🍥｜anime

[lyric zealot]

[wooden coral]

Is that with XL tho

[oak spoke]

I didn't even prompt the wing

[oak spoke]

look like 1.5 to me

[lyric zealot]

are you talking to me?

[wooden coral]

Yeah

[lyric zealot]

yup, it's 1.5 based model

[wooden coral]

Since hes trying to achieve it on XL

[lyric zealot]

I also cheated, using regional prompter

[wooden coral]

Im slowly starting to understand animatediff

@lyric zealot

[oak spoke]

animatediff just hurt my eyes

with it flashing gif

it just switch multiple images really fast

[wooden coral]

Im trying to get it more consistent

[lyric zealot]

that's how animation works

that's how video works

[wooden coral]

Background is still meh

[oak spoke]

https://media.discordapp.net/attachments/755070259907657769/1154473890987397211/00000-2740738347.gif?ex=6513c1d8&is=65127058&hm=6d4a24bad5f621288c9a96ff3a17670663df250955cc04d2b9bd7c716b8e0135&

this's what it did to me

[lyric zealot]

on some frames, her eyes looks like it's gonna pop out

[lyric zealot]

you are definitely not using the animation module

[oak spoke]

we need tutorial

[lyric zealot]

it's not inferring the previous generated frame for next frame

[oak spoke]

please teach me Boto sensei

[wispy canopy]

i don't know how any of this works, is all black magic! (got red skin/mask tho)

[lyric zealot]

okay, back to dinner

[wooden coral]

I can give you this

[oak spoke]

thank you I'll try to decipher it

[wispy canopy]

looks like my aaaaaaa prompt

[wooden coral]

Messy workflow for now, but will do the job

[wooden coral]

Its pretty simple once you get behind it

[wispy canopy]

the prompt: anime, aaaaaaaaaaaaaaaaaaaaa
the result:

[oak spoke]

lemme try that

[wooden coral]

I need some deflicker tool for comfy

[oak spoke]

here goes nothing

[wispy canopy]

THAT IS GOING TO BE KOOKED

[wooden coral]

Her arm is doing things

[oak spoke]

o oh it froze at 49% for a bit

[wispy canopy]

anime, (aaaaaaaaaaaaaaaaaaaaa:3.0)i wonder why these are les... uhh.. AAAAAA

[oak spoke]

yea disappointing result

[wispy canopy]

let me bring out the big aaaaaaa guns

just switch back to darkalfa 😄

there. EZPZ

[oak spoke]

3rd one got witch fingers

[wispy canopy]

anime, (aaaaaaaaaaaaaaaaaa:1.5)

nr 2 tho 😮

holy shit

[oak spoke]

her middile finger's like 70% the width of her face

here's the aaaaaaaaaaaaaaaaaa

with no weight

[wispy canopy]

sdxl too clean at times

[oak spoke]

can't show lower cuz she was naked for some reason

[wispy canopy]

she was screaming AAAA for a reason.

[oak spoke]

🤣

[wispy canopy]

lets see if i can make darkalfa survive 2:0

nope, complete disintegration

but... i've got a better prompt for that

if i just want the most ridiculous aaaaaa

SAY AAAAAAAAAAAAAAAAAAAAAAAAAA

ah. even sdxl can make this prompt work 😄

i see this, next prompts gets the junji ito in there

it's still generating, and i'm already regretting my choices

[silent kraken]

Flusturd

[nova remnant]

@wispy canopy aaaa

[wispy canopy]

just keep on smiling bois.

[nova remnant]

Send me

[wispy canopy]

reminded me of this prompt, had to anime-fy it

[nova remnant]

Am I crazy or does SDXL images look darker

Warmer would be a better way of saying it

[gusty palm]

meew

[dawn jasper]

Miku Hatsune, is Magical.

[magic wave]

Does anyone know what this insanity is or anyone that might know how its done? :o

[lyric zealot]

[nova remnant]

that tiny cat

[nova remnant]

https://media.discordapp.net/attachments/866165925815386153/932940832485568622/853894086041927730.gif

[wispy canopy]

she's got a sephiroth exprission going on

[nova remnant]

(needs face fix but cute)

two door handles, whatever

[magic wave]

Its a removable door, to enter you need to succeed a strength roll

[nova remnant]

doow

[nova remnant]

sounds like something crazy lmao. this is just a 3d vtuber model, but they took screenshots and made a Lora for each angle at least that's what I think they're saying? translation could be off. "created Lora for each" and idk what "extracted the difference" means. probably something like taking the loras and merging them together to create one eye lora. But I don't see why you couldn't just use all of the angles and make one complete lora.

the only reason I can think for doing that is you need enough images or data to make a set for the Lora, like if it's an original character

looking closer at it and the 2 images maybe they made a Lora for each set, ex: starry eyes and normal eyes.

Sometimes loras can be trained on multiple concepts, but you need a lot of data and it depends on the optimizer to prevent "concept bleed". lots of experimenting

[magic wave]

I mean, I think they're probably doing block merge between all the LoRAs for individual images to make the AI able to only focus on that and not get anything else, but that sounds insane

[magic wave]

Maybe

[nova remnant]

It sounds like something very specific and trained for something intense, like they're making an animation or making a LOT of images. Even loras with multiple outfits or specific things like that are pretty simple preparation in comparison.

[magic wave]

Yeah it seems like they are making a lot of images, I'm a big fan of this person, their LoRAs have like 0 effect on your image, you can generate, slap one of them on, and it will only really change what its made to change, so they're amazing and I wanna do that, but I don't know what kind of insanity they're doing

[nova remnant]

Ahhhhh

Do they have a civitai page

[magic wave]

https://civitai.com/user/2vXpSwA7/models They do yeah

[nova remnant]

You might wanna talk to someone who has more experience with loras then. nubby or someone outside that's a developer of a Lora training program (stuff like everydream2)

The only thing I can think of is having a dataset that wouldn't affect anything else, so tagging it properly or making it very simple (like using vroid models) OR the opposite, making the images have enough variety where the model knows what to do with specific generations. either way i have little idea

or maybe I'm thinking too hard about it

could also be something to do with text encoder training... which wouldn't really train the clip / prompts but Im too unsure to say completely

i know kindof that you can adjust "layers" with actual model merging but this is just lora

sorry im rambling at this point

[supple raptor]

I think they're using this process https://discord.com/channels/930499730843250783/1113466741671084182 (Touhou Project AI server link)

[magic wave]

I mean, I would ask the person themselves, but I don't think they have open DMs :o

[magic wave]

I mean it should be the same, but its still really complex

[magic wave]

Lemme see...

