🍥｜anime

[nova remnant]

extremely smol

[zinc flax]

Hella cute

[gloomy scroll]

[steady grail]

the fox girl is so charming

[steady grail]

what's the prompt, i want to test it on the model

[nova remnant]

grabby

[near elbow]

healthy glow

[steady grail]

with part of the prompt of funya

[steady grail]

derpy hands, here we are again lol

[zinc flax]

Pretty much expected ig if you don't prompt hand related stuff

[steady grail]

fair enough

not that i can't fix it easily

just testing stuff

[keen hollow]

Need more Tifa and Aerith in here...

[steady grail]

agree, but tomorrow bro

going to sleep

see ya guys

[unique mortar]

some desktop and cellphone wallpapers

[brittle rampart]

Guys, tell me what could be the problem with this error occurring when using ControlNet in Forge?

I'm currently using this version

[unique mortar]

Ms. Devil

[steady grail]

wow, intriguing

[raw estuary]

/dreamprompt

[median thunder]

nai clip
that's not a thing that exists though. it's just clip

[vital raptor]

he's talking about the clip from nai

the old leaked one

[median thunder]

like i said, it's just clip

[steady grail]

not sure what he mean by just clip

like saying, a model from a model or another makes no difference

[median thunder]

it's the original clip trained by openai

no modifications

so calling it nai clip is a bit funny, since we didn't do anything to it

[steady grail]

that's the point, every merge, improper training or improper stuff done after that point, just ruined it

mutations and weird results just increase by a lot

trying to improve it little by little can be a funny experiment

[supple raptor]

i always just assumed you guys trained the te and never bothered to check for differences myself. the main point there was just ensuring that any poorly performing te's from attempting to merge models (often that were of unknown provenance themselves) are gone.

[median thunder]

it's kind of funny how apparently a lot of people assumed that and never bothered to check haha

[supple raptor]

i guess the real lesson here is that training TE was probably a mistake in a lot of the SD1.x finetunes that we have

[keen hollow]

Yes! Well done.

[magic wave]

Mushroom Wife™️

[modest fog]

Luuna look in dm

[bitter latch]

2048 x 480

[supple raptor]

https://tenor.com/view/suwako-post-this-frog-touhou-fumo-touhou-fumo-gif-26420843

[modest fog]

[green temple]

big 🥲

[brittle rampart]

she looks like a real Waifu

for the last one left, I’m afraid they’ll spank my ass 🤣

As for me, there’s nothing like that there... but still..

so the white neko- musicians won! 🤣

[brittle rampart]

Brand - "BigBoo" 😂

[brittle rampart]

No! still a "draw" 🤣

[green temple]

[brittle rampart]

why so - ?

[green temple]

beacuse she is big

i should try make her smoler

[brittle rampart]

comforting gift from Yue

[green temple]

👍

[near elbow]

for contrast

[brittle rampart]

wwwhat is thit?

[desert island]

https://tenor.com/view/trogdor-gif-4762062

[steady grail]

eheheh

[desert island]

🎵 burninating the countryside burninating the peasants! AND TROGDOR COMES AT NIGHT!🎵

[green temple]

👌

[desert island]

trying to get that american animated cartoon vibe with a touch more anime sprinkled in. I think it's coming along nicely 😄

[green temple]

yeah! very nice

[desert island]

I don't know what it's about but I'd watch it.

[frosty canopy]

hi

[supple raptor]

https://tenor.com/view/touhou-fumo-fumofumo-chen-cheeeeeeeen-gif-10838945564327708541

[desert island]

the longer I watched it the funnier it got.

[supple raptor]

I couldn't pass up the perfect opportunity for that gif

[bitter latch]

VELMA!

[bitter latch]

ok where does tht key thing keep coming from?

[modest fog]

[bitter latch]

goes back to watching Terminator Zero anime

[near elbow]

going tall (2048 native)

[modest fog]

@dull jackal what about this stuff ^

sounds like a definitely scam just bonk I guess, its spammed in all channels

[desert island]

https://tenor.com/view/what-confused-persian-room-cat-guardian-gif-11044457

Your name suggests that your not a cat but your bio says that you are... I'm confused...

