🍥｜anime

[nova remnant]

definitely not going to fit

its done lol

the oven has big opening to its right

True

[nova remnant]

Top part looks awfully round >.>

btw im using a model that i merged with 3 different other anime models :p

[nova remnant]

How do I apply to become a kit kat

just be

dont think

once you know that secret you can be the kitkat

i wonder when nubby will finally start playing with sdxl

im not going back to sd1.5 now

in fact this is my a1111 checkpoint now

[supple raptor]

I have a bit. I'll do it more once there's a model that's as good as 1.5 models at existing characters.

[nova remnant]

you know yo8u can make a model thats good

especially with the speed you get now using sdxl you can train w/ far less effort

[supple raptor]

Quality wise, sure. But if I can't make all my touhous without needing a bunch of character LoRAs then I'm just gonna end up back on 1.5.

And I don't have the energy to figure out training XL.

It's way harder to train than 1.5

And I barely even train 1.5

[nova remnant]

you're looking at sdxl from dev pov, i was looking at it from user pov, and i kept going back and forth with gunuine dislike for sdxl for how slow and lacking it was but that all changed with lcm lora and turbo

[supple raptor]

At the resolutions I gen at it's never been much slower than 1.5 for me. XL slows down less at higher resolutions than 1.5 does and basically everything I gen is 1080p after highres fix.

[native halo]

because u only use it for generic characters,for anime characters theres no loras for XL

[nova remnant]

i think i'd be honest to say i now get better outputs with sdxl in less time than with sd1.5

[native halo]

now make Momonga on XL without a lora

[nova remnant]

idk what momonga is and why should i?

[native halo]

prob cause u are in an anime channel?

[nova remnant]

so?

i dont have to make all that anime character but given the curiousity im sure people can do it

[native halo]

sure,when they can train the thousands of loras that 1.5 has im sure they will switch

[nova remnant]

sd1.5 didnt happen over night

and you'll be ignoring the dev process of sdxl if you dont acknowledge the progress its making

[nova remnant]

Booba is all you need

[native halo]

yea they been sayin that for what? 6 months maybe they are waitin for A100's to drop to $600 for them to be able to train

[nova remnant]

so you are upset cause of the cards?

i know i can't have a100

[supple raptor]

I hate needing character LoRAs for moderately popular characters. Even if every character I wanted to gen had an excellent XL LoRA, I wouldn't be happy.

[native halo]

no im actually happy a100'S are worth like 5k

[nova remnant]

haha

ok well i think sdxl has taken a turn towards the right direction with all thats going on lately

[nova remnant]

Why

[native halo]

maybe SDXL turbo is cheaper to train ?

[nova remnant]

idk about training nubby should get involved as a dev

[supple raptor]

No, you don't train a distilled model. You train the normal model, then distill it.

I'm no dev, I just hang around a lot of people who are.

[nova remnant]

@native halo where are the body suit gens

itd be nice to encourage nubby to get on sdxl and we might see more progress coming

[native halo]

no bodysuits today 😔

[native halo]

he would probably join once we are able to train on a consumer card🌚

[nova remnant]

i like how you feel for the average user, im one too

[native halo]

thats why we need a new img contest with a couple of A100 as price for the winners

[nova remnant]

sure but i'll stay away from that, i dont like competing

these anime characters are cuter than from sd1.5

i tried generating girls playing tennis on sd1.5... i got that horror of hands

Hands

look at these aesthetiic hands and legs on sdxl

all the sd1.5 models are now idling in my backup storage

Floating net x3

the pole is invisible

but its there

you just haev to touch it

High tech stuff

[nova remnant]

Very safe for a sport

the hula hoop is bit tricky

you know i was attempting for sd1.5 quality with sdxl cause the overly glossy sdxl images didnt feel natural, now with the options to tweak it i think sdxl is giving lot better results

the sdxl coherence is a big and major point over sd1.5

which got even more fine tuned with lcm lora

and at a speed you dont get excuse to not use sdxl

[nova remnant]

💤

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1179757468507721799/00036-3832953082.png?ex=657af1c3&is=65687cc3&hm=af7579706fd8f3ec396c491d0352439bf4e05888a5d1dc1a338bfd85723a9ef4&=&format=webp&quality=lossless&width=447&height=662

[gloomy scroll]

seems like some else generated something similar to what I got from wildcard

[hearty raven]

https://www.phixiv.net/en/artworks/111756708

[gritty spruce]

made of light works

[nova remnant]

the light was weird for me, i wanted a solid plastic circle

[lyric zealot]

everyone

[nova remnant]

smol sus i made the cross over

no longer using sd1.5 lol

[lyric zealot]

[nova remnant]

im seriously glued to sdxl 🙂

[lyric zealot]

is sdxl-turbo really that good?

[nova remnant]

well im using lcm lora more

but yes they are both great

[lyric zealot]

on auto?

[nova remnant]

render times is very short and great output quality

yes auto

[lyric zealot]

you have 12gb vram right?

[nova remnant]

yes rtx 3060

[lyric zealot]

that's why

[nova remnant]

i was trying for nubby to play with sdxl too

[lyric zealot]

i only have 8gb vram, and sdxl is bad on auto with 8gb

[nova remnant]

with his knowledge he can do some nifty stuff with it

oh ouch

i thought you had 12gb vram too

[lyric zealot]

oh nubby fully swapped over to SDXL as well?

