#🍥|anime
1 messages · Page 6 of 1
seems on par with previous xformers speeds
meanwhile lol
which gpu do you have?
same old 3090TI
"old"
but running on the new pytorch as you suggested
but you have a100s at work, cry me a river >.>
i haven't been able to use them for over a month
GODDAMN the people who actually work with them
time to "repair" one of them 😼
my 2080 is doing fine!
hum, I wanted to share what a ComfyUI console looked like for a generation but it's not pretty to look at
Ooof
🙏
that reminds me
any idea of new controlnets that would be fun ?
Like, it's possible to invent new types of controlnet
i'll try doing the perfect hands thing, gl for me
does anyone wanna test a lora
sure
ohmi
can't guarantee hands yet, still downloading the models
amber (genshin impact series) *doubt*
am i the only one here who renames loras to be a single word?
uhh... if topless girls are forbidden here... are topless men?
finally smth decent
i dont think so
:(
@sharp bluff weight?
im trying out my prompt i spent 3 days tuning with this model
i hope it comes out nice
lol
ive been using 0.8
clyde blocked the men 😮
too beefy
i think 1 actually works fine too tho
cockless...
but i prefer 0.8
yep, excessive nudity is caught too, whatever the sex of the person
have the girls instead
One hell of a team - Hazbin Hotel
even 6 packs tend to trigger clyde sometimes
her shoulder morphed into her boob
ok yeah 0.8 is a lot better
Booblder
0.8 / 0.4 / disabled
it didnt botch up the character but i think is cause my weights were for that model in tuned my prompt for
came out nice still
when u want 2 cats but you get girl with cat ears 🤣
HOLY FUCK FLOPPA AI
i'm still wondering how the hearing works with furries who have double ears
ikr xD
double headpats 😄
that the image counter on the lower left?
yes, apparently it's an option now
@sharp bluff
ever since my reinstall
but that ruins the images 😮
@wispy canopy only in the preview and grids, the originals are not marked
1.2 / 1.0 / 0.8 @sharp bluff
weird how 0.8 to 1.0 changes so much
imma try to find the sweet spot
Lemme know how it goes!!
I'm still experimenting with it
It used to be an embedding
Before Lora's were a thing
help
lower your expectations ? 
what are your settings
depends what you were trying to do
i was going to generate 1080p pic 
around here i think is a good spot (0.95)
Woo!!
you dont, you generate something like 768x768 and then upscale
yeah, it's a pretty nice artstyle
noob symptom sorry
props to the artist
generate at half or a quarter of your final resolution, then use upscaling tools to make it x2 or x4
what settings ?
idk how to upscale but ok
these things
i'll show you
Extras tab
Good evening
i'll be back later and dm you if thats ok
ok
I wonder how the Lora looks on more realistic models
uberchad 😄
lol i found an old server of mine
the extension thingy?
guess i'll delete it now that midjourney blew the shit out of free trials lmao
you can make your own bot with SD now
nah i prefer not to depend on that
i have good results on a1111 already
the bot fetches from your automatic1111 backend... lol
why not use automatic1111 gui?
basically the same as using the web-ui, but with discord
so you can use it remotely and let your friends play with it too
how used were people to midjourney
don't have to sit in front of the PC
hmm i mean, sure
yeah but not all extensions, and less custumizable overall. I would use the bot when I have the UI
but yeah, I do sit on PC so...
it was nice to do waifu roulette for my friends in my personal discord server
and it had a queue system as well, made new threads for every query
kinda polished, but less UI options ofc
brb, smoke time
good morning
anyone is more knowledgeable of controlnet?
bbl
It doesnt want to give me two normal cats... but I like it xD
controlnet help anyone?
if I look at that screenshot I think you forgot to set the preprocessor
they're already preprocessed
afaik
this left one,
with a depth map preprocessor, this would get turned into a depth map
but i already have the depth map
ok, got the idea you wanted to change hand position, never mind
im gonna try restarting webui
i dont remember restarting after installing the controlnet
good yield
holy fuck i'm not using controlnet
i ran out of vram before even loading the 2nd controlnet model
how that's some epic shit