[brittle jay]

just came to the realization that every single one on the internet is ahuman (I mean except bots/ai) and are actually real.

[supple raptor]

You can block merge LoRA too. The blocks are a little different though.

[supple raptor]

nah, you're just some bot that lives inside my discord.

[brittle jay]

like I could be walking past someone in the street and idk that I heard abiut thwm online

[brittle jay]

how did you know

[supple raptor]

i know all

[brittle jay]

oh right

[supple raptor]

https://cdn.discordapp.com/emojis/1001217158430339103.webp?size=64&name=SakuyaSmug&quality=lossless

[brittle jay]

i forgot that property about you

[supple raptor]

wait.... what if I'm the bot?

https://tenor.com/view/ohnaruto-muni-d4dj-d4dj-first-mix-anime-not-like-this-gif-22650857

[magic wave]

Holy shit this is absurd but it makes so much sense

[brittle jay]

there is yet a way to distinguish between an extremely complicated ai and a real human

perhaps the real ai was the friends we made along the way

[supple raptor]

or perhaps the real friends was the ai we made along the way

[brittle jay]

yes

thos one so cool

[magic wave]

WAIT, NUBBERS

@supple raptor How do I merge LoRAs at -1 weight

To eachother?

[nova remnant]

nubby to the rescue once again

[nova remnant]

You type too slow to be one

[brittle jay]

isnt there an extensions for merging loras

[brittle jay]

haha fr

[supple raptor]

I run on an AMD gpu

https://cdn.discordapp.com/emojis/971633476837183608.webp?size=64&name=trollface&quality=lossless

[nova remnant]

https://tenor.com/view/cat-lava-cat-minecraft-cat-gif-13440479435726322249

bye

[nova remnant]

holy shit it does make so much sense

[supple raptor]

this little guy right here https://github.com/kohya-ss/sd-scripts/blob/main/networks/merge_lora.py (you have this already inside derrian's repo.

[brittle jay]

almost laughed out loud

[magic wave]

Ohhh alrighty

Also, do you know if I need to like... use the tagger on the images?

I guess I dont?

[magic wave]

ITS DOES, ITS DUMB, BUT IT MAKES SO MUCH SENSE

[nova remnant]

like I actually wanna try it now

because it's not rocket science like I thought it was

yeah I'm curious too
like how do you train it for one image? just don't tag it? increase the lr to an absurd amount?

[magic wave]

I'm going to try it right now xD

[supple raptor]

I'd probably ask Ao to be sure, but I would think it would be best to not train the TE at all so tagging wouldn't really matter.

[magic wave]

Will do

And then only Unet?

[supple raptor]

That's what I'm picturing. I don't really train LoRAs though so I could definitely be wrong.

[nova remnant]

so tenc lr is 0 then

on my branch kohya stop text encoder training doesn't work so I'll usually set it to 0

yeah they put the unet lr in the thread I think

[supple raptor]

Oh, there are unet only and unet+tenc versions in his HF repo so you can download both and try them https://huggingface.co/Aotsuyu/LoRA/tree/main/diffLora/8dim

[nova remnant]

stopping text encoder training would prevent the model from being trained on any sort of prompt. so it doesn't need a prompt to activate, and won't remember anything. so yeah it makes sense why you'd stop

oh awesome

training tenc could affect composition, but yeah this is interesting

whats this for in controlnet >.>

[brittle jay]

it makes it pixel perfect

you know how it takes your images and then processes it and then magic?

this makes process use higher resolution

if you want better explsnstion ask nubby

[supple raptor]

@magic wave I was wrong about which script it uses to merge. This is the one you want, which you should also have since it's part of the same repo and the repo is included in derrian's https://github.com/kohya-ss/sd-scripts/blob/main/networks/svd_merge_lora.py

[magic wave]

OH, What is SVD?

[supple raptor]

I have no idea aside from that it's a different merging method, but there's some discussion in that thread from someone who was doing the merging in bmaltais (which uses the normal merge script) and it wasn't working right.

[magic wave]

Ohhh damn fair

Also in what folders are those scripts in, and how do I use them?

[supple raptor]

.\sd-scripts\networks

[magic wave]

Found em! How do I run em? :o

[bitter latch]

with your computer

https://giphy.com/gifs/travisband-Bng9nsAhSaDVxWsSLh

[nova remnant]

https://github.com/derrian-distro/LoRA_Easy_Training_Scripts this is "derrians" repo right

[supple raptor]

from .\LoRA_Easy_Training_Scripts\sd_scripts in CMD
venv\scripts\activate
then
python .\networks\svd_merge_lora.py --save_precision "fp16" --precision "float" --save_to "new8dimLoRA.safetensors" --ratios -1 1 --device "cuda" --new_rank 8 --models .\LORAofBASEIMAGE.safetensors .\LORAofEDITEDIMAGE.safetensors change the --save_to and --models values accordingly.

[magic wave]

Alrighty, Imma train the 2 LoRAs which should only take a few minutes each, and then I do that

[supple raptor]

yep, if you use bmaltais or any other kohya based trainer, then you should have it too in essentially the same location. If you have an sd_scripts folder, it will be in there.

[nova remnant]

no, i think kohya just has them in the base directory? not sure

[supple raptor]

they'll be in .\networks for regular kohya since that repo is sd-scripts

[nova remnant]

oh awesime ty

[supple raptor]

The other trainers are pretty much all just wrappers for the sd-scripts repo

[nova remnant]

so i just..replace?]