[nova remnant]

no he hasnt lol

[lyric zealot]

sadly not

but soon

[nova remnant]

nice 🙂

[lyric zealot]

probably getting 4090

[nova remnant]

oh that'd be very good

is that 16gb?

[lyric zealot]

24gb

[nova remnant]

oh wow awesome

[lyric zealot]

then i can probably gen SDXL stuff with speed like how im genning SD1.5 models

[nova remnant]

yep absolutely

youlll love it

[lyric zealot]

for sure for sure

[nova remnant]

i had some difficulty adjusting to sdxl before

[gritty spruce]

turbo on my 4090 is doing like 3 frames x second.. at 1 step, doesn't look great like that but merging better models and going 5-10 steps you are fast and looks really cool

[nova remnant]

1. slow render time
2. outputs are kinda generally too glossy

[lyric zealot]

people are now getting live generation on 4090 and it's crazy

[nova remnant]

after merging some anime models i've overcome the glossy texture

[nova remnant]

wyz thats why im using lcm lora more .. 8 steps, 2 cfg

[lyric zealot]

tbh, the greatest perk sdxl has, it's the context and knowledge

[nova remnant]

totally thats the reason i kept switching back and forth and finally settled with it lol

nice

great place to heal minds

[lyric zealot]

im actually surprise that they did a SDXL turbo and not SD1.5

[gritty spruce]

they finetuned sdxl at 512px

[lyric zealot]

wait, then maybe i can finally run sdxl on my auto

[nova remnant]

you get better results with 768 and 1024

actually emad made a post saying 512 would give a bit weird effect for turbo

[lyric zealot]

hmmm, for sdxl model, it's usually using refiner and not hires fix right?

[nova remnant]

nope

[lyric zealot]

[nova remnant]

no refiner needed unless you want to train a model with base and refiner

and i cant do hires with sdxl since i need more ram

[lyric zealot]

that means you are getting pics at like 768x768

[nova remnant]

but even with 1024 or 1500 images are good w/o hires fix

[lyric zealot]

or 1024x1024?

[nova remnant]

both

i stay around 768 to 1500

[lyric zealot]

[nova remnant]

but usually 768x1024

something along this 768x1344
832x1216
896x1152

[lyric zealot]

i will stick with sd1.5 for now

and wait for gpu upgrade before trying

[nova remnant]

yeah better that way

that way you will get the full impact w/o getting slowed down

[lyric zealot]

[nova remnant]

a lush island with cascading waterfall inside a glass jar with tiny king kong inside

lol how do you like the context

[lyric zealot]

it's great, it's so much better than 1.5

[nova remnant]

yep

[lyric zealot]

i think dalle3 is on another level tho

i think

[nova remnant]

dalle has great context too but far less control to certain aspect

i feel i can use sdxl any way i like unlike dalle3

[lyric zealot]

true, and the restriction on dalle3

[unreal ridge]

No monke 😔

[nova remnant]

token limitations then some silly censoring

[lyric zealot]

from my very limited testing of dalle3, it's context knowledge is crazy tho

[nova remnant]

neat 🙂

[unreal ridge]

seems like 1.5 doesn't know what king kong is...?

[lyric zealot]

kinda cursed but here with sd1.5

[nova remnant]

some sd 1.5 models can do it

i think nubby used one last time we were rendering jars lol

[unreal ridge]

Man, novel AI new model looks neat

[nova remnant]

how do you access it?

subscription?

[unreal ridge]

yes

[nova remnant]

bleh

[unreal ridge]

Cursed

[nova remnant]

oh man .. king kong back there

and look at thos tiny people below on the shore

smol sus what you rendered and this one with jile .. sd1.5 has its perks

and with better coherence with sdxl you can be more creative

btw i was using t2i controlnet models an hour ago

they are very small in size and works great

compared to the 4gb or 2gb CN models

i made extra 10gb space on my ssd after removing all the sd1.5 files

[lyric zealot]

[nova remnant]

an elephant standing on its hind legs, and reaching out for an apple with its trunk from a tree above

i had <lora:lcm-lora-sdxl:1> (anime:1.2)

or i could make it more realistic

[nova remnant]

lcm lora weight set to 1 was also bit high

[lyric zealot]

injuries

[nova remnant]

oh no

I'll bring her to my house

to treat those injuries

👍

[lyric zealot]

surely it's to treat those injuries

@wooden coral @gritty spruce ^ #🍥｜anime message

@dull jackal @finite herald #🍥｜anime message

[gritty spruce]

OUT! thank you!

[lyric zealot]

[nova remnant]

@gritty spruce another one

[lyric zealot]

[nova remnant]

[finite herald]

ty!

[gritty spruce]

the spambots are crazy today

[lyric zealot]

[lyric zealot]

they are slowly activating their script

[gritty spruce]

thank you

[lyric zealot]

more inc, maybe

[pure oyster]

Tactical girl time?

[lyric zealot]

i literally just changed my prompt

[nova remnant]

cute

[unreal ridge]

is it just me or the png files are kinda blurry on discord now?

[lyric zealot]

can't really tell if there's a difference

[dull jackal]

I missed ye old sammich, didn't I?

