🍥｜anime

[modest fog]

You are smol

@green temple she'll treat, you'll become better at noob, don't worry

[vital raptor]

finding a style/artist you like in pony/noob models can be rough, there's so many and if you don't really nail it down then the models tend to shift styles and you don't get consistent results

takes a lot of x/y grids, and even on faster hardware takes so much time

they do both work with natural language, but are so heavily trained on tags that it's just better to use them. I think tags actually work better (for me) because I'm not trying to do anything super specific and make up my own characters

[supple raptor]

tbh neither of them are any good at nl prompting. you can add small bits of it among tags, but you're not going to get good results even remotely consistently unless your prompts are primarily made up of tags.
artist styles on the latest noob models are also much more stable than on typical pony based models, to the point you can use a single artist and will generally get fairly consistent styles across different prompts.

[keen hollow]

yeah the artist consistency is pretty neat

So I've been using the "comfy" release of the flux dev model (flux dev fp8 with the text encoders built-in). And I use SwarmUI.
Is there now support for a faster workflow versus the original flux dev model? Anyone know?

[unkempt cape]

What kinda style u want tho thats the problem

I mean i suck too but i discovered something funny

[green temple]

any cute style but not chibbi ( the pastel kinda one)

"anmi" is the closest to what i want so far

maybe there`s a good artist combo for it ( still trying )

[green temple]

i like some of the styles u posted too

[unkempt cape]

Some of them achieved thru funny means

[green temple]

oh really 😂

[unkempt cape]

Check my lastest one and guess how i got it

[green temple]

ok

[supple raptor]

yeah, having a properly trained model is a very nice change

negative prompt is quite important for noobai. the wrong negative will destroy your gens

[bitter latch]

DMT artstyle interesting

[modest fog]

Grabby

Black smol

[green temple]

Evol

[modest fog]

Nooo

[vital raptor]

smol cant be evol

[keen hollow]

I feel like maybe those are gonna be too well-done. 😄

[unkempt cape]

Ok last pic i go sleep

so i can catch my morning Sunday peak fiction, absolute cinema tomorrow

[modest fog]

No

Free candy is over

No look at me like that

[green temple]

is she late ?

[modest fog]

Only grabby left

[green temple]

grabby what

[modest fog]

runs away

[supple raptor]

https://tenor.com/view/raiden-shogun-bonk-raiden-ei-small-gif-12242670327846551222

[trim current]

beep boop uploading new style lora soon

[modest fog]

cookie

[green temple]

tried to keep it safe it`s safe i guess

[trim current]

well this lora is also a success

and goes really well with my other one

[modest fog]

Unsafe with me

[green temple]

🥲

[modest fog]

Ugly emote

[green temple]

have wheat

[modest fog]

Tired

[modest fog]

eats wheat

dies

[green temple]

why she look big-smol

[modest fog]

For bigger grabbies

[green temple]

no grab

[modest fog]

@green temple she angy

[green temple]

why angy ?

[modest fog]

You didn't post enough smols

[bitter latch]

haha

using the animal wildcard for onesie worked differently

[modest fog]

@green temple smol hater!

Angry smol

[bitter latch]

nah, emo smol

[vestal ferry]

https://cdn.discordapp.com/attachments/1271917327494942793/1313844750448721940/eca3df0457da64629c0afd35eea2c373.gif

[green temple]

oh yeah , the smol hater

[summer night]

the LCM sampler is weird

[near elbow]

more flowers 🙂

[modest fog]

Smoking is bad

[modest fog]

Flower

[steel zephyr]

got new ears

couldn't do what I wanted

she's supposed to touch tip of the ear while looking to the side

but SD said no

[trim current]

pointy ears, elven ears is what i use to get decent ears

[green temple]

try hand to own ear ( single hand), looking away, looking at side, front view ( or whatever view you want),

[steel zephyr]

i did put all possible variants

but even at :1.4 it almost does nothing besides hand behind head

[modest gyro]

@steel zephyr send me a dm for assistance mate

[steel zephyr]

btw what causes those glitches? high cfg?

it happens out of nowhere like 1/100

[trim current]

lora?

also could be high cfg yea

[green temple]

ok this is bad , but something like this ?

[steel zephyr]

more something like this

but hands are a bit cursed

[steel zephyr]

to my surprise, I couldn't even find a reference to my idea

[modest fog]

[gloomy scroll]

Yue is looking for body pillow

[modest fog]

I ate it

[gloomy scroll]

Then there isn't other choice than borrow that black haired friend

[modest fog]

She's too skinny to be soft

More pillow

@gloomy scroll

[green temple]

i tried to delete the tail is my smol inpainting ok ?

[vestal ferry]

https://cdn.discordapp.com/attachments/1113950544143335525/1308019922349195284/BadDoro.gif

[trim current]

can you guess the anime!?