[magic wave]

Also you should send me this and the Ao guide link in DMs so I can pin it and not forget it XD

[nova remnant]

nvm me being an idiot its right there

[magic wave]

I think I did my settings right

[bitter latch]

playing with the model mixer merge extension

[wispy canopy]

don't do it!

you'll lose your sanity trying to mix

[bitter latch]

to late

lost that decades ago

Miku Hatsune , full body portrait, classical anime background

[brittle jay]

Wait until you find out about the Chinese tutorials for block merging

Very legit

Man I need someone to send a picture of it

[bitter latch]

interesting style

[brittle jay]

https://rentry.org/Merge_Block_Weight_-china-_v1_Beta

This whole rentry feels like just a big meme

Fr

With all the arrows

And chinese

Which button??

Ok I now shut up and go slep

Goodnight

[magic wave]

O, good nite Kazoo

[wispy canopy]

sleep well kazoo 🙂

[nova remnant]

@lyric zealot HELLO

smooool

what detailer lora do you use

and does it keep the same art style ?

[lyric zealot]

face_yolov8n.pt

wait

[nova remnant]

not adetailer u dumdum

[lyric zealot]

I thought you were asking about adetailer

hmmmm

depends

because sometimes i use two loras for face

sometimes just based

so it depends

i can get metadata for you if you wanna point me to which image you trying to recreate

[nova remnant]

errr

which one do you think is good for general purpose

[lyric zealot]

i think it depends

i usually use aoi with taf

if i wanted more "beautiful" face

and kinda semi realistic anime

if not it's based

or even this

that can help with some cinematic background

[lyric zealot]

but usually comes with grass, bushes, garden

[bitter latch]

cross anime with realistic 😄

[nova remnant]

does this do the detailing on things

[lyric zealot]

nope

the one you want is probably the lora called more_details?

[nova remnant]

I'm lost

[lyric zealot]

what do you even mean by detailing on things tho?

like keep the style, but more details?

[wispy canopy]

ugh. why am i downloading models so i can start merging again

am i sick?

[lyric zealot]

if yes, then this maybe ?

[nova remnant]

this one

Did you use anything like that for those

[lyric zealot]

literally just this lora

[nova remnant]

hmmmmm

oki

[lyric zealot]

metadata

[nova remnant]

I'll try it too

[lyric zealot]

you will almost always get garden/bush with that lora, since the activation keyword involves garden

[nova remnant]

bushes...

But people can be hiding in those

is it safe

[bitter latch]

on new merg without lora

[lyric zealot]

might be

[nova remnant]

ohh

this is what I'm talking about

it makes it like pencil drawn

[bitter latch]

replaced with abby from NCIS, lora

face is sorta messed up

[lyric zealot]

pencil drawn is another lora

im not sure i have that one tho

[nova remnant]

no, I'm saying that more details one makes it look like a drawing sort of

[lyric zealot]

you don't want to usually use it at weight 1, since it can get weird really fast

[nova remnant]

wtf i just aopplied my own lora and it turned it into a drawing too LOL

maybe a1111 weights are different

[nova remnant]

i thought mine was overtrained or something

[bitter latch]

yeah I play around with weights till it works

nuElement model works good with this prompt

wait it put abby in corner.. huh ok

and restart sampler is a gawd send

hmmms

there we go

[bitter latch]

see now this is a place I would expect to find weds adams

[bitter latch]

not inpaited

there thats better

ok i'll quit running it into the ground

[bitter latch]

ok not done 😄

finally got a lora

[magic wave]

Do you have a VAE on?

[bitter latch]

yes

[lyric zealot]

that's kinda uncanny

[magic wave]

Reminds me of my negisaray LoRA, I should retrain that one

[nova remnant]

something's not right

[bitter latch]

changed it up a bit " I may look sweat but I will cut your heart out and eat it. try me "

in an alt universe

hmm is it me or does that look like a boy Weds Adams?

hmm no adams apple

[nova remnant]

maybe she's just flat

rip

[bitter latch]

well she is a child vampire so , prob

umm oops

"never gonna grow up , She's Fangs R Us Kid"

runs

Now they are a boy

I think.... shrugs

tired my brains wondering

[bitter latch]

being sick sucks

sick in body . i'm always sick in my mind 😄

as anyone else

[nova remnant]

Oki

[bitter latch]

now this is uncanny

[brittle rampart]

@nova remnant hello can tell me your arguments in Bat file for AMD card?

[nova remnant]

[nova remnant]

I use
--medvram --opt-sub-quad-attention --opt-split-attention-v1

[brittle rampart]

thanks. try it

[nova remnant]

That chair looks very good

[bitter latch]

vampire pugsley adams

tho he lost weight it seems

[nova remnant]

piercing eyes 👀

[nova remnant]

[nova remnant]

Did you do anything special to get them bright?

[nova remnant]

i 100% did not, it hurts my eyes just looking at it as i type this lol

get something by pure luck
Doesn't know how to replicate trying

Me some days

the joy of not using loras or controlnet

[bitter latch]

ok came out to dark

o...k then

yeah not sure why it gave him boobs. i put 1boy.. shrugs

stupid model

here have this.. thing . i'm gonna lay down

[nova remnant]

AI loves girls

[nova remnant]

Definitely one of those statues that will attack you

nah, she's super friendly (she isn't)

[nova remnant]

I'll ask

Why do you screenshot the images to send

the screenshot gets saved in screenshot folder automatically because i forget a lot lol

oh

You're a light user I see

Just here and there for a bit, few images a day

so many random images in there

[nova remnant]

what an upscale (other than the bad eyelash gen)

[nova remnant]

those are pretty nice wings o:

[brittle rampart]

yes

[quiet bramble]

What are the best ways to hide hands?

[nova remnant]

Have you tried putting hands in the negative >.>

or showing hands in negative

arms behind

idk >.>

[brittle rampart]

hands in pocket

[quiet bramble]

Yeah. I was wondering about creative ways to hide them rather than always using 'arms behind back' and 'hands in pocket.'