Wanted to make something else but that came out
happy accident
Prompt was devourer of gods iirc
calmaity reoference!??!??
oh mah gawsh
Yeah but when i add mechanical worm it just breaks
Now it looks like a gurren lagann mech
what are these numbers?
filename number
The numbers mason
lef the pc on to generate while i was doing shit in the house. husbando storm incoming.
they all kinda need fixing 😛
still looking forward to the day that someone will make an extension to autosave ALL settings
you mean setting stuff up in ui-config and config?
mostly extension configs that don't save
also, good gen 👍
hmm, i think most sliders are in ui-config.json
i wonder what would happen if i don't use my lora
i'll do for the same seed, but i'll try a cherrypicked without lora later
idk what she's holding.
maybe trying to commit kms
dude i love img2img upscale, it makes it so much better (original)
are you not using hires fix?
hell no
well i'm not everyone
Not really
Mayday - Coldrain
had some weirdo asking me how to nsfw prompt from here smh
Multidiff is a thing
I'm not using highresfix right now 😮 this model is pretty solid tho
step one: civitai
done.
😛
original vs img2img upscaled
fixes basically everything
also did that bush turn into green water lmfao
or just say dumb shit in the prompt box till it works lmao
or use the models that don't need dumb shit .
^^^^^^
I feel like its harder at times to sfw prompt than to nsfw prompt
thats model dependent in my experience, one model was giving me naked moon boobies when i was trying to make acrylic moon paint stuff 😅
why no hiresfix?
Sussy model
pretty sure my model isn't even the most thirsty one yet
can't use over 1.5 without my pc exploding (6GB vram)
also, am i the only one who uses a1 with 6gb vram without using medvram?
because your brain becomes too horny to notice the faults 
Sussy LoCon atm for me, using a LoCon that is like 95% NSFW, but there is such little images that this time its actually horny
didnt give me straight tits without asking except the odd gen here and there so yeah its less horny than most considering, ironic isnt it XD
Nah, some people like a fap challenge xD
Wha XD
never had problems with horny gens, idk what you guys are talking about
btw, what does [something] do?
OxT - CLattanoia
random gen
this seems like a pretty fun prompt
Also inpaint sketch is pretty nice
Lyric prompting again?
get the right or wrong model and any excuse the model throw titties at your face
as long as my prompts are drenched with RNG, i'll prompt!
Let me join
only sometimes i put nsfw in my negative but still never had issues other than a single incident
again, i ask
[this does nothing]
[this however:does a thing:0.2]
well the incident i said is because of this
it's prompt editing
i thought it reduced weight like () increases
so i put nsfw on it
[] is negative, the opposite of ()
[this reduces weight to 0.9]
(but you can also do this:0.9)
what about {}. ive seen t hose but unsure what they do tbh
{these are for novelAI site}
ah okie
-0.1? damn
{this is for dynamic promps|or any other extensions| but usually dynamic prompts}
or novelai (site)
yes
ive been mixing ([]) on some words, been getting good results when things become too much/not enough
for clarity, i'd wrap them the other way around
but that's just me
[this alternates this|and this]
[()]?
i just (something:0.9), (another:1.1)
any reasoning behind this?
and never do (this) or ((this))
damn this generated horribly
there's no harm, sometimes it's faster
can img2img fix?
i want [(Rain:1.6):0.7] -> i want it to weigh very heavy, but introduce it to the prompt very late
you can also use Ctrl + up/down to quickly modify weight
yeah i do this
i also put the amount to modify to 0.05
i see what youre saying okay, this gives me a new approach at prompting thank you
man, masculine, {medium|short} hair, {stubble|3:: },, [__medium*__|__medium*__] by ([__artists__:0.3]:1.3) , [anime line art:0.2], [{1-3$$__nouns__|__nouns__|__nouns__|__nouns__}:0.01], [[__movements__|__movements__]:0.1],desktop wallpaper , {abstract|surrealism|realistic} back to the true random prompt
it's a bit annoying that it doesn't select the current word if you don't have a selection already
what THE FUCK is that prompt
how did you get this outfit 😮
that's true chaotic random 😄
honestly looks like saber's outfit
kind of

