🍥｜anime

[native halo]

yea i remember that one its like watching white album but without the rape

[nova remnant]

hmm it wasn't doing anything but now it looks normal >.>

[lyric zealot]

omg white album 2 was great tho

[lyric zealot]

[native halo]

yea the 2 was great but first one 😔

[supple raptor]

Haven't watched it but I'll add it to the list

[lyric zealot]

it's like the closest anime that's borderline ||h||

[nova remnant]

tried typing cd to see if it was working, took a few seconds to come back

[lyric zealot]

story is great tho, don't get me wrong

now try running it again, the webui

[lyric zealot]

@nova remnant

[supple raptor]

I'll move it up the list a bit

[brittle jay]

Scum's wish?

[native halo]

speakin of borderline h but also shonen try with seikon no qwaser

[brittle jay]

I watched like most of the first episode

So bad

Omg

Like idk

Didn't even finish the first ep

[nova remnant]

O: @lyric zealot installing torch and torchvision

[lyric zealot]

never heard of that..

[brittle jay]

Any error or something?

[lyric zealot]

you meant scum's wish?

[nova remnant]

has that in red, not sure what it is

[brittle jay]

It is no problem

Very good man

[lyric zealot]

well, i think the story is great. I like how there ||isn't a great ending for both MCs||

[lyric zealot]

||not really not great ending tho, just kinda not how people would expect it to go||

[brittle jay]

@supple raptor is it supposed to download cuda?

On amd

[brittle jay]

Me when I spoiler myself on a series I do not plan to watch

[supple raptor]

I don't think so. https://cdn.discordapp.com/emojis/1093689156280000524.webp?size=64&name=Anime_Think&quality=lossless

[brittle jay]

We will see

Maybe he need to do something specifically for amd

@nova remnant go to automatic1111 repo of the webui and search for amd something

Probably says something about how to do it

I am too lazy to search myself you see

[brittle jay]

I think I just understood

You are using betterdiscord fake emojis?

[supple raptor]

nope

[brittle jay]

Manually?

[supple raptor]

I don't use betterdiscord

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

[brittle jay]

Bruh

Wdym

[supple raptor]

VC

[brittle jay]

What vc?

Oh

Yeah that one

Same same

[nova remnant]

we don't have the cuda power

[lyric zealot]

[brittle jay]

Do what I said above

[supple raptor]

you can ROCm™ all night long though

[brittle jay]

W

[native halo]

its about a guy who draws power from drinkin breast milk

[supple raptor]

that's the one with booba lasers right?

[nova remnant]

a what

you anime people are weird

[native halo]

yes

[native halo]

still not the weirdest out there

[supple raptor]

not even close to the weirdest

[nova remnant]

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

I don't wanna know, goodbye cruel world

I think the thing is installing stuff

I'll get foodies and come back later >.>

[native halo]

but now u know it https://www.animenewsnetwork.com/encyclopedia/anime.php?id=7439

[lyric zealot]

https://i.imgur.com/6tZWZOV.gif

[nova remnant]

Is this still the one where I asked if you could make with backgrounds?

[supple raptor]

yep, same model+lora combo

[supple raptor]

btw, i post all of my images with PNG info intact unless they're edited so you can almost always click the image, open in browser, and drag from the browser to the PNG Info tab in your WebUI to get the full prompt and parameters.

[lyric zealot]

unlike me

[nova remnant]

so what do I do now

[nova remnant]

Why don't you do it

[lyric zealot]

because im too lazy, i usually just copy and paste from browser

[gusty palm]

same

[supple raptor]

from stable-diffusion-webui folder run these three commands one after another venv/bin/activate
pip uninstall torch, torchaudio, torchvision
pip install torch torchvision --extra-index-url https://download.pytorch.org/whl/rocm5.1.1
then try launching again.

That should give you the AMD compatible version of torch. Right now it's looking for an Nvidia GPU.

[lyric zealot]

cursed

[nova remnant]

@supple raptor ...

[nova remnant]

It did a bunch of things and ended like that

is it fine, doesn't look fine

[supple raptor]

wtf. why is it still insisting on cuda bullshit?

[nova remnant]

like it did a bunch of cuda downloading

[supple raptor]

which gpu specifically do you have?

[nova remnant]

Asus RX 580

[nova remnant]

Blood moon!

This appeared right after the command

[supple raptor]

hmmm... that might be the problem. I know ROCm is supported for RX 5xxx+ but i'm not positive about RX 5xx series. Since it's installing ROCm stuff now, I'd say to let it finish and see what happens. If that doesn't work then I'm a bit lost.

[nova remnant]

It finished, should I try running the webui.sh again ?

[supple raptor]

as long as you got the line that says Running on local URL: http://0.0.0.0.7860 you should be good to just open the WebUI in your browser. If not, then go ahead and run the webui.sh again.

[nova remnant]

and do I do the skip cuda test argument ? >.>

[lyric zealot]

try without

[nova remnant]

same thing
"Torch is not able to use gpu add --skip..."

With it

The ./webui-user.sh doesn't do anything still, the commandline argument is in the webui-user should I put it in the other one instead ?

