetuneSD3Boobs

[wide beacon]

Someone made an anime LoRA but that's about it

in a week there will be more things

[olive onyx]

was it any good ?

[wide beacon]

clearly no

[heady echo]

@floral mulch reinviting you to the thread

[wide beacon]

he was already here

your server preferences are saved for a while before they are cleared for leaving

[heady echo]

but when i added him he showed up in the list

[wide beacon]

maybe because you pinged him?

[heady echo]

ping will add to thread

[wide beacon]

he did react to this when rejoining the server

[heady echo]

oh

then idk

[wide beacon]

@floral mulch please back me up

[floral mulch]

you can react without joining

[wide beacon]

why would you not be in the post when rejoining?

[floral mulch]

dunno

if no one sends msg it closes auto

[plucky dove]

Whale cum back

[wide beacon]

[floral mulch]

I'll merge latest aid and 218xl weakly into this test to maybe make the artist tags a bit stronger

[floral mulch]

found out why the model was unmergeable

token_embeddings had a lot of inf

gotta fix it

also gonna do something about shitty backgrounds

[plucky dove]

Looking forward to it 🫡🫡

[floral mulch]

looks like the pose is merged from b to a in these

c is not used other than b-c so no influence on the looks

[floral mulch]

[open thunder]

I love this

[languid cliff]

Waggy waggy mofu heard theres very rarely floofy furballs here

[floral mulch]

https://civitai.com/models/282341?modelVersionId=629577 @frail zealot

In 0.4.5 500k pictures from dataset have been captioned with Claude 3 OpusVision, then captions were pruned and converted to hybrid with booru tags. In combinations with accurate approach to TE training, this significantly improves prompt control and allows some funny/useful things like prompting multiple characters without extensions (much stable then before), referring to their traits or outfits, using natural text for more accurate description and so on

[jaunty vortex]

looks pretty accurate to the artist styles

[plucky dove]

😳

Can someone check quasarcake

Oh wait they have the artists on rentry

[wide beacon]

@thorny ridge give it a try too

[thorny ridge]

Oh this is a new release? I haven’t even tried the previous one properly yet, looks really good

[floral mulch]

best pony model I tried so far is one of the mix jems from gem knight

shit goes hard

Epoch 2/2, Step 1500, Loss: -2040.0000

[floral mulch]

OI'm merging tghat

fuck tponynai

shits ass

looks so bad in comparison

[jaunty vortex]

tponynai is kinda too round

[gritty lynx]

holy let's go

ah sorry for ping

[floral mulch]

[plucky dove]

Which one u like best?

[floral mulch]

the latest one

qxsm or something

[jaunty vortex]

wtf is qxsm

oh I found it

[floral mulch]

mymixGJem_xlsQromew

[jaunty vortex]

https://civitai.com/models/138257?modelVersionId=589294

I wish there was an sdxl model like the T2 model shown there

[floral mulch]

like detailed bgs?

prob just need to merge niji/bg loras and then turn up the noise

but wont be that easy since it's 1.5

[shadow trail]

Hello

[wide beacon]

just tried and it's kinda mid

[floral mulch]

f

I've been trying to make merge methods for 2 months now

[wide beacon]

@floral mulch This looks pretty cool from the samples https://civitai.com/models/564727/leekai

[jaunty vortex]

merge of animagine finetune kohaku and anime finetune of pony, I see

https://x.com/kranny_el/status/1800027046284042244/photo/1 hmm

Kranny (open comms!) (@kranny_el) on X

#inART

Twitter · Jun 10, 2024, 4:47 AM

[floral mulch]

easist shit to larp as an artist in 2024

[jaunty vortex]

can draw him upside down, pretty crazy for AI

leek

chimera vs leek

[wide beacon]

I think we could take some parts of leek

[jaunty vortex]

leek is gonna be better at kanae bc of holodayo

[floral mulch]

wtf is leek

[hidden matrix]

Yeah, artist styles don't seem that strong.

[wide beacon]

So many artists are missing too

It’s not worth using

I’ll try LeekAI later

[floral mulch]

did some change for bayesian merger (does anyone even use this still) so it saves best score merge always

[nocturne cosmos]

Neta v2 is out

[floral mulch]

[languid cliff]

Noone will notice mofu wagging around to spread happi!

[floral mulch]

did you try it yet, and is it good

[nocturne cosmos]

not really. POE new league out, don't have time to generate from now on.

[floral mulch]

path of exile andy

[floral mulch]

the noise on neta

[wide beacon]

Yep they trained the old AidXL way

[floral mulch]

spent the last 1+ month

maybe 2

worknig with a scuffed merge code

like critical error

[wide beacon]

@floral mulch what's the merge you did on that leekai test?

bbwm-testrotatetraindiff11-leekai-sd_xl_base_1.0-best-16bit

[floral mulch]

just some random test

has glitchy faces

last pic shows

[wide beacon]

neat

[plucky dove]

looks nice despite the glitchy face, do u plan on uploading this or gonna wait til u get something u like?