[modest fog]

Avatar the last Airbender

[trim current]

[modest fog]

don't look at me

I'm trying to use my AI history memory to see if I know this character i don't

[vital raptor]

doesn't look like anime to me, more like a bad american tv show

[trim current]

oh its a joke cos its a family guy lora i made

[quiet turtle]

family guy

[trim current]

yes lol

[quiet turtle]

awesomeeee

btw do u mind if i dm u for a little advice based on that?

[trim current]

if you have questions you can ask here idm

also @magic wave would be a better person to ask, shes way better than me and i always annoy her for help

[quiet turtle]

bwahahah

way to let ur friend under the bus 😭

[trim current]

[quiet turtle]

i just wanna know which are some good models to use

im absolute beginner just installed it 💀

[trim current]

like to generate stuff? depends, first id need to know what kind gpu yhou got

how much vram it has

amd or nvidia

[quiet turtle]

okay so

its nvidia

16 gb

[trim current]

i use a 3060 with 12 gb, and im able to use pretty much all the fancy stuff,
what kind of images will you be wanting to make, anime stuff (guessing cos we in anime channel), realism, etc

[quiet turtle]

RAM-->20GB
VRAM-->16GB
GPU-->NVIDIA RTX A4000
CPU-->INTEL CORE BROADWELL

im using a cloud pc

[quiet turtle]

okay gotchu gotchu i think anime stuff is a good start ?

[trim current]

yeah, look into PonyXL models, Illustrious Models, those are the base models on SDXL (1024res image gen), i recommend you start out with a basic UI so you can learn as you like Invoke or Focuuus, but if you want to jump into what most of the people are using, A1111 Forge and Comfyui/Swarmui. Youll wanna make sure you find and download sdxl VAE to use alongside the models.

[vital raptor]

most models have vae baked in, haven't seen one that actually needed a vae in awhile

but it doesn't hurt to have it

[quiet turtle]

okay thank uuuu i'll post updates here

can i seee some more of ur stuff 🙃

[trim current]

oh rly? ive been w sd since 1.5 so just force of hasbit to use one lol

[modest fog]

summoned luna how dare you

[trim current]

luna is my homie ofc i did

[magic wave]

I reccomend NoobAI, one of the Epsilon Pred which are the straight forwards ones to use or the latest V-Pred model which are fancier but take a lil getting used to, they're getting better though, not too difficult nowadays :o

[magic wave]

True :o

[trim current]

Comfyui is the best tho totally not biased

[magic wave]

I personally use reForge cause it looks nice and its easy to use

[trim current]

for some weird reason that i dont understand is reForge doesnt let my trained loras work

its ONLY in reForge i cant use my loras

and idk why

nothing in console

just

wont apply

[modest fog]

cursed lora

[summer night]

i wanted to move from old forge to reforge but the extensions not sorting properly on the generate page is a dealbraker for me
on old forge they sort alphabetically but on reforge there doesn't seem to be any method of sorting them

these are what i'm talking about

on reforge i can't get the aspect ratio selector to be above all the other extensions, i tried adding like 20 letter As to the start of the folder name but it just didn't budge in the order

so this whole time i've just been using the same version of forge from before they switched to a newer gradio version because a1111 generates slower and reforge doesn't sort things right

and comfy is too complicated for me

[magic wave]

You could use this, not a fix, but may help, its a different UI that looks fancier and sorts extension settings on a different tab

But I can respect the dedication if you stick with regular Forge, I basically already memorized aspect ratio sizes xD

[summer night]

am i really missing out on anything other than SD3 and flux

[magic wave]

Uhhh, iirc better performance, and I think V-Pred support

[summer night]

noobai is a v prediction model and it was working properly on this version of forge when i tried it

[magic wave]

Well, not all NoobAI models are V-Pred, and if you got these 2 extensions then yeah it should work fine, just missing out on the improved performance side

[summer night]

the v-pred one is the one that i was using

it works on this april 2023 build of forge

here's an example

idk if the fuzzy details and messed up hands are just how noobai is, but i switched to using one of the illustrous models because i thought it worked better

noobai, at least the version i tried, couldn't really do hands well at all compared to pony and illustrous

i prefer illustrous over pony now simply because it recognizes more characters
i feel like pony deliberately excluded character tags from the training data or something

[trim current]

i just like the fidelity difference

my tism picks up the small details

[summer night]

ayo pony

what's going on bud?

