#💬｜general-chat

1 messages · Page 32 of 1

vestal dew
#

no, that strength helps avoid conflicts

#

what do you think, in their private offices, the Chinese are saying as their balloon floats unmolested over America right now?

astral goblet
#

i dont think it went thhat way. people probably suffered his rule

vestal dew
#

the balloon they claim is a lost weather balloon that was sent up by a student?

astral goblet
#

you might be dippin to much q anon

vestal dew
#

you claim there is no balloon?

#

it's in the news right now

astral goblet
#

think the balloon is out to reveal the flat earth once and for all?

#

i think its a slow news cycle

vestal dew
#

a strong President would help the chinese by capturing the balloon, testing it to see why it went off course

#

our President lets it float on its merry way while proclaiming there is no inflation: eggs from 89 cents a dozen to now 4 dollars a dozen

#

Vlad's direct action, meeting power with greater power, was a good part of him

#

his tyranny was a bad part of him

astral goblet
#

they probably don't have to capture it to determine anything and its fine

#

electronic countermeasures do exist afterall

vestal dew
#

they haven't intercepted it, examined it, deployed countermeasures,

astral goblet
#

i didn't even hear about the balloon tbh. i just don't watch the news. propaganda addiction is nutso

vestal dew
#

they've done nothing

astral goblet
#

follow real information feeds

vestal dew
#

I see

#

you just wait to be told when to wear a mask muzzle, when to take a vaccine, when to leave your home, when to stay inside

astral goblet
#

think the news will tell you when the spacex starship test flight is? or are they just going to tell you afterwards

#

they sure aren't covering the cool ass comet in the skies right now

vestal dew
#

nobody told anyone about the balloon until after it was here

astral goblet
#

real shit doesn't matter in propaganda land. it's a delusional addiction

vestal dew
#

it's reporting on what's happening now

#

the balloon is real

astral goblet
#

the baloon is stupid

vestal dew
#

you say you heard nothing about it, but claim to know it's perfectly innocent, safe,

astral goblet
#

took me a min to figure it out

vestal dew
#

where do you live?

astral goblet
#

not in america. indoctrination doesn't catch me

vestal dew
#

I see

#

It's hard to discuss these things with non Americans

astral goblet
#

because of no indoctrination

vestal dew
#

ignorance abounds outside the US of GOD's A

astral goblet
#

i'll give you a hint

#

my country uses the metric system

vestal dew
#

when you need help, however, Americans will send the most

astral goblet
#

Canada has never been succesfully invaded. not even by americans

vestal dew
#

Um

astral goblet
#

i think what would happen is if America tried to invade canada, they'd have a war on all fronts

vestal dew
#

there's a balloon in Canada too

astral goblet
#

so?

vestal dew
#

likely launched from a canadian location

astral goblet
#

thats how high altitude works

vestal dew
#

America wouldn't invade Canada

astral goblet
#

no shit. it'd be dumb

vestal dew
#

If America did, Canada would fall in the blink of an eye

astral goblet
#

we'd just go burn the white house down, again

vestal dew
#

you might now, with what is being done to the US military

astral goblet
#

only takes about 6 crazy canucks going commando. we practically invented it

shrewd elk
#

No point. JD has already destroyed it.

vestal dew
#

An American redneck would singlehandedly wipe out all the Canucks you have

shrewd elk
#

JT*

vestal dew
#

JT is much like JB, just younger

astral goblet
#

release a flock of sheep an the redneck would be busy for days to care

vestal dew
#

they'd kill and skin those sheep in no time

astral goblet
#

yeah thats what theyd do lol

vestal dew
#

anyway, SD can't make action images

#

can't make good hands or feet reliably

astral goblet
#

more to your inquiry before you spiraled into conspiracy, no. stabliity probably doesn't train the new models on nazi worship imagery

#

you will be censored

vestal dew
#

and if you rely on it for inspiration, your inspiration can be taken away

astral goblet
#

i dont think you have a concept for how open source software works

vestal dew
#

Once again, opposing nazis isn't worshipping them

shrewd elk
#

Lol Flow why do you keep pushing the toxic chat? Just let it go

vestal dew
#

anti nazi isn't woshipping them

#

Flow is adding the toxin

astral goblet
#

but your imagery needs nazi worship to create a raised by hitler nazi superman, right? so your idea is being censored. you are right

vestal dew
#

no

#

it's anti nazi

astral goblet
#

even though you intend good, since nazi imagery is avoided, thats censorship against you

#

😦

vestal dew
#

anti the lefist ideals held by the nazis. Anti Hitler. Anti tyranny

fervent thunder
#

The fuck did i just join

vestal dew
#

Your claim is that a movie about hitler showing how evil he is would be nazi worship?