maybe if i increase denoising?
I just put dark souls in the prompt


oh okay
Also just me or AI is better at doing robot hands than human hands? 
interesting meta
This kind of prompt can kinda go anywhere
damn ok
Like these hands are pretty good
seems that prompting style doesn't like big resolutions 😦

almost perfect
Exactly 
for me, i found generating hands with glove easier
They're biased ofc
oh my gosh settings
Angy
what did you do to get this
what lora do you suggest for eyes?
das so cool 
Fair fair :O
1girl on an acid trip, trippy perspective, chaos, universe, timeline,
Negative prompt: (low quality, worst quality:1.4), (bad anatomy), extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000, loli, nude
Steps: 20, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 3726090474, Size: 512x768, Model hash: 1b43df1916, Denoising strength: 0.53, Clip skip: 2, ENSD: 31337, Hires upscale: 2, Hires steps: 10, Hires upscaler: 4x-AnimeSharp
Template: detailed background, best quality, masterpiece,
1girl on an acid trip, trippy perspective, chaos, universe, timeline,
Negative Template: (low quality, worst quality:1.4), (bad anatomy), extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000, loli, nude```
The beginning - One OK Rock
vs inflated token prompt
Borger
what are the top 3 models for anime in your opinion?
I like meinapastel and camelliaMix a lot
I like Holokuki, Aom3, and Kenshi, mostly use Holokuki now though
now time to do magic!
ill check those out
originally:
i feel like it's going back to bad hands
might have to down the denoising
my top 3 most used will probably be counterfeit, holokukiv2 and expmixline
in no particular order
if only it wasnt for the weird leg re


wd 1.5
sd bout to give so many weebs a deformed fetish
I like
Hows counterfeit? I would use expmixline more, but it doesn't seem like the type of model to be too versalite with style LoRAs 
improved a bit more
rap god 💀
tbh, counterfeit is really bad at following prompts. The images generated has a very "counterfeit" feel that you can definitely tell.
Lil Jesus™️
I like abyssorange
cool trick is to upscale and then inpaint the hands with inpaint area: only masked and a resolution of like 768x768
Oh, fair then xD
not super sure about following lora, but yeah, i'm past the stage of using counterfeit tbh
i've never got inpaint to work nicely
but will still probably mix it in when doing my own merge
someone recommended https://civitai.com/models/17598/holokuki to me, i use it ever since

and most importantly, copy other people
Fair fair, I just vibe with Holokuki, it listen and it easy to change style 
Think I did 
true true, and also, if you are bored of the "anime" style, then put in realistic, or (realistic:0.6) and the style changes
my lora doesn't let me
tsk tsk, salutemix is the true best model smh
I mean, I usually do (photorealistic:1.4) 
well, it really depends on the lora, so yeah
I use void brush
I use cetus to make goofy stuff
haven't try photorealistic with holo yet, gonna try that later 
it's not bad also, sometimes give worse quality but it'll depend on the prompt
Its pretty neat 
ive been putting too much photo realistic, cubism in my art lol
Does salute do better with full body?
I do believe so
Salute does better with salute
I was getting pretty mixed results yesterday with full body
this is factual
true, tested, can confirm
That reminds me, haven't made any military related gens with it yet
inpainting my hand is only worsening it
Might include the salute
this one is anihex, but salute works still
with deep ng instead of badprompt 
@lyric zealot (photorealistic:1.4) 
Is there a neuro sama lora
this one is salutemix
Pretty sure there is 
@lyric zealot (photorealistic:1.4) + Holokuki noise offset = Hotness x2 
@umbral depot
kinda weird that i'm genning her, while watching her 
genshin sticker
use the script to remove background 
It big brain 
Honestly... novel has stepped up their game it feels
what script
you need this extension first, then use it under script
tbh, it's kinda cute
i agree
there's limited free trial right?
it legit looks like some japanese artists work
like it's kinda disney like, not gonna lie
No idea
the top left is like, "hi"
sup
Bottom left mood
top right, "tea bag"
the paint texture is nice here

I'd buy that
looks cool
actually perfect
Gives me Owari No Seraph vibes
actually true
I want a Shinoa lora or smth
omg, there is actually a lora for that

oh my gosh i need to squish it
Oh fuck i found it
best not link that here
👀
because there's nsfw preview images
didnt even look oop
cute things are always welcome 😄
das so cute omg, jsut a basic prompt with the lora on?