[floral mulch]

I meant glitchy literally

[wide beacon]

oof

can you do a new umbra mecha with updated models?

[floral mulch]

like which ones

from the ones I used only heart of apple had a new versino I think

but it was whatever

if not worse

[wide beacon]

holodayo leekai and stuff

[floral mulch]

yeah

good idea

and maybe some loras for uniqueness

[wide beacon]

can do an umbra mecha lora extract

[plucky dove]

oh dang

[wide beacon]

There are a few more extracts now https://huggingface.co/AshtakaOOf/lora-extract

[wide beacon]

#ai-art message

sink recipe

merge animagine and umbra

yep, just put 0.4269, 0.4269 instead of 0.5's in the MBW
I use the supermerge extension auto111 for merging

idk if it’s real

[floral mulch]

💀

[floral mulch]

meh

[jaunty vortex]

We came a long way from 1.5

[wide beacon]

#ai-art message

[wide beacon]

Oh ye in classic sink fashion the recipe was fake

[olive onyx]

I did not lie, the merge was animagine + umbra

I merged in some lora's afterwards

[wide beacon]

Which means you didn’t share the recipe

Except if you count the first line as sharing the recipe

I’m still tilted tbh

[floral mulch]

just do the same thing

[feral elbow]

anyone recreated sinks model yet

[wide beacon]

he said it was the recipe but he merged LoRAs on top

[wide beacon]

.

[floral mulch]

just merge loras to it

[plucky dove]

when will sink share recipe

[olive onyx]

I need someone to help me with making the model better

then I will release

maybe\

[plucky dove]

🫡 good luck

[olive onyx]

flux will be better than anything currently available in 2 months anyways

[plucky dove]

i have no idea how to deal with models

[gritty lynx]

what does better mean

[olive onyx]

it can't do furry futa yet so that means it's incomplete

[nocturne cosmos]

can I use flux locally or no?

[floral mulch]

if u use my model for furry I'll send you a c4

[floral mulch]

24gb

gpu

[nocturne cosmos]

i guess I have

[floral mulch]

same

[gritty lynx]

we need 1bit quantize flux

[feral elbow]

nice

[floral mulch]

he left

[floral mulch]

you do it

[feral elbow]

why

[gritty lynx]

i aint qualified

[plucky dove]

Where did bro go

[gritty lynx]

he went for milk

[feral elbow]

day 367 of asking @olive onyx to share his model

[olive onyx]

if I leave the server sometimes, just assume I needed a break

that's usually what it is

[plucky dove]

welcome back

[wide beacon]

replace o with u

[plucky dove]

Welcume back

[floral mulch]

break of 1 day

[olive onyx]

I am goated like that

[wide beacon]

leaving this here https://rentry.org/noiseinjection

The LeekAI dude used wdv's noise to enhance stuff

[floral mulch]

Does it work in a1111

Nope

[feral elbow]

day 27373 of asking @olive onyx to share his model

[floral mulch]

just make ur own

[heady echo]

and never release it so ppl can ask you

[feral elbow]

i dont have a gpu good enough to merge models

not to mention we dont know what models he actually merged

[feral elbow]

ok i dont care anymore

hes not going to share it

[floral mulch]

you dont need a good gpu to merge...

that's like the one thing you can do even on the worst gpu

or without one

he merges loras

just find good interesting loras and you can make unique shit

kinda defeats the purpose of a general model for me but I'm also interested

[plucky dove]

Merging requires more cpu ram right

[floral mulch]

ideallly

depends on what merger you use

the better ones use like a few gb at max

[plucky dove]

Very nice

[floral mulch]

https://huggingface.co/yodayo-ai/clandestine-xl-1.0

I'll try this

current test is a bit too far from animagine outside of artist and characters

[wide beacon]

looks very weird

[floral mulch]

looks fine to me

[wide beacon]

from the images in the readme

it seems very saturated and there's weird stuff in the composition

Misato has a cross on her chest?
Kana has a heart inside her heart shaped fingers?
Bloom is weird?

[floral mulch]

hmm

I think its a finetune on

nai images

from pixiv

[plucky dove]

it definitely looks like it

seen that artist mix an absolute ton

https://aibooru.online/posts/72134

artist:j.k., artist:blushyspicy, {{{artist: yd orange maru}}}, artist:Cutesexyrobutts, artist:redrop, realistic, 2023,

this

[floral mulch]

doesnt look too bad

[plucky dove]

https://aibooru.online/posts/69327

[floral mulch]

tough to say which one I like more

my fav is probably the smooth texture ones

[plucky dove]

all the same mix

[plucky dove]

so this doesnt have artist tags?