#

Imagry should be banned?

astral goblet
#

well, stability the company probably doesn't want that in their revolutionary media release

vestal dew
#

can't have a nazi soldier, that's nazi worship

astral goblet
#

so i geuss you're being censored. you aren't wrong

#

your idea has been stamped out. sucks

vestal dew
#

they don't want anti nazi?

astral goblet
#

no they just didn't use nazi imagery most likely. total censorship

vestal dew
#

I see

fervent thunder
#

Personally i do not believe ideas should be stamped outright because of the danger of them.
It sounds counter intuitive but supressing such things causes things to fester and become locked in place for a person, thus potentially creating a extremist.

vestal dew
#

that's the point

astral goblet
#

their intentions are mostly that nazis shouldn't be part of their revolutionary media release though, reather than personally stamping out your well intended story about hitler's son

vestal dew
#

if you censor talk of evil, people can't be shown evil, prepared for it, taught to fight it

astral goblet
#

i welcome people to reveal how deluded and hitler worshiping they are, but i can also completely understand how a company miight not want that under their banner

vestal dew
#

Remember when there was a nationwide manhunt for some teenagers who escaped from a "voluntary" quarantine camp in Australia?

#

how do you escape from something that's voluntary>?

#

Do you recall somewhere else where people were forced into camps?

astral goblet
#

sucks but what does that have to do with a hero story about hitler's youth?

fervent thunder
#

By the way i dont know what this conversation was originally about

#

i literally just joined lol

astral goblet
#

how a prompt about super nazi soldiers won't work but it's bad because the prompt is for a good reason

vestal dew
#

people who didn't know the history of Nazi Germany wouldn't recognize what was going on with those "voluntary" camps

shrewd elk
#

My original question has been lost now 😅

astral goblet
#

i don't think pandemic quaruntine is anywhere near what the jewish people suffered under the nazi regime. it's weird to compare that

vestal dew
#

I want a guy in a black and red Superman suit with an eagle on the chest

astral goblet
#

the nationwide manhunt was probably 6 volunteers and a constable on phones

vestal dew
#

do you know the history of nazi germany?

#

do you know it didn't start with concentration camps?

astral goblet
#

it was all canadian propaganda teaching it. history goes to the victors you know

#

it was all a sham right?

#

(cliche)

vestal dew
#

so you don't know

empty barn
#

Oh man. Just stop it.

fervent thunder
#

My main fear in the modern age is we seem to have simply said with nazis and extremist ideologies "they are bad because they committed genocide" yet wish to never look into why they came to those conclusions, leading towards false ends what actually just potentially worryingly push us down similar pipelines, whilst thinking we are fighting against it.

vestal dew
#

You seem to be quite an expert on many things you don't know

fervent thunder
#

I am trying to break this debate up by throwing out interesting tid bits of stuff. anyways hows everyone doing

astral goblet
#

lol "it starts smaller" is just a deluded way to compare everything to "i have it as bad as the jewish people!" and it's always so q anon and weird

vestal dew
#

Before Jews were sent to camps, and many others as well, they were forced first to identify themselves

astral goblet
#

it's so cliche

shrewd elk
#

I spoke to an old guy who was in the camps. Said the lockdowns had the same origins as what preluded the Nazi camps. So there’s that..

vestal dew
#

then there were strict curfews - Jews could only be out of their homes at certain times

astral goblet
#

same conversation is basically boiler plate

vestal dew
#

then their businesses were taken from them

fervent thunder
astral goblet
vestal dew
#

In America, some people in congress suggested forced quarantine camps for people who would not take the vaccine

fervent thunder
vestal dew
#

others suggested prison, even the death penalty

astral goblet
#

you probably could get an image you want still, but he's stuck on using hitler in his story and prompts

vestal dew
#

I don't even want a nazi soldier. I want a guy in a black and red superman suit with an eagle on the chest