[vital raptor]

pony tends to be more artist/style focused and not as heavily trained on specific characters

[summer night]

i feel like being reliant on loras for characters is a big negative due to how they take up more space and influence the style

[vital raptor]

i don't use loras or generate specific characters, so it's fine

[modest fog]

More purple girl

[vital raptor]

more smol

[summer night]

are you sure
also these are on 2 separate grids so i could give the pony models their score tags

pony models never seem to be able to do any of the artist styles i try

[vital raptor]

i didn't say it was trained on every artist

the pony model was trained mostly on danbooru tags, some of that will include characters and some will include artists, some get trained better than others

noob/illustrious on the other hand, had specific attention paid to training characters, and maybe artists, i don't know exactly what they did

the main point is that i stuck with pony instead of moving to noob because it sticks to the styles/artists I like better and it doesn't get contaminated with character information because it's not heavily/specifically trained

but, like everything else, if pony doesn't do it for you, don't use it 🤷‍♂️

[supple raptor]

That's just not how any of this works. Including character tags doesn't make a model learn styles worse or "contaminate" them.

[vital raptor]

i mean, it maybe over simplified, but if a certain tag is trained more than another it works better than others

[supple raptor]

Noob and illustrious are good at characters and styles because they utilized better training techniques and much larger datasets

[supple raptor]

Not when trained properly

[vital raptor]

and if that certain tag is also associated with other specific tags, those tags will bleed into the results just by using something associated

[supple raptor]

See: noob/illustrious's ability to create character cosplays even when there is no cosplay tag for the associated character.

[vital raptor]

usually the easiest test is something like 'blonde hair' if it always generates blue eyes, that's what i consider 'contamination'

and maybe noob/illustrious models dont have any of that, i haven't really tried much, but most models do

[trim current]

[supple raptor]

That's bias. And proper training goes a long way towards addressing it.

[trim current]

sounds like no regulation pics

[supple raptor]

Pony didn't even use caption dropout, nevermind "anti-miku engine"

[summer night]

i feel like anime characters with blonde hair hardly ever have any eye color other than blue, green, yellow, or sometimes red

irl people with blonde hair more commonly have blue or green eyes simply because of genetics

[vital raptor]

which is why a lot of models tend to generate blue eyes when you only prompt blonde hair, it's a popular stereotype

[summer night]

i mean, just prompt the eye color you want

it's as shrimple as that

[vital raptor]

whatever source images are in the dataset generally have small associations like that, i use ponyXL based models because for the images i generate i run into very few of these 'bias' tag associations

and it adheres to the styles i use very well with very little drift

[modest fog]

Ugly green monster

[modest fog]

runs

[green temple]

i tried 🥲

[gloomy scroll]

Yue is willing to offer helping hand with massive backpains

[modest fog]

I need help with something too

[gloomy scroll]

Here's your headpats

[modest fog]

nurse I feel pain everywhere

[keen hollow]

looks like Elsa

[humble current]

any curly hair enjoyers?

[bitter latch]

psst @lone got a new piece of .. whatever this thing is for ya

[humble current]

I'll send her after that guy

[glacial coral]

NoobAI V-PRED V0.9R goes hard

im still shit at prompting

also question anyone got a workflow for V-PRED versions of models that i could use that would be helpful

[vestal ferry]

https://cdn.discordapp.com/attachments/1271917327494942793/1313844750448721940/eca3df0457da64629c0afd35eea2c373.gif

[trim current]

https://civitai.com/models/1045095?modelVersionId=1173476
sharing in case it helps someone

[steady grail]

genning with ip adapter makes genning weird stuff easier lol

[modest fog]

Fake elf

[jagged axle]

Very very beautiful picture anime and very very beautiful dress.💖💖💖💖💖💖💖💖💖

[humble current]

another earlier gen. I love the sleeves

[bitter latch]

I always go back to this style somehow 😄

Hello Mister Bond!

miku go !

live action miku?

hmm interesting...

[terse ore]

For noobXL, is it good to almost always use aesthetic quality modifiers lora? I'm looking at the examples on civitai and it is used heavily

[supple raptor]

the only loras i've used on any of my ~5k gens on noob models have been artist loras for styles that aren't found on danbooru. loras in general are generally unnecessary for noob. the one i see all the time on civitai is the one that's just trained on a some pony gens that seems to give a more 3d look. usually it's paired with about 4-5 other loras so hard to say what any of them are actually doing.

[slate narwhal]

nice one. The neon lines/line art effect in the prompt is really nice

[near elbow]

thanks - flux fusion, not even an anime model 🙂
what are you using?

love the more abstract drawn style

[slate narwhal]

I'm really liking this mix.

Another one with the neon lines/line art in the prompt:

[slate narwhal]

I'm using Flux + a mod of my own lora (my lora is on civitai as I_Pomo_Mix, civitai.com/models/652317 ) , and the excellent Ivan Beslic Lora by Chronoknight, get it at civitai.com/models/920923) the data with everything I'm using should be in the image

[slate narwhal]

Ok, the other one is the Ivan Beslic Lora (civitai.com/models/920923), it helps with the abstract lines.