[floral mulch]

no clue

havnet tried it yet

[plucky dove]

interesting

[floral mulch]

its based on kivotos so it probably does

[plucky dove]

will check it out

i just wonder what its going to do since the base style looks like that

[floral mulch]

this one

[jaunty vortex]

not bad

[plucky dove]

nice coloring on the second one

those from pixiv?

[floral mulch]

ye

the 2nd one is a very popular style

maybe I should just train it

[plucky dove]

Can you send account of second one. Gonna try to see if can replicate on local

[floral mulch]

might do a style training for it, tons of solo images cuz of how much they are farming

[jaunty vortex]

Damn

It's a good style, although I can't name any artists that do it even though it's so popular 😂

[plucky dove]

its definitely a mix so unless they accidentally upload without scraping the data we wont get it

[floral mulch]

they didn't, these ppl alll spam them for selling

its ok lora training will work

[plucky dove]

😂😂

Do people really pay for this stuff

Crazy

[floral mulch]

yep

[plucky dove]

Time to get nai sub

And make bank

[floral mulch]

I made lik40 bucks last year and the pics looked like shit

💀

[plucky dove]

Wow

Just open a patreon or just on pixiv fanbox

You can get nai sub to pay for itself

[floral mulch]

ofc

but who the fuck wants to spend the entire day genning images for other ppll

sounds miserable

[plucky dove]

Cant u automate it somehow

Or will nai ban u

Def not worth it lol

Just go do something productive

Or hang with friends

[snow elk]

who's that on pixiv?

[floral mulch]

user/joemama

qaq or some shit

[floral mulch]

[plucky dove]

You can do it! Fighting!

[gritty lynx]

you dont need a lot of ram nor gpu to merge models

1gb ram is more than enough if the input models are safetensors

you can merge flux with 1gb ram only

[floral mulch]

[nocturne cosmos]

I hear that sanaexl is good

[floral mulch]

Link

[nocturne cosmos]

https://civitai.com/models/647664/sanaexl-anime-v10?modelVersionId=724572

its mid

[wide beacon]

new version https://civitai.com/models/286574/swampmachine

[plucky dove]

Firefly!!

[floral mulch]

ye got alrleady

hgavent trried tho

tell me if its good

[hidden matrix]

mignon is a good artist to test. Don't know of any models better than https://civitai.com/models/611079/nai3likediffusion
Those two models didn't do so well.

[nocturne cosmos]

aid does that well

[hidden matrix]

which model is that again?

[nocturne cosmos]

https://civitai.com/models/124189/anime-illust-diffusion-xl

[hidden matrix]

not that accurate.

[nocturne cosmos]

no way

[nocturne cosmos]

this is very accurate to me

[hidden matrix]

This is what mignon is supposed to look like.

[nocturne cosmos]

that is very close, the skin texture.

[floral mulch]

and face

[plucky dove]

Seems close to me as well

[nocturne cosmos]

that is actually a bad example, V8 is doing better than V61

[plucky dove]

@hidden matrix

4th tail

[hidden matrix]

Yeah, 4th Tail is great (and also has those artists), but needs a few anatomy improvements. Falls behind compared to newer stuff.

[olive onyx]

not really accurate

[wide beacon]

imo 4th tail is a big waste of time

[floral mulch]

it has animal ears and tails tho

did a 8 hour flux lora train, shits not worth localyl

[frail zealot]

u can just download the 4th tail lora and pray

[plucky dove]

Idk it's pretty good at recreating artist styles

And it's only at v0.45

[frail zealot]

nah

ash is just hating on pony

[plucky dove]

lol

[wide beacon]

I was training SDv1 LoRAs for 12 hours on my laptop

The heck you mean not worth it

That’s just forge not having the fp8 lora fix

[wide beacon]

I haven’t gotten anything aesthetic or good stylistically out of it

[plucky dove]

I don't think that's the main point of it (it's literally called hentai model)

Hopefully it will get better with future versions

[wide beacon]

Yeah which is why it’s a waste of time

Hentai is not what image gen should be about

[plucky dove]

Porn advances

[wide beacon]

It doesn’t

[plucky dove]

Isnt there that thing where blender literally came to be because people wanted to animate porn of their favorite character

Feel like that can apply here

Nsfw loras for flux came super fast

I get that you want to use it for nice images though

I just don't think it's fair to call it a waste of time

[floral mulch]

what fix

I meant the training is not worth it, makes ur room hot and I don't like having to wait that long to see if my settings work

[wide beacon]

To make LoRAs in fp8 have as much strength as on fp16 by adding a bit of gaussian noise

[floral mulch]

yeah thhat would help

[wide beacon]

Fried looking ahh