astral goblet
#

i told him he probably shouldn't lead with hitler and instead use warhammer 40k

shrewd elk
#

Flows very left. For censoring/banning what left considers offensive etc and Frank doesn’t know why SD won’t produce a picture of a nazi in the context of being anti-nazi for him 😂 haha

vestal dew
#

he keeps projecting his love of nazis onto me - after again and again I've mentioned my images are anti nazi

astral goblet
#

its weird though. he lead with the "nazi super soldier bit" and as i got more disgusted, he got more righteous and good with it

vestal dew
#

No AI guy

#

can you read"

fervent thunder
vestal dew
#

I want a guy in a black and red superman suit with an eagle on the chest

shrewd elk
#

Oh sorry, a superhero to take on the nazis?

vestal dew
#

I can get a nazi style uniform, add the swastika

fervent thunder
#

Why do you want that if i may ask?

vestal dew
#

Ubermensch - raised by Hitler, but turns against the Nazis. That's the story

astral goblet
#

oh yeah. he'd still have the swastika but it represents good again, but as the nazi swasitka this time

#

of course

vestal dew
#

No

#

there is no swastika on Ubermensch

shrewd elk
#

Passive aggression is always strong with the left. As you can see by Flow

vestal dew
#

there is a golden eagle, which is the symbol of his family

astral goblet
#

lol sure

#

sure sure

vestal dew
#

exactly

astral goblet
#

i'm super left lol

vestal dew
#

flow is projecting his beliefs onto me

fervent thunder
#

I don't really mind nor care for that but it would very much worry me if nazi symbology was not allowed but soviet was.

astral goblet
#

i'm lololol i'm one of the most conservative voters i know, but okay, im a lefty lol

vestal dew
#

that's why I continue the discussion. Evil needs to be rebuked hard

astral goblet
#

omg you guys

#

omg omg

#

i'm dying

#

these methed out supremacists are nutzo

shrewd elk
#

Why are you so for not allowing Frank’s idea then?

astral goblet
#

oh yeah. totally. i'll allow the nazi super soldier idea. its such a good idea lololol

#

go for it

fervent thunder
#

If I may be honest, it sounds like at least two people here need to take 5 minuets away from the computer to chill out and perhaps stop thinking its a personal attack on them.

#

no offence meant of course

empty barn
#

I agree.

fervent thunder
#

Calm down because at the end of the day, you are both just going to entrench your views and become furthered in them.

#

that is not targeted at one side

#

Anyways, hi im new here. hello im here to do some ai funny haha art

vestal dew
#

The idea is that Ubermensch's father was a tyrant who conquered and destroyed worlds. He and his race were trapped on a planet where they had none of their powers, so they could not escape. U's dad found a way to secretly create a ship to carry him off the planet, but it was discovered. He placed his son on the ship as the planet was destroyed, sent him to Earth to conquer it, restart the empire. On Earth, Ubermensch was raised by Hitler to lead his armies to conquer the Earth. But at his core, Ubermensch was good rather than evil. He wouldn't become ruthless as he was raised to be. He eventually kills Hitler, ends the Nazi regime, restores democracy to Germany.

#

I am entrenched in good, against evil

fervent thunder
#

Ok sure and cool but dude take 5 minuets as im not trying to be an ass, im saying because we all been through it

#

Things get heated, its best to cool off

vestal dew
#

No

#

Evil must be rebuked, hard

#

look up what FRANK means

#

my name was well chosen

fervent thunder
#

Ok but you realise where we are right now. We are in a discord server for AI generation

vestal dew
#

that's how evil grows

#

by infesting small corners of society, so that it seems to be the norm everywhere you go

fervent thunder
#

Sure and i do agree subversion is there but by the fact i am able to talk right now with you about such things, perhaps shows you not everyone is that

vestal dew
#

why, there're even saying it on this discord I visit, so it must be true, it must be right

#

I didn't claim you are

#

flowolf is

#

read above. No matter how many times I mentioned anti - nazi, anti-HITLER, AN the F&CK TI, he claimed I woship nazis and hitler

#

he was desperate to get that as the last word, so people would say "OMG, that franktutor guy woships nazis and hitler"

frigid trail
vestal dew
#

Someone earlier mentioned image view sending to img2img wasn't working

#

it's not on mine now

#

the image goes over, but the prom;pt and settings dont

#

reloaded the web page...then it worked

astral goblet
#

aahaahaha now the evil guy the super nazi is coming after omgomgomg hahaaaaaaaaaahahaha

fervent thunder
#

Ok so im here and dont know how to fucking start doing the prompts. do i dm a bot or is their a channel?