gonna try that now 

I'll drive your seat
I should make an Adult™️ LoRa 
do it 😏
are you sure your previous lora isn't lewd? 
It shouldn't be hard, but it'll probably be good to pick the best pics
picking best pics is always hard 😄
I mean, somehow it isn't, the dataset is like 95% NSFW tho
I guess its just LoRA training magic
Same with the Sabu ones
cries in slow amd
Also by adult I meant like, no baby face LoRA 
it's capable of nsfw, that's all that matters 
OH thats neat to knoq
she's too tall here 
Tol
i like these fat outlines
it's been missing from anime imo
Celshading isnt it
might be the model i'm using 
Real Cyan is big hot, the two in the wide generations suck 
I mean, that adds up
xD
Which is 
I'm using ExpMix_LineV2 atm
Oooooo i think o has
OOF!
sure hope you do, because as far as i know, it's taken down from civitai and maybe even hf
https://drive.google.com/drive/folders/1otqqXc0JVA0AlIfIgkWoMctgKXGJ1yyf Camellia family here
can't read korean sadly
oh wait
that's the folder
lmao
i didn't know they have a nsfw section 
😏
link saved
whats a prompt so that my chracter isnt so close to the canvas?
I have saved it too for future updates 😄
try close-up in negative
full body ?
Eldritch abomination
half body might work as well
ohh thanks
trying this one too haha
Lol
what's the reason for taking those models down from everywhere tho? 
oo which sampler works best with this you find?
was used by others without permission probably
I'm almost always using DPM++ 2M Karras at 20 steps, with Hires fix at 10 steps, denoising strength at 0.53
canny or depth
i need to learn contrlnet one of these days
Cyan with aqua eyes or with black eyes? 
Is it a good idea to use both? Do you know how I should configure the Weight, Guidance Start and Guidance End?
left looks better imo
Hmm
i can't choose, they are both great
ok I decided. I am gonna start making a visual novel with ai
2 votes for left, 2 votes for right
first, should I use koboldai or oogabooga?

perfect
Let's go!! and good luck!
Oh damn that is perfect :o
you can use both if you want, but probably not that useful with both canny and depth. Not sure tbh. Start meaning when does it start influencing the image generated. And same concept to end
ya i should prolly hires lol
for weight just use at 1, and test from there is my best guess
keep the 80s vibe hahaha
i feel like the lora is a little overtrained, not sure tho
Good way of looking at it yea
baby
That was at 0.5 at 768x768 but keeping in mind im on doodoo rx 6600
But i also had lime 35 steps on euler a 
is it better to txt to img with hires + extra, or txt2img + image2image+extra?

i see, at what upscale level do you usually use?
i usually just hires fix all images, and if it's something i really like, or for a challenge, then i will upscale it even more using "extra"
model?
depends, for me usually around 2
ohh ok, i did notice that with hires it usually is more crisp, other than using image2image
salutemix
damn
amazing
i had that for breakfast 
I want a pizza
sweet
leftovers pz
anyone here uses abyssorangemix model ?
yes
any tips for beautiful gen
had it for dinner like 5 hours ago, 
(((((((((((masterpiece:999999))))))))))))
wrong reeply but yea
Do these symbols (+, : ) do anything in the webui?
bombastic side eye
(beautiful:69420)
nah fr i need some tips my pics turn out blurry or ugly
das like 6 am
tried my best to fix hands
looks great!
use proper negative prompts, depending on the model you are using, you can try mimic the author's negative prompt
and learn from there
This style is cool 
i once used negative prompts that had 2k tokens, was so long discord turned it into a txt file kek
damn, there is lots of activity here today, and quite clear from the NSFW line. Thanks a lot mates
it's a crazy prompt

(I'm here to say when it's bad so don't forget to say thanks too 😉 )
time to hold your horses guys 
anyone have ideas for this?
oh shit, forgot to reply to yours. So yeah, "+" means something in prompt, what it means, not sure
What about :?
What is that btw
what style is it
tbh, i have no idea why i downloaded that, because i have no idea what it does


JWST model 
How do you remove the information from the image? There is information in the name of the image but when I pass it through png info nothing comes out
dz spitzer ?? and inpunk 
discord started striping all metadata. Unless the both of yall have the silent metadata extension, you won't be able to see metadata from discord now
finally got pretty eyes 
hello, bottom moment.
how do i disable a lora? it still affects my design
good fingers as well
I don't. All my images have information/exif in the png. Discord updated and it removes all the info by itself now, I will try downloading this later to help: https://github.com/ashen-sensored/sd_webui_stealth_pnginfo
Photoshop does magic 😄
but yea, wanted to fix it too but got crappy result
so photoshop was faster
how are you your lora? is it from the "Show extra networks" tab?
sounds like hard work 
naaah, I just had to remove 1 additional finger 😄
Exactly 
if only i can say "removing 1 finger is easy in photoshop" 
it is! 😄
heard people saying there's problem with that, never really used it that way, so can't advise you on anything 
not to me tho
i see might aswell remove it and use it manually
can this be made better ?
at 0.1 and some weighting, get 80s outta it lol
this almost doesn't looks like her
but PS has AI to remove things, it's called content-aware fill
for starters
try using the extra network tab way