Certain perspectives and outfits don't work well with those two prompts.

[brittle rampart]

inpaint!

creative way i think/

[nova remnant]

well

your best bet would be hiding them in the negative

and AI would try its best

in "creative ways" whatever they might be depends on your gen :p

[slate narwhal]

Sorry for the long delay, I was offline for days. These are some variation (Comfy workflow type of variation) tests of direct artist images; variations tested in SDXL with a lot of upscaling, SAG, ADM, eye detail Loras etc. in a (kinda) unsuccessful attempt to get good eyes without Adetailer

[nova remnant]

no problemmm

Since then I found a good model :D

with colors like that

[slate narwhal]

nice, what model are you using?

[native halo]

idk why she was crucified

[native halo]

https://tenor.com/view/jesus-is-my-shepherd-gif-5499790

[nova remnant]

kakigori

our friend nubby's model

and also found some good loras for styles

which are very good to use I should do more

[slate narwhal]

Yeah, kakigori is a really nice model, we need that type of quality in SDXL. SD 1.5 is still winning in the anime department.

[nova remnant]

https://tenor.com/view/squishy-cat-gif-25192306

[nova remnant]

yellow

sdxl

[dawn jasper]

Who is this woman?

[native halo]

michelle obama

[supple raptor]

I don't think we'll see a truly good and flexible anime model in XL until there's a decent huge scale finetune. Waifu Diffusion is the only current attempt at it afaik. NAI has made some comments that sound like they're working on it too, but I doubt that will ever find it's way into the public.

[nova remnant]

What is NAI exactly

also I'm not really sure what is wrong with SDXL doing anime

[supple raptor]

NovelAI

[nova remnant]

To me it all looks anime

I don't watch that stuff

I'm anime blind

[native halo]

exactly

u dont know it

[nova remnant]

is that a dog

devil dog

[native halo]

yes it tried to draw a dog idk why 🐶

[supple raptor]

XL just doesn't know anime concepts and characters. Some of the finetunes have gotten the style pretty well, but that only goes so far on it's own.

Even Miku is not very good on XL models and she is basically the easiest character to gen.

[nova remnant]

https://tenor.com/view/twin-peaks-old-people-typing-one-finger-typing-technology-out-of-touch-gif-9186182

[nova remnant]

ohh

can you show me some miku o:

how sdxl makes her

[supple raptor]

I don't use XL models because they can't do what I want. Xd

[nova remnant]

I'll ask eface later

[supple raptor]

I don't want to have to find/train a LoRA for every character I want to gen.

[nova remnant]

hey thats me

trying to find weird game characters to gen

whats worse, they are probably in chinese so good luck me

[supple raptor]

Here's a couple of Eface's miku gens, they're not bad quality, but her hair ornaments are all wrong and her outfit is wrong #✨｜sdxl message

[nova remnant]

https://media.discordapp.net/attachments/931925914470711346/1061076535525785680/321666B2-CC36-467F-B254-56055DEB1C8C.gif

you went searching for some xd

I had to open google and search miku cause I couldn't remember what she looked like, they looked fine at first

[supple raptor]

Just had to search from:eface miku

[nova remnant]

smol

[nova remnant]

https://media.discordapp.net/attachments/194772750957674497/1039982173618438164/1034723981825159228-1.gif

apple looking monitor

[supple raptor]

smol

[nova remnant]

path of exile bricked 2b for me

I cant not think of her ||butt||

[supple raptor]

did they do a PoE crossover?

[nova remnant]

No, like...

I'll explain briefly

nikke gacha game has 2b crossover right now

When you go trade with people, you usually go into their homes (hideouts), and trade with them
You can customize your hideout in a lottt of ways, and one of the most famous hideout is a big ass on your screen

@supple raptor

[nova remnant]

and...

NARAKA

(naraka its a game I really like, my friend plays with the 2b skin lmao)

[supple raptor]

there's a reason she's named 2bestgirl

https://cdn.discordapp.com/emojis/1098308726307426374.webp?size=64&name=SagiriSmart&quality=lossless

[nova remnant]

is she holding an apple or... is she an apple?

[supple raptor]

yes

[nova remnant]

my loras folder just grows

too many to use

[gusty palm]

hewo

[nova remnant]

me thinking of prompts (very real)

lmao she looks paralyzed

[gusty palm]

hows uall

[nova remnant]

im having messy gens one after the other, either i messed up my negative prompts by trying to fine tune them, or Ai just came alive and messing with me

[nova remnant]

not bad, but dealing with messy fingers :/

[nova remnant]

I am meow

gimme some good meow

I can't upscale everything

have more thinking amano hina

any suggestion on this prompt....

Watermark -,,"(watermark, signature, text, logo:1.1)"
or (watermark:1.1) is enough?

[nova remnant]

I've been using this
negative_hand-neg, (KHFB, AuroraNegative:1.2), (Worst Quality, Low Quality:1.2), border, grayscale, (watermark, multiple views:1.2)

I've never once again seen watermarks, white bars or AI trying multiple views

ahh ok thanks, i see you only used watermark

i've done a few renders with just watermark and it seems to work but i needed to be sure if other words for it were necessary

this model I like using is really bad with multiple views

but I don't think it hurts other models

i had multiple views as generic prompt but i took that off few days ago

this is my default template

Default -,,"(bad anatomy, bad proportions, malformed, bad hands), (fused fingers, interlocked fingers:1.2), (asymmetric, conjoined, distorted, cloned), (low quality, worst quality:1.5), (low contrast, oversaturated), (monochrome, grainy, pixelated, borders)"

guess i'll add watermark at the end

@nova remnant 👀

big catto

any use for this?

Bad Art -,,"(bad artist, bad art, bad fan art:1.1)"

no idea, you should try a x/y/z plot real fast to see if you notice difference

maybe its good

uhh i dunno what x/y/z plot is or works

can you explain

i was avoiding a lot of extra features

hmm

not xyz for that sorry

but a prompt matrix lmao

let me screenshot it 1 sec

i run that with default setting?