astral goblet
#

guys i'm dying

vestal dew
#

SD gets pretty close to Nazi imagry. Just posted one in general with images

vestal dew
fervent thunder
#

where though?

vestal dew
#

download images. If using A1111, open in the PNG tab. It shows all the settings and the prompt

#

you don't have SD on your computer?

fervent thunder
#

ah my most humblest apologies i do be dumb

astral goblet
#

there is no bot no mo

vestal dew
wise arrow
#

good afternoon every body!

brave matrix
#

hello!

#

I am wondering if it is possible to do a batch inpaint job in auto?

warm tiger
#

ACK!

#

you cant train an embedding with --lowvram enabled QQ

#

and instant crash from lack of memory trying run on --medvram and train the embedding XD

placid hatch
#

Not sure if this is the right channel for this, but how safe is the distyapps picklescanner to use?

spark sigil
#

Anyone suggest some best stable diffusion models

brave matrix
broken spruce
#

when launching the sd webui i get this message in the console : "No module 'xformers'. Proceeding without it.".
where can i find this module ?

warm junco
broken spruce
#

okok

wraith sky
#

xformers doesn't work on Mac m1 does it? Needs cuda ?

left knot
#

anyone here a Lora expert? just need a quick consultation as to why my image looks great in previews, but on last step it takes a 180 and decides to look like shit.

#

Restore faces off and tried with VAE (86000) + no vae.

elder cradle
#

Now my "New python version" Kaggle literally stopped working

unreal cargo
hearty fox
#

this is a dumb question but in the models > stable-diffusion folder, can the models be called anything? model.cpkt, model.safetensor can be model1.cpkt, model2.safetensor?

#

if you want to have multiple in the same directory

warm junco
#

you can call them whatever you like only .ckpt or .safetensor has to be the filename

fervent thunder
#

I got an old install of sd which runs via cli like : ```
activate ldm
python optimized.py --prompt "123"