[nova remnant]

lol sorry had to do it

[twilit portal]

angelic bride 🤍 🤍

[nova remnant]

looks like realistic game world 🙂

[lyric zealot]

[lyric zealot]

hopefully this isn't nsfw

[nova remnant]

very cool, i wonder how you got to do that

[lyric zealot]

here, with metadata

i should just delete the old one since it's so close

[nova remnant]

thanks going to look into the data

[lyric zealot]

[nova remnant]

i thought that involved some complex lora and controlnet trick

[lyric zealot]

i'm a simple man, i type my prompt and press generate

?

[nova remnant]

used this line white hair, long hair, japanese yakuza, bare back, full back dragon tattoo, piercing, head tilt, traditional house, upper body,

and got this

[lyric zealot]

lmao

i guess the yakuza, they take it as the game?

[nova remnant]

lol idk

[lyric zealot]

or maybe cos didn't specify gender

[nova remnant]

a girl with bare back, she wears a dragon tattoo on her upper back.

[nova remnant]

I'll take it

2 please

[lyric zealot]

[nova remnant]

[lyric zealot]

take what

[nova remnant]

Her shirt

And the beach vacation

👍

[lyric zealot]

[nova remnant]

Balloons

[primal wharf]

at 0:14: Pop these balloons like a silly ||REDACTED|| https://youtu.be/mPMwrnmjJGo?si=Qg-GF9Eed-zyULIO&t=14

Vietnam flashback

[nova remnant]

[primal wharf]

1:49 shout out to my dad but I'm fucking adopted

The sentence mixing is nice

and 1:22 - Moo moo, cents 50

xD

[zinc flax]

huge balloons

[gloomy lodge]

things in jars big fan

[dawn jasper]

Those women are quite strong. 😮

[primal wharf]

When you time travel and everyone looks at you:

[tulip mist]

Does anybody have an idea how can I take any anime art and change character’s face similar to the real person?

I’m using web ui

[waxen plaza]

loras, make or find a lora of said face if possible

then use the lora and inpaint the face onto the character, or try to generate the face naturally, but id rather generate what i want and replace the face for guaranteed results

[tulip mist]

Thanks skelly. What if I want to use my own selfi and make a like anime photo out of lora. Can I do that?

[waxen plaza]

well lora are mini models basically, so you can use an anime based model, and use a lora filled with a bunch of your faces and that should do the trick, youll probably have to tinker some things as anime style means a loss in details im pretty sure

editing your faces on the anime images would jusrt reinforce the style, so if you wanted the specific body of a char then perhaps

[tulip mist]

Basically saying, I wanna generate naruto for example from Lora and then change face structure so it will look like me referencing my selfie

I tried doing it with controlnet, doesn’t work well or I’m doing it wrong

[waxen plaza]

im not sure on the specifics, im still trying to learn how to make loras, id ask in #🤝｜tech-support probably

i just know basic stuff

[tulip mist]

Got it, anyways thanks for the help

[stable crane]

heyo

[stable crane]

so cute

[native halo]

speakin of shirts 👕

[tulip mist]

what model did you use for that?

[supple raptor]

He used this one https://civitai.com/models/211628?modelVersionId=238399

[dawn jasper]

That is a witch lady.

[supple raptor]

yep, he used one of my prompts for the cover image.

I did lightning and fire themed ones too.

[dawn jasper]

Witch I like, really.

[twilit portal]

Yeah 😂

[dawn jasper]

@twilit portal If they are well-endowed, they are definitely strong in a physical and/or magical sense.

[nova remnant]

Slurp

[native halo]

needs (open mouth,tongue out:1.5)

[steady grail]

so cursed

[twin river]

swx

sex

[vivid creek]

Greetings, i am new to SD und i created a image with 904*512 - but i want it to be in a 21:9 format (at the End 3440x1440)... but when i change the format of the image, a hole new image appears... is there a way to get the same image but in 21:9 ?

[nova remnant]

fingers getting so good you actually have nail polish on her nails, now make bare foot girls...

you'll be making proper fashion models

[native halo]

ill prob try with feet later once i done with the 100 open mouth imgs

[supple raptor]

I don't remember if I ever posted this one here

[native halo]

perfect feet

[supple raptor]

It had minimal edits to get to that point.

[nova remnant]

editing is not allowed, i mean sure but the point is getting the sd gen to a level of perfection

[nova remnant]

those open mouth are bit scary

like what if she bites?

[native halo]

its ok it doesnt hurt

[supple raptor]

Both feet were already perfectly formed. Her right foot just had an extra blob of flesh below the toes and her left ankle had an extra bump.

[nova remnant]

thats what i mean

in my sd1.5 saga i rarely got perfect foot, i haven't bothered much with sdxl

but sneakers were always perfect

the sdxl model i use doesn't go below knees

[supple raptor]

Feet are significantly tougher than hands on every model. Many images are portraits which don't show feet, and outside of specifically foot focused images the ones that do show them are full body shots which means the feet are small and then on top of all of that, most images that would have feet, have shoes or at least socks/stockings on.