[lyric zealot]

I don't know why your torch isn't able to detect your gpu

[native halo]

maybe try to put a space here commandline_args=--skip-torch-cuda-test so it looks like this commandline_args= --skip-torch-cuda-test

[brittle jay]

You are running webui.sh

Oh wait nvm

My bad

It should be that way since webui.sh imports the stuff from webui-user.sh

[nova remnant]

still the same

[supple raptor]

this is webui-user.bat which we aren't using here. for comandline args in linux you add them to the launch command instead like ./webui-user.sh --skip-torch-cuda-test

[nova remnant]

10/10 beanie would use

[bitter latch]

there now she's more anime 😄

thats with the realcartoon model

lets try an anime one

😄

[wispy canopy]

having to do my merge again 😦

something went wonky

[bitter latch]

ok this is s a xl model gonna take longer. joy

[bitter latch]

oh no

[wispy canopy]

the bias is extremely towards female portraits

which generally would not be a problem, unless you really don't want to 😦

the strange thing is, that neither of the original models had that bias

[nova remnant]

@supple raptor it worked till this point

it really wants the nvidia card it seems, any idea ?

[bitter latch]

ok maybe not

[supple raptor]

I'm lost at this point, only thing I can think of is that the prebuilt ROCm torch wheel might just not like the RX 5xx series or the combination of that and WSL. It's probably possible to build it manually but that isn't something I know how to do.

[nova remnant]

Maybe these didn't work ?

[supple raptor]

You can run pip list to check the installed packages

[bitter latch]

oooh (non XL)

[bitter latch]

nice

[supple raptor]

The last thing I can think of is to run pip uninstall for all of the cuda stuff as well as torch and torchvision, then try to follow these instructions under Option 2 starting at step 3 https://rocm.docs.amd.com/en/latest/how_to/pytorch_install/pytorch_install.html

[nova remnant]

@supple raptor

[magic wave]

I'm cooking

[analog dagger]

Funky style

[wispy canopy]

welp, here we go again. huge merging attempt, to get the model to do what i want 😮

[fringe drum]

sry again got questions - finaly managed to move to D particion and got a lot of free space - can you maybe explain to me how to get lora samples ?

[lyric zealot]

i think you meant getting lora right?

[fringe drum]

yes i think? the data for it

[lyric zealot]

okay, so basically lora is basically a much much much smaller model that's an addon for your model

you can find lora at civitai

or huggingface

i recommend civitai since it's easier to find what you want

[fringe drum]

so when i download what will should i do ? just extract to my local Lora folder?

[lyric zealot]

download the lora that you want, and drop it into your models/lora folder

[fringe drum]

last question (i hope) after instalation i shut the computer down and turned it back up - how do i start SD back again?

[wispy canopy]

getting close the kind of outputs i want it to be able to make

[lyric zealot]

run the webui.bat or whatever it's called again

[wispy canopy]

isn't it webui-user.bat?

(in case of a1111)

[lyric zealot]

yup that

I don't really remember what it's called

[magic wave]

Do any of yall know how to get perfect lines in AI pics?

[glacial bane]

you're the best 2d generator ik here dawg fr

one of the best

[lyric zealot]

(perfect lines:999)

[magic wave]

[magic wave]

XD

[glacial bane]

results is one straight line

[magic wave]

Well, not perfect lines

But like

Stable lines

[lyric zealot]

i appreciate it, but i think it's mostly because i just dump whatever i gen here, but people here usually nit pick

[glacial bane]

nit pick about what?

[lyric zealot]

like people usually pick like few of their better generation before posting

[glacial bane]

so mostly you post your generates that arent you best ones?

[magic wave]

I usually do that xD

[glacial bane]

its obvs cuz you use your best generates as a pfp

[lyric zealot]

no more like, i just post all my generation, so chances of people finding stuff they like is higher

[magic wave]

True

[glacial bane]

i did get inspiration from you FR

well it is time to dilly dally for tea time tally how

[lyric zealot]

[nova remnant]

Can someone slap me

I was using SD without any commandline arguments

[brittle jay]

This very cool and good

[brittle jay]

This more

[brittle jay]

Man you tryna make her seem suicidal?

That's why blue colors?

And rain?

And top floors

[summer shuttle]

nah she's just liking the view

[brittle jay]

Bro

[brittle jay]

When

How tf do you spell exaturate

Exsaturate

To exaggerate

Omg

Is that it?

Damn

Let's go

Finally

[nova remnant]

all this time

on my normal windows one

[brittle jay]

Bruh no way

Not even xformers?

How much time was it per image? 20 years?

[summer shuttle]

kinda mysterious though

[nova remnant]

AMD doesn't have that

[brittle jay]

Wait fr?

[nova remnant]

yup

I have --lowvram --opt-sub-quad-attention --upcast-sampling now

[brittle jay]

Not even something similar?

Lowvram?

How much vram you have?

[nova remnant]

Guess ^^ 👆

[brittle jay]

2

Omg man

I am fr stupid

Took me time

Like even after you edited the message

Bruh

Not trolling

Anyways tho

[nova remnant]

300 iq touhou player confirmed

[brittle jay]

With 4gb I would use medvram tbh

[nova remnant]

I could try

[brittle jay]

Unless like you can't make images big enough to your taste

Tho upscaling is op

[brittle jay]

I am no touhou player

Maybe that's why

[nova remnant]

What difference would it make

[brittle jay]

It makes it faster

But less maximum high resolution

And less multiple controlnets at the same time

And some other stuff that depends on vram

Try making a 768x512 image

[nova remnant]

I can make them and hires to near 700x1000

[dawn jasper]

This is powerful.

So is this.

[steady grail]

using auto1111?

[nova remnant]

yes

[fringe drum]

are there any bots or services like midjourney?

[nova remnant]

the bots in this discord can do image like in #1100170312106127410

[fringe drum]

are there any limits? or rules ?

[nova remnant]

#✍🏼｜rules-and-tos probably these apply to the bot, mainly nothing NSFW

[fringe drum]

so i need midjourney or something like that

[fringe drum]

and ideas?