and now i tried the webui and some newer models in it and they work but if i try to plug in the models into the old install they dont work?

```magic_number = pickle_module.load(f, **pickle_load_args)
_pickle.UnpicklingError: invalid load key, '\xdf'.```
hearty fox
warm junco
#

But you can keep it for the "Add Difference" Merge, then you can strike out all the 1.5 images, if the SD1.5 is at C (Only needed when you know what you do)

hearty fox
#

@warm junco if the SD1.5 is at C (Only needed when you know what you do)

Thanks. What does C refer to.

warm junco
hearty fox
#

Ah, you mean basically the tertiary model to merge with

fervent thunder
#

AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA

hearty fox
#

_>

light folio
peak orchid
sonic bridge
#

anyone seen the never ending seinfeld show on twitch

brazen fox
#

@thorny depot PES_AngeryPolice

hearty fox
#

Hey does anyone konw why when I upscale it makes like... 8 different images into one weird collage image?

#

I only have batch size 1

thorny depot
#

Got him. Thanks for tagging me in.

neat condor
#

guys anyone uploads models here on ko-fi for supporters? need to ask a few questions if you have time

jade atlas
#

مرحبا انا جديد واريد اعرف كيف طريقه العمل وانشاء فيديوهات بهذا الموقع

dusky basin
#

i see this protogen model a lot lately

#

is it better than sd 2.1?

#

what is it good for

neat condor
dusky basin
#

would you say its better than 2.1

neat condor
#

but with loras and TI's vanilla can do anything - any other model can.

dusky basin
#

but like are they better at stuff in general

neat condor
dusky basin
grizzled rock
#

does anyone remember carson? the actual owner of this community?

neat condor
dusky basin
#

or those you mentioned

neat condor
#

protogen

dusky basin
#

aight

#

is it good at ai art and stuff?

#

i see its based on 1.5

neat condor
#

man i think you need to be more specific about what you want 😄

#

if you want anime for example - it won't do. if you want realism - it's better than 1.5 but not the best and so on.

dusky basin
#

i dont want anything specifically i just wanna know

dusky basin
warm junco
neat condor
#

you need to know exactly what you want then find according model which does that the best.

astral goblet
#

You can ask it about an image. "Green like what?"

dusky basin
#

gate guardian

neat condor
#

can you show me the image what u consider to be good art?

#

i'll find it easier to give advice like that

#

drop it in general with images

astral goblet
#

Vanilla 2.x is a lot better at many of the merges at zero shot images. Where the training data has never seen that concept, and is able to put it together still. Merges of 1.x models tend to get very focused on one particular style and are great at churning that out. A noticeable example is how most merges can only replicate a handful of faces. A single seed with multiple characters will often all share the same face. This is because of over training which is good for the "aesthetic" goals, but not great for generalized work.

There is a lot of hate for 2x, but it's a base model that hasn't caught a lot of momentum like 1x has. 1.5 went through a few iterations before it got to where it is. 2.1 is rough around the edges but is very versatile. I'll also go ahead and say that using 2.x with embeds IS using vanilla 2x. Embeds are just very emphasized prompting.

#

I've been running "inspiration generation" script from the inspiration extension, on the 2.1 768 model. I've got nearly 8 gigabytes of quality art across a range of styles. Artists name's work, just not "rutkowski", which barely worked like it should've before it was removed. Most Rutkowski images coming out of 1.5 don't look like his work. The clip for his name just happened to tie to lots of concept art under the old model

warm tiger
#

yeah... still trying to work out how to get multiple characters on the same picture without them being the same. though if you use 2 embeddings of different characters it blends them together into 1 for some pretty comedit results

#

comedic*

astral goblet
#

generally speaking, SD is really poor at getting two different characters interacting. I think that comes from the limited clip models, the pretraining

warm tiger
#

so far my only solution is to run the thing to generate 2 characters and th ebackground you want then use inpainting and change the prompts to change each individual character after the fact

astral goblet
warm tiger
#

the what?

#

i didn't hear about this... i'll need to look into it

astral goblet
#

i thought that was the thing you meant

warm tiger
#

no. using inpainting and the black marker

#

i black otu one character then select it to change only the marked spot

#

edit the prompts to say what i want, generate a few options and picj the best one

#

then take that result and do the same for the next character

astral goblet
#

OHh yeah the manual way. That's a good thing too. The extension is pretty cool. i intend to play with it more.

#

gotta go. have a great day

warm tiger
#

i'll need to mess with that as well now. You too!

#

I'll go make tea while this batch of 100 finishes generating XD

glacial frigate
slim relic
#

what's a good software/app for upscaling my images?

#

it looks quite blurry if I zoom in

dusky basin
