🍥｜anime

[magic wave]

Yes

XD

I haven't slept

I got allergy

So I can't sleep XD

[wispy canopy]

pollen allergy?

[magic wave]

No clue, cause it happens randomly :o

[wispy canopy]

even worse

[magic wave]

Yeah, I'm pretty sure its not cause of season changes, not cause of pollen either, or dust

[wispy canopy]

@magic wave be blessed by the goddess of thiccness (pretty sure clyde will block me)

wait... clyde... WHAT?!

[magic wave]

Smash :O

Clyde is sleeping I guess

[subtle crown]

clyde has a thing for thic 😛

[wispy canopy]

i really thought it'd get triggerered on #2 though

#4 was the most fun in terms of pose... alas, camels were detected

[magic wave]

Clyde worrying about camels at this point?

[wispy canopy]

nah, that'd be me

[magic wave]

Oh, fair then

[wispy canopy]

these better? 😄 pocket size thiccness goddess!

[subtle crown]

cute

[wispy canopy]

that's a huge apple

[subtle crown]

huge

[wispy canopy]

MelonMancer is back!

almost got it right on #3 😄

[nova remnant]

does increasing the training batch size of LORAs significantly decrease the final quality?

[tawny helm]

Man, I'm having way too much fun with that new ControlNet feature....

[nova remnant]

the reference model?

[tawny helm]

Mhm.

[nova remnant]

I haven't fired that up yet but it looks insanely accurate

[tawny helm]

Yeah, for sure. So far I've been throwing stuff at it and it replicates characters, places, etc pretty damn good.

Only thing is I haven't been able to make it work on any model that has NoiseOffset inside of it.

[wispy canopy]

lol radishmancer

[tawny helm]

Oh no....Eface has discovered cuteness.

Turn away now before it's too late!

[wispy canopy]

here, have a picklemancer

[tawny helm]

Welp, I guess she doesn't need oxygen when she has pickles...

[wispy canopy]

catmancer!

[wispy canopy]

Foxmancer!

[wispy canopy]

thx @runic vault, you inpsired clownmancers!

[wispy canopy]

going for chibi (random noun)mancers now 😄

graphmancer and grapefruitmancer? 😮

[nova remnant]

aww lol the one where shes standing on the little podium is cute

[wispy canopy]

the chibis need more chibis

[wispy canopy]

Absolute unit

[cyan burrow]

I think it's great that soldier to the right brought his legless friend with him into battle, good character there in him. Well done chap

[tepid geyser]

nah he's just riding side saddle!

[cyan burrow]

that seems very unintuitive

[tepid geyser]

this is how i see it: theyre doubling up on horses cause there aren't enough for everyone to get one right? but that's not ideal in combat, so the back one is sitting side saddle so that he can easily slip off once they get into melee combat and the one sitting normally will keep fighting horseback

wow the AI was so smart to come up with that kind of realism!!!

[cyan burrow]

yes Im stunned

[tepid geyser]

wow three cheers for AI

[cyan burrow]

Started working on DG4

my goal is to up quality and make it easier to get pseudorealistic images and to maintain and improve natural language prompts but keep booru

[wispy canopy]

hi boto ^^ 👋

[wooden coral]

[wispy canopy]

i'm still sooooo fuckn satisfied my model does such nice transparant fabric ❤️

[wispy canopy]

my model forgot what astronaut helmets look like, it seems

[wispy canopy]

well... i ddi prompt for the goddess of partying hard!

[wooden coral]

Yeah I had that issue too lol, or they lack I visor at least

[tepid geyser]

same here, my astronauts also lost their visors without some extra prompting targetting that specific problem

like here

[wispy canopy]

the one real pro about working from home, and being the only one in your household? -> the music is on full blast the whole day ❤️

Livening up the party!

[wispy canopy]

i feel the series did her injustice

went from badass waifu to permanent damsel in distress

[tepid geyser]

yeah a little bit

but then kirito also became in distress later i guess

or like

a vegetable

[wispy canopy]

vegetable kirito was series at it's best...

up until he woke up, and then it was... yeah, shit tier again

animation is really good

[tepid geyser]

the model gave her elf ears

maybe alfheim online influence

[wispy canopy]

probably

[tepid geyser]

i do agree with that

alicization was very pretty to look at

[wispy canopy]

yea, but that's the thing... it only was pretty

i mean, any actual depth in the story was just absent

i liked the twist of the immortal king copy though

[tepid geyser]

beautiful

[wispy canopy]

one of my favorite prompts

Grace the morning sky
Comfort Luna as she wanes from Terra's might
Veiled in amber haze
Fueled by violent rage
Dancing, discordant, spinning in retrograde

Born from the sea and the waves
From a cosmic array
Of the colors in space
Nothingness erased
Come dance with her in the stars
Watch her transit afar
Crossing Sol's burning gaze
Venus sets ablaze

Welcome us with your light
Born from the sea and the waves
From a cosmic array
Of the colors in space
Nothingness erased
Come dance with her in the stars
Watch her transit afar
Crossing Sol's burning gaze
Venus sets a-

Born from the sea and the waves
From a cosmic array
Of the colors in space
Nothingness erased
Come dance with her in the stars
Watch her transit afar
Crossing Sol's burning gaze
Venus sets ablaze
BREAK character concept, promotional art```

[tepid geyser]

your prompts are poems?

is it using your custom model?

[wispy canopy]

it's a song, actually 😄

and yeah, on my own model

it's Cepheid - Goddess

[tepid geyser]

what's your model?