[modest fog]

random cat invaded my house and typed that I swear

[desert island]

sus

[modest fog]

@green temple tell him in innocent

[green temple]

don`t worry we all know you are [innocent]

[modest fog]

👍

Literally me

[green temple]

is that you

[modest fog]

The prompt was "cute and innocent"

So definitely me

[steady grail]

hi not cat

[modest fog]

@green temple she demands your loot

[green temple]

no

[green temple]

i would give her my loot

[modest fog]

Big angry pirate

[green temple]

can`t see exploding anger

so she is not

[unique mortar]

Ms. bone bone

[unique mortar]

Miss bone bone wallpaper 1920x1080

other bone bone girls

[steady grail]

strangely dark and cute at the same time lol

this what is called gap moe?

XD

[unique mortar]

it comes out unexpectedly. I want some Halloween stuff cute girls with skeleton outfits but turns out some ms. bone bone

[modest fog]

Try "cute skulls outfit"

Or "bone outfit"

and nothing else about bones I guess

To be safe

I tried that

And she came out wearing nothing lmao

nai

[lyric zealot]

surely it's nai fault and not yours

[modest fog]

NAI is dumb

You need to tell every piece of clothing you want

I guess the artist doesn't help

[lyric zealot]

[unique mortar]

never mind , I combine the girl and previous Halloween environment settings and get a better result like this

one more

[bitter latch]

finishes some pizza, some 5 cheese mac and cheese with bacon and cookie brownies for desert

bounces off walls

now for some art before bed

[green temple]

smol bone

[modest fog]

runs away

[bitter latch]

slight horror?

shoots them with a shrink ray

hmmm

little voodoo priestess

hello.... !

3 Diff pictures, same prompt, added to one pic in Gimp

waits

zoom?

good night

fAINLLY

thuds

[near elbow]

morning

[near elbow]

big protector

[modest fog]

Grabby

[modest fog]

2 grabbies

[green temple]

grabby what

[modest fog]

[modest fog]

[brittle rampart]

oh yes! I'm a bloodthirsty little loli today! 🤣

@green temple but he still loves me even like this.🤣

[green temple]

🤣🤣🤣

tall nose

[brittle rampart]

at the dog's?

[green temple]

yup

[brittle rampart]

Well, it seems to me that it shouldn’t be visible from this angle.

[brittle rampart]

then just delete it in I2I

[green temple]

dunno how 😂

[brittle rampart]

something like this

[green temple]

can`t see any changes i dunno

[brittle rampart]

We're talking about a dog's tail, aren't we?

[green temple]

no

the nose

[brittle rampart]

Ouch

[green temple]

is it ok ? 😂

[brittle rampart]

and I removed the tail

but the tail was also superfluous

[green temple]

yup i noticed

[keen hollow]

Still a complete lack of Aeriths and Tifas...

disappointed

[brittle rampart]

then you need to fix it I think.

[keen hollow]

always up to me? sheesh

firing up forge/stable diffusion is such a chore now lol

I need a 5090.

[terse ore]

Anyone tried making 3d anime stuff too? Like for blender or just to rotate it around or whatever

[terse ore]

Using 3d text/image 2 3d mesh

[modest fog]

Fu

[brittle rampart]

the second for some reason reminded me of the story of “The Little Prince” by Antoine de Saint-Exupéry 🤔

[brittle rampart]

Nya

[modest fog]

[brittle rampart]

Is this a lazy cat?

[brittle rampart]

Good morning! 🤣

[zinc flax]

she sleepy

[modest fog]

who's lazy? I'm active, just looked at those 2

Grabby

[brittle rampart]

They picked her up but didn’t wake her up!

[modest fog]

[brittle rampart]

your “activity” is growing! 😂

[modest fog]

I'm innocent

[brittle rampart]

we all know that! 😉

[zinc flax]

Why'd you make an alt

[brittle rampart]

and on mine too

[desert island]

I need more VRAM... I can't run all the models I want running simultaneously. I want to be able to gen images/video, do autocaptioning, talk to my llm, generate music, and voices all at the same time! 💢

[brittle rampart]

you are not a cat! you are an octopus!

It’s impossible to do that!

[modest fog]

soon it will make coffee for you

[desert island]

That'll be nice ☕

[modest fog]

[steady grail]

cute colors

what's that?

[desert island]

there's a joke here... but I'm not going to make it... I'll let you make it for me.
"Would you like _____ with your coffee?"