[near elbow]

sounds great - i'm hesitant with loras, always feels like a crutch, but the results are cool

"anime style, minimalist art line drawing, abstract, neon color scheme, girl face with rainbow hair, ink splat, rough drawing"

"anime style, minimalist art line drawing, abstract, neon color scheme, girl face with rainbow hair, ink splat, rough drawing, pixie cut, black uniform, color explosion radiant background"

"anime style, rough charcoal sketch, flat 2d cel shaded, minimalist art line drawing, abstract, neon color scheme, girl face with rainbow hair, ink splat, rough drawing, pixie cut, black uniform, color explosion radiant background"

[near elbow]

ok, now we're getting fancy 🙂
"biomechanical cyberpunk anime style, rough charcoal sketch, flat 2d cel shaded, minimalist art line drawing, abstract, neon color scheme, girl face with rainbow hair, ink splat, rough drawing, pixie cut, black uniform, color explosion radiant background . cybernetics, human-machine fusion, dystopian, organic meets artificial, dark, intricate, highly detailed"

last one, don't want to spam the channel 🙂

[steady grail]

i don't think anyone will get tired

[slate narwhal]

I like it. And flux LoRa merging expands the possibilities even further.

[steady grail]

see ya

[supple raptor]

Nothing about LoRAs makes them "a crutch" it's just a far more accessible method of finetuning that's also far more economical to distribute and store.

[near elbow]

well, if i can reach very similar results via prompting, i'd rather skip a hundreds of megabytes download and the additional processing time 😇

[supple raptor]

If you can get similar results through prompting then the person who trained the LoRA waisted their time

[near elbow]

and i totally get it that LoRAs make sense when the AI doesn't know about a concept or a specific character, but most style-only LoRAs i've come across are quite close to what a more elaborate prompt brings.
well, whatever works best for everyone, just my personal mindset 😉

[summer night]

i'm looking forward to the 1.0 release of noobai-xl v-pred

we finally got something on local that's on par with/better than novelai for now

[modest iron]

testing some flux1 + Retro anime flux LORA... pseudo Afro Samurai (anime)

[modest iron]

20 steps

50 steps

[modest iron]

loading 100steps

trying to understand the maximum capability of my new gpu XD

[modest fog]

Smoking

Bad

https://tenor.com/view/rrane-battal-rrane-battal-kızgın-battal-kızgın-gif-25355985

[modest fog]

Is this text to image or something else

[modest iron]

maryjane warriors united

[modest fog]

Blocked and reported

[modest iron]

@modest fog😂

changer ur name to order... youre not in the chaos side of the force

[keen hollow]

pure text to image. Dall3 is my cheat code 😉

[modest fog]

@green temple

[zinc flax]

She :D

[near elbow]

step count is boring - increase the resolution 😉

Negative prompt: photo, deformed, black and white, realism, disfigured, low contrast
Steps: 7, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Seed: 394081776, Size: 2048x1024, Model hash: a40a0a48fb, Model: fluxFusionV24StepsGGUFNF4_V2NF4AIO, Version: f2.0.1v1.10.1-previous-633-ge073e4ec
Time taken: 11.3 sec.
A: 15.82 GB, R: 16.47 GB, Sys: 18.7/23.9883 GB (78.0%)```

[glacial coral]

Does that say V24Step does that mean V2 4 step?

[near elbow]

yeah, but i run it at 7

[glacial coral]

Ok thx it does anime well or does it do better on other styles?

[near elbow]

no more need for a special model, anime and photoreal and expressionist and noir and comic and whatever else

just "a cat" with anime prefix

no prefix, default is photo, same seed

expressionist style, still same seed

comic style

biomechanical

etc etc 🙂

[glacial coral]

Ah thanks one more thing with text how well does it do can it more than one sentence and still have the text fairly accurate like how the caption at the bottom and then another text or does it only do one sentence well and then anything longer I may mess up? (Sorry if my explaining is bad)

[modest fog]

but where blue

[green temple]

blue became gray

[modest fog]

Blue blue

[modest iron]

2k x 1k res... almost uhd, 4k always oom

[modest iron]

any tips to control net with flux in comfy ui?

[green temple]

@modest fog look , 4k smoking

[modest iron]

@green temple call him order, chaos is too much

[modest fog]

Blocked him

Idk

[modest iron]

LOL

[green temple]

no , smol is chaos that`s why he is chaos

[modest iron]

sorry im high, cya

[modest fog]

Hariy

[modest fog]

Just block this trash

[modest iron]

It's funny that the guy blocked me for a simple joint, but he calls me trash, I would like to see a hand drawing by this pseudo artist, apparently he must be just another rich kid who thinks he knows something because he knows half a dozen codes that any idiot, like me and him, can learn in 7 days.

clowns

[vital raptor]

https://tenor.com/view/first-time-meme-first-time-the-ballad-of-buster-scruggs-gif-24656975

[supple raptor]

i don't use flux so can't help very much, but iirc flux's controlnets aren't actually controlnets so you have to use the differently than normal. there was some discussion about it in the SwarmUI server. you might check over there if you don't get a response from someone who knows about it here

[keen hollow]

this is neat, how'd you make this one?