[wispy canopy]

https://civitai.com/models/64843/dark-alfa

be sure to use my negative prompt, that makes the images pop off 🙂

[tawny helm]

We got a ControlNet update guys. 2 new models.

[wispy canopy]

again a new update?1

[tawny helm]

Yeah, 2 new models for the reference feature.

The reference_adain+attn model can get even more accurate than the one we've already got now.

[wispy canopy]

where are the models then?

can't seem to find them on hf

[tawny helm]

You're not using Auto's repo right now are you? Otherwise I'd say go for the extensions tab.

I'll see if I can get you a link, if there is one....

For the models, of course.

Damn....

Okay...there is no link YET.

Wait a sec, hold up. I updated myself on the extensions tab, I should be able to just throw you the Github link.

https://github.com/Mikubill/sd-webui-controlnet

Hopefully this fixes the situation, I updated and I have the models already.

[wispy canopy]

i can't see they use models for those?

[cyan burrow]

mmm Zodiac of all Trades alpha 0.2 achieved

[tepid geyser]

what's controlnet? is it a well known model?

[cyan burrow]

adorable kitty wants to play

[tepid geyser]

what resolution do you generate your images at?

[wispy canopy]

around 640x640 upscale 2x

or 512x768 in either direction

[tepid geyser]

alrighty thanks!

[wispy canopy]

upscale with 4x ultrasharp

denoise 0.5, unipc

[tepid geyser]

my gpu is a bit underpowered to do all that upscaling, but i think it turned out pretty neat!

cool model!

[wispy canopy]

when your reality shatters around you

[tepid geyser]

do you always use long-form song lyric like prompts?

[wispy canopy]

i use whatever comes to mind :), the only thing i'm not using is quality prompts

masterpiece, high res, that kinda shit

so i can prompt lyrics, LLM bot answers, just some random words

Consistency is not my strong suit

I just love experimenting with the prompts, and hope something interesting happens, the prompt 'glitching' (simply not understanding) is the most fun imho

[tepid geyser]

well they definitely seem to work!

the original output kinda messed up the socks merging them with the legs but running the output through img 2 img and upscaling it a bit fixed it

neat

[wispy canopy]

illuminati texts on waifus

[tepid geyser]

I tried using song lyrics too

beautiful

[wispy canopy]

i kinda tuned the model to be VERY good at that 🙂

[wispy canopy]

SUNAFORMARS

just like the manga terraformars... damn, what a ride that was

wait, it's still going!?

oh... still hiatus

[wooden coral]

Not sure what you're talking about

[wispy canopy]

a... strange manga called terraformars

your sunaformars tabled reminded me of that

sunafornnars?

[wooden coral]

Now i see it lol

[wispy canopy]

dang, these are good

[wooden coral]

The details this model produces... dont think ive seen that before

[wispy canopy]

which model is that?

[wooden coral]

https://civitai.com/models/54917/breakdomainrealistic

[wispy canopy]

lol, i have that one, couldn't get it to work for me

ooff.. blursed kitty

[wooden coral]

I found a prompt to set the style and now im palying with wildcards

[wispy canopy]

a psychiatrist with superpowers?! yes please, make this an anime!

[wooden coral]

ill be using this for this weeks POW for sure

[wispy canopy]

you already know the pow? i call admin abuse!

[wooden coral]

I dont lol

I wish

[wispy canopy]

ah, you're just gonna submit that

pretty damn fine picture though

[wooden coral]

nah i meant the model/style

[wispy canopy]

you got png info embedded?

[wooden coral]

should have it yes

[wispy canopy]

imma try that prompt too ^^

[wooden coral]

Sure has a lot of potential i think

[wispy canopy]

, A stunning binary star system, with two stars orbiting each other, A planet with a massive, glowing crystal at its core mission success? 😄

[wooden coral]

[wispy canopy]

well, mine is done on an anime model...

[wooden coral]

Mine too

[wispy canopy]

hmm... technically, we shoudl be allowed to post this in #🍥｜anime then, right? XD

[wooden coral]

I dont see why not

[wispy canopy]

imma go back to character posting tho 😮

[stone onyx]

yes pls

where are the waifus

I'm not seeing ANY

[wispy canopy]

got husbando

[wooden coral]

This has a character

[stone onyx]

[wispy canopy]

i'm just trying old prompts on my new model

there's... over 50000 images in that directory...

[wispy canopy]

pretty clothes! used UniPC sampler?

[tepid geyser]

eface i really like playing around with your model putting phrases, concepts, and poetic lines

[wispy canopy]

^^thx, that makes me very happy to hear 😄

[stone onyx]

fineeee

I'll post the waifus

[wispy canopy]

post 'absolute thiccness'

[stone onyx]

she has THICC hair I guess

and a THICC outline as well

forgor to change the upscaler

[stuck aurora]

I used this #🍥｜anime message

[stuck aurora]

What model is great for chibis? Can anyone please recommend?

[wispy canopy]

just add chibi to your prompt 🙂

[stone onyx]

^

[wispy canopy]

i'm not sure though, if there's models that are specifically trained for chibi, i suppose you'd best use a lora in that case

[stone onyx]

I only ever used the chibi prompt on yesmix, seems to work fine

[wispy canopy]

my model, only added chibi to prompt

[stuck aurora]

How come I've never thought of doing that? Thanks! lmao my 2MB brain cries

[wispy canopy]

my 1mb brain is default glitching -> i just add random stuff to my prompt!

[wispy canopy]

WHAT?! clyde blocked eyes?!