[desert island]

this reminds me of something... lol 😄

[zinc flax]

brain freeze miku

[brittle rampart]

https://www.youtube.com/watch?v=FtutLA63Cp8&pp=ygUJYmFkIGFwcGxl

[brittle rampart]

I should add an evil grin here!

[wooden coral]

Flux with anime lora

[steady grail]

looks so alike cute lunar mix

[wooden coral]

It has some cool styles id say

[steady grail]

i can see it

[wooden coral]

https://civitai.com/models/832858/anime-art?modelVersionId=951509
Its this one for the curious

[desert island]

damn it 🤦‍♂️ now I have to go back to flux 😭

[wooden coral]

Been playing with it for the past few days

It has some cool stuff going for it

[desert island]

I saw I just didn't want to switch while I was working on trying get this new workflow working. The images are too good and with what I'm trying to do it makes more sense to switch back to flux now... I just didn't want to.

[wooden coral]

My workflow if anyone needs it

[wooden coral]

[vital raptor]

Oh yeah, we're cooking now..

[desert island]

just now got around to looking over your workflow. I love mistral-nemo too btw 😄

[vital raptor]

hmm, 9m 32s on cpu only, 4m using a 1050TI.. I think I need another gpu

could try forge or comfy and see if I can get better performance from the 1050TI, but it's probably not worth the effort

[gloomy scroll]

[vital raptor]

I was literally cooking, 32 cores of my server were maxed for almost 10 minutes, it got hot 😄

[wooden coral]

I didnt use mistral

[vital raptor]

I'm debating replacing my AI server (old 4 core cpu with 3080TI and 32gb ram) with my main desktop (13900K w/64gb ram and 4090) and add the 3080TI to it. Not sure how well I can get image generation, LLM, and TTS to play nicely together, but maybe with the LLM having access to both gpus i'll be able to get away with only 2 gpu instead of adding a third somewhere

Unfortunately my actual server (dual 28 core cpus w/128gb ram) doesn't have room for real gpus, currently has the 1050TI, might be able to squeeze a 4060TI in it

It's hard to tell what GPU is the best price to performance for different AI applications, I can get 24gb gpus for the server, but they're old so they'll probably be really slow anyway :/

[modest fog]

lick

[near elbow]

for LLM that might be ok-ish (eats VRAM like crazy for the better models) but for image generation that fits within 24GB nicely, a single 4090 will blow away a farm of old server GPUs with 24GB each...

[near elbow]

adding cinematic

semireal

[brittle rampart]

@near elbow Painting!? 😁 😉

[near elbow]

trying to add a bit of expressionism on top 🙂

hm, also nice

[brittle rampart]

better impressionism. hmm. yes it's better!

[near elbow]

structured oil painting

[brittle rampart]

impressionism

[near elbow]

graffiti and pop art 🙂

ok, enough from me for now, or we need a separate teddy angel channel 😉

[brittle rampart]

as the Cat says - I am innocent! 🤣

[vital raptor]

the model I'm currently using is only 13b and runs on my 3080ti just fine. For my needs it might even be excessive, but I don't want to cut it down too far. I'm thinking of trying to run that model plus sdxl image generation on the 4090 at the same time and see if it has any problems

[near elbow]

hm, i have been running a quantized 13b model LLM and SD 1.5 at the same time, but that was already cutting it close... ok, there's flux derivates that work with 6GB VRAM so it might be possible. and then also a multimodal model so you can feed it images and it can also generate its own, that would be neat

plus text to speech AI, but that would be a bit much for 24gb all in one

[vital raptor]

yeah, realistically, the tts i setup actually runs fairly well on just cpu since i have a pile of cores to throw at it, but if I can get it on one of the gpus it would be nice

maybe I just run the tts and image gen on the 4090 and leave the llm on the 3080TI

not sure yet, I have to decide if I want to turn my desktop machine into a server and figure out something lower power for my desktop, or if there's a better answer. I'm trying not to spend another $500+ on a gpu, but I don't want to have a PC that can't play games on the occasion I play them, but I also can't turn off any of the AI stuff just to play a game once I have it all setup

oh well, off to work

[near elbow]

have fun 🙂

for me the 4090 does the job, and also runs current games at max res and setting, no complainting there (since it came out in 2022)

[green temple]

is this "finger heart " ?

[brittle rampart]

Yes!

you did it! Now you can't justify yourself!

[green temple]

yay

[near elbow]

ouch

[green temple]

[brittle rampart]

I don’t know why she has face in her blood.. probably it should be like that 🤣 She’s very fighting...