#

2.1 doesn't wanna work for me for some reason

upbeat birch
#

Article: "CEO of Stability AI, an OpenAI rival, reportedly told employees they were 'all going to die in 2023' as competition heats up"

#

@wise stratus ?

maiden crystal
#

How do i change some small detail of a image? Without effecting the rest of it?

neat condor
maiden crystal
neat condor
maiden crystal
#

No, I'm just using the demo. I'm not sure how to install . And doubt my 4GB Graphics card can use it.

#

Using online demo.

neat condor
#

it will probly choke on it

#

not impossible to run but

#

you'll just struggle

#

okay so inpainting is a function in A1111

warm junco
neat condor
#

check out tutorial about how to use inpainting with A1111

maiden crystal
#

Ive tried. I installed Stable D, using a bat file from github, worked a little bit. But nowhere near what Im getting online. sadly.

#

The auto bat install i used, was the only one that worked for me

#

But its low quality

#

because of my system/

neat condor
# warm junco It will be fine with 4gb

i have 8gb 1070 and it generates 512x512 in 11 seconds, on 4gb that probly be x3 times longer, at that point i wouldn't advice anyone to experience that struggle.

warm junco
maiden crystal
#

but that guy dont ever make good tutorials for beginners.

neat condor
warm junco
warm junco
neat condor
maiden crystal
#

i mean, i had some issues with the bat install version, and that installed everything even the hugging face libraries. its just i uninstalled it because it was not working up to a decent standard on my low memory pc

warm junco
#

Which install bat did you used ?

neat condor
#

as far as i know/remember only mistake in aiterpreneurs tutorial was python version, u need 3.10.6 other than that if you follow it you should have no problem

maiden crystal
warm junco
maiden crystal
#

i remember i had to be signed into hugging face, and confirm some stuff there.

#

and the bat would connect and do all that

warm junco
neat condor
#

after that i install it from memory.

#

but i remember i struggled for whole day only to find out he just provided wrong info.

warm junco
#

It works fine with any 3.10.X Version

neat condor
#

i remember it didn't work with 3.10.7 (the one he installs in the video)

warm junco
#

Okay i startet with 3.10.8

#

It worked

#

Then i tought hey i can upgrade to 3.10.9 xD

#

Still works

neat condor
#

i see, so 3.10.7 might be a problem or they just improved A1111 which is highly likely.

warm junco
#

Both can be possible

maiden crystal
#

ok i installed python

#

again. this is a new ssd drive so

#

i had to install it again

warm junco
#

Did you checked the "add python to path" option while installing?

maiden crystal
#

ah ill install again

#

okay so that guys video? thats it?

warm junco
#

Yes

#

@maiden crystal Reproduce every step and when your done or have Problems you can ask us.
At the end of the Video he edit the webui-user.bat to add --xformers
You have to use --xformers --medvram
For your 4gb card

warm tiger
#

blending anime style lartwork with RL photos gives..... interesting results

#

makes it look almost CGI

warm junco
#

Yea gives great results

maiden crystal
#

i think that address he says to add to cmd, but getting a error

#

that its not recognized

#

maybe a restart after installing git and python ?

warm junco
#

when you are in the git bash

maiden crystal
#

he said, make a folder, go to folder, cmd at top of address, then enter that

warm junco
#

ah okay

#

yea if it dont work then make a right click in the folder, select "git bash here" and then paste the command then hit enter

maiden crystal
#

i cant copy paste into that window?

warm junco
#

right click is paste

#

in the window

maiden crystal
#

it says that

#

it never works like in the videos.

warm junco
#

which command did you paste in ?

maiden crystal
#

that

warm junco
maiden crystal
#

oh i gotta add git clone at the start

warm junco
#

its also shown at 3:36 in his video

maiden crystal
#

$ git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
Cloning into 'stable-diffusion-webui'...
remote: Enumerating objects: 16043, done.
remote: Counting objects: 100% (65/65), done.
remote: Compressing objects: 100% (41/41), done.
remote: Total 16043 (delta 33), reused 39 (delta 24), pack-reused 15978
Receiving objects: 100% (16043/16043), 27.25 MiB | 46.04 MiB/s, done.
Resolving deltas: 100% (11253/11253), done.

warm junco
#

if its done you can close the window

#

then continue the video

maiden crystal
#

he says the newer models are not as good as the old, for ease of use?

warm junco
#

he uses 1.4 official and 2.1 official and shows them both because newer models need the yaml files and older ones not

#

you can use any model you want

#

the new models are fine, there was just a little controversy with 2.0 model but now there is 2.1

neat rapids
#

Does anyone have an idea on how to produce that kind of "warping transitions"? it seems he even holds for 2 or 3 seconds 2 or 3 different images and then "warps" again to the next seed.

maiden crystal
warm junco
#

in 2.1 these issues got fixed mostly

maiden crystal
#

not sure where they are pulling from artists by name.

#

sounds tough.

warm junco
#

it was just tags from images that include artist names, so they removed them

#

is your install working?