[stone onyx]

eye contact

haram

[wispy canopy]

LOL there's a song playing which has a natural resonancy frequency of my room 😮

this bass is intense all of a sudden

[wispy canopy]

VERY haram

[stone onyx]

[compact warren]

How did u manage so many coherent people

[wispy canopy]

tile resample perhaps

[storm ivy]

that's just luck

[compact warren]

[stone onyx]

sometimes I get crazy images without using latent couple

[slow needle]

'

[wooden coral]

50k? I thought im hoarding a lot

[slow needle]

same here

[stone onyx]

do u guys delete ur images after a while

I'm too lazy to delete anything

[wispy canopy]

https://tenor.com/view/lied-liar-i-lied-not-true-fiction-gif-5544606

it's 120k files lol

[stone onyx]

[wooden coral]

I delete all the crap, im only keeping decent stuff

Im only at 3.5k

[tepid geyser]

this is addicting

[wispy canopy]

these prompts were made by using random nouns... but wtf is up with this combination LOL

[wispy canopy]

welcome to purchasing more expensive pc parts hell

i didn't upgrade my pc for games, but i've invested more money ever since i started SD lol

[wispy canopy]

he/she/it has seen ... stuff

used the eye cleanser a bit too intensively

[gritty patrol]

did you just avoid misgendering a cosmic horror

this is getting too far

[wispy canopy]

you never know man, they might be sensitiver about it...

[gritty patrol]

also, iirc empty eyes is the danbooru tag for dead eyes

14C my fingers are freezing i can't type

cold.

[wispy canopy]

damn, this one is cute ❤️

she be casting a spell of some kind

a water blade?

[wooden coral]

The bleacher

[wooden coral]

Did you have any success with the prompt?

[wispy canopy]

the canyon one?

it just gave pretty pictures but nothing really special, imho

[wooden coral]

I meant like the base prompt

[wispy canopy]

ah, the part before the canyon?

hmm, didn't try that one

can always give it a try

[wooden coral]

Since I use dynamic prompt, I have like a base that I keep and the rest changes dynamically

[wispy canopy]

yeah, dynamic prompts are the best ^^

[wooden coral]

💯

[tepid geyser]

I'm trying to make it write words on the blackboard but that's really difficult

[wispy canopy]

dont'worry, that's just about impossible

with 900m parameters, SD simply doesn't know what words are or how to write them

it might be able to output characters, but it will never be able make something coherent of them

@wooden coral your base prompt made my desktop waifu prompt even more real man 💯

[wooden coral]

Ignore the

[wispy canopy]

sick! damn nice ❤️

and it's SD...we all know to ignore

[wispy canopy]

although, it has conditioned me in this 8 months to look at hands in EVERY picture i see

[wooden coral]

[wispy canopy]

what is up this afternoon 😮 such great art being posted all over the place ❤️

[wispy canopy]

honestly, i like stuff like thsi better than default waifus

default waifus are by no means unpleasant to look at, but stuff that gets the imagination going, now that's the true definition of art!

[wispy canopy]

damn @nova remnant u on a roll ^^

saved up some good gens?

[nova remnant]

have not slept yet so i guess im on a roll LOL

[trim current]

[wooden coral]

Yeah I enjoy the more unique stuff more as well

[trim current]

I love how youre going for like a info sheet

[stone onyx]

This time I wasn't prompting for it

but it's slowly starting to do it by itself

[trim current]

Still very neat tho

[wispy canopy]

all of a sudden, vehicle promotional art 😮

[ocean reef]

looks cool tho!

[somber oasis]

Your art is amazing keep it up

[wispy canopy]

this prompt hella cool ^^

[stone onyx]

TY

[wooden coral]

[nova remnant]

wow that anime girl is next level

[teal pivot]

Heyy

[foggy minnow]

Anyone have suggestions to make hands look better? I get them ALMOST perfect, and then SD starts making haunted looking ones again

[trim current]

textual inversions, the one locon

[rapid shuttle]

I usually do a neg prompt that says "more than 5 fingers, less than 5 fingers"

[trim current]

badhandv4 is nice

deformityv6 aswell

i use those allot

[foggy minnow]

Ive tried a few but they just dont help. They make it worse even

[foggy minnow]

I will check this one out!

I might just not be using bafhandv4 well

[foggy minnow]

How would I go about?

[trim current]

i inovoke them from the side

[foggy minnow]

Just "more than 5 fingers, less than 5 fingers"?

[foggy minnow]

Maybe I am using too many negatives?

[nova remnant]

may i ask what model boto?

[foggy minnow]

I have like 378/450 on the negative prompts lol

[nova remnant]

im defo a fan of your work now lol very inspiring and makes me want to get better

[trim current]

christ

notice how i only have like 2 ng inversions?

[foggy minnow]

Yes

[trim current]

thats all you really need, or like 1 inversion and a few words

thats too much imo lol

[foggy minnow]

Thats probably breaking it

[trim current]

https://tenor.com/view/spongebob-panicking-panic-run-stressed-gif-5303378

ai tryna do all the ngs

[foggy minnow]

How many would you say I should max out at?

I love your pfp btw

[wooden coral]

holy

[trim current]

yeeeis ty i love it

[trim current]

those are nice, i like the color

[foggy minnow]

Omggggg

All of these are soooo cute

[foggy minnow]

I'll try to limit the neg prompts to like 5-10 for now and I will let you know how it goes!