[nova remnant]

it may not hurt only cause she looks cute

[nova remnant]

yep

[native halo]

ahh

[nova remnant]

also cute

if she bites it'll feel like a butterfly landing on your shoulder

[native halo]

this one came out with bad teeth 😔

[nova remnant]

come to the other side... SDXL

[native halo]

https://tenor.com/view/the-what-gif-26849967

[nova remnant]

i say that with considerations for low end gpu btw

[vestal crypt]

https://media.discordapp.net/attachments/1041682785418629125/1180060639851462677/00183-1430382811.png?ex=657c0c1d&is=6569971d&hm=53df0ba833c2b49bd999288441638be9ae33f75f1d55aefb41949d333b147b1f&=&format=webp&quality=lossless&width=447&height=662

[nova remnant]

there are people who may need to use sd1.5

i dont wanna be insensitive to their fun

[native halo]

i probably would but i need my characters 😔

[nova remnant]

i miss some of the sd1.5 models

cause of the character facial look

[nova remnant]

perfect lil cute vampire

you'd actually want her to bite you i think

she is so cute if she can't bite you'd wanna give her fangs

anwyays lol... im just goofing around

[native halo]

we have to check fangs first

[nova remnant]

do you have a gpu for sdxl, just asking

i think nubby has one

[native halo]

just a 1060 i can run the turbo one and i dont mind waiting but i dont use it cause it doesnt have all the anime characters that i prompt

[nova remnant]

but he likes sd1.5 girls more atm

[nova remnant]

yes thats a bit of lacking with models now, which is why turbo and lcm lora are important to encourage developers to contribute on models

i got some pretty epic faces with realistic photos using basic sdxl model i merged

and it works both on anime and realism

[native halo]

dont look at neck

[nova remnant]

but you know what.... the coherence with sd1.5 and sdxl are vastly different

sdxl adheres to context so much better

i think with sd1.5 there are lot of accidental beauty that makes up for randomness of the image

but with sdxl we are getting to point to only speak our thoughts

[nova remnant]

beautiful pair of hands to match her face

[native halo]

of the 100 imgs this model got like 65% good hands

[nova remnant]

sounds fair lol

actually that percentage is way bit more fair compared to what i got with most sd1.5

[native halo]

probably depends on the model and also pose,the hands on own mouth pose is prob too hard for sd and causes deformities

[nova remnant]

surely cause even with letting sd generate pose loosely i got maybe 30% good images out of 100

that image you posted just now, even her knuckles look cute

[native halo]

ill post some of the bad ones

[nova remnant]

im wanting to do some image gens but im too distracted by youtube and asmr videos

[native halo]

there are more but ill only post a few

[nova remnant]

yeah most of my pic looked like that on sd1.5 ..

constant delete was part of my workflow lol

and some of the images the face came out so cute it was difficult to delete the bad hands

[native halo]

yea when the img is cute but hands are deformed 😔

[nova remnant]

come to the other side .. 🙂

you guys need to populate the other side

[native halo]

try to do a few ones with hands on mouth on XL

[nova remnant]

cute miku with her mouth open

im using untrained model, its generic

but the model you are using is fine tuned

for anime

[gloomy scroll]

I saw marisa feet I already knew who posted it before looking user name

[nova remnant]

btw with sdxl you may need to narrow in on the context some times but its almost less likely to get deformities

for example here im not getting what i'd want and i'd have to fine tune the prompt, but the image itself in general isn't bad

[native halo]

try with (own hands on mouth : 1.3)

[nova remnant]

seductive girl with side ponytail with her mouth open try with (own hands on mouth : 1.3)

first one

let me gen a couple more to see how it goes

slight anomaly but nothing to harsh

3rd image is better of those

about fingers i mean

[native halo]

yea all 5 fingers

[nova remnant]

the other thing about sdxl base is that its not too focused on anime and the resources it takes is hard to train anime models

but so far with recent developments things will likely change 🙂

try this cute miku playing electric guitar, she is excited with joy.

i'd guess the fingers will go nuts on sd1.5 when playing guitar lol

i had such difficult time msyelf when i was using sd1.5

[native halo]

im doing the 100 imgs for (feet,from below:1.3)

[nova remnant]

lol thats hard work given sd1.5

[native halo]

🦶

[nova remnant]

hmm that actually isn't bad at all

except a few more fingers

toes

[native halo]

one feet came out with 6 fingers 😔

[nova remnant]

yeah

what was your prompt?

[native halo]

deletes her

[nova remnant]

let me run that on sdxl

[native halo]

from below,1girl,solo,hatsune miku, soles,barefoot, feet, looking at viewer,

[nova remnant]

ooh booru tags, but lemme try

first one is nsfw, let me put nsfw in neg

thats with your exact prompt.. 2nd attempt

[native halo]

at least feet were perfect

[nova remnant]

her hand got messed up btw lol

[native halo]

more feet = better

[nova remnant]

im surprised how well her nails came out

[native halo]

pov u are an ant

[nova remnant]

6

[toxic dragon]

Got new computer testing new settings

[nova remnant]

@native halo forgot it

[nova remnant]

big settings, its looking good

[native halo]

i like the sneakers

[supple raptor]

Try adding foreshortening. Should help to make their feet look more like they're bigger because of perspective rather than just looking like they're too big. foot focus might also help.

[nova remnant]

Nubby knows everything about feet

[native halo]

feet expert

[steady grail]

feeticist

[supple raptor]

I might have genned a few of them

[nova remnant]

@lyric zealot go gen feet too

(I would like to review them, just to make sure its good)

[lyric zealot]

[supple raptor]

The vast majority of them aren't even remotely lewd. Just touhous soaking their feet in the stream.

[lyric zealot]

imma need proof of that

[nova remnant]

beautiful feet isn't lewd lol

[lyric zealot]

just so i can, like @nova remnant said, review them

[nova remnant]

everyone goes to horny jail

my sd been awfully slow

so much its timing me out I'll ask support if somethings up

timing out ? like broken?