[green temple]

niice 😍

[gloomy scroll]

Yue found snack

[desert island]

lol I litterally just said this yesterday 😆

[brittle rampart]

by the way. what tags are responsible for the effect of pressing on the glass?

[gloomy scroll]

xianyudian_laoban, sandcasks, khyle, best quality, absurdres, 1girl, chibi, {loli, petite}, {white hair, long straight hair, long sidelocks, blunt bangs, blue eyes}, elf ears, choker, {oversized white hoodie}, {licking screen, pov, other side of glass, hands against glass}, sticker, portrait, simple background, [blush], best quality, amazing quality, very aesthetic, absurdres

[brittle rampart]

oookaaay! thanks

[green temple]

some cat "activities"

[brittle rampart]

@green temple this is Booo...ba?!?!?

[green temple]

sorry

[brittle rampart]

Guilty! of inciting the Cat to "activity"! 🤣

[green temple]

nice 😍

[green temple]

what`s the style prompt ?

[brittle rampart]

(konya_karasue:1.376),impressionism painting illustration,(age_down, preteen:1.3),(Depth of field, Cinematic angle, Serene expression, backlighting:1.25),(dynamic angle shot, intricate dynamic pose),
BREAK
in a judge's suit,pointing,angry,v-shaped eyebrows,brave pose``` i think it's more of an artist tag

[modest fog]

im innocent

[lyric zealot]

who is this cat bot looking person

return cat bot to me

[modest fog]

I already have another alt

https://tenor.com/view/cat-running-away-escape-getaway-bye-gif-16631286

[lyric zealot]

you are not the chosen one

[lyric zealot]

this is so cute, i need metadata

[brittle rampart]

very simple

BREAK
blue shirt,blue skirt,blush,braid,closed mouth,dash b,gloves,green neckerchief,halo,shy,blush,looking away,outdoors```

[modest fog]

[lyric zealot]

https://media.discordapp.net/attachments/1291041773266796575/1291337505160953981/nuero_.gif

[keen hollow]

evenin'

[modest fog]

Ugly

[lyric zealot]

as ugly as you

[green temple]

mob da goat !

[lyric zealot]

https://tenor.com/view/ziya-gif-26244868

[vital raptor]

I still need to watch that

[zinc flax]

https://tenor.com/view/blue-archive-miyu-trash-trash-can-yuko-goto-gif-25174661

[modest fog]

why she eating a cucumber

[lyric zealot]

https://tenor.com/view/cari-gatitos-oreja-mordida-gatocari-gif-21622974

[modest fog]

me at birthday parties

[lyric zealot]

you are the one on the far right, going

[modest fog]

no

I'm the cake

eat me

[lyric zealot]

https://tenor.com/view/vocaloid-hatsune-miku-anime-animated-2d-gif-9912548670923473158

[zinc flax]

Actually rare anime posting

[lyric zealot]

I'm just listening to anime opening on youtube, and going from one recommended to another

[zinc flax]

Charlotte opening is a banger for sure

[lyric zealot]

https://tenor.com/view/anime-boy-eyes-glowing-eye-yuu-otosaka-gif-17663321

[modest fog]

Hentai songs

[modest fog]

What are those tentacles

[lyric zealot]

[modest fog]

Hmmm

lick

runs

[lyric zealot]

https://cdn.discordapp.com/attachments/1067638175478071307/1292911100060373002/harpoon_him.gif

[zinc flax]

Which illust model is this btw

[lyric zealot]

https://civitai.com/models/833294?modelVersionId=932238

there's a new version, im still on early access version

[zinc flax]

Oo it's noob

I have tried that one

Haven't tried the epsilon version thoufh

[supple raptor]

epsilon is the newer version of the early access version. v-pred is a continued training of the early access version on a much smaller dataset that converted the model from epsilon prediction (like virtually all SD1.5 and SDXL models are) to v-pred which produces a wider color range, primarily noticable in truer blacks and whites. from the little testing that i did, the vpred version suffered a lot from the small dataset and became a quite a bit worse at many concepts. i haven't tried the newer epsilon version enough to confidently say that it's better than early access, but it seems like it is from the few gens i've done.

this is a pretty good example of the kind of difference you should expect from vpred models.
left is EarlyAccessVersion, right is vpred

also, vpred is broken on Forge currently and will only preduce noise. they apparently work on ReForge and I can confirm they work on Swarm and Comfy.