[nova remnant]

I don't know what you mean by that, just have common sense to not be

[nova remnant]

is it faster than nothing ?

I mean not having any commandline arguments

[brittle jay]

No

Makes it slower

Compared to nothing

But then you will be extremely limited in what resolutions and stuff you can use

With 4gb vram that's very few

[nova remnant]

I can't seem to be able to hires above 1000 with --lowvram

and is slower

So not really sure what it is doing

[brittle jay]

Then switch back to lowvram

[brittle jay]

Should be faster

@supple raptor right?

[nova remnant]

I wasn't using lowvram

[brittle jay]

Oh

[nova remnant]

but now I'm trying it and its just making it slower >.>

[brittle jay]

Then it should be slower

[nova remnant]

but can't seem to be able to do higher res

[brittle jay]

Because medvram is not enough probably

Do lowvram and you should be able to

But it will be even slower

[nova remnant]

Okay crazy

Listen

I wasn't using anything before

and I could do like 500x700 and upscale by 1.3 or 1.4

but now with lowvram is still the same resolution at twice the time

[brittle jay]

Huh?

[nova remnant]

idk amd life

[brittle jay]

By before you mean windows?

[nova remnant]

without any commandline arguments

I've always used windows

[brittle jay]

What are the other command line arguments

Oh right you sen them

Sent

[brittle jay]

Idk what upcast sampling is

What is it?

[nova remnant]

github says "Upcast sampling. No effect with --no-half. Usually produces similar results to --no-half with better performance while using less memory."

I was just recommended it by someone in this discord

[brittle jay]

Remove it ig

You don't need that no half shit

I think this should fix stuff

Probably

Guys you know how I bought a steam deck like a couple of days ago?

Turns out like a day later valve decided to celebrate their 20th anniversary by doing a huge sale on the steam deck

Bruh

I could have saved SOO much money

Bro

This is unfair

I should send gaben a message or something

[nova remnant]

Now nothing wants to work

Like I turn on my PC, open SD and try a 512x512 image

It gives me out of memory

[brittle jay]

Bruh

[brittle jay]

Just checked the dates

It was literally an exact day later

Bruh

[nova remnant]

Skill issue

[brittle jay]

How would I have known

I sent gaben an email

Not requesting anything

Just wanted to tell em

[magic wave]

I try for nearly 3 hours to get a anime style (like a anime screenshot), only to reach failure

[brittle jay]

I wonder why so many different models exist out there

Might be a reason why

lol

[magic wave]

Anime screenshot style is so hard to do

[nova remnant]

[brittle jay]

Unless you use the right modwl

Then it's hard not to do

[magic wave]

Well, you see

I aint for the streets

Am loyal to Aurora

[brittle jay]

Bruh

[magic wave]

Best model

XD

[brittle jay]

I see

You got too used to it

And now you are bad with other models

Hehe

Jk

[nova remnant]

I am bad at everything

[brittle jay]

Nah man

[untold glacier]

Yeah... I see, lol. What's the story there?

[brittle jay]

About me

Goodnight everyone

[magic wave]

Good night Kazoo

[untold glacier]

GN!

[nova remnant]

I am dangerous cat

I put funny bio in my discord :p

@brittle jay

Now it wants to work

With lowvram

pls AI gods stop bullying me

[nova remnant]

512x768 is going at 14s/it

Why does this thing hate me

It was 3-5 before

[magic wave]

AI discriminates low VRAM

[nova remnant]

I'm installing 1.6 to see if it gets any better

Might just give up if not

[nova remnant]

one 512x512 and the other 512x768

[lyric zealot]

can't you refund this, then buy the sale version?

[lyric zealot]

if you intend to gen higher res, maybe get multidiffusion(?)

[untold glacier]

You sure nothing else is taking up your VRAM?

[untold glacier]

Nice pfp

[lyric zealot]

i was shocked that you are getting 13it/s

then realise it's the otherway around

[nova remnant]

https://tenor.com/view/copium-cat-gif-27161395

Too fast, I don't like it

[lyric zealot]

this is my 1 generation at 512x768 for 8gb vram

[nova remnant]

Nvidia card?

[lyric zealot]

[nova remnant]

@lyric zealot

I think I fixed it

On 1.6 was working!

I reinstalled a bunch of stuff

[lyric zealot]

nice nice nice

[magic wave]

Me, though I should go the same as you, cause I'm using Restart, and it goes twice like SDE, so I don't need tis many xD

[lyric zealot]

restart?

[magic wave]

Yeee

[bitter latch]

there now it's more anime

[magic wave]

New sampler

Its gigachad

[nova remnant]

@lyric zealot on 1.5

[magic wave]

Better than any other imo

[nova remnant]

praise be

[lyric zealot]

oh shit, i didn't update yet

do i update and risk breaking everything?

[magic wave]

I didn't break anything :o

[nova remnant]

the UI on 1.6

[lyric zealot]

I'm scared

[magic wave]

Its fine for me :o

[bitter latch]

anime Ahsoka

[nova remnant]

That gen I sent was "red lipstick, cyan eyes, black eyeliner"

Is there like a way to make ai do colors on specific stuff well

Through text2 image

[magic wave]

Prayers

[lyric zealot]

i was sure there's an extension for that, but can't remember the name

and if im not wrong, it didn't work very well

[nova remnant]

Okay Google download Mario

[magic wave]

https://github.com/hnmr293/sd-webui-cutoff#how-it-works

That

@nova remnant It can help

[wispy canopy]

can someone generate me a waifu to cheer me up?

[wispy canopy]

my mixing is starting to become depressing

[lyric zealot]

wtf where my buttons?