[nova remnant]

for example

if you do this |(watermark:1.2)|(multiple views:1.2) at the end of your negatives

it will gen one image without them, one with only multiple view, one with only watermark and then one with both

easy way to notice if something is helping

ahh interesting

and I think you have to select negative in the option there >.>

so i use something like regular prompts here|(bad artist, bad art, bad fan art:1.1) ?

hmm

each prompt inside the | will become one full prompt in an image

ahh i see

you'd kinda want
Your normal prompt |bad artist|bad art|bad fan art

so regular prompts here, (bad artist, bad art, bad fan art:1.1)|(bad artist, bad art, bad fan art:1.1)

[nova remnant]

i have other prompts i want to put in before that

Check how I typed it

anything after the first | will become a new version of that whole prompt

adding it together

ok but if i wanted to evaluate the entire combination of (bad artist, bad art, bad fan art:1.1) wouldn't i used them together?

It will add them together in the end

Just do it once, you'll see

it'll be like 9 images

ok thanks time to test out

[nova remnant]

🤷‍♀️ no clue what im doing in this corner, how about you ?

eeeesh, i got a run time error

i had ...

(bad anatomy, bad proportions, malformed, bad hands), (fused fingers, interlocked fingers:1.2), (asymmetric, conjoined, distorted, cloned), (low quality, worst quality:1.5), (low contrast, oversaturated), (monochrome, grainy, pixelated, borders), (watermark:1.1), (bad artist, bad art, bad fan art:1.1)|(bad artist, bad art, bad fan art:1.1)

[native halo]

(good 5.6) (bad:4.5)

[nova remnant]

hi steven, what's that?

[nova remnant]

do like
(bad anatomy, bad proportions, malformed, bad hands), (fused fingers, interlocked fingers:1.2), (asymmetric, conjoined, distorted, cloned), (low quality, worst quality:1.5), (low contrast, oversaturated), (monochrome, grainy, pixelated, borders), (watermark:1.1), (bad artist)|(bad art)|(bad fan art)

[nova remnant]

I think only one | isnt enough

[nova remnant]

ok copy / pasting that

should i be check marking the negative prompt?

[native halo]

just dont fill negative with all of those bad proportions malformed fused word salads

[nova remnant]

yes >.>

[nova remnant]

hmm lol that was the usual convention with sd 1.5 tho

[native halo]

try only with low quality,lowres,watermark,low contrast,monochrome

[nova remnant]

ok, so no references to bad anatomy of any kind?

[native halo]

no need for that just use a negative embedding those have the bad anatomy deformed references inside

[nova remnant]

you mean embeddings like Textual inversions?

i think TI arent comprised of words, they are trained on images

yes

yes to embeddings like textual inversion idk what they do

I just use stuff and prompt

yeah those are image trained

TI can influence the model output in a big way

and sometimes render outputs that a mod weren't intended for

this is an interesting comparison ... but i dont see too much difference on those bad art choices

[nova remnant]

Was this with the one i sent ?

yes

I thought there would be 3

did you reduce

i had batch size set to 2 if that had anything to do with it

naa, i copied and pasted the line

(bad anatomy, bad proportions, malformed, bad hands), (fused fingers, interlocked fingers:1.2), (asymmetric, conjoined, distorted, cloned), (low quality, worst quality:1.5), (low contrast, oversaturated), (monochrome, grainy, pixelated, borders), (watermark:1.1), (bad artist)|(bad art)|(bad fan art)

hmm

Its including bad artist into those

because it isn't separated with | at first

a mistake

so (bad anatomy, bad proportions, malformed, bad hands), (fused fingers, interlocked fingers:1.2), (asymmetric, conjoined, distorted, cloned), (low quality, worst quality:1.5), (low contrast, oversaturated), (monochrome, grainy, pixelated, borders), (watermark:1.1),|(bad artist)|(bad art)|(bad fan art)

?

[nova remnant]

this is going into geeky territory ... im not good with programming languages

I wanted to show you :D

These TIs I'm using influence very little with the model

its really good

ok

for me it feels like if i can't control the output with my word prompts, they are illusive

elusive i mean

But with stuff like this

comparing them side by side

but hey.. the prompt i wrote has less than 75 tokens used

you get those mysteries off :p

so nothing in that prompt getting cut off

From what i know things don't get cut off but is paid less attention to

and its like a generic template that i'd use with every images

(bad anatomy, bad proportions, malformed, bad hands), (fused fingers, interlocked fingers:1.2), (asymmetric, conjoined, distorted, cloned), (low quality, worst quality:1.5), (low contrast, oversaturated), (monochrome, grainy, pixelated, borders), (watermark:1.1)

yeah no reduction on emphasis

you should try TIs

They are great :D

i did lol

Maybe you got some bad ones/overtrained

after learning how TI works i chose to not use them much cause they can alter a model's intended output

I had some that would completely change the look of the model

and I went back to using negatives like yours

[nova remnant]

without negatives she looks pale lmao

yes

doing elves to see if something can help those bad ears

[native halo]

model issue

[nova remnant]

wait

what if I... (elf:4.3)

surely that helps

[native halo]

(elf,steven seagal :4.3)

[bitter latch]

you mean these?

pinted vulcan ears doesn't work either

pointed.. hmm takes that out

[native halo]

fake pointy ears

[bitter latch]

confused vulcan elf?