[zinc flax]

I have tried vpred too

Though I'm still a bit skeptical about how the model will look cuz of the e621 dataset itself

Afraid that it'll look a bit cartoony

[supple raptor]

the early access model is basically Illustrious but a little more stable and and some styles are a bit cleaner. in my usage i haven't noticed any negative effects from the e621 dataset

i also tried a couple e621 prompts and they turned out badly. i suspect their lr was quite low for how minimally it learned e621 at 5 epochs

it definitely doesn't have that signature pony look that i suspect you're probably worried about

[zinc flax]

Basically this lol

Don't want another pony

[supple raptor]

same. pony made me basically quit AI for months until I got NAI, then after that I didn't really gen again until Illustrious and Noob came out

kxl were the only anime sdxl models that i considered somewhat usable and even those had some serious issues. animagine had good styles but unless you wanted to gen 1girl, or some blue archive character it was pretty bad and pony was..... pony.

[zinc flax]

Illustrious is already good enough fortunately

Idk if local has plateaued because of this

[supple raptor]

it had plateaued for a while, but i'm confident that Illustrious' release is our NAID1 leak equivalent for SDXL. even some low effort test runs of finetunes have shown noticeable improvements in aesthetics. now that we have a solid base anime model i'm confident that we will see some actual improvements in the coming months. it will probably be a bit slower than sd1.5 improvements were since the training requirements for sdxl are higher, but they aren't unattainable

if/when wdv team gets the compute that they've been promised that model also shows some serious promise and would be a vpred zsnr base model which would be kind of a big deal.

[zinc flax]

dang, they still haven't gotten the compute?

[supple raptor]

Advokat has already shown a proof of concept merge with illustrious and a pony based model trained on photos that showed some promise for improving backgrounds, though the pony component kind of screwed over the styles. I think once someone manages to train illustrious for photorealism so we can merge it with either base illustrious or a finetune of it we'll finally be able to compete with SD1.x models on backgrounds.

[zinc flax]

Please don't merge anything with pony lol

[supple raptor]

nope, it was promised a while ago, but hasn't happened yet. unfortunately they're reliant on donated compute because they don't want to take donations since that opens a can of worms if the model doesn't turn out good

[zinc flax]

Won't it mess with the TE or something

[zinc flax]

Cascade be like

[supple raptor]

you can merge the unet without merging TEs i dont remember what she said she did exactly though

[zinc flax]

Will wdv shoot straight for vpred?

Shouldn't they test it first or something, I also heard that they'll be using coxl or something

[supple raptor]

yep, it'll be vpred zsnr and trained on cosXL which is already both of those things.

[modest fog]

Cartoon and anime are all the same

runs

[supple raptor]

if nothing else the color range of wdv should be SoTA for SDXL models since basically nobody is doing vpred+zsnr

[zinc flax]

Hoping for the best

[supple raptor]

https://tenor.com/view/sinner-cat-kitty-cute-kitten-gif-17576170

[zinc flax]

step 1: wait for the compute ig lol

[supple raptor]

yep, at some point it should happen. it's just a matter of time at this point. knowing a couple people on the team, it would still be trained just out of principle once compute is available even if a much better base model comes out before then

whether it will be better than illustrious and it's finetunes is anyone's guess though

[zinc flax]

We'll see

[supple raptor]

pretty sure Cascade's real problem is that red team screwed it even harder than they screwed SD3

[modest fog]

All ugly

@lyric zealot you ugly too

[lyric zealot]

You the ugliest

[modest fog]

No

I can't hear you

I'm inside my pyramid

Everyone ugly but me

:<

[lyric zealot]

Hiding inside cos you ugly

[modest fog]

[modest fog]

No, she's cute...

[lyric zealot]

https://media.discordapp.net/attachments/1072697081443131476/1173827110335434792/ezgif.com-gif-maker-1.gif

[brittle rampart]

I liked the epsilon version. Although I am not a very sophisticated user, I really like the result. I really have to get used to the conditions for constructing a prompt. and especially keep an eye on the weight, but you did a great job. Thanks a lot

[bitter latch]

.

[bitter latch]

ok thats creepy

o.O

[modest fog]

Chess

[brittle rampart]

no game - no life

[bitter latch]

hmm I think she's on wrong side of glass

[modest fog]

Anime

[bitter latch]

.

hmm "Otute= braids"?