[lyric zealot]

yup that

[magic wave]

Its all down here now

[wispy canopy]

i've seen racing tracks with less curves

[lyric zealot]

is the button gone or i fucked something up?

[nova remnant]

You're now left with 2 sad buttons alone

[lyric zealot]

i wanna paste my style down tho

[gloomy lodge]

all the more fun to ride

[lyric zealot]

how dafug do you paste your style now

[nova remnant]

Isn't it the pencil button located on the side

[lyric zealot]

okay they hidden it inside that pencil button

that clipboard button is what im looking for

[magic wave]

Ye I was gonna say that

[lyric zealot]

i like how you can remove or even edit style now from webui

[nova remnant]

Wait

It doesn't go to your prompt area anymore?

[lyric zealot]

you can get it to paste there using that clipboard button

[nova remnant]

Why is the clipboard not in the same place too

There's space kek

[lyric zealot]

exactly

[magic wave]

Restart

[nova remnant]

How long did both take

[lyric zealot]

looks pretty good

[nova remnant]

restart I want

[lyric zealot]

[magic wave]

Restart all the way

[lyric zealot]

https://tenor.com/view/galaxyroleplay-restart-gif-27103273

[nova remnant]

👍

Calm down smile

[wispy canopy]

i'm always voting for restart

even restart at 10 steps usually beats all the others

[magic wave]

Peachy never stops

[lyric zealot]

let's test that

[wispy canopy]

ON SALE NOW!

[wispy canopy]

it's a good thing latex stretches, but i think you missed the (skindentation:11)

[lyric zealot]

[wispy canopy]

in terms of details, restart is not fucking around

and also -> already more image elments already in the correct shape

[magic wave]

True, and in terms of smartness

[magic wave]

True

[wispy canopy]

once you go restart , you hardly can go bck 😦

[magic wave]

True and real

[wispy canopy]

it's not even close...

[lyric zealot]

now im trying to see if there's even a point in having 10 steps for hires fix

[magic wave]

There be

I do 20 tho

[lyric zealot]

i think 5 is enough for mine

[magic wave]

Well, for me 20 looks really clean

[wispy canopy]

right now running my tests on 30

[lyric zealot]

the changes from 10 onwards, in my case, is probably due to xformers

[untold glacier]

Here's a sampler test I ran with all the new samplers and some of my favs when I upgraded Auto to 1.6 awhile back. I wasn't a big fan of Restart compared to 3M SDE.

Be forewarned, the image is quite large.

[magic wave]

I didn't really like 3M SDE :o

[untold glacier]

3M SDE is a heck of a lot smoother than 2M SDE, yet it keeps the detail. If I want super smooth though, I still finish things off with Euler A. Usually by using 3M SDE on a gen, with the Hires Fix going Euler A.

[lyric zealot]

can you do a plot for hires fix steps 10 and 20?

curious to see if it's really that much differences

[magic wave]

I'm gonna do 1, 5, 10, 15, and 20 :o

[lyric zealot]

[magic wave]

THOUGH, bear in mind, all of those have ADetailer at 40 steps, I forgor to turn it off

Though if you look at the sleeve you can notice the difference between 1 and 20

[lyric zealot]

yup, can even tell some differences from 10 to 20

[magic wave]

Yeee, 20 is a goated amount

[lyric zealot]

this reminds me that there isn't a need to test all samplers

[magic wave]

XD

[untold glacier]

A single seed?

[lyric zealot]

yup

[untold glacier]

[magic wave]

You don't need to test them

Because Restart is the best

[untold glacier]

[magic wave]

It literally is tho

xD

[untold glacier]

Not in the testing that I've done and posted up above, but granted--it wasn't in depth.

[lyric zealot]

well, i will probably go with restart or stick with what i usually used

[untold glacier]

Regardless, different samplers do different things, and should be used for their best fitted purpose.

[magic wave]

I mean, the only reason to not use Restart is if you're using UniPC or Euler a

Because those behave differently

Euler a and UniPC have wild effects over style

[lyric zealot]

[magic wave]

Specially UniPC, Euler a is softer, and UniPC is more sketchy and messy

[untold glacier]

If you want more texture, 3M SDE or 2M Karras. More smoothness? Euler A. A solid starting point for most Anime models? UniPC. Etc., etc.

Knowing the quirks of each sampler is a large part the battle when genning.

[untold glacier]

Seems you know some quirks pretty well. 🤝

[magic wave]

I mean, I think

UniPC

Is good for sketchy styles like ExpMix_Line

But bad for a neutral style

Because sometimes it can be really strong

[lyric zealot]

https://tenor.com/view/aimoto-rinku-d4dj-first-mix-anime-eating-popcorn-gif-19850531

[untold glacier]

Definitely depends on the model. DarkAlfa works great with it in many styles, for instance.

[magic wave]

Euler a vs UniPC

[untold glacier]

Also, different samplers have different convergence points, so I wouldn't use euler A at 10 steps, but I could use SDE Karras at that many steps.

[magic wave]

Well

Basically

There is 2 kinds

If we ignore the weird ones

There is 2 kinds

The single action ones, and the double wammy ones

Single action, 40 steps, double wammy, 20 steps

Euler A does 40 steps at 40 steps

SDE Karras does 80 steps at 40 steps

Because it goes twice per step

It also takes twice as long

[untold glacier]

This is what you're talking about.

From: https://stable-diffusion-art.com/samplers/#Image_Convergence

[magic wave]

Yeah that

Single action boys, and double wammy boys

[lyric zealot]

[magic wave]

Also

Fun thing about Restart

And why its so good

[untold glacier]

I get that. But even within those groups we have different lengths of convergence. It's hard to directly compare samplers that way.