unreal cargo
#

Or more so, they just used a different CLIP and not the one from OpenAI(?)

#

not sure what the names are alicatKEK

#

openCLIP (LAION used in 2.0) vs alicatHm2 CLIP (OpenAI used in 1.5)? mmm

#

with one having the artist names and the other having captions that better match the images

maiden crystal
#

Now i want the protogen model he said is better than all of them, but that is another video? still not sure how to install the model yet

#

he still talkin about it

warm junco
#

then search for models

maiden crystal
#

i need a video

warm junco
#

at 6:33 he goes forward how to download models and use them

#

you should watch the video twice 😄

#

its all explained

sonic tundra
#

check ur dm's @warm junco

maiden crystal
#

he just says download it.

warm junco
#

He linked it under his Video and shows it

hexed bramble
#

Hi guys, are there any checkpoints other than 512-base-ema and 768-v-ema?

maiden crystal
#

got some errors, and it wont install.

#

maybe his video is outdate . python version he uses is older. i installed latest.

mellow canopy
#

Stable diffusion distilled when?

old marten
steady arrow
#

stabilityai has been so quiet lately

#

whats it doing

hexed bramble
orchid wadi
steady arrow
#

we invented human-level text to speech ai in the time they were on vacation 😄

maiden crystal
#

whys the loading take so long? is it even going to produce a image

old marten
#

The TTS cloner I use is TorToiSe btw. And it works like 🤌

steady arrow
orchid wadi
#

And considering that training the model takes about 30 days, 200.000 machine hours so...

old marten
#

It's sexier when it's nonconsensual.

orchid wadi
#

2.1 came out like 2 months ago.

#

Assuming they use the same cluster setup thats like 720hours per GPU. So 32 Days.

#

We know that they developed new datasets and worked on dealing with things like SEO spam from Amazon/Wish/Alibaba...

#

SO if they did that during december, then put the machine to work during vacation. I'd give that mid- to late this month we get something.

crude comet
#

Hey- does anyone else have problems downloading checkpoints from civitai?

orchid wadi
#

Go to civitai. First thing on the list "Hyperbreasts". AI art community loves to make it hard to defend in polite company.

crude comet
#

what?

old marten
#

Bro, only chicks use stable diffusion to generate anything other than big boobies and hot chicks. That's my hot take

#

It's also proven by the popularity of NSFW models

orchid wadi
#

You have no idea how hard it was to come up with presentation about AI image generation and trying to present variety of models....

#

At an university

empty barn
#

lol

orchid wadi
#

When 9/10 for everything was big tits, mecha, anime and loli.

#

I did embedding examples myself because jesus wept I thought that was the safer alternative

#

Because all this shit is of public record for 10 years so I chose to keep the contents clear of... potentially controversial things.

#

As in the presentations and reports we had in. Anyone can ask them for the next 10 years as part of my academic record.

#

Considering I am a mechanical engineer so this would stand out to begin with.

#

Like I am about as sex positive as they come. I made a god damn Tom of Finland embedding.

#

But fuck me I can't with straight face go do a presentation about "Big titty goth cat girl waifu model". I'm sure some of my course mates would been... Considering the kind of people attending. BUt I just wanted easy points towards my degree on AI course which had no coding at all.

#

Just writing about theory and ethics and shit.

#

Funnily enough that was also the place I first heard about "This Dick Pick Doesn't Exist" which is exactly what you think it is.

#

Not images of people named Richard you filthy sods.

empty barn
clever saddle
#

We're do I prompt for an image?

cold siren
#

not here

clever saddle
#

So where?

fervent thunder
cold siren
#

there used to be a bunch of dreambot channels but now they're offline 🤷

clever saddle
thorn elk
#

hello everyone. I am Monen, I am new to this Stable Diffusion. I love this and hoping I learn more from everyone. Nice to meet everyone😵‍💫

brazen fox
light folio
#

this is me avoiding nsfw models

brazen fox
#

(after the gif cuts off I mean)

orchid wadi
#

But talking of like NSFW models. What I enjoy is just going through them. I have seen models for kinks and fetishes I didn't even knew were a thing.

#

Haven't dared to test many of them. I tried few and...

#

Well...

#

In engineering we have this concept of "Name the things exactly what it is".

#

And it seems like that has gained popularity.

cloud granite
orchid wadi
#

It rescales the noise it generates.

#

First the image is noised, then noise is scaled.

#

And then it starts the normal latent iteration on to the noise.

#

The "Denoising amount" is basically "how much noise is added"

#

Very much like if you did it on Photoshop with "add noise" filter.

#

Except that it is multi dimensional but lets not get to that at 2am.

elder cradle
#

Aw man, if the anti-ai community has the "SD is collage" misconception, then the SD community has the "NSFW images in training help improving anatomy"

old marten
cloud granite
old marten
#

I have a feeling you'll eventually end up blocking because I can't help but to take the piss out of you sabrina lmao