I'm using a service

4gb isnt awefully bad

if the thing can't generate in like 10 minutes it aborts

you should be ok with 4gb and --lowvram

not sure why its like this

also my pc cannot do hires

i ran my 4gb with --medvram

I gave up on it

do hires of 1.5

[lyric zealot]

your free trial is done

[nova remnant]

Theres a big difference between anything nvidia and old amd

[supple raptor]

What, you haven't seen enough variations of this prompt yet?

[nova remnant]

ohhh damn.. i forgot you have amd

[lyric zealot]

actually i have

[supple raptor]

I haven't. I should probably make some more

[lyric zealot]

forgive the other leg

[nova remnant]

woes of sd1.5

but prompt wise it came out good

even for 1 foot

[supple raptor]

That's one that you just crop off the bottom few pixels to hide the bad toes

[nova remnant]

steven made a post with perfect nails

i have some issues with sdxl model im using cause its not too well trained, but i think with several tries i can get good nails

[supple raptor]

There's an Adetailer model for feet. Believe it or not, I do not have it.

[nova remnant]

really?

[supple raptor]

Really to which part?

[nova remnant]

part of adetailer feature or separate extension?

the ext lol

[supple raptor]

It's just a custom model you toss in the adetailer folder. Uses the same extension.

[lyric zealot]

where do i find that, so that i can avoid it?

[nova remnant]

sounds interesting, do you have the link you could share?

[supple raptor]

They're apparently pretty easy to train and can be trained on basically anything.

[supple raptor]

On civitai I think.

[toxic dragon]

finally I'm up from 6gb vram to 12

[lyric zealot]

disgusting, imma avoid that

[toxic dragon]

I've yet to find a way to make use of the extra space adequately still

[lyric zealot]

more vram, means you can probably train stuff now, and also gen at higher res

[nova remnant]

google isn't giving me any relevant result on adetailer feet

[toxic dragon]

I heard that over 768 is messy in 1.5

I've always been doing 512x768 x2 upscaler with my 2060

But I'm not sure how to take advantage of my new 4080 in upscaling without it looking a little funky

[nova remnant]

you should use sdxl with that new card

[supple raptor]

Link is at the bottom of the guide https://civitai.com/articles/1224/training-a-custom-adetailer-model

[nova remnant]

thanks, checking it now

[zinc flax]

Are we posting feet

[supple raptor]

Much like the hands model, it's probably not going to help you get more feet that are good, but should help with getting more fine details on the already good feet, such as toenails.

[supple raptor]

Always

[toxic dragon]

Haven't touched sdxl mainly because I have like

50 presets saved on sd 1.5 but if the interface is the same I might try iti

The only pain is finding some good new models

Since I've been relying on cetusmix3.5 forever

[nova remnant]

this is an amazing guide, it starts with training method too, but i could also just use the model linked below?

[supple raptor]

Most modern models can handle 544x960 without breaking a sweat, which is convenient because 2x upscale puts that at 1088x1920 which is as close as you can get to a perfect 1080p image from SD.

[nova remnant]

UI should be same if you are using either a1111 or comfy, but model is that crucial part, there aren't as many as sd1.5 due to the level resources it takes to train sdxl

[toxic dragon]

Yea

A1111 is what I use so that's nice

[supple raptor]

Yep they made the guide for people to train whatever they might want and posted the foot model that they used as an example.

[toxic dragon]

But its the models I guess that gets to me

[nova remnant]

great... i will have to try that for aesthetics

[supple raptor]

I wish for you to receive many very nice toes in the future

[toxic dragon]

A lot of what I used to do with my 2060 was just post edit. (Example)

upscale x3 gives me the dead inside eyes lol.

[nova remnant]

i looked at their samples in the link, you know sdxl can generate pretty decent toes in comparison, although im sure training would make a big difference, maybe they have basic samples uploaded atm

[supple raptor]

The adetailer model is just identifying them. It doesn't affect the quality at all.

[toxic dragon]

Any mature looking models that you'd recommend? I've been using my current cetusmix (3.5) because it gives a somewhat more mature look than whalefall does (at the cost of some consistency)

[nova remnant]

ahh ok, i think you should use sdxl 🙂

@toxic dragon if you are doing production level work you'd have to put some extra effort getting the right images, and with realistic photos its very obvious to tell if images are AI generated or real. but that changes the game alot with sdxl

[supple raptor]

This is currently one of my favorite models. It might fit your needs. https://civitai.com/models/211628/soushiki?modelVersionId=238399

[toxic dragon]

Definitely not going for realistic but definitely anime

[nova remnant]

sure but even with anime you'd have to put extra work on sd1.5 which is not as good as sdxl in coherence

[toxic dragon]

Tried the recommended 544x960 its a bit more chaotic than 512x768 on this one

[toxic dragon]

If you mean spending way too mcuh time for good images then yea I get you on that

[nova remnant]

basic coherence

those two engines are vastly different regardless of time you put in

[toxic dragon]

?

[nova remnant]

poster quality image !

[supple raptor]

I've had better from that same prompt but with scythe instead of polearm.