[magic wave]

It Restarts while you generate, so it re-adds some noise onto the image, and basically adapts to the image you're generating, so its extra smart because its like adaptive

[magic wave]

Most of them in the group have the same point

Only weird ones have either a limit

Well

They all have a bare minimum I guess

But

Some of them have limits

Like 2S a, and PLMS

[untold glacier]

If you're using a complicated prompt, prompt editing, or composable diffusion, that really doesn't help things--hence having issues like these.

[magic wave]

I mean

[untold glacier]

But that just falls into the category of proper prompting with the tools that you have--model, sampler, and steps.

[magic wave]

Latent Couple, Controlnet, Restart

So it might be something else giving you issues

[untold glacier]

SD Is certainly a complicated beast that we wield. With so many factors having to work together in unison, if one part is off, then the end result can suffer. The sampler is definitely a part of that, and why the same prompt and settings (such as I posted) can work great with a different sampler.

[magic wave]

I mean, maybe, I haven't had any issues with it so far

[untold glacier]

We need a fang reaction.

[supple raptor]

same image with Restart came out a bit cursed.

[magic wave]

She just has bad aim with lipstick

[untold glacier]

Lol

...and that.

[supple raptor]

https://tenor.com/view/dog-lipstick-make-over-gif-5218367

[magic wave]

One bad hand, one good hand

[untold glacier]

So I just started watchin Jujutsu Kaisen for the last couple of days.

I wish I knew just how much of a gut punch this thing would be.

[magic wave]

The first season?

[supple raptor]

I can't relate

[magic wave]

XD

[untold glacier]

I've now watched the first and first few eps of the 2nd season.

[magic wave]

Shimakaze is fun to generate

[magic wave]

O, fair, first time? XD

[untold glacier]

Yes

"Here, let me give you some great new characters that you'll love and now watch as I brutally rip them away from you."

[magic wave]

Oh you poor soul

And I would say

[untold glacier]

If I wanted that, I would've rewatched Made in Abyss.

[magic wave]

Jujutsu is pretty mild at being mean

I feel like Made in Abyss is way more brutal with it

[untold glacier]

Perhaps, but it's the same analogue.

Here I am wanting to watch an anime for its awesome chars and amazing animation and it takes out my heart and crushes it repeatedly.

Still a great show though.

[magic wave]

It do be

Also, Undead Girl Murder Farce is a fun show, less mean, and its really cool

[untold glacier]

If I didn't know better I'd say that was a made up title. Lol.

[magic wave]

Fair XD

The original title seems to be in english tho, not even translated title

[untold glacier]

I can hear the katakana pronunciation in my head, lol.

[magic wave]

XD

Fair xD

[supple raptor]

sometimes you get one that you just have to fix

[lyric zealot]

should i use opt-sdp-attention or xformers?

[supple raptor]

xformers gives me lower vram usage and faster speeds despite all the claims I've seen of them being equivalent.

[lyric zealot]

what's gonna happen if i enable both

[supple raptor]

it'll use SDP because Automatic1111 doesn't like xformers.

[lyric zealot]

ahhh alright

so that means im getting a little speed boost from SDP

maybe

oh my, the latest version, if you interrupt, it doesnt save?

good feature

[supple raptor]

It uses one or the other, never both. SDP will give a speed boost over nothing (which uses Doggettx's cross-attention optimizations on Cuda devises.)

[lyric zealot]

well

guess i will have to figure it out tomorrow

or should i say later when i wake up

[supple raptor]

yeah, that's a setting now which is set to not save by default.

[lyric zealot]

cool cool, been wanting that for like forever

funfact, my pnginfo is still fucked

[supple raptor]

if you use --xformers in commandline args, you can freely switch between them without restarting in the Settings > Optimizations menu.

[lyric zealot]

ahhhh

[supple raptor]

--xformers flag is required for WebUI to even acknowledge the existence of Xformers, but the setting overrides which one is actually used.

[lyric zealot]

so i should run sdp with commandline but using settings

got it

new thing to test tomorrow later

time to take a nap since it's 6.15am for me

imma wait for Nubby's

[supple raptor]

If you don't have xformers 0.0.21 already, you should also update to that as it gives Flash Attention 2 if your card is compatible which is a little faster than the Flash Attention in 0.0.20

update with pip install -U xformers in cmd with venv activated.

[lyric zealot]

@lyric zealot If you don't have xformers 0.0.21 already, you should also update to that as it gives Flash Attention 2 if your card is compatible which is a little faster than the Flash Attention in 0.0.20 update with pip install -U xformers in cmd with venv activated.
Credit @supple raptor

so that i can get back to this message

[supple raptor]

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

[lyric zealot]

okay time for bed

[supple raptor]

night Smol Sus

https://cdn.discordapp.com/emojis/998672574345121855.gif?size=64&name=ChenWave&quality=lossless

[lyric zealot]

goodnight nubby

[supple raptor]

those are pretty good feet.... wait a minute

[obtuse sluice]

some cuties for y'all

[brittle jay]

1. I threw away the original packaging in a previous hotel and by now they probably threw the trash

2. My steamdeck is stuck in the beginning. You cannot install all the stuff it needs to function in the first time installation process thingy.

[brittle jay]

Unless you restart

[wispy canopy]

here, have some output from my newest model merge, and the cursed lyrics!

but in al seriousness, i think my new merge is a succes!

about fucking time!

welp, time to got to bed, 3 am

[dawn jasper]

Powerful images above, there.

[summer night]

do you guys generate with hires fix on all the time or do you only turn it on when you want to re-generate the same seed in high quality?