[trim current]

that feel when finally get the lora results you wanted after 2 days of messing with the same dataset

[gloomy scroll]

https://hailuoai.video/create

[modest fog]

What lora were you cooking

👀

[near elbow]

text quality and features highly depend on the model used. this is with flux fusion and simply
"speech bubble text "Depends on the Model", a friendly artist pondering, atelier background"
in abstract expressionist style. older models like SDXL or the old Stable Diffusion 1.x totally struggle with text, and the longer the text the more luck (or tries) you need to get it completely and correctly

"gothic headline text "Just play" subline text "it's fun""

[glacial coral]

Ty

[near elbow]

well, close enough 🙂

[green temple]

[trim current]

personalami style

[trim current]

m akes me think of the have you ever had a dream kid meme

[near elbow]

yeah, i love it when the AI comes up with stuff. the prompt was super simple

Steps: 7, Sampler: Euler, Schedule type: Simple, CFG scale: 1, Seed: 1056506656, Size: 896x1152, Model hash: a40a0a48fb, Model: fluxFusionV24StepsGGUFNF4_V2NF4AIO, Version: f2.0.1v1.10.1-previous-633-ge073e4ec
Time taken: 4.9 sec.
A: 13.73 GB, R: 14.62 GB, Sys: 16.3/23.9883 GB (67.9%)```

[trim current]

wow that was super simple lol

i plan on trying to see if i can train a flux lora at somepoint today, saw fluxgym supporta 12 gb so curious how long itll take

for 2k steps takes me about 2 hours,

which isnt great but couldnt be worse

for XL

[green temple]

smol prompt

[near elbow]

giving room to the AI is more fun than describing an image in mind-numbing detail (at least for me)

[modest iron]

can u tag me?

[modest fog]

Smol

[keen hollow]

Interesting. Maybe I'll make a burner google account and check it out. I hate these sites that want all my other platform information. 😛

[bitter latch]

ahh better Miku BOT-mech-thing?

cross-eyed

ok.. umm.. accidental furry miku kitty

[humble current]

AGH EXTRA FINGER RUINED MY NEKO GIRL

[slate narwhal]

Melting lava hands...o.O

[modest fog]

Bloo

[modest fog]

Wyt

[modest fog]

Purp

[green temple]

[green temple]

woite

[green temple]

porp

[modest fog]

@green temple bloos

[green temple]

left

[modest fog]

Florwa

[modest fog]

eep

[zinc flax]

cuute

[vital raptor]

i usually change characters every few days, but i've been stuck on her for a bit, so cute 😄

[zinc flax]

she a cute

[supple raptor]

https://tenor.com/view/dungeon-meshi-delicious-in-dungeon-marcille-donato-senshi-dungeon-meshi-dungeon-meshi-senshi-gif-15862516973591178775

[humble current]

Sometimes you just need to sleep in

[green temple]

my eyes

[rough oxide]

[modest fog]

where nai 4

[rough oxide]

it's here

literally

they just announced it

that's naiv4

[modest fog]

my eyes

too much text

I didn't ask for this

just wanted to see some gens from it

I know you'd have

ugly

[rough oxide]

pfff. i like my anime screenshot style.

i can try one of my old prompts that used a more artistic style if i can find it, one i used in noobai a lot

seems i have to modify it a bit

prompting style feels a lil diff than v3, getting different results

[rough oxide]

how's this

kinda messed up on one hand tho uhhh

i might have to tinker with prompt a bit more xD

bettr hand.

[modest fog]

@rough oxide is it smorter than the current one

[rough oxide]

wdym current one

[modest fog]

v3

[rough oxide]

i think so, it feels more like noobai but like more effective

[modest fog]

ugly pink blue girl

[wooden sail]

Illustrious Pony Mix 3 + style transfer for "texture"

[rough oxide]

someone mentioned a neat artist sumwhere and it made iono more badass xD

[modest fog]

looks cool

[modest fog]

no feet allowed

[supple raptor]

why ride broom with shoes?

gotta feel the wind betwen ur toes

[vital raptor]

smol feet are allowed

[modest fog]

Angy at everyone, even @green temple

[bitter latch]

ha finally got the eyes

changes sampler, add's (chibi:2) at beginning , generate

waits

gah nitemare

warning nitemare fuel

haha, peeka boo

[vital raptor]

Is that hard?