[rapid shuttle]

You guys just keep +1uping each other

I hope I can get my art looking this controlled

[cyan burrow]

I'm just gonna drop this for a game over

[safe bane]

https://giphy.com/gifs/ShalitaGrant-girl-hair-nug4Bc3oAghGg

[foggy minnow]

Im already seeing a mass improvement with WAY less neg prompts....

I thought I HAD to go overboard lol

[wispy canopy]

yeah, i use somethign similar. AND the embedding

are we doing horror inspired now?!

[cyan burrow]

they must've spilled some tomato soup on the floor

here's some nightmare generations that didn't go as expected

[terse ore]

Has anyone tried the new reference ControlNet? It can use img2img on a nonanime image with an anime reference and then change the style to anime like so:

[cyan burrow]

cool

[rapid shuttle]

i tried but it crashed my python. Had to reinstall SD. Gonna try again later this week.

[terse ore]

It doesn't even need a model anymore apparently

[cyan burrow]

might not need one, but it's probably good to have

[rapid shuttle]

When I did it for the first time it was giving me completely different people

was not working as expected

[wispy canopy]

all 3 reference preprocessors don't need a model

they just get 'basic' information from the feeding image

[terse ore]

I was thinking, if you can get a basic 3d scene in blender, this method could be used as a style change to make it anime quickly and easily for webcomic

As a lot of artists use 3d to save time and do paint overs. This could save huge time

[sage kiln]

Anyone here got some experience with Tiled Diffusion and Regional Prompts?

[cyan burrow]

oh shit right, in an hour Swe vs Finland in Ice Hockey World Championship

muhahahahaha

[magic wave]

Gmornin lads

[cyan burrow]

I want to train a LLM based on woke twitter and another based on maga twitter. Then make them have a conversation and see how long it takes before they start arguing and before someone says "BLOCKED"

[wispy canopy]

LOL

[supple raptor]

There's a 134m parameter llama-lite model on HF. You could probably just use that. Probably around the same intelligence and coherency as the average twitter user. Might still be a little too good so you might have to prompt it to act dumb.

[trim current]

Thats not very promising if thats thst level of intelligent

Twitter is a cesspool

[supple raptor]

😅

[supple raptor]

@wispy canopy did you see that https://huggingface.co/TheBloke/Wizard-Vicuna-13B-Uncensored-GPTQ/tree/main is out?

[wispy canopy]

yeh xD

still doesn't allow me to kick babies

sadge

possessed babies are ok now, so that's an improvement

[supple raptor]

There's also this one that came out 2 hours ago. I'm not familiar with the mega model though so I have no idea if this is good or not. https://huggingface.co/TheBloke/wizard-mega-13B-GPTQ

[wispy canopy]

yeah, i saw it, when i check, it only had the model created, but files not uploaded yet lol

only one thing left to do

[supple raptor]

Let me know what you think. I'll probably grab it later unless it's bad.

[wispy canopy]

sure, i'll give it some tests

i'm primarily testing on reasoning and common sense though

[supple raptor]

Fair enough

[wispy canopy]

with enough primer material the 30b models are extremely good, but they need a bit of priming before they actually become 'context aware'

as in, throw them into tavernai for a roleplay, and they nail it. best roleplay model atm (openassistant 30b sft-7)

[supple raptor]

Nice. For now, 30b models are out of reach for me

[wispy canopy]

but yeah... 30b model eh.

[magic wave]

O, bois, either of ya tried WD 1.5 beta 3?

[pseudo condor]

[supple raptor]

Yeah, I've tried it out all 4 of the aesthetic models at least. It's a lot better than beta 2 was for sure. I still want to test it some more, but I think it still needs more training. It finally proves that the 2.x text encoder architecture is actually better than the 1.x one at least.

@magic wave

[pseudo condor]

this one look good

[supple raptor]

hopefully the next epoch of WD won't take 2 months like this one did. I think they had to sort out a problem with training the text encoder and so I assume the next one should be a bit faster.

[magic wave]

Oh damn, that is pretty sweet then

Damn that is really good yeah :o

Though I think I like the aesthetic of the radiance and illusion models the most so far :o

[wispy canopy]

@supple raptor seems imageassistant can make coherent descriptions at least 🙂 so it's fine for prompt generation 🙂 (mega 13b)

[wispy canopy]

next up: a goddess of thiccness, invading a drug den . at least it didn't refuse (censored models either refuse drug den or thiccness, because of drugs or sexualisation)

[supple raptor]

When mofu gets it right it can look nice. I think ink is my favorite. It's the most coherent from my testing so far and I think just the fact that it's a significantly different style from most of our current models appeals to me.

[magic wave]

Fair fair, I like illusion cause its simple, and radiance cause of like the funky glitch/chromatic aberration effect it has

[wispy canopy]

and it delivered!

for me wd1.5b3 doesn't want to load

dunno why

[supple raptor]

If you're still on an old commit of WebUI, you might need to add the SD2.x .yaml and rename that to match the model filename.

[wispy canopy]

she be slaying!

[rich tartan]

Eface you got dual boot?

[wispy canopy]

both latest git pulled vladmandic and auto

[wispy canopy]

wsl2

[rich tartan]

tf is that

[wispy canopy]

linux + windows at the same time, at kernel level

wsl2 enables windows to load the linux kernel as native

[rich tartan]

hm, interesting

[supple raptor]

That's how I run Oobabooga.

[rich tartan]

i want to get linux bc amd gpu, but i don't want to format a drive for that

[wispy canopy]

@supple raptor same -> GPTQ repository, right?