then inpaint after removing?
depends how well it's removed. Usually I do it that way so I dont have to inpaint
But it will work with inpaint too
how may i do that? i am not aware haha
cute
you did it 
when i was trying to make stickers with midjourney
kinda not for them since finding out being able to run this shit locally >.>
damn you made that?
Not with SD
by hand?
oh didnt notice
maybe i should give it a try
just inpaint the face, ez
you can make full body while keeping the high quality? damn
I mean if you want to pay 40$ a month when you can use SD go for it lol
hires fix 🤝 me
there is discord bot for it i think its the same results? havent tried it yet
@lyric zealot
oh so that's what it's called, thanks, been using that before aswell
I did 😎
try higher weight, might be better, but will be hard to prompt in other style
the angle is tilting more and more 
very noice image indeed
all credits to Nerfgun Aoi lora 
waifu manual
low resolution is indeed bad for quality, but looking at the image, the other factors look more important here, like model prompt etc
I will share the credits evenly
can you make a character sheet in that style with front, left, back orographic views ?
not on an anime model i think
starting to look better
your bottom left text keeps misleading me, making me think it's a gif
evening 😄

Lovecraftian prompts go hard, but what even is this
it's a thing.
obviously
ugh. don't you just hate it when you're trial prompting and you're moving further away from your intended result?
Yis
Yup
ofc no1 likes it 😄
for stuff liek that you could try to use cutoff
whas datt?
instead of assigning colors, you can try to establish relations
a controlnet thing?
it's YAE: Yet Another Extension https://github.com/hnmr293/sd-webui-cutoff
interesting, ill give it a try
i used it for the main pciture of my model page
it helped me establish what each person should be doing in the picture
instead of mixing up prompt elements
again, i'm pretty sure it wasn't the intended way to use that tech. but as the song goes: "The internet is for porn..."
805 my fav
no comments? ok 
so good, the semi realistic vibe
gotta love it~
i installed cutoff and now my ui is broken .-.
update v4
If having problems after installing an extension, always close SD completely and relaunch it and be sure to refresh the WebUI webpage.
Made a kikuo thumbnail 💀
FR
ah okie illl have to do that
was was the command to add to the bat file to pull updates
i did have git pull origin master but now i get this error when running the bat fatal: not a git repository (or any of the parent directories): .git
how can i make controlnet stronger?
its not doing exactly what i wanted with the hands
Yeah, I don't remember where in the .bat is the correct place to add it though.
@proper trellis, native , this model doesn't know about orthographic. i would have to use latent coupling to be able to get it done
noice thing going on here
Pretty sure it was here
it knows ortographic has something to do with perspectives, but not versed enough to know what it exactly is
my model has 200 downloads? 😮
who are these sick people 😮
Ayyy, lets go 
if you installed it rn, you need a full webui restart
That seems right. Thinking about it more, it could probably go anywhere before call webui.bat and still work.
ive had it installed for a while 😔
True too :o
well idk, i only have 6GB vram, i can't use that
usually even with controlnet, if the hands are not super close-up, it's hard to get it right as well
Also do you still need to make a new folder for the new update? :o
ah alright
@lyric zealot MotHER Kill ME
I'm not sure. I had updated and had problems, then reverted and was having problems so I just did a full fresh pull and that fixed everything for me. It's also possible that my problems were coming from torch 2.0 which I had upgraded to but didn't after the fresh pull, or maybe it was the AutoMBW extension that I didn't reinstall after the fresh pull.
I see, and whats up with torch 2.0? I think I remember you said it was faster than xformers, but can you have like xformers + torch 2.0? :O
hmm maybe ill need to go find character sheets with front/side/backviews and create a lora or embedding to get the orographic views, i tied the charturner lora but can't seem to get it to work right either 😦
i think it's actually a challenge without a lora
you'd need to get it to learn 4 views: normal, top, side, front
AND be consistent
consistency is not my forte...
i tried with latent coupling, but i can't seperate the areas strong enough 😦
It was faster for me for about a day, then I started having the problems that caused all the above. My gens were getting slower and slower each day and vram utilization was going up. Normally I can gen 512x768 with 2x highres fix at batch size 2 no problem, but at the worst point I was getting occasional OoM at those settings with batch size 1.
You can have torch2.0 + xformers now, you just have to run the command to manually add xformers as the --xformers commandline arg will pull the wrong version.
xformers + pytorch 2? 😮
does that even give an advantage?