[nova remnant]

omg anime startrek

[bitter latch]

those ears omg

[nova remnant]

those are pretty cool interiors @bitter latch

[nova remnant]

can you give me the prompt for this one

[bitter latch]

Elf-Vulcan Girl in black and green Star trek uniform, (Small Vulcan ears), short hair, star trek inspired background, Below decks tv show, fine art, soft colors, black outline style, elegant, very detailed Cel-Shaded With shaded detail's, soft shadows, masterpiece, epic detail, cinematic lighting, sharp detailed, focused, vivid colors, multi-tonal shading, great shadows, hi-res, HDR, 3.5D, Unreal render,
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, one tail, no extra legs,
Steps: 20, Sampler: DPM++ 2M SDE Heun Exponential, CFG scale: 7, Seed: 308038884, Size: 760x512, Model: CounterfeitV30_25, VAE: counterfeitV25VAE_v25.pt, Clip skip: 2, Version: v1.6.0

cool but I used Anime star trek vulcan in uniformed, tipped ears

I think it got confused

[nova remnant]

I have no idea what vulcan is, thanks

[bitter latch]

https://giphy.com/gifs/spock-star-trek-836gLmVVZgJEI

spock

[nova remnant]

I never watched any of those

[gusty palm]

star trak so good

[native halo]

https://tenor.com/view/skill-issue-club-penguin-club-penguin-dance-skill-issue-gif-24610458

[nova remnant]

EasyNegativeV2, negative_hand-neg, (low quality, worst quality:1.5), (low contrast, oversaturated), (monochrome, grainy, pixelated), borders, (watermark:1.2)

[dawn jasper]

Star fleet?

[native halo]

a beautiful hand 😍

[bitter latch]

Star trek

[gusty palm]

ikr thats the best hand ive ever gotten lol

she tried to give the llap

but shes new

lower decks

[native halo]

idk what happened here

[nova remnant]

interesting way of sitting in a spaceship

[nova remnant]

https://tenor.com/view/stranger-things-fire-flame-gif-14920648

thats so cursed oh mygod

[native halo]

🤣

[gusty palm]

the consequences of a inproperly calibrated warp core

[bitter latch]

no I think thats a break down between realities

[native halo]

this one was better but still bad hands 😔

[bitter latch]

and of course

[nova remnant]

style

[native halo]

thicc ears

[nova remnant]

look like plastic

[dawn jasper]

I know, Star Fleet is from Star Trek.

[gusty palm]

any development on what the freeu thing does

[bitter latch]

ponders making a star trek punk anime style

[nova remnant]

this is with regular prompts

(bad anatomy, bad proportions, malformed), (bad hands, fused fingers, interlocked fingers:1.2), (asymmetric, conjoined, distorted, cloned), (low quality, worst quality:1.5), (low contrast, oversaturated), (monochrome, grainy, pixelated), borders, (watermark:1.2)

[bitter latch]

wow

[nova remnant]

token used 56/75

[nova remnant]

🤨 but ears

[nova remnant]

nice butterfly effect !

[bitter latch]

Miku In Miku world, Starfleet, science fiction, sci-fi-punk, star-trek Punk, fine art, soft colors, black outline style, elegant, very detailed Cel-Shaded, sharp focus, Anime lighting,
Negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, one tail, no extra legs,
Steps: 40, Sampler: Restart, CFG scale: 4.5, Seed: 397894241, Size: 760x512, Model: CounterfeitV30_25, VAE: counterfeitV25VAE_v25.pt, Clip skip: 2, Version: v1.6.0

miku in miku world is prompt, the rest b4 neg is style

[nova remnant]

Very confused trying to think why it wasn't generating an enclosed room for a few too many minutes

beach was the second thing in my prompt forgot it

[bitter latch]

how?

[bitter latch]

hmm prob better model....

[nova remnant]

i'm using calicomix

oh no

I tried that prompt

my miku is in danger

🙂

[nova remnant]

sending help ....

[bitter latch]

ahh better

[gusty palm]

i had to use a startrek tendi lora to get starfleet stuff to work

[nova remnant]

cool ~

[nova remnant]

so seerious

[bitter latch]

Envydream model

[nova remnant]

what is happeningg

theres a hand coming out of who knows what monitor control thing

[bitter latch]

StARRRR TREkkkkk

[nova remnant]

(tight clothes:3.0)

[bitter latch]

huh trying this thing ai comic factory

[nova remnant]

Miku In Miku world, Starfleet, sci-fi-punk, star-trek, elegant, sharp focus, (highly detailed:1.2), aesthetic, (vibrant colors:1.1), cinematic, hyper realistic, (full body:1.2)

[nova remnant]

im gonna try bring Miku in amazon forest setting...

[bitter latch]

oh boy

[gusty palm]

make miku commando squad carrying aks through the jungle

[bitter latch]

added aesthetic, (vibrant colors:1.1), cinematic, hyper realistic, (full body:1.2) to end of my style 😄

[nova remnant]

lol .. that cant cause such high color contrast tho

[nova remnant]

I want to generate stuff faster

Can't keep up with the ideas

bringning miku in snowy amazon forest

[bitter latch]

I think it's the vibrant color

[nova remnant]

you using LoRA too?

miku having fun

[bitter latch]

nope

[nova remnant]

hmm i dunno then, i dont get such strong color tones with vibrant color

[bitter latch]

prob the model then

[nova remnant]

nope

i tested with several mods

[bitter latch]

shrugs

[bitter latch]

.. hmmm?

[nova remnant]

This one looked cool

she can walk on water

[nova remnant]

Now... Underwater

i need to prompt it better

Interesting place to sit

Miku In Miku world, swimming underwater, star-trek, sci-fi, elegant

she's even got a sci fi fish in there

[bitter latch]

ok then

this thing is pretty good

yep it was the model

hmmm ok then

[nova remnant]

Miku In Miku world, studying in her room, sci-fi, elegant, (highly detailed:1.2), aesthetic, (vibrant colors:1.1), cinematic, hyper realistic, (full body:1.2)

[nova remnant]

That shark is very fake

[gusty palm]

i like that one

[nova remnant]

holds laser with hands, your opinion is invalid

[gusty palm]

ever gottten sanrio or hello kitty results

i havent had good luck with those terms but ive gottan close

i got some stunning bedroom interior with no characters though

[bitter latch]

she does not look happy

[gusty palm]

that reminds me iw atched idiocracy the other day

looks like the environment from that movie sorta

[bitter latch]

hmm yeah, I see it

found the comic book lora manhua

not bad

tho not sure whats happening on side panel

[nova remnant]

[gusty palm]

https://cyberes.github.io/stable-diffusion-textual-inversion-models/

has anyone tried any of these

i downlaoded a couple but couldnt get any really noticable results

sygil had a thing called concept library that i never figured out

i feel like this is part of that but idk

[nova remnant]

Miku In Miku world, studying in her room, sci-fi, elegant,
but Miku is missing......