[bitter latch]

ooh

[vital raptor]

heterochromia, (grey eye:1.2), (red eye:1.1),

[bitter latch]

here have data

hmm I thnk using three vae's at once might be messing with it 😄

clip skip 12, even the woman is wtf

[vital raptor]

well you're basically 'skipping' the last 12 steps, i wouldn't expect much unless you're doing 30+ steps at least

[bitter latch]

hmm almost looks final fantsy

ooh idea

[bitter latch]

hmm (rounded eyes) umm... well umm.. ok

[humble current]

an eye in a text box...

[bitter latch]

😄

[bitter latch]

better

finally got small pointed ears--> small point human ears

[bitter latch]

takes a break from art and rocks out to some Shamisen × Rock

[near elbow]

have some miku action thanks to local hunyuan 🙂

[humble current]

lets put on some suits

[limpid coyote]

Hey anyone know of other discords that are AnimeAI focused?

[supple raptor]

東方Project AI (Touhou Project AI) is popular. despite the server being touhou oriented, the ai-art channel is assorted anime

there's also the AIBooru server

[rough oxide]

seconding touhou projact ai server. best anime focused AI server imo.

and then each like state of the art model has its own server like the laxhar dream lab sdxl server for noobai, anlatan for novelai, purple something for pony (i forgot name of pony server)

[slate narwhal]

PurpleSmart.ai is the pony server

[supple raptor]

LyCORIS is another one. That one covers all of kohakublueleaf's projects

[unkempt cape]

Thirding touhou ai

U MAY get early access to best model like illustrious

Like test that stuff via gradio sometimes

[stable crane]

join my server too

[vital raptor]

is it fun?

[stable crane]

I think so

[vital raptor]

invite?

[stable crane]

link in my profile

[modest fog]

lotta server is bestest

👍

[heavy portal]

hey

[bitter latch]

hmm tried running SD3.. nope nope nope.. BSOD

[tepid forge]

I made this amazing video with CapCut. Tap the link to try it out! https://www.capcut.com/t/Zs8A7jkUL/

[wooden sail]

Space Ritsu

[trim current]

if you have knly 4gb vram idk why you would try when it makes my 12gb 3060 cry

[wooden sail]

Abstract muffy

[bitter latch]

oops to much diet

[analog dagger]

idk if this count as anime

[rancid remnant]

[vestal ferry]

https://cdn.discordapp.com/attachments/1271917327494942793/1313844750448721940/eca3df0457da64629c0afd35eea2c373.gif

[modest fog]

are these vee four

[rough oxide]

yeh

[rough oxide]

vee four is gud.

[modest fog]

what kinda of prompt do you do for text stuff in it

[rough oxide]

the same as noobai basically

but then there's a new character prompting thing

so instead

all general tags can go in one spot

and then characters can have their own spot

like i did for this one

that way char details don't get mixed up as often, and you can give each individual char a specific pose & such

& also their position lol

[modest fog]

show me an example

[rough oxide]

looks like this

but u can add as many chars as u want

result for that one

[bitter latch]

you said what.. say it again.. please?

[bitter latch]

hmm loooks like shes holding a card against humanity card

[supple raptor]

Here's an example with 2 characters and the new target#/source# syntax

[modest fog]

Looks like that regional thing

[supple raptor]

The use case is similar but the nai devs claim it doesn't function similarly.

And it's far more user-friendly and flexible because you don't have to specify the regions that characters should be in (though you also can do that if you want)

[modest fog]

closed source

[supple raptor]

Yesh

[rancid remnant]

why isnt there a way to do this in Stable diffusion yet

[modest fog]

Cause someone needs to create it first

[steady grail]

oh, so cute

[supple raptor]

According to the devs, these functions are built into their model, not add-on inference tricks like basically everything that we use.

[rancid remnant]

Well shit then

[supple raptor]

Maybe someone in the open source space will figure out how to do it now that they've done it. I'm not holding my breath though because I can only assume it requires much higher quality captions than anyone else is using currently

[rancid remnant]

hope someone bits the bullet and just figures it out

i kinda need that

i want to prompt two people but doing the whole

img2img thing or inpainting or any other method ive seen is just ehhhh

i can prompt two people without any of that it just takes forever to find the specific format but once i do that i can generate a bunch of really good images of that set of characters but every time i start a new set i have to do this all over again

[supple raptor]

at least with semi-popular characters, noobai/illustrious can often just do it out of the box. if you need to describe each character you're stuck using regional prompting though. in that case, InvokeAI is easily the best local UI to use for regional prompting

[rancid remnant]

nah no one im making is popular ngl

i had to make my own loras for them

probably why im having trouble

one lora might have just been better, i think

i just dont want to retrain this lora for an 8th time because im not sure whats messing it up

is it my tags

[supple raptor]

id suggest trying invokeAI for regional prompting then. it's the only regional prompting i've found to be worth the time at all

[rancid remnant]

invokeai?