[native halo]

i generate a batch of like 50 to 100 images and only regenerate the good ones

[summer night]

i have a GTX 1070 so it takes like 20-30 seconds just to generate one 512px image at 30 steps with no loras or hires fix

i think with 15 steps of hires fix it can take like 1 minute

[supple raptor]

I use hires fix all the time. If I don't get a pretty good image out of a batch of 4 then my prompt needs work and I adjust accordingly before the next batch.

[summer night]

it took 1 minute and 4 seconds

to generate with these parameters

i can't generate at x2 because my gpu just runs out of memory

unless the image is 512x512

if i want the image any higher resolution after 1.5 through hires fix, i run it through the upscaler in extras

[native halo]

u can because i generate up to 2k ( 1k at 2x with high res fix) with a 1060

u prob dont have -xformers enabled or medvram

[summer night]

i have xformers but not medvram

it's an 8gb card

[native halo]

try with medvram

[summer night]

doesn't it slow down the generation

[native halo]

yea but u wont get oom if u go up to 2k

[summer night]

will generating at a base of 1k even work well for a model trained on 512px images

[native halo]

depends on the model some 1.5 models were finetuned to work on up to 1024 check the model page to see max resolution it supports

[summer night]

i'm just usin like random civitai models most of them don't list things like that

i guess the best way to find out is to try

[brittle jay]

All the time

[summer night]

i wish xyz plot had an option for hires "upscale by"

[native halo]

yea thats how it is with models u have to download a lot and test them out until u find the ones that u like

[brittle jay]

True and real

[summer night]

i like the model "meinahentai v4" just for regular SFW use, it has a nice style

[brittle jay]

Models that work in 768x768 or 1024x1024, except they are sdxl are extremely rare

Most of the good models we use are 512x512

[native halo]

we use or you use?

[summer night]

[REDACTED] has had a detrimental effect to the advancement of anime SD models, pretty much every "new" anime model or mix relies on it as a backbone

which is understandable because when it "released" all we had was waifu diffusion

[brittle jay]

Everyone in here

[native halo]

everyone? damn quite a big sample size

[brittle jay]

Yes

That just shows how rare it is

[native halo]

such a sad day indeed

[brittle jay]

I mean I am pretty sure you could finetune it with a different resolution but it might have to relearn a ton of stuff

Or something

[supple raptor]

I'd argue we wouldn't be anywhere near the quality level we currently are without that.

[native halo]

like this one for example it says in description it is best to use at high res but depends on the model https://civitai.com/models/12262

[brittle jay]

Man redacted literally changed my life

[supple raptor]

Most finetunes are done at 768x768 with bucketing these days btw

[brittle jay]

Wait what

I thought 512x512 with bucketing

[native halo]

well its good to see we still gettin 1.5 releases

[brittle jay]

How do people have enough vram

Jk

[supple raptor]

People that are finetuning actual models are using at least a 3090.

[brittle jay]

I mean aurora for example is 512x512 right?

[native halo]

heres another one that recomends a high resolution https://civitai.com/models/55306

[supple raptor]

People that have less than that and aren't willing to rent compute are training LoRA and merging those into the model for the most part.

[supple raptor]

Aurora isn't a finetune. It's a merge, but most of them were 768 models as far as I can tell.

[brittle jay]

I mean most people use merges

[summer night]

i wish there were more neutral models

[brittle jay]

Me be like

[supple raptor]

Yes, actual finetunes are rare.

[summer night]

like a model with no real style so style loras work better

[native halo]

you should do another survey about what models ppl use

[brittle jay]

Anylora?

Jk this thing doesn't exist

[brittle jay]

Best tip ever: copy what people over at projectai use

[supple raptor]

That's pretty much Aurora.

[summer night]

so i just put it there

and i should be able to generate in higher resolutions?

[native halo]

remove no half vae

[summer night]

i cannot for a reason that i cannot specify

[brittle jay]

Yes

[brittle jay]

Brug

[native halo]

oh ok well then try it now to see how high you can go in resolution until u get oom

[supple raptor]

You don't want the model to have no inherent style in my opinion because that leads to randomness that then messes with style LoRAs. Rather, you want it to have a predictable style that is also flexible through prompting. (things like graphite \(medium\)).

[summer night]

i just think it looks better than kl-f8-anime2

[brittle jay]

What looks

[summer night]

kl-f8-anime2 adds a redness that i do not like

[supple raptor]

try this version of it https://huggingface.co/nubby/kl-f8-anime2-blessed/blob/main/WD1-4-kl-f8-anime2-bless09.ckpt

[brittle jay]

Less saturation?

[supple raptor]

The contrast of kl-f8-anime2 is too high, I lowered it which fixed the problems that so far basically everyone I've talked to has had with the original.

[brittle jay]

Oh

Contrast

[summer shuttle]

img2img test

[summer night]

i was using this one for a while which i think combines the best aspects of both
https://civitai.com/models/22354?modelVersionId=88156

[brittle jay]

I think the creator of the blessing of vaes should make more options for it

[supple raptor]

that would be salt

[brittle jay]

Wait what

Fr?

[native halo]

you could try with blessed vae too or cute vae theres a lot on civitai

[brittle jay]

Salt is a genius fr

[summer night]

huh, i thought the only options were kl-f8-anime2 and [REDACTED]

[native halo]

no theres a lot of vaes out there

[summer night]

so if i stop using [REDACTED] then i can remove --no-half-vae which will speed up generation?

[native halo]

theres orangemix,cute,blessed,basil,cetus,counterfit,expline,grapefruit etc...