[modest fog]

booba

[near elbow]

it's only magic if it has runes and glows

[modest fog]

grabby

[modest fog]

Would eat

Tomatoranges

[green temple]

[green temple]

are you in my head 😂 i tried to do something similar

[lyric zealot]

I'm in your wall watching

[green temple]

what`s the prompt ?

[mental marten]

#🍥｜anime message```yaml
(ningenmame:0.95), (ciloranko:0.95), (shosho_lwlw:0.95), (tianliang_duohe_fangdongye, as109:0.9), (maccha (mochancc):0.4), (konya karasue:0.9),

best quality, amazing quality, very aesthetic, masterpiece, absurdres, backlighting, depth of field, impressionism, painting (medium), illustration,

1girl, fubuki (one-punch man), glowing hand, aura, glowing eyes, telekinesis, outstretched hand, force, forcefield, motion blur,
Negative prompt: bad anatomy, bad hands, error, missing fingers, extra digit, fewer digits, (worst quality, low quality:1.1), jpeg artifacts, signature, watermark, username, sketch
Steps: 24, Sampler: Euler a, CFG scale: 7, Seed: 4234947764, Size: 768x1024, Model hash: baeaeff491, Model: noobaiXLNAIXL_earlyAccessVersion, VAE hash: 66f78ad136, VAE: sdxl_vae-fp16fix-blessed.safetensors, Denoising strength: 0.4, Clip skip: 2, Hires upscale: 1.5, Hires steps: 10, Hires upscaler: 4x_fatal_Anime_500000_G, Version: f0.0.17v1.8.0rc-latest-269-gef35383b```

[lyric zealot]

here you go ^

[green temple]

thx

[lyric zealot]

btw, if you want, you can add this bot to your discord user, and you can just right click any image on any server and get their prompt (99% of what im posting should have metadata this way)

[modest fog]

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

Anya worst character ever

[green temple]

indeed , miri is better

[modest fog]

Anything is better

[green temple]

noo

[modest fog]

I kept miss clicking her lora instead

Cause they have the same name

Stupid name

[green temple]

how ?

[modest fog]

Another anya lora

:<

[green temple]

🤣

[modest fog]

@green temple big ear smol eep

big big

[modest fog]

No grabby

[near elbow]

splash

[modest fog]

@green temple you didn't pat her right, she's coming back

[supple raptor]

marisa discovered that people just leave boxes filled with all sorts of interesting things sitting on their porch

[vital raptor]

wait a few days and they'll be full of candy 😄

[unique mortar]

some pics using controlnet and real human for action model. I want some motion like pic for swinging under the tree, but all looks sit still.......

[terse ore]

Anyone try making 3d anime mesh characters?

[modest fog]

What model is that

I saw a lora like that but for flux lmao

It looks stooooopid ( like @lyric zealot )

You do leave png info
Looked at it
How did you get the camera lens thing effect

[lyric zealot]

[brittle rampart]

Run Cat

[modest fog]

@green temple

[modest fog]

lick

@green temple

in innocent

[supple raptor]

it was just inpainted into a real picture from a ring camera

[modest fog]

cheating

banned

[steady grail]

he is gaslighting us, Risa is real and that's the proof

[supple raptor]

i've been caught. this is a real photo from my front door.

[steady grail]

shame on you, you will not take the girls all for yourself lol

thinking that most of them have very sad stories <.<

[green temple]

supposed to be sweets

[modest fog]

Looks like those balls that grow in water

You know?

[green temple]

ohh yeah 🤣

[modest fog]

Where smol

no more smols for minutes

[green temple]

in her way

[modest fog]

She's the girl that says no to me

[brittle rampart]

https://tenor.com/view/nyanpasu-gif-21749467

[keen hollow]

I need a new model or something to mess around with.

Whatcha all usin' right now?

FluxDev + Lora?

[modest fog]

@green temple evil smol

[modest fog]

Pony or nai

[keen hollow]

There's a new nai?

[modest fog]

Nai on phone

No

[steady grail]

playing with comfy as usual

[green temple]

illustrious xl

[steady grail]

just made a new workflow to change style of an image lol

[keen hollow]

what's that one?

[steady grail]

with this krita upgrade i'm totally going nuts

[keen hollow]

what do they have in krita? isn't that the web-based paint program?