[rich tartan]

and i don't know if it's either possible

[wispy canopy]

it is. you just need hypervisor capabilities from your cpu

if your CPU supports wsl2, you can run linux as a native linux machine in your windows environment

[rich tartan]

and also use rocm?

[wispy canopy]

can native linux can use rocm? ^^

i'm using repositories that are only made for linux in that WSL2 machine, there are none for windows

(there are fringe cases, but i believe you can run rocm on wsl2)

[wispy canopy]

it does come with a memory footprint though. i had to upgrade my memory to fit both kernels + all AI software in memory and vram at the same time

[supple raptor]

Yeah, though I don't remember if I used the "Old CUDA branch" or the "Up-to-date CUDA branch" off the top of my head. I think I used the old one. https://github.com/oobabooga/text-generation-webui/blob/main/docs/GPTQ-models-(4-bit-mode).md

[rich tartan]

interesting new thing, i am gonna do some research on that, i thank you very much for that information

[wispy canopy]

old cuda works on windows, up-to-date doesn't. up2date works with the act-order safetensor models, which as such, will not work on windows

[rich tartan]

24GB enough?

[wispy canopy]

you'll be rebooting a couple of times 😛

[rich tartan]

fk

[wispy canopy]

fitting a 30b model + SD just barely 😄

13b model + sd? ezpz

[rich tartan]

Dam, didn't know that i need to deal now with ram, while i try to deal with vram

: (

[wispy canopy]

i REALLY hope stability AI releases a wizard uncensored vicuna 1.1 model of their model though. 😦 their vicuna model is the best when it came out, and it's already superceded 😦

[wispy canopy]

i upgraded my ram from 32 to 64gb ... else simply impossible or your system is just unresponsive all the time (not to mention swapping)

[supple raptor]

At a glance at filenames it looks like I used the old one. Since I'm missing a few files that the new one has. Maybe I should switch over to the new one.

[rich tartan]

dam, thanks, at least ram does not cost as much as a gpu

[wispy canopy]

i think the up2date branch uses some triton libraries, which do not exist for windows, in any case, oobabooga git pull just loads all models just fine on linux

on windows you get gibberish on the act-order models

[supple raptor]

I run it all through wsl2 anyways.

[wispy canopy]

ooof... first hard fail on complex task on mega-13

[supple raptor]

The one click installer didn't work for me when I tried to use it, so when I had to manually install I figured I might as well just go with wsl

[wispy canopy]

i had it running in windows, but couldn't get the act-order to work, so i switch to wsl, and now it's just... permanently active lol

2nd attempt, got it right 🙂

[brittle jay]

man I should fr try training a vits model already

[wispy canopy]

no lora's... just pure prompt and it works out of the gate? what black magic was this?!

[brittle jay]

sd changed order of channels and the logo of the server

[compact warren]

Again?

[rich tartan]

is that a custom model?

[magic wave]

Can I do that with a i5 13600kf?

[wispy canopy]

it should. i know 13 series dhould be new enough, but i don't know if they support hypervisor (i suppose so?) , if your motherboard and cpu support it, you can use wsl2

[pseudo condor]

no!

[rich tartan]

literally yes, as long you can enable the feature Hyper-V in the windows features

[pseudo condor]

but i thinking about made a custom

[rich tartan]

oh cool, i wish you luck

[magic wave]

Seems like I can

@wispy canopy You'll have to teach me

[brittle jay]

there was such a thing as wsl2?

bruh

https://cdn.discordapp.com/attachments/807809192537882647/1091905253739331584/attachment.gif

this is me rn

[wispy canopy]

yeah, there was(for some time now) and that simply gives you a linux machine inside your windows machine

[wispy canopy]

ugh... I STOPPED TECH SUPPORT BECAUSE OF THIS

[brittle jay]

yeah wsl = windows subsystem for linux

[wispy canopy]

LOL

[magic wave]

And that means speedy AI ye?

[wispy canopy]

wsl2 => it actually can connect to your gpu in a native way

[magic wave]

WELL, TOO BAD, I'M PULLIN YOU BACK TO HELL

[brittle jay]

oh shit

[wispy canopy]

not more speedy. more possible repositories

[brittle jay]

https://tenor.com/view/ah-shit-here-we-go-again-ah-shit-cj-gta-gta-san-andreas-gif-13933485

[upbeat wasp]

Sometimes. Depends on if you have any other bottlenecks, but a powerful enough system can reach 40+it/s on single image batches.

[magic wave]

Ohhh, thats fair then, I thought it was more speedy, cause for @trim current its more speedy

[magic wave]

I think I dont? Maybe?... I'm not too sure

[upbeat wasp]

You have to run PyTorch nightly and torch.compile() or just compile your own PyTorch from source to get that speed though.

Torch.compile() makes models take a long time to load (since they have to compile each time) but if you normally stick to one model it might be worth it.

[magic wave]

I usually do

[trim current]

I am the speed

What i miss

[magic wave]

Linux is faster for you ye?

[trim current]

10x cos it uses rocm instead of vulkan for my amd

[upbeat wasp]

Allegedly the easy ways to get that speed are built into Vlad fork of A1111, so I'd try there and check out the torch settings first if you intend to run on WSL2.

[magic wave]

Ohhh, so its cause of the AMDussy?

[trim current]

Yus

[rich tartan]

ah

[trim current]

But am getting 3060 in a month or so

[pine wasp]

.