[supple raptor]

blessed is just NAI VAE with increased contrast. Still has all the same problems as NAI VAE just with better colors. Texture still sucks and still needs --no-half-vae. Cute i believe is the same but with brightness also adjusted.

[nova remnant]

like too many at this point

but I also say that about models...

[supple raptor]

all of these except cute and blessed are just NAI vae renamed

[native halo]

no it wont speed it up but some vaes give u NAN's so i just go with clearvae no nans here

[nova remnant]

wait I'm kindof a newb do vaes give texture too? is that why all of my gens have been weirdly scratchy??

not all but most hi res

[supple raptor]

for a while it was a trend to upload a renamed copy of a vae with your model because people didn't know how to select their VAE manually and placing a VAE with the same name as the model inside the model folder would make WebUI load that VAE

[native halo]

yea i remember that

now we have it even easier with models and their baked vaes

[supple raptor]

yes, vae is responsible for translating between pixels and latents (latents being 8x8 blocks of pixels which are represented by a single number) which means the VAE has heavy influence over texture.

[nova remnant]

alright I'll try another vae then

it's fine but with loras it tends to get wonky

so it's probably just the vae

[supple raptor]

Try the one I linked just a little ways up

[nova remnant]

blessed?

ok

[supple raptor]

https://huggingface.co/nubby/kl-f8-anime2-blessed/blob/main/WD1-4-kl-f8-anime2-bless09.ckpt this one

[native halo]

i tried blessed but gives me nan's with some loras

[nova remnant]

i use comfy, does that need a half vae arguement too

[summer night]

this mf is probably gonna take like 10 minutes

[supple raptor]

it has blessed because that is just the name of the script used to modify vaes, but is not the same as the Blessed VAE that everyone talks about. Blessed VAE is a modified version of NAI vae. Mine is a modified version of the WD1.4 vae, often called kl-f8-anime2

[native halo]

try with 540 x 960

[nova remnant]

ah i see

yeah i usually use kl-f8

[brittle jay]

That's such a good explanation, as expected from nubby-sensei

Or is it senpai?

[supple raptor]

you'll probably like mine better then. it fixes the color bleeding and high contrast that kl-f8-anime-2 has

[nova remnant]

nice

[summer night]

[supple raptor]

I'm quite a bit older than you

https://cdn.discordapp.com/emojis/755556110966063255.webp?size=64&name=oldfrog&quality=lossless

[native halo]

idk why u gettin that cause i get 11s/it

[summer night]

i still have no-half-vae on

if that could cause that big of a dent in s/it

[supple raptor]

yep

[summer night]

i'm just generating it with the vae i used for the control

after i'll try out that vae nubby sent me

[nova remnant]

also i wanted to ask about training an anime lora, i asked in finetuning but nothing so idk

[nova remnant]

holy s* u fixed it 💀

[summer night]

i generated a control image at like 512x680 with 1.5x upscale, and i'm generating this one at 680x1024 with 2x upscale

otherwise the same parameters and seed

[native halo]

did it came out deformed?

[summer night]

hold on i'll upload both

[native halo]

yea 2nd resolution is no good try with 540 x 960

or 768 x 1024

[summer night]

see you in another 6 minutes

i'm using this lora as well
https://civitai.com/models/123633

cul is tragic because she has such a nice voice but she never took off in popularity

[native halo]

😔

[nova remnant]

also (this is a lora at 1) does this look overtrained? hires, dont mind the bg artifacting idfk it does that

its on like abyss 3

i think its fine but its also my first lora

i wouldnt know

[summer night]

i like when lora authors go through the extra effort to make it balanced on a weight of 1 instead of being like "yeah just put it on like 0.7 or something idk"

[nova remnant]

this one is still recoknizable from 0.6-1.2 which is nice

[summer night]

the auto1111 feature to set a default weight and tags to load with a lora has come a bit late

since i've already forgotten where i got most of my loras

some i can find on civitai but some i also got from like MEGA dumps

[nova remnant]

another question im gonna throw out there - this is a style lora but it only works at like weight 2.0. does it just need more steps? its already on like 3600 i feel like im doing something wrong. adaptive optimizer, both learning rates are 1 idk if anyone could help with it

[summer night]

most of my artstyle loras are from mega dumps

[nova remnant]

like for dataset?

this was like 700~ images

[summer night]

no for use

[nova remnant]

ohhh

[summer night]

i kind of miss when dreambooth was the top dog for training for a specific style or character because it was so much easier to do

but loras allow more combinations and you can use them on any model

i feel like sometimes dreambooth models can still do specific characters better than loras though

[nova remnant]

most of my characters are ocs which is already hard

with like random cropped doodles or images

[summer night]

i wonder if you could take a bunch of screenshots of a character in koikatsu from different angles and in different poses, and if that would work good for a character lora

[supple raptor]

hard disagree on dreambooth. I do not miss 2+GB checkpoints for a single character or artstyle which were almost always worse quality than loras and required learning how to prompt a new model.

[supple raptor]

you could do it, but it would be difficult to not train the style as well if you didn't add illustrations as well or at least use regularization images.

[summer night]

maybe you could run the screenshots through img2img first? you think that could help?

[supple raptor]

controlnet would probably work better, but that is an option.

[summer night]

it would probably be better to filter the image like that to make it look like an illustration and not 3D

but don't you have to like use hundreds of images for loras?