[toxic dragon]

Better results with 544x960 ig

maybe I'll try some other models to see

[nova remnant]

i think all well trained sd1.5 models have achieved a commonly likeable sweet heart face

[toxic dragon]

Nah that face is mostly deliberate because it doesn't rock the boat while testing other features

Since I'm testing outfit, hair, bow/ribbon placement

[nova remnant]

not talking about that one altho that face isn't bad either

im checking out posts on civitai

[supple raptor]

Grab PNG info off this image and try my parameters and negative prompt. A lot of people overdo their negative which hurts models that aren't really overfitted.

[toxic dragon]

How do I do that again?

OH

I definitely overdo my negative.

[nova remnant]

would read a story for her

[supple raptor]

Most people do. Try that one out for a little while and see what you think. If something is popping up that you don't want with a specific prompt just add that thing at the beginning while running it.

[gloomy scroll]

Wildcards going wild

[native halo]

where feet tho

[toxic dragon]

There.

[wispy canopy]

improtant for SDXL users!!! all checkpoints can be used with turbo tech
https://www.reddit.com/r/StableDiffusion/comments/1888j8a/all_sdxl_checkpoints_can_be_used_with_sdxl_turbo/

[lyric zealot]

bad fingers but

[nova remnant]

good symmetry

and her cute face makes it even more noticeable

[native halo]

🦶

[toxic dragon]

Tried Whalefall with some tweaks

And now ribbons act up sadly.

At least it looks more mature now.

[nova remnant]

nubby you should get into poster production, those are some pretty amazing images

whats the point of having highres in prompt?

is that supposed to have effect on canvas

or image detail

[native halo]

idk i dont use it neither masterpice,best quality etc.. but some models recomend u use either of those quality tags

[nova remnant]

yes masterpiece may be necessary for anime models

i use (highres, highly detailed:1.1), aesthetic, masterpiece

thinking of removing highres

[nova remnant]

lol you spoilered that, wonder why

[native halo]

it didnt came out with the denim shorts 😔

[nova remnant]

:p

[native halo]

this one came out with deformed hand 😔

[nova remnant]

her face is the winner 🙂

[nova remnant]

i'd like to have cute anime face like that on my sdxl model

to elaborate on the cute part ... she is innocent and hot

[native halo]

i think she needs a hand

[nova remnant]

sd1.5 🙂

[nova remnant]

face like sunshine in the sky, are you using LoRA?

[native halo]

no almost all faces are like that with this model

[nova remnant]

nice.. thought you were messing with LoRA, i have to check if i can have LoRA on sdxl for similar look

[karmic heath]

merging a few realistic models usually gets you close enough

[native halo]

yea i think this one is a mix of the realistic asian model with an anime mix

[nova remnant]

heck thats interesting, didnt know merging like that could do that

i think i have a asian realistic model for that

xxrealstic or something

@native halo What about some magic

[native halo]

i like that big ball

[karmic heath]

ah yes, the big ball...

just to make sure, you're talking about the one in the top left, right?

[nova remnant]

do you suppose 60% anime and 40% realistic a good try or .. much less on realistic model?

[karmic heath]

sounds good, give it a go

[nova remnant]

cool, time to download that model

[karmic heath]

merging won't take that long anyway

[karmic heath]

try a few different combos and see what you like

[karmic heath]

something's wrong with that pinky toe

[native halo]

it tried rly hard to morph into a 6th toe

take this one instead

[nova remnant]

small

[native halo]

here take this big one instead

[nova remnant]

https://tenor.com/view/cheesed-to-meet-you-gif-21409625

hello feet

[native halo]

hello there

[gloomy scroll]

@supple raptor

[gloomy scroll]

[lyric zealot]

[slender kernel]

https://tenor.com/view/hi-deniz-deniz-gif-23653159

[nova remnant]

differnt mod but they didnt focus much on the face

that one is with reproduction

this one is with bluepencil which also have reproduciton in their models merge among countless others

face looks bit better on bluepencil

think i will cross mix bluepencil with asian model

how do you merge LoRA with model on a1111?

or is that even possible?

[native halo]

with supermerger

[nova remnant]

whats supermerger?

[gritty spruce]

degens! 😂 we're going to have to call the police 😜

[native halo]

[gritty spruce]

keep it safe please... blur ; )

[nova remnant]

so supermerger is an extension and usable by a1111?

[gritty spruce]

nobody fear da police keep posting! just spoiler blur the sexy/horny stuff 😜

[native halo]

yes its @ a1111 extension

[nova remnant]

hi wzy, i hope you are having a pleasant day and like the beauty of life 🙂

[gritty spruce]

hey.. yeah day going fine, tons of work

[nova remnant]

gonna look into it ty

[native halo]

i think u need 32gb ram or more to load a sdxl checkpoint on supermerger but idk if its just for merging checkpoints or it works with less ram if u just wanna merge loras but give it a try to see

[nova remnant]

oh gee, ok, i have to check, i have 16gb at present... so no other choice

[native halo]

yea i dont remember if its 32gb or 64

[elfin valley]

how to install that ? guide me pls

[native halo]

just go to extensions>install from url > then paste this https://github.com/hako-mikan/sd-webui-supermerger and install

[elfin valley]

I'm a new SD, so Idk how to use @@

[native halo]

read these two guides ||https://stable-diffusion-art.com/install-windows/ || and ||https://civitai.com/articles/904/settings-recommendations-for-novices-a-guide-for-understanding-the-settings-of-txt2img ||

[elfin valley]

got it

[nova remnant]

hmm the realistic asian model i went for cause of great quality with sd1.5 their sdxl version came out poorly when rendering images with it

and i don't see many xl realistic asian model on civitai

blemishes always dont mean real lol

[native halo]

😔

[nova remnant]

her eyes speak

[dawn jasper]

Thanks so much.