[rich tartan]

good that you came

[trim current]

So ditching linux when i do

[magic wave]

Ive heard Vlad's repository is kinda wonky tho

[magic wave]

Fair xD

[trim current]

And can train loras

[rich tartan]

when you installed ubuntu, did you need to format your drive?

[upbeat wasp]

You want speed or you want stability? Pick one and don't complain 😛

[trim current]

Manjaro and i used rufus to prep my d drive

I dual boot

[magic wave]

Fair

I mean, I use xformers even though it removes determinism cause I wan speed

So I think speeed is the way to go

[trim current]

Ill give something quick

A hug

[quiet bramble]

Any new banger anime models drop within the past month or two?

[magic wave]

[upbeat wasp]

Remove xformers, use --opt-sdp-no-mem-attention with PyTorch 2.0 will get your determinism back.

[magic wave]

Aurora maybe? I forgor when it came out

[trim current]

Did chu have time for lora? O;

[brittle jay]

man you making me remember the role playing server I briefly joined a couple years ago cause my friend forced me

[magic wave]

Non-determinism (or what ever the antonym of determinism is) is kind of a baddie tho

[supple raptor]

I've tried Vlad's 4 times now and have not been able to get more than a dozen images off of an install before something breaks XD

[magic wave]

[brittle jay]

lol

[stable nimbus]

U guys seeing sdxl anime gens? Its already comparable to base NAI i feel, the finetunes gonna be wild

[magic wave]

Sounds way too unstable

[brittle jay]

wiat what

[brittle jay]

no way

[magic wave]

Not yet :o

[brittle jay]

show example

[stable nimbus]

just search "anime" in the bot-1 channel

[trim current]

I tried using wd1.5 to merge

Kept breaking

[upbeat wasp]

The good kind or bad kind of baddie? I don't understand.

The point is, Torch 2 has the xformers built in as a standard feature called sdp. --opt-sdp-attention activates it. No-mem is just in case you want to keep determinism in your results at the cost of 3% less speed.

[trim current]

Ree

[magic wave]

#🍥｜anime

[quiet bramble]

I'll check it out, thank you.

[magic wave]

The good kinda baddie, cause I mean, it can be kinda good

[magic wave]

Yee

[wispy canopy]

because of ROCM

[supple raptor]

xformers 0.0.19 is deterministic at least during a session and has been faster and more memory efficient for me than --opt-sdp-no-mem-attention. All of the people I've seen show sdp to be equal in speed are on 40 series cards so maybe that's the difference.

[wispy canopy]

AMD has great benefits for rocm, but that's only available on linux platforms

[upbeat wasp]

Nah, it only messes things up for me when I want to go back and run high-res fix on a group. But just use the regular sdp and you'll still get your cocked up random results that you like.

[magic wave]

Fair then

[trim current]

Kachow

[rich tartan]

you think ubuntu would work also?

[trim current]

Prolly

Best to bug @wheat falcon about it

[rich tartan]

thanks

[magic wave]

Fair, I like the smol changes xformers does sometimes tho :o

[upbeat wasp]

Probably. On 4090, so your mileage may vary. Xformers determinism still varies based on sampler in the latest version for me, but euler_a is a wash anyway.

[brittle jay]

brh

[stable nimbus]

pretty good right

[magic wave]

That is pretty sweet

[supple raptor]

Oh, it isn't deterministic on certain samplers even with 0.0.19? I mostly use DPM++ 2M Karras these days, so I might have overlooked testing it on other samplers.

[magic wave]

I mean

On DPM++ 2M Karras I feel like its barely noticeable

[wispy canopy]

i'm working ubuntu on WSL2

[upbeat wasp]

Last I checked euler_a was still messy, and DDIM was only static if I used 30+ steps.

[rich tartan]

ah cool, so it works clean i guess

[magic wave]

Swap to 2M Chadrras

[supple raptor]

I'll have to try them out and compare. I haven't used euler_a in ages and only use DDIM for Hires fix with UniPC because I've been to lazy to swap that back to 2M Karras.

[upbeat wasp]

I choose based on what I'm trying to make and how fast I want it to go when I'm testing. There's no point trying to declare that any single sampler is always going to be the best.

[magic wave]

2M Chadrras

https://tenor.com/view/gigachad-chad-gif-20773266

[wispy canopy]

UniPC masterrace here

[magic wave]

Same

[wispy canopy]

i like transparant cloth

[magic wave]

2M Chadrras for upscaling tho

[wispy canopy]

yeah, install wls2, then you have to do some shit so it can actually use your gpu native, and then you just install ubuntu from store (the abreviated version)

[upbeat wasp]

I did huge comparison grids and legit could not find any benefit for UniPC over DDIM. They were exactly the same speed, had the same quality, converged to a good image at the same time.

[wispy canopy]

i disagree, did huge comparison grids, and when using low cfg (because my model can't handle anything over 6 properly) unipc is a workable image at 5 steps... ddim not until 10

unipc got 90% resolved at 25 steps, ddim not by far

[magic wave]

If you don't change the upsclaer, UniPC uses DDIM as the upscaler if you use hires fix

[wispy canopy]

ddim however, does get MORE of the prompt right

[upbeat wasp]

Sounds like a model problem more than a sampler issue 😛

[magic wave]

UniPC is also pretty similar to 2M Chadrras, and 2M Chadrras is good

[upbeat wasp]

Was just doing raw outputs for comparisons.