[steady grail]

krita diffusion, a plug in

[keen hollow]

I see

[steady grail]

but now i can interface my own comfy workflows with it

[modest fog]

What are you using krita for

[green temple]

https://civitai.com/models/795765/illustrious-xl

[modest fog]

What's the difference

[keen hollow]

whatcha like about thi sone?

[steady grail]

other than genning directly high res without mutations, and inpainting or region genning without a problem

the usual you do with comfy, experimantal workflows and stuff you can't do with a normal gui

[green temple]

cuteness

[brittle rampart]

I'm afraid to imagine the number of nodes and out links!

[steady grail]

eye that don't see, heart that doesn't feel

[brittle rampart]

Well, this is somewhat offensive. because I tried to work in ComfyUI

[steady grail]

i'll leave you to your immagination 🙂

my last workflow only have 52 nodes

not very much

[brittle rampart]

only 52... not much at all.

[modest fog]

@green temple she's an angel

[steady grail]

this workflow for now is very much straight forward

load an image, tag it, concat with the promp, add the image in the noise and gen

[brittle rampart]

and why the incoming image?

[steady grail]

it's just a restyler, for fun

instead of genning from 0, i'm using another image as a base model

[brittle rampart]

trying to do VT?

[steady grail]

i can change it, inpaint on it ecc.

yup

[modest fog]

Take one of my Korean girl images

And do that

[steady grail]

regenning the image of sooma

gimme that chaos, i'll try next

[keen hollow]

so this Illustrious one uses tags, yes?

I'll give it a whirl

[green temple]

yes

[keen hollow]

ok, trying it now

first attempt was really bad... adjusting prompt 😛

hmm

might need to restart the generator... being weird

[modest fog]

Show it

[keen hollow]

too late, already restarting forge ;P

[modest fog]

[keen hollow]

i know...

[modest fog]

I want bad gens

[keen hollow]

maybe restarting will be no help and it will still be trash 😄

generating....

[modest fog]

Did you look how to prompt for it quality wise

[steady grail]

punching the pc may get things better, or worse

[keen hollow]

sigh...

[modest fog]

Get a prompt from the prompt master @green temple

[keen hollow]

man this is so fast

after using flux dev, this is refreshing lol

[green temple]

funnya is

[keen hollow]

wow this model is ... clearly I'm not doing something right lol

[steady grail]

gimme the prompt, let me see

[keen hollow]

LOL

[modest fog]

10/10

[keen hollow]

well honestly,, I revered to very basic prompts with just a couple of keywords

[modest fog]

She has the lara croft Booba

[steady grail]

this is literally 1girl lol

[steady grail]

when at first i tested the model XD

[keen hollow]

good grief

[modest fog]

steals her heart

[keen hollow]

they're getting worse LOL

"1girl, solo, long hair, black hair, red eyes, pretty, attractive, pleated skirt, hands on hips"

[steady grail]

what model is that?

[keen hollow]

that's some quality output

[steady grail]

it's weird

what vae are you using?

[modest fog]

Try adding high quality masterpiece

[vital raptor]

illustrious needs a bunch of prompting and negatives, reminds me of 1.5

[steady grail]

even without that stuff, models should be biased toward good images

[modest fog]

Don't think it is

[green temple]

yup

[steady grail]

if it need a bunch of negatives, it must have very terrible mutations

[modest fog]

It's a base model

Kinda

[steady grail]

a proof of concept?

[keen hollow]

either no vae, or the sdxl one

[vital raptor]

I think it's because of the massive amounts of wide range training so it can 'do more' but I'm not really sure. I don't like needing 300 tokens to try and get something good, so I stopped using it

[modest fog]

It's just a base model

If you want quality then you merge with the styles and stuff you want

[steady grail]

after 75 tokens, things get resetted

[keen hollow]

that seems more in line with the standard output of this model

but I do not understand why people prefer this to blue pencil/animagine/anima-pencil/pony/several others

[modest fog]

You just listed a bunch of models

[keen hollow]

I did indeed, what do you mean?

[modest fog]

What's wrong with them

Pony is fun

[keen hollow]

nothing, they're great

I am saying I don't understand the Illustrious model over those.