[nova remnant]

not character

style yes

for character can be 20- however many

as long as it has variety

[summer night]

if there is a colab for training loras i might try it for one of my koikatsu characters

[nova remnant]

there is, well theres 2 but i find one easier

hollowstrawberry

theres also uhh

its just called kohya lora finetuner

[summer night]

is thelastben's colab still the best for stable diffusion on colab?

it says it's for colab pro now whatever that means

[nova remnant]

you need to pay for more vram

or you can pay as you use it

[supple raptor]

afaik training on colab is basically dead unless you get lucky

[summer night]

i looked at the runpod website but i have no idea how any of it works

[nova remnant]

why? im training both my loras on it

[summer night]

it looks like a gpu renting service and the prices seem decent, but i just have no idea how to work it

[nova remnant]

i could possibly do locally but i have a 2070 and its a bit limited

[summer night]

does runpod connect to your google drive like the colab does?

i got the $2 plan for google drive so i could have 100 GB to store all the stable diffusion colab stuff on

i've had it for a while but i might cancel it and delete all the stuff because the colab constantly seems to break each time i run it

[nova remnant]

break as in what

im curious

if mine has just been broken but i didnt realize

[summer night]

it just sometimes doesn't work

idk how to really explain it

but when it does i have to move all the model files out and delete everything else and have it reinstall it on my drive

[nova remnant]

ahhhh

weird

[summer night]

i might look into runpod

because 27 cents an hour doesn't sound bad for 16 GB of vram

but i don't really want to mess with it if it just deletes everything after i'm done with it

i don't want to have to bother putting all my settings back in

[nova remnant]

i think thats how cloud things work iirc but im not super techy, ex for colab i have to reinstall everything if its disconnected for too long. it keeps my config in a file though that it finds when i reinstall

[summer night]

when i used the colab it installed everything to my google drive

so i could just start it back up and it would have all my settings and extensions and stuff

[nova remnant]

"Our Cloud Sync feature allows seamless data transfer to any cloud storage." youd have to try

probably

[summer night]

the colab basically treats your google drive like the file system of a local install

[nova remnant]

I can try runpod and then report to you idk

I was gonna try something like it anyways

[summer night]

for some reason my SD installation can't allocate more than 6.50 GB of memory

i turned off hardware acceleration on like every program i have that uses it to try to squeeze more performance out of SD but it just won't allocate it

is the IP for the webui the same for everyone?

i know it's probably like a local server or something

[supple raptor]

yeah, the IP is just the localhost IP. It's reserved by default for communicating itself basically

[summer night]

could someone just make a chromium based app for the webui?

it would be nice to have it on my taskbar and not lost in my browser tabs

[supple raptor]

Yeah. SirVeggie actually did that already. Let me look for it.

[nova remnant]

same seed, different vae, left to right kl-f8, blessed, anything v3.0

tones are very noticeable

[native halo]

[summer night]

do you think she has enough lamps

also is she playing starfield on the TV

[supple raptor]

Here's what SirVeggie made https://github.com/SirVeggie/sd-browser
it's an image browser that also has a fly out window that can be used for SD.

[summer night]

i seeee

[nova remnant]

oh yeah, i think high tech interior design must include lots and lots of lamps :p

[summer night]

it looks like it has mobile support so if i do the share thing in the commandline args i could use it from my phone

[brittle jay]

SirVeggie developer?

[supple raptor]

yesh

[brittle jay]

Damn

[summer night]

do i uhh
have to worry about all the "gradio depreciation warnings" in the stable diffusion console
or is that something for automatic1111 to worry about

[nova remnant]

People

[brittle jay]

Yeah but it's only an image browser

[nova remnant]

How do you do that steps comparison

[brittle jay]

Btw can you make chromium based apps on phone like on PC?

[supple raptor]

I use his Windocker app every day. It's very nice having VoiceMeeter hidden so I can just flick my mouse over the side where it's docked to manipulate it and have it autohide when I'm done.

[native halo]

xyz plot

[supple raptor]

https://github.com/SirVeggie/sd-browser#flyout-sidebar

[brittle jay]

Oh damn

[supple raptor]

Has an iframe for loading the webui

[brittle jay]

That's very cool

[supple raptor]

yep, I've been meaning to try it out, just haven't gotten around to it because I haven't spent that much time on SD lately.

[summer night]

@supple raptor is this for me to worry about? or is it for automatic1111 to worry about

[supple raptor]

BG3 more important.

[nova remnant]

can you take a screenshot of it on a simple prompt, so I can see the settings and stuff

[summer night]

me but with this

[brittle jay]

Omg veggie was youtuber

[supple raptor]

looks like you may have some outdated extensions there. I see the images-browser, artists-to-study, and system-info extensions at a quick look.

[supple raptor]

so is Nerf

[brittle jay]

Btuh

I am so out of touch or something

[native halo]

like this https://www.youtube.com/watch?v=y7rDdamvuAM

[brittle jay]

Just noticed he didn't talk for the past like 2 months

Do you have his YouTube channel?

Actually I think I saw it once before

[summer night]

ok i have both of these
i only know how to use the 2nd one

but the 2nd one is apparently depreciated

the first one looks complicated and scary

[supple raptor]

yeah, nobody really uses latent couple any more. for regional prompter you'll want to read the documentation in the github because there's a lot there.

[summer night]

i'll try to figure out how to use it

i know latent couple still works for now because i just used it yesterday and it worked, so i'll keep it around just in case

[nova remnant]

runpod is working fine so far, i think its easier to use colab and all files are stored in a "pod" for next time

if you were still curious

[summer night]

i might look into it later

[nova remnant]

I think its working

o:

[summer night]

i'm a bit wary because it looks like i have to give it a minimum of $10 to start using it

[nova remnant]

it takes a while but that means I can do those too :D

How do you make the sampler appear

I used euler a I thought it would type euler a on the side