[nova remnant]

This looks pretty good

[dawn jasper]

BEAUTIFUL!!!!!

[elfin valley]

save my pc 💀

[lyric zealot]

while genning or?

[elfin valley]

yes, but I gen in 90 step and the image resolution at max 2048x2048 and it broke at 90/90 💀

[lyric zealot]

are you using multidiffusion or something to aid with genning such high res?

[elfin valley]

are u meaning this ?
or what ?

[lyric zealot]

nope, i'm definitely not talking about that

[elfin valley]

sorry but this

:v

[lyric zealot]

ahhh, okay, first, are you using auto1111?

[elfin valley]

no, im using SuperMerger

[lyric zealot]

I'm so confused

Do you want to generate art or do you want to merge model?

[elfin valley]

I think both

[lyric zealot]

okay, firstly, if you are just getting started, i think try getting something to generate art first, like automatic1111 webui

unless you already have other ways of genning art

[elfin valley]

I used to use txt2img before. Im just try using SuperMerger

[lyric zealot]

ahhh okay, i thought you were trying to use SuperMerger to generate art

lmao

then i can't help you with that, since i never use that before

good luck tho

[elfin valley]

I did 💀

[lyric zealot]

[elfin valley]

It seems like I'm not used to using it, so my arts are bad

[lyric zealot]

ahhh okay, i get what it's doing now, it's merging and genning art at the same time, so you can save the recipe of models you want

[gloomy scroll]

SuperMerger is meant to create new model mixes and do gen previews and reason why your ram was so high was probably because you loaded 2 or more models at the same time because you were doing merge

SuperMerger is like a1111 checkpoint merger but more options

@elfin valley If you are new I wouldn't recommend using SuperMerger

[elfin valley]

honestly, I'd adjusted the max resolution into 3840x2160 and gen it that make my ram was so high

💀

[elfin valley]

yes sir

[gloomy scroll]

Ofc you are free to use it but it's meant for creating new model mixes and not generating images

[elfin valley]

This art had used the purchased lora ?

[lyric zealot]

what do you mean?

[gloomy scroll]

You can buy free loras here
https://civitai.com/models

[elfin valley]

are u using paid lora ? or just use prompts ?

[gloomy scroll]

I don't think anyone here uses paid loras

[lyric zealot]

here's the pic with metadata

the lora i was using can be found on > #🍥｜anime message

[elfin valley]

of course, but I dont get it💀

what is metadata ?

[elfin valley]

many prompts ?

[lyric zealot]

Vil, I'm counting on you

[gloomy scroll]

1. Save the image
2. open pnginfo tab in webui
3. Locate Smol Sus's image and drop it to pnginfo are

Should see something like this

[lyric zealot]

not all my pic have metadata tho

I should just stop being lazy and copy from folder

from now on, I should try my best to upload all pics with metadata

If found any without, please lmk

[elfin valley]

I got it🐧

[lyric zealot]

[elfin valley]

look wild

[nova remnant]

what does copy config from imply in checkpoint merger?

i mean there is 0-1 multiplier for model mix

[elfin valley]

@lyric zealot when I use Hires Fix and upscale by 2
my res of art is improved by a problem happen is. it created another one 💀
how to fix ?

[lyric zealot]

what's your generation res?

[lyric zealot]

try changing your denoising strength to 0.51 or something much lower than 0.8

[lyric zealot]

and also, not sure about upscaler LDSR, but DPM++ 2M Karras is a good one to start with

wait wrong, that's the sampling method

lmao

[elfin valley]

🐧

I'll try

[lyric zealot]

https://openmodeldb.info/ you will wanna get some good upscaler here

[lyric zealot]

and, you can turn off refiner, since you are using the same models for both, so it's actually the same thing

[lyric zealot]

when using this lora lora:SIC_outline_v1.01:0.6, it should just be in your prompt, and not in negative prompt

[lyric zealot]

your cfg 12 might be too high, without extra extension, i won't recommend going over 10

[elfin valley]

I forgot remove it :v

[lyric zealot]

usually you don't need to have that many steps for hires fix. I usually use it at 10 steps for hires fix, and 20 steps for generation

I should probably consolidate all those messages to a single message but oh well

[elfin valley]

@@

[lyric zealot]

[nova remnant]

slaps

[elfin valley]

🐧

[nova remnant]

can someone explain this equation for tertiary checkpoint merger?

A + (B - C) * M

M is the multiplier of value 0 to 1

A is the primary model and B is the secondary

[lyric zealot]

A: This represents the primary checkpoint, which is the base model that you want to modify.
B: This represents the secondary checkpoint, which contains the additional information you want to incorporate into A.
C: This represents a tertiary checkpoint, which is used to remove unwanted information from B before adding it to A.
M: This is the merge multiplier, which controls the strength of the influence of B on the final merged model.

Here's how the equation works:

Difference Calculation: We first calculate the difference between B and C. This essentially removes any information that is present in both B and C, leaving only the unique information in B.
Weighted Difference: This difference is then multiplied by the merge multiplier (M). This allows you to control how much influence B has on the final model. A higher M value will result in the final model being more similar to B, while a lower M value will have a smaller impact.
Adding to Base Model: Finally, the weighted difference is added to the base model (A). This results in a new model that combines the information from all three checkpoints.