[modest fog]

Oohh

[keen hollow]

outputs seem derpy

[modest fog]

It's more like novel ai

Very anime and artist focused

I believe

[modest fog]

I believe you really want to type some artists in your prompt for it to be decent

[vital raptor]

I still use pony autismmix because i found styles I like and it does everything I want, but I'm not trying to make actual anime characters

[green temple]

cause of cuteness < ok i`m kidding

[keen hollow]

Getting more stylized and interesting....

[modest fog]

Big @green temple

[vital raptor]

then again, if i wanted a specific character from an anime I could just train a lora on pony, i probably still wouldn't change models

[brittle rampart]

@green temple saving

[green temple]

[keen hollow]

sketchy/splotchy - by design?

[brittle rampart]

Yes. I would argue about Sketch.

[keen hollow]

i'm dropping quality again...

tough model to use

[vital raptor]

smol is lost

[steady grail]

this is tifa catart!

[modest fog]

The legs are like SpongeBob's

Cx

[steady grail]

if spongibob was an anime character

[green temple]

🥲

[keen hollow]

wow there are a lot of illustrious mix models on civit right now

ok... let's do a comparison test.

[steady grail]

another experiment, just for fun

[keen hollow]

had one output that i had to censor a little bit there, but what do we think? which one wins?

a couple "non-illustrious" ones snuck in there...

[green temple]

illustrious pencil xl ( the style not the image itself ) 🌚

[keen hollow]

running another set

Goes quick except switching models takes forever 😐

[modest fog]

only cause she big

[green temple]

yes

[keen hollow]

trying to do some widescreens with this

not working so far lol

[supple raptor]

base models should not be. that is what aesthetic finetunes are for, which makes the model worse for further training

[unique mortar]

How to make proper/desired words appear on the pic

[brittle rampart]

hope for luck when using the text, label, sign, print tokens. or use Flux. or just like in the good old days, ask for help in post-processing.

[zinc flax]

bad nana

[supple raptor]

aside from a few extremely common words like "happy halloween", "happy birthday", or "merry christmas" it simply isn't going to work even remotely reliably since you're using an SD1.5 based model. SDXL models are a little better at text, but even then it's very unreliable. like funya said, Flux is your best option for text. if using a larger model isn't an option due to hardware, sometimes using controlnet on an image that already has text will also work or you can always just add text in post.

[supple raptor]

1girlyesyes

[bitter latch]

can't sleep brain on random x1200!

[bitter latch]

hmm what style is that?

and yes i'm in ooh shiney distracted mode 😄

finally getting tired tho

[wooden coral]

Flux and 2 Loras

[bitter latch]

O.o what?! da funk?

[wooden coral]

Anime and Crayon

[bitter latch]

ahh

[bitter latch]

goes to bed thud

[wooden coral]

After months of testing I finalised my Flux workflow

[wooden sail]

And this one is default img2img

[near elbow]

using the right model helps a lot with text 🙂

[gloomy scroll]

Yue is ready for halloween party

[supple raptor]

I know there's some Forge users here. In case anyone needs it, I uploaded a version of the noobai vpred test model to HF with the keys to load as vpred and ztsnr added, this will allow the model to be run in the latest version of forge as well as A1111 dev branch. Figured I'd throw it up on HF since catboxanon only uploaded to mega. https://huggingface.co/nubby/noobaiXLNAIXL_vPredTestVersion_with_keys/tree/main

[modest fog]

Forge > others

[keen hollow]

I use Forge. But not sure what this is. 😄

[modest fog]

its a model like illustrious

[keen hollow]

Has anyone run a comparison grid?

[modest fog]

I mean whats a comparison grid gonna do anyway

You just change the looks of the model by prompting

its a good model, it looks cool

[supple raptor]

v-pred and zsnr are training techniques that improve the color range of the model, however when you run the model, you have to tell the software that those features were used otherwise you will just get noise. i just added the keys to the model to allow them to be loaded properly in forge and a1111 dev branch since the original upload of the model didn't have them which meant it could only be used in software that allows the user to specify such as Comfy, Swarm, ReForge, and Diffusers

here's the original upload of the model that is missing those keys https://civitai.com/models/833294?modelVersionId=962003

[keen hollow]

ah cool.

nice work

maybe I should look into reforge...

[supple raptor]

i haven't really looked much at ReForge, i know it had a ton of custom samplers added that aren't in Forge, but in other regards is pretty far being the latest version of Forge. it was mainly made when Illysaviel had said that he wasn't really going to work on Forge anymore but now he's working on it again.