is it an extension or is that a site

[supple raptor]

it's a completely separate UI https://github.com/invoke-ai/InvokeAI

[rancid remnant]

[supple raptor]

they have very good inpainting/outpainting and regional prompting

[rancid remnant]

wdym its its own UI

hmm downloading it

lets see what it does

is it basically like forge

i just throw models and loras in and make stuff

[rough oxide]

if the chars you wanna make have at least 20-30 tags (with some good high quality solo art) novelaiv4 might be able to do it. (but it's not local)

also it can do prompting for more than 2 people very easily

has a whole interface for it

(i usually just do 2 chars tho)

[rancid remnant]

hmmmm

[rough oxide]

but on left, for instance, char details would never merge or anything

[rancid remnant]

i dont think nai has the characters im making/plan on making ngl

[rough oxide]

u have a name for the char?

i could test it for u

[rancid remnant]

idia shroud

[rough oxide]

um

that'd 100% work no problem

this is the char baseline

no artists

[rancid remnant]

okay to be fair i threw out the least weird character i was thinking of

ill try nai then

[rough oxide]

nai is paid tho

[rancid remnant]

i know

[rough oxide]

just making shur. as long as you're fine with that then yea looks like they've got plans to update it for a while cuz currently

[rancid remnant]

i likely wont use it permanently just try it

but

[rough oxide]

this is all on a "preview" according to the site/blog

o the preview has no nsfw right now tho, nsfw tags don't seem to work.

but i think it will work in full based on how it was for naiv1 curated/full

[rancid remnant]

i still have to find out how to do this on a1111

[rough oxide]

yea so uhhh

for local

i would NOT recommend a1111

if you want a ui

i'd recommend swarmui or reforge, otherwise comfyui

cuz noobai uses a v-pred model

which essentially means you need different settings

& you might wanna use the MaHiRo CFG fix node on comfy too to make gens better

[rancid remnant]

im using uhhh

stability matrix, stable diffusion forge it seems

[rough oxide]

reforge is a fork of forge i think, but from what i hear it's better

[rancid remnant]

define better

what does it do better

[rough oxide]

also here's with a random wildcard artist lol

[rancid remnant]

oh wild

[rough oxide]

u can customize crazily

with artist tags

[rancid remnant]

yeah but nai doesnt have nsfw does it

[rough oxide]

the curated, currently, does not.

but blog post says they're priming a "full" version

based on prev releases, the full will have nsfw again. the char probly also works in nai v3 which does have nsfw.

[rancid remnant]

they just need to make forge capable of making multiple characters easily

[rough oxide]

yeh

so uhh

i can show u an example 1 sec

of how it works here

not the exact prompt i used

but u can do something like this on nai

then add more chars

[rancid remnant]

i had to make a lora for broski

its alright

[rough oxide]

and either choose their positions

or let them go left to right

[rancid remnant]

it just doesnt do multiple characters it seems

the quality drops hard when you add a second character + i cant seem to prompt in a way that keeps the characters separate

[rough oxide]

yea, v4 does that 100%

[rancid remnant]

sometimes it just mixes the two characters together

[rough oxide]

even if the other chars aren't humanoids

[rancid remnant]

wtf nai

[rancid remnant]

whys nai gotta be PAIIDDDDD

[rough oxide]

probly how they can afford the funding for all their gpus xD

their v3 paper said they trained on an absurd amount of gpus

for an absurd amount of gpu hours like 75000

i gotta find paper so i dun write wrong num xD

256 h100s

[rancid remnant]

trying to think of something theres no way it can do

not yet at least

[rough oxide]

u can suggest lesser known chars. but the criteria seems to be around like 20+ tags on danbooru with good quality pictures/data

[rancid remnant]

oh my god, can it do cartoons

[rough oxide]

it's not like an omni model like flux

but it can do toon style?

danbooru has toon style

[rancid remnant]

i mean like characters from cartoons

worded that badly

[rough oxide]

not reaaallly? it's pure anime

[rancid remnant]

ah tragic

[rough oxide]

but u can get figures

and fumos

[rancid remnant]

where my issue lies

im doing a lot of cartoon characters and game characters

[rough oxide]

u can get a lot of game chars

if those games are popular in jp

[rancid remnant]

ye

but not the cartoon characters probably

[rough oxide]

u can get...