[magic wave]

Oh, fair then

[wispy canopy]

nah, model works pretty fine, it's just very sensitive 😛

[brittle jay]

chadrras?

bruh

gotta say 2S a > 2S a Karras

[magic wave]

Fair, I mean, I don't really care about the 2S a samplers, I've tried em, they can be good, but 2S a seems to be usually more unstable

[supple raptor]

I don't like karras for 2S a or SDE

[brittle jay]

Cranberry juice

bruh my hand looks ai generated

[magic wave]

It kinda do XD

That glass did you dirty xD

[steady grail]

wow, how did you do this? it's exactly what i'm looking for

[brittle jay]

lora

[sly rune]

Damn, that's some nice text on that monitor. Are you using DeepFloyd?

[brittle jay]

frfr

[steady grail]

the colors and the lights are all part of the lora?

[steady grail]

i love this style, so close to stains gate, but warmer. how did you do it?

[sly rune]

Will take a bit of time to type out. Please be patient.

[steady grail]

yeah, sure

[brittle jay]

no

[brittle jay]

got the same camera as this one singular image from steins gate 0 visual novel

[steady grail]

oh goood

[brittle jay]

how tf is this like steins gate

[sly rune]

First I changed ALunarDreamV2.0's base from animefull-final-pruned to AnythingV5PrtRE AnythingV5PrtRE + (ALunarDreamV2.0 - animefull-final-pruned) * 1
Then I used the Model Toolkit extension to put the CLIP from ALunarDreamV2.0 back on the model.
The style used in the image itself is a combination of artists used in ALunarDreamV2.0. mokunoki matcha Hiten fajyobore, PROMPT_HERE

[steady grail]

never played the game?

[brittle jay]

also you know how I rewatched it

for 3 days in a row I watched the entirety of steins gate, steins gate 0 and the movie

now I wanna play the games

probably gonna crack them though

maybe I'll buy in the future

[steady grail]

look at this glorious style https://i.ytimg.com/vi/Wj5AD97TV9g/maxresdefault.jpg

[brittle jay]

why is everyone trashing steins gate 0 anime though?>

I thought it was good

[steady grail]

it's very good

[brittle jay]

not as good as s1 but good nontheless

[steady grail]

deserve to be part of the series

[brittle jay]

though people say steins gate 0 vn >>>>> the steins gate 0 anime

that's why I wanna play it

[steady grail]

especially the ending, gave me goose bumps

[brittle jay]

should I also play steins gate vn first?

[steady grail]

the vn is always better

more details lost in the anime

but the anime is really well done

[sly rune]

Here's what the setup should look like. Add ALunarDreamV2.0's data to AnythingV5, then apply the CLIP.

[steady grail]

wow, it's your personal model

[brittle jay]

I mean, do I need to play steins gate vn to play steins gate 0 vn?

[steady grail]

the anythingv5 you don'0t get weird mutations? or the 5 is better?

3 and 4 were thrash

[brittle jay]

anything is trash in g eneral

[steady grail]

they pretty much fully covered the first game with the series+-spin offs and films

[brittle jay]

so I only play steins gate 0 then

but first I watch spinoffs

[steady grail]

well, yeah, but i still love the first vn, the feeling it's just different, more intense

but in story, i don't think you lose too much

yep

[brittle jay]

only story wise

I am too lazy to play it if it's just for a slightly better experience

maybe when I will inevitably forget everything again

[steady grail]

the story of mr. brown broke my soul >.<

SUZUHAAAAA!!!! >.<

Shippaishita... Shippaishita!! SHIPPAISHITAAAAA!!!!

ok sorry

i feel better

[sullen hollow]

maybe SD was right all along and this is the true form of hands

[trim current]

after fighting this merge for days, its finally okay™️

[magic wave]

Ayyy

[trim current]

[trim current]

Faces kept coming out the same. Finally got some nice faces goin

add detail lora does werk

[magic wave]

That is pretty good

[trim current]

will upload it later

[cyan burrow]

eeeey so Sweden wins Eurovision and then Ice Hockey game vs Finland.

As it should be

You should check out the other VNs/LNs/Animes that take place in the same multiverse as Steins;Gate

You can recognize them easily because they have a ; in their name

[trim current]

dang im liking this

[cyan burrow]

chaos;head for example

and then chaos;child sequel

occult;nine

[sullen hollow]

definitely looks like really good merge

[magic wave]

It do

[trim current]

adding my lora without it keywords

with keywords

[trim current]

aaaaaaaaaaaaaaaaaa

[sullen hollow]

wow thats pretty

[trim current]

Thank you! Im making preview pics before i upload

[sullen hollow]

only thing thats a bit odd is the little deformation on the eyes but other than that its hard to find anything that stands out.

preview pics in terms of controlnet preview?

[trim current]

If you expect perfection from a model youll be searching awhile lol

Adetailer exist

[sullen hollow]

that's true 😆

it does happen tho from time to time

its like the onion ring in fries of AI

[brittle jay]

Yo @wispy canopy congrats

[wispy canopy]

thx 😮

wasn't expecting a win there, i really thought red riding hood would win

that one magical!

[pseudo condor]

dragon fairy

[magic wave]

Is that with WD 1.5? :o

[fallen sundial]

hello how to use this extention ?https://github.com/AbyszOne/Abysz-LAB-Ext

I can't find any tutorial

[obsidian flame]

Serendipitous euphoria cascades through my being, engulfing me in a symphony of blissful harmony

[wispy canopy]

just trying old prompts

hmmm... this one looks nice ^^

[sullen hollow]

serendi what

[teal pivot]

there we go xd

[cyan burrow]

I wonder when AI game making tools come out. I mean like real solid ones. So you just tell it the type of game you want and the engine to use and it codes everything for you. Maybe asking you questions

[cyan burrow]

How do you get so many likes, downloads and ratings? I struggle like crazy to get just a few. Do you advertise it elsewhere?