[native halo]

shes under the bed

[nova remnant]

whats she doing there? with who?

[bitter latch]

Anime harry potter using that lora and the modified animetrek style 😄

[bitter latch]

ok gonna set a big batch to run while I sleep.. so to bed I got. nite

thuds

[nova remnant]

later, btw a big batch can cause memorry issue

[bitter latch]

ok only 40

😄

ok ok i'll reduce it by half

[nova remnant]

sd actually created a transparent bg?

[native halo]

no i removed the bg because deformed twin

[nova remnant]

oh

i hardly use photoshop anymore lol

[brittle jay]

good mormijg everynyan

man i wish sd on cpu only would be as good as llms on cpu only

i have 1 vram on the steamdeck

[nova remnant]

Hi!

[brittle jay]

:(

[nova remnant]

What is some other famous anime character stable diffusion might know

Like Miku

[native halo]

asuna

asuka

[nova remnant]

https://tenor.com/view/strive-guilty-gear-guilty-gear-strive-asuka-that-man-gif-4249384552291477030

This is asuka

He's difficult

[native halo]

asuka from eva

[nova remnant]

I only know guilty gear

Dunno what that is

Skill issue

[nova remnant]

Huh

[nova remnant]

Is that chat gpt

[nova remnant]

no .. its bard lol

Hmm

[nova remnant]

I'll try some later

They seem so random to me

But cool if it works xd

[native halo]

try with sasasegawa sasami

[nova remnant]

Did you have a stroke typing that

(are you okay:6.2)

[native halo]

https://tenor.com/view/anime-little-busters-riki-naoe-komari-kamikita-sasami-sasasegawa-gif-23564639

real char

[gusty palm]

anyone tried that freedomgpt thingy

i hate things with the word freedom in the name but

[bitter latch]

ok happy now I goto bed.. thud (again)

[gusty palm]

it looks like the download isnt massive and the vram was like 6gb

[nova remnant]

They get to near the rock, girl is disappearing, they are shocked, my hair is black what is life

[nova remnant]

yeah, that model hallucinates a lot

[gusty palm]

im ok with that i think

is it like a ui that u can load any model into

[nova remnant]

Is it the leaked Facebook thing model?

[gusty palm]

i think its an official release

[nova remnant]

the one i tried a few months ago, you download the files and you run the commands on cmd window

[gusty palm]

llama

[nova remnant]

name sounds very not official

[gusty palm]

https://github.com/facebookresearch/llama

[nova remnant]

llama is official, but freedom gpt came out from a leaked verson of chat gpt i think

[gusty palm]

i needa to research i think nubby recomended some good ones

i want basically a ui like auto11 but for chat

or text

and also low ram

[nova remnant]

why katie?

And then you stop taking to people

you can use Bard for free without any wait time at all

[gusty palm]

i need a locally hosted uncensored ai chatbot

[nova remnant]

it sets no limitations either except maybe pornographic

[gusty palm]

for research purposes

i do like bard tho

and c.ai

brainstormer is very smart

[nova remnant]

my fascination to run language model locally shifted when i discovered SD.

[gusty palm]

https://github.com/oobabooga/text-generation-webui is the one nubby recommended i think

[nova remnant]

that oobabooga has online guide and runs on UI like A1111

Time for me to go

Gud nit

nite!

[gusty palm]

goonight

[gusty palm]

cofeeeee

im about to make sum

[nova remnant]

the girl is missing again

[gusty palm]

its amazing how good that rooms looks

[gusty palm]

my mom would love that style

[nova remnant]

yeah texture is nice but w/o a girl that looks empty

[gusty palm]

i wanna live here

[nova remnant]

nice kitty

do you remember watching this video? https://youtu.be/mTTwcCVajAc

i thin thats' the closest how i feel about girls

[gusty palm]

i think most dating websites hire girls to be on there

so id take w pinch of salkt

[nova remnant]

oooh, i dont use those apps

[gusty palm]

loooll

'i really love cats'

[nova remnant]

lol yes

[gusty palm]

i didnt knwo what this was gonna be

[nova remnant]

she wants to hug every cats but she can't hug every cats

exactly how i feel about girls

[gusty palm]

aaaa feels

[nova remnant]

but the guy is calling her a troll lol

oh its a different video my youtube went on with

the next video some dude calling her a troll

[brittle jay]

this is the good one yes

[gusty palm]

nice i am downloading the install.sh thingy now

i dont have high hopes i have enough ram and stuff but im gonna try it

[nova remnant]

6gb is enough for it prolly

[gusty palm]

hoping these come in handyu

[gusty palm]

what model is that btw

[native halo]

is it breakdomain?

[gusty palm]

super clean looking

[native halo]

or maybe meinaunreal

[nova remnant]

oh... its https://civitai.com/models/107842

was trying to find a decent video guide for oobabooga

most of the video guides have too much humor in them, kinda distracting

the fingers ruined it

[slate narwhal]

The good old days of NAI were magical, and I agree that the public will not see another anime model from them anytime soon, or maybe ever again. Technology always solves one problem and creates more, or several more. In this sense, the upgrade to SDXL made some things worse. SDXL, in some aspects right now, is losing even in photography. RealvisXL v2 is the best, but only for portraits; if you get distances, it fails, and so it still loses to the great SD 1.5 photography models that already exist and don't have this problem with a nice setup. Crooked teeth and distorted eyes even in portraits are fairly common, whether using the refiner or not, and this happens in base too. The refiner thing could solve a lot of problems, but people don't know or don't have the right tools to create their own refiner models. So just like before, everyone is going to use the fixes like Adetailer, and this should not be happening in SDXL as we were supposed to be free of the fixes for basic things. I think RealvisXL will get better and solve the problems with photography soon, but it's still not there.