In simpler terms, this equation allows you to selectively add information from a secondary checkpoint (B) to a base model (A) while removing any unwanted information using a tertiary checkpoint (C). The merge multiplier (M) allows you to control the balance between the base and the secondary model.

Here are some additional points to note:

This equation is a simplified representation of the actual merging process, which may involve additional steps and calculations.
The specific values for B, C, and M will depend on the specific models you are using and the desired outcome.
Experimenting with different values of M and the choice of B and C can help you achieve the desired balance for your specific needs.

I hope this explanation clarifies the equation and its role in tertiary checkpoint merging. If you have any further questions, feel free to ask.```

i asked bard for that

[nova remnant]

i wanted human touch lol.. but ty

[lyric zealot]

probably wrong tho

[nova remnant]

i have bard open in one tab but was feeling bored to ask bard lol

was messing with checkpoint merges for couple of hours now :p

[elfin valley]

where the folder which I put in ?\

[lyric zealot]

\stable-diffusion-webui\models\ESRGAN most of the time here

[lyric zealot]

sorry, i should have recommended some for you

[lyric zealot]

all of these are pretty good for anime

[elfin valley]

How to use or just Upscale ?

[lyric zealot]

in hires fix

[elfin valley]

[supple raptor]

[nova remnant]

bluepencil xl has like 100s of model merged into it

crazy

i dont see the point to merge anything with it

only thing that makes sense to merge is the turbo xl

and that just for the sake of better coherence

is there a limit to how much variation a model can have?

if there is no algorithmic limit a model can contain all kinds of possible variation

then its just a matter of poking that model with the right trigger prompts

[lyric zealot]

possibly, but then i think all concepts that the model knows will be a little diluted?

[nova remnant]

not diluted i think

it can pretty much cover the expanse

but that would come at a cost of less generic prompt

[lyric zealot]

[nova remnant]

i think its good

you can have a paramater of generic prompt not just overly generalizing them

look at some of the sd1.5 prompt patterns, some of them are completely off the generated image

meaning those models are not coherent specific which sdlxl does better

regardless of possible variation in this case

[lyric zealot]

[nova remnant]

i made a merge between two realism focused sdxl model and managed to generate fairly decent anime image with that new model

all i had to do is put in (anime style) as prompt

what seems to be the core reasons for removing one from the other in this sampling dataset?

if it's just showing the difference that doesnt make sense to exclude one set of imagery, cause neither of the images are discardable

[nova remnant]

btw, the images im generating there seems to be difference using anime theme, anime style, anime 👀

using the same seed and rest of the prompt

visible difference in subject's posture

as if word tokens are affecting the output

what if i put made up gibberish for words

yeah every meaningful or meaningless token of words influence the output

booru tag kind of made people less descriptive which is limiting the users getting the most out of sdxl

using a prompt cat doesn't describe the cat

[gusty palm]

i think theres a sd non xl turbo model

its for sd2.1 i think?

not sure if 2.1 works or if it will take 1.5 loras and stuff

[supple raptor]

2.x models can't take 1.x loras. The models were generally also worse than 1.5 models despite the architecture being better on paper.

[gusty palm]

thats the vibe i got

[nova remnant]

adetailer is ruining some faces with SDXL models

[steady grail]

what setting do you use on adetailer?

[nova remnant]

only the denoise strength that i set to 0.3 intead of 0.4, but that might not be the reason, i think i got some XL TI that could be conflicting

[steady grail]

oh, auto1111

[nova remnant]

does increasing the detection confidence threshold affect closeup portraits?

default was 0.3 which i changed to 0.82

[steady grail]

well, a portraits require a bigger mask and so, a bigger hook

[nova remnant]

going to switch it up and see what i get

[neon harness]

Are there other good discords dedicated to ai anime?

[steady grail]

dunno, pay attention to channel sharing porn and virus

[supple raptor]

detection threshold only affects whether the faces are found or not. If the faces aren't being found you turn it up a little. If rocks or other objects are being detected as faces then you turn it down a little.

[nova remnant]

something's definitely messed up, when i enable adetailer with sdxl its messing up the face with jagged lines

[supple raptor]

Touhou Project AI is good (and not only focused on Touhou)

[steady grail]

with a too little threshold, you even get partial hooks, sometimes i got only hair or half of the face hooked

[steady grail]

can share link in pvt?

[nova remnant]

let me give an example what's going on ... i will delete them since they are not anime, but just for sampling

the settings i changed with adetailer are only detection threshold and denoise strength which SHOULDN'T mess up like this

[steady grail]

that's left over noise

[nova remnant]

left over noise? how do i sort that out?

its obviously from adetailer

[steady grail]

yeah, adding more steps maybe?

[nova remnant]

you mean steps in adetailer?

[steady grail]

yep

if you can do it in auto

[nova remnant]

i didnt have that problem with all the other images i was generating but let me try again

yes i can

adetailer has entirely separate section to custom set all the detail

im using LCM with 8 sampling steps and 2 cfg, which you saw the good image i posted, and when i set separate adetailer steps and cfg by increasing them to 30 and cfg 2 that gets even badly distorted

[steady grail]

share the result