[nova remnant]

does your model have a really cute anime girl as the cover?

because that's how you catch eyes on civitai

[teal pivot]

only on civitai and 1 russian-language server about stable diffusion, that's all

[teal pivot]

maybe 😄

first of all I make a model for myself to like it, and then I publish it for other people, maybe someone will like it

no vested interest

[nova remnant]

props for getting to release-worthy quality

I'm 7.5 versions in still feeling like my model isn't ready for the big stage lol

[teal pivot]

oh me oh my 7.5 ver 😄

[hearty heart]

@teal pivot

[teal pivot]

good job!

[tepid geyser]

asuka

[hearty heart]

shes my favourite

[tepid geyser]

good choice

my astronaut became a ball

[queen void]

Hey 👋

[sly rune]

See the original message as well.
#🍥｜anime message

[trim current]

https://civitai.com/models/67131?modelVersionId=71762
weee uploaded

[magic wave]

Ohhh, fair then :o

[tepid geyser]

what even is going on here

looks kind of cool but confusing

is it a reflection?

it's also a planet

a moon

[stuck aurora]

What model is it?

[queen void]

Nice

[umbral depot]

NewTestMix

I see my negatives are doing nothing

[tepid geyser]

kinda cute

[polar sable]

Anyone happen to have Old Sketch Style - v1 lora laying around somewhere?
was originally here: https://civitai.com/models/41882, https://webcache.googleusercontent.com/search?q=cache:MPkQWaygmG8J:https://civitai.com/models/41882/ol&cd=12&hl=en&ct=clnk&gl=il&client=firefox-b-d
Author seems to have deleted it for some reason ...

[polar sable]

it gets the style pretty much spot on, shame about the hand tho

I wonder if it's causing the stick wobble because of the upscaler or model is just borked 🤔

Could be the same reason with the extra hand

[tepid geyser]

what model/lora?

also why does interrupt only sometimes work

[polar sable]

based on this https://huggingface.co/MyneFactory/MF-KonoSuba but I've merged a lot of stuff into it since then

I don't remember the recipe, apologies

[tepid geyser]

i went too far

[tepid geyser]

thank you!

[polar sable]

It's giving her the aqua hairstyle because of the konosuba tag

[tepid geyser]

hahahah that's pretty funny actually

how can i convert a .bin VAE to .ckpt?

[tepid geyser]

do you actually go the recommended 70+ steps for yours?

it worked :D

[queen void]

[cunning gyro]

Vesper, a harpy tech expert

[merry tendon]

why is everyone using reference only? reference_adain+attn is way better

[wispy canopy]

they have their uses

[south sable]

made her a song in music-lm cause why not?

there should be a music channel

[wispy canopy]

ngl, this one is very cool ❤️

[steady grail]

cool that dong, done all by ai?

[south sable]

yes, all by Google's music-LM model

[steady grail]

my god

[south sable]

I've been making a few really good ones...

[steady grail]

is full of noise, but the idea is very good and can be truly played

[south sable]

yeah lol it always sounds like it was recorded live with a nokia or something

[steady grail]

well, it all started with noise

like in images

[south sable]

this is my fav so far

[steady grail]

the timing is wrong, put it's pretty good

[south sable]

its cool, like tape warble

feels pirated lol

[merry tendon]

wow, that sounds like an actual song xd

[steady grail]

lol

[south sable]

you know?

it can really chug too. the ai is too good already lol

[merry tendon]

can you try making breakcore?

[south sable]

lets try

[steady grail]

testing models.
different models, same settings

[south sable]

[steady grail]

i'm in love with that warm light in the 2°

[south sable]

@merry tendon

[steady grail]

the last one can be used

i would expect to hear it in fall guys

[merry tendon]

ok, sounds ??. but music industry is f too if this continues

[south sable]

it just came out so imagine it in like 6 mo

its too fun

[steady grail]

weird, same exact setting, but lightly different results

[south sable]

using xformers?

[steady grail]

yeah

[south sable]

it's possible it makes things harder to reproduce. not 100% sure but I think I heard that

[steady grail]

weird

so, there is something random in xformers

[south sable]

I'm not sure how it all works >.>

I like making a song to go with the image :p

[wispy canopy]

i've gone in the deep end of model merging... i'm merging specific blocks xD lol

[supple raptor]

Unless you're using xformers 0.0.19, yes it will make your images non-deterministic, meaning that your results will vary slightly every time you generate, even with the same seed, prompt, and parameters.

[wispy canopy]

remixing my model yet again, did drop the boob size to more sane levels XD

[wheat falcon]

Yes it makes images non deterministic. The same process happens when you restart sd even without xformers.

[wheat falcon]

So 0.0.19 fixed that ?

[wispy canopy]

@wheat falcon this one acceptable, or take it down?

[wheat falcon]

Hmm thats on the edge i would say

[wooden coral]

Good morning

[wispy canopy]

yeah, that's for sure, that's whyi asked 🙂

i thought the effect i got was just pretty damn cool

[wheat falcon]

It is, reminds me of a skyrim char

[glacial bane]

@wheat falcon

New problem

[wispy canopy]

and i just love transparancy ^^ though, my reforged model has other issues again... meh, merged block weights is hard 😮

[wheat falcon]

Did you tried the transparent skin lora ?