Let's hope WD gets back and keeps the project alive. There are some similar efforts, like the newly released HorizonXL (released today), which they've trained and merged with Kohaku-XL, a huge XL model. Both models are at around 400k+ datasets. I'm testing it, and it still has problems and is still far from the dataset of WDXL. And from what I've been reading, SDXL seems to have some training problems when the dataset is too large. It seems that the quality gets worse and worse the larger the dataset gets, making training more difficult. I haven't tested this myself; I'm waiting for Everydream 2 with SDXL support. I think this is the reason why we see people training the smaller "good" checkpoints and then merging everything after into one final model. And about Loras, well, I use 5 Loras only for fixes for eyes and quality stuff; there is not much space, memory, or compatibility for..

[slate narwhal]

Loras that could create different styles.

I think a lot of people are noticing all of this, and the way to go is probably merging a lot to get to SD 1.5 levels (which will take a lot of time and effort) and training good bases like the old NAI. Someone did a 400 GB or 500 GB Civitai SDXL Lora mega merge; I haven't tested it, but maybe it does solve some of the problems.

I miss kl-f8-anime a lot. In SDXL, when you get out of the portraits and close-up shots, the eyes are bad no matter what you do. You can get better eyes, but the quality is not the same as it was with kl-f8.

There is still hope in Comfy; the Ipadapter "instant Lora" thing or image prompting in real time solves some of the problems; the merging capabilities and tools people are developing show great promise; but Comfy is not for everyone. (I know Ipadapter is in auto1111, but Comfy and others that derive from it with all of the memory and quality optimizations are so much better that I don't want to use auto ever again.)

[gusty palm]

im just following the github

i made it to where im downlaoding the model i kindof picked at random .. idk what gguf means but pi chose that over fp16 it seems to be almost interchangable

theyre all like 20+gb and multiple files so idk im using the ui and hoping for the best

my hd has 80gb free so i am gonna delete it and try again if i chose wrong

[nova remnant]

i have a strong feeling that a lot of the technological discrepancy comes from a corrupt economic model that in most cases focuses on profit rather than actual progress

[gusty palm]

10000%

[nova remnant]

and a lot of the engineered cirsis are used to further capitalize

[gusty palm]

thats why i love foss

blender has come so far and im just dreading for the day they account theyre going propriatory

[nova remnant]

its systemic on a collective scale ... no one person is at fault

[gusty palm]

ive noticed alot of the products in the 3d printing industry are just gimics with marketing to attract sales rather than meaningfuly improve the industry

not that the industry hasnt advanced but also i dont own a printer but from watchign videos

[gusty palm]

and true

i get depressed if i focus on the fact that its everything society-wide

[nova remnant]

it will change eventually cause this is not sustainable trajectory and can only go on for so long

[gusty palm]

we are feregies 😭

[runic vault]

innovation is hard/costly/takes time all three things venture capitalist hate.
gimmicks cost less and are more easy to hyped up.

[nova remnant]

yep

the economic model we have in the world is debt centric and inherently scarcity driven

if you can engineer scarcity you can make profit .. thats the economic model of today

[gusty palm]

ironic when we . produce more food and housing than we nequire

[nova remnant]

yeah statistics show the earth has enough resources to meet demands of 10 billion, the world population currently is much less

[runic vault]

those under mammon's spell tend to not see/care about the true value of accelerated progress.

[summer night]

what do you guys think is better for generating, RTX 5000 or A4000?
the cloud gpu service i'm using has them and i'm just wondering which is better

[gusty palm]

I think a certain economic system selects for anti-empathy /psychopathic tendency , playing nice isn't as effective strategy for board game monopoly and even worse irl

i would check a benchmark or compare cuda core count/clock speed

[nova remnant]

century old mindset from the old economy has shaped the present day social structure, its gonna take a while to change those mindsets

[gusty palm]

they probably both have adequate vram

i hope the logical mindset takes root and isn't stomped out by dictator class

[summer night]

apparently the RTX 5000 has 3072 cuda cores while the A4000 has 6144

[gusty palm]

since karma and justice are not guaranteed

that's a large difference, check core speeds

if 4 is slightly slower but has almost 2x core count it might be better

[summer night]

RTX 5000 operates at 1620 MHz, A4000 operates at 735 MHz

they also both have 16 GB of vram

[gusty palm]

oof.

that's close

[summer night]

they aren't always both available so i usually have to take what i get anyways

[gusty palm]

twice cores an half speed or half cores at double speed

[summer night]

i have been picking the RTX 5000 when both are available though

[gusty palm]

I'm curious to know

gonna 3d print the world's biggest guillotine

rocket powered

[summer night]

make it electromagnet powered instead

railgun guillotine

[gusty palm]

which service do u use I've never tried hosting a l
clouduld sd ui

I used to run remote Minecraft and various other servers though

[gusty palm]

pulse laser guillotine

rail guillotine sounds cooler

it's cool what u can do with a mega cheap 1gb virtual server

ten years ago anyway

[nova remnant]

btw katie, the change to shift away from this scarcity driven economy is necessary on a global scale, the climate crisis affects everyone due to resource exploits and mismanagements

the dictators over lord cant afford not to change

[gusty palm]

climate change could potentially cause actual food scarcity

[nova remnant]

it just has to hit their senses and is inevitable

[gusty palm]

i figure they will just stockpile decades of rations and live in hermetically sealed bunkers

[native halo]

you should keep an eye on the titan rtx ( 2000 series) it also has 24gb of ram i only saw one being sold at 450$ and that should be the correct price since its older than a 3090 and 3090 is at 700$

[gusty palm]

like how cockroaches survived the dinosaur extinction

[summer night]

would catgirl makima be called nyakima or makinya

[nova remnant]

not a long term option in an unliveable environment

[gusty palm]

hmmm

[summer night]

i'm trying to figure out how to do that img2img thing where people make scenery look like patterns

[gusty palm]

like the QR code ones?

[supple raptor]

I vote for Nyakima

[summer night]

maybe?

These

i don't know what settings to use

[summer night]

are you using aniverse?