SOME cartoon

does south park count as cartoon?

and any like anime-ized versions of some cartoon chars

like finn the human

[rancid remnant]

BAHAHAHAHAHA

i mean yeah those count

[rough oxide]

okay gud

glad this counts

xD

[rancid remnant]

LMAO

thats crazy 😭

[rough oxide]

we're talkin

how ppl on danbooru draw these chars

lmao

[rancid remnant]

im wheezing

[rough oxide]

(danbooru sourced from like, pixiv/twitter and only high quality-ish stuff)

[rancid remnant]

try uhhhhh

[rancid remnant]

think it could do like

2012 spiderman

characters

[rough oxide]

before that

i wanna try harley quinn from suicide squad

xD

cuz u reminded me i haven't tested that yet

[rancid remnant]

fair enough

shes popular

i bet she could be made

oh my god

my

my

my lora just made the weirdest shit

but its so clean too?

what is this though

[rancid remnant]

i-i dont know if i can send this

[rough oxide]

she's popular but u'd be surprised

[rancid remnant]

its not nsfw

[rough oxide]

no other model can do this

lol

[rancid remnant]

but

wtf is this

[rancid remnant]

not bad tbh

[rancid remnant]

spoiler cuz its just weird LOL

whys he back there 😭

whys he got better shoes than me

[rough oxide]

sort of

even got the purple/red eyeliner now

[rancid remnant]

they turned her into a kid

bring my harley back

[rough oxide]

so uh wuts this about 2012 spiderman

xD

[rancid remnant]

see if it can make any of the characters from ultimate spiderman

its a cartoon from 2012

[rough oxide]

i think no

my wildcard just landed on superwoman tho

as we're talking about marvel

xD

to tell if it'll work the #1 test is to search the tag on danbooru

some stuff might still work cuz they might have data from elsewhere?

cuz anos voldigoad worked

but hmm

[rancid remnant]

tragiiccc

worth a try though

[rough oxide]

my met transformed

well

for nai too

they have something that shows you if that tag was in the dataset

and u can see it doesn't pop up

on the auto complete either

these ones do

more filled in/bright = almost guaranteed to work

[rough oxide]

2099

[rancid remnant]

ayyeee thats cool af tho

[rough oxide]

can get suuuper close for ssi harley

lol

not sure how to describe the eyes tho

but even got the zipper bra thing xD

nai is pretty decent at more natural language

[rancid remnant]

being able to prompt in natural language all the time would be great

[rancid remnant]

installed how do i use this thing

oh neat found out how to install the models and loras from my local files

its only needs to install 673 of them

rookie numbers

[rough oxide]

make sure u even like the ui first b4 u go thru all that effort xD

but

otherwise if the loras r saved in a folder somewhere on ur comp

shouldn't invoke just have a way to plop em in like other uis do?

[rancid remnant]

oh yeah theyre all saved

i just had to go to involes settings and tell it to scan the folder

[rough oxide]

like a models/lora folder

or something

o

yeh

i never used invoke tho so idk what it can/can't do

[rancid remnant]

idk maybe but i just copied my lora path and pasted it

and its installing everything from it rn

i need broski who recommended this to come back

[rough oxide]

i'm so glad we're effectively in the lora-less age of ai

i had 200 gb of pony loras xD

and was still rising

[rancid remnant]

i HAD over 400 gigs but

i deleted loras and models from older generation like sd 1.5 recently

[rough oxide]

if ur using noob u likely don't need a lot of loras.

[rancid remnant]

so its like half that rn

[rancid remnant]

im not.

[rough oxide]

noob can do anything like 80+ tags

....then i'd recommend if ur doing local to use noobai xD

[rancid remnant]

plus i like loras.

[rough oxide]

cuz it's the top anime model right now

[rancid remnant]

for specific stuff.

no

i use a specific model

cuz its what i train my stuff on

[rough oxide]

mfw

[rancid remnant]

im not retraining loras so i can switch models

that costs MONEY

[rough oxide]

i mean

that's the thing

noobai literally for the most part

doesn't need loras

xD

[rancid remnant]

thats cool

still no

i like the model i use

[rough oxide]

fair

lol

if it was the nsfw u were worried about, noobai can 100% do that too

i mean, i'd recommend at least checking it out

and comparing to ur model xD

[rancid remnant]

nah its the styles and characters i want

that i know it wont have

since its trained on danbooru

and i already have loras trained for some of the styles and characters i need and i dont wanna go back and have to retrain them after finding out noobai has better quality but 0 data on what im looking for

[rough oxide]

i can refute the point about styles too; there's roughly 35,000 artist tags that work, some combination/weight of those can usually get you a style you're looking for, but fair. it's just new-gen models like noobai/novelai v4 far surpass anything that came before them xD

[rancid remnant]

[rough oxide]

but yeh, i can sort of relate to that, when pony was best local i kept using anima pencil 'cause i liked the styles way more xD

going between anima pencil/naiv3 vibe transfer

but i had to eventually switch to pony cuz every lora was made for pony

but now errything is lora-less and my ssd thanks me xD

[rancid remnant]

[supple raptor]

it has to index and hash all of them. took ages to do with my collection of stuff, but i have a ton of models and loras and they're all stored on a slow hdd

[rough oxide]

....kinda glad i didn't try invoke during my era of loras then.

[rancid remnant]

my stuff is loading quick af ngl