Oh damn, so like for a time it had like a massive memory leak? But now its fine? :o
it's not a leak, it's memory fragmentation
What Eface said, Eface smorter :o
make sure to have this: set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.9,max_split_size_mb:512
Hmm
nice
I'll have to do that if I update :o
yep, SDP gives most of the speed benefits of xformers, but doesn't seem to give the VRAM benefits.
uhh...
Just started trying HexaW that was one of the first images I made with it.
Ohhh, so pytorch is for the speed, xformers is for the vrams? :o
elaborate, which command is it you mean ?
i'm on torch 2 currently
how can i use a .bin
Just throw it into your embeddings folder :o
KETCHUP. ACCIDENT.
Ketchup
I followed these instructions https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8691#discussioncomment-5403723
makes sense, thanks :3
lemme try that rq
sweet
Do either Anihex or HexaW have offset noise mixed in?
why is it that landscape sometimes is more lower quality than a portrait orientation?
If you're generating people its because the relative size of the person is smaller in landscape resolution than portrait. Less pixels means less quality.
Is there any embedding or lora to have detailed eyes?
I see, would increasing the pixels in the settings fix that?
You can try any negative embedding, and see if they work for you
not by me at least
thx @supple raptor ->adding Xformers 0.18 seems to have added 0.1 it/s and reduced memory pressure . That helps. a lot.
it's xformers tho, i lost determinism again 
oh wait. i don't deal in consistency anyway
I mean, by actually how much do xformers & pytorch 2.0 actually speed up generations?
significantly, and reduced memory pressure
kinda blew it because xformers depends on torch=2.0, but the installs currently are nightlies 2.1, so i suppose i gotta reinstall this
I mean, I got xformers
I feel like the best would be to have both tho, if they work together
Possibly but you may run into limitations of the model. increasing either dimension over 768 will generally start to decrease the coherency of your generations. Above 1024 is when this usually starts to get really noticeable. The better option is usually to just use Hires fix or Multi-Diffusion upscaling to increase the total resolution from a smaller starting resolution while adding more details.
To do that buff stacking style things
i think eface told it best earlier : you get the speed of xformers while preserving deterministic results
throwing xformers in there you will lose that deterministic for sure
Is deterministic better? 
it means you can always reproduce exact results
if you want to reproduce the same results consistently, yes
But also, super speed does sound nice 
xformers prevents you from getting exact results again
tested on 3080 -> xformers + sdp does not give a significant speed boost (none), but does help alleviate my memory issues
as in, i can do 8x 640x640 and have memory left over
@somber nova here is an example of one I just did in landscape using Hires fix and then multi-diffusion. This started as a 768x512 image, hires fix increased it to 1536x1024, and multi-diffusion increased it again to 3072x2048
and batch size 8x > batch size 4x batch count 2x
This two the same seed
And there is a difference, but I feel like its too mild to matter
amazing
welll..... in dutch there is a word for it: "mierenneuken"
"fucking ants"
people who nitpick the really small details
I mean, weird flex, but fair, I'm not nitpicky xD
uhh... who splattered ketchup on this ... soundbar? console?
nah, it's a common expression here, when someone is nitpicking at insignificant details: that person is a "mierenneuker" -> Person who fucks ants
what do you mean by multi-diffusion? Is that similar to 4x-animesharp?
Fair XD
Hello
cutee
ohh isee ill give it a try thanks again!
do we still need to type --xformers in the CMD arguments ?!
@narrow nebula if you value your time, leave this channel. this is an infinite maze -> infinite waifu generation slavery
Why is your cmd aesthetic 
Yes, otherwise you won't get any benefits from installing xformers.
it does look for it tho
i'll take the xformers, even if it doesn't give speed advantage, it drops memory usage by a LOT
I'll just stay on the update I'm on, if I can't do buff stacking, then I'll just keep making waifus the same way

i use Terminal
Wha 
Ohhh, fair then 
did you the pip install -U --pre xformers
yes
it insists on going back to pytorch2.0
yes
i know i'm a sysadmin, but honestly, SD is kinda black magic to me tho
not devops or programmer
🔥
not husbando tho ❤️
version conflicts
not today 😄
uninstall torhc 2.0 and install 2.1? 😮
i dunno, tha'ts what i would do
that's what i'm doing already 😛
done it twice now
i hate it even more that it worked out of the box for me 😦
I don't know of anyone using torch 2.1 for SD. I thought there were some incompatibilities that haven't been sorted yet. I could be wrong though.
Can you not run multi controlnet on 8gb vram?
apaprently i would need to compile --xformers myself i suppose (again)
I keep getting out of memory errors and all I have is Firefox open
did you set it to software mode?
nope






) pic