#

It's hard to resist

cloud granite
#

Done

old marten
#

Noooooo

#

🤪

orchid wadi
#

These are bookstore stuff... Some of these are still sold and reprinted regularly. YOu don't need porn to teach anatomy.

#

However I doubt that google (which LAION scrapes) can find any fucking good anatomy that ain't porn.

#

But 2.0 had a big issue with and 2.1 still to some degree of "Surprise jeans" and it is ALWAYS blue levi's jeans when trying to get more naked forms. Well not "naked" but like swimwear and underwear.

#

Althought 2.1 can prompt naked. Just... Ken Doll naked.

#

Which IMO is just fine.

old marten
orchid wadi
#

civit had few. They had a diaper model also. I think that is for 1.5 tho.

#

Then there was this blog post.

#

Lemme see if I can find it

old marten
#

immediately checks to see if they have my fetish

orchid wadi
#

Oh and then there is the rentry post /sdmodels. Which I wont link directly because... Well...

#

It has stuff.

#

It is really outdated tho.

slate crescent
#

I'm being really impressed with instruct-pix2pix but am not sure I understand the magic

orchid wadi
#

Basically you are engaging with a AI that understand phrases. Like... customer service bot or OpenAI GPT chatbot.

#

Then it also understand what is on the image by interrogation.

#

So if you have sunny picture and ask it to turn it to rainy. It knows that the picture is sunny, and it should be rainy.

#

However you deal with limitations of both the text AI model, interrogation system and the image model. But as long as you keep it simple it seems to work REALLY fucking well.

#

But it is even stricter about "Good things in - good things out".

fervent thunder
#

des français ici?

orchid wadi
#

?

fervent thunder
#

sorry, i go to french channel

slate crescent
orchid wadi
#

Well... It isn't and it is. The text encoder is GPT3.

#

But all it is used for is to do tokenising.

slate crescent
#

and this ignorance causes me to have questions like, "is this specific pix2pix model based on a specific SD base model? (and resolution?)"

orchid wadi
#

It doesn't understand "Young man wearing a hoodie" in other way that it fetches things from its training that had those tokens in them.

orchid wadi
#

"young man wearing a hoodie" appears to the SD ai as: [35465, 582, 5762, 257, 14263, 494]

#

And the task of the ai is to basically do this math: prompt tokens - interrogate tokens (from latent space) ≈ 0

slate crescent
#

right, I generally appreciate how tokens and parsing work, and it's intuitive to me why standard img2img can't take in "put snow on the ground" or "make it night"--those are (more) full concepts, beyond mere literal tokens, they map to lots of associated tokens that are contextual with the scene provided

orchid wadi
#

There is no AI components that understands what that means.

slate crescent
#

I don't understand how pix2pix models enable it to go from those concepts to a token manifest it can work with

#

right, I'm attributing too much abstract conceptualization here

orchid wadi
#

All it does is to find things from the model that werein the training set with most meaningful tokens of that as in "Snow on the ground".

slate crescent
#

right, but I'm failing to understand the distinction between standard img2img generation and instruct-pix2pix in this regard

orchid wadi
#

This is the average idea it pulls up. It doesn't understand anything beyond that.

lament warren
#

hi hi

orchid wadi
#

This is text: "banana donald ass pinkey purple potato".

#

That ain't language.

slate crescent
#

right, it is the language component I'm failing to understand, or rather it's integration into the typical model pipeline--it's somehow been "baked in"

orchid wadi
#

Pix2pix is basically if you'd take Gobold, use NovelAI or AIdungeon to do text with.

slate crescent
#

right, I understand that

orchid wadi
#

Now... It quite literally is an additional AI.

#

It is an AI that understands language

#

That then interacts with SD.

#

Think it as... Having GPT3 Chatbot that writes the correct prompts for your img2img so you get what you want.

slate crescent
#

maybe I'm misunderstanding the nature of how pix2pix was integrated with Auto1111

orchid wadi
#

It has nothing to do with auto.

slate crescent
orchid wadi
#

The system itself is from like 2017

#

That is how it works

#

IT is all there.

slate crescent
#

if that info had been on the huggingface card or academic publication, I probably would never have asked about it

#

the trick I was looking for all this time was, they trained the model on triplets: prompt/tokens before, edit instructions, and prompt/tokens after

nocturne fox
#

Anyone know when the bots will be back up?

orchid wadi
#

If you read the paper it is citation 27.

#

Isola et al has 3 mentions 😄

slate crescent
#

right, but I was never asking how Isola et als stuff worked, but rather how my machine, in automatic1111, is capable of executing it locally and on what dataset(s)

orchid wadi
#

et al just means "and others".

slate crescent
#

....yes, that's why I used it just now

orchid wadi
#

Since Phillip Iso is not the only person who made that.

slate crescent
#

........right

orchid wadi
#

Well... It raelly is just additional AI.

#

I don't know what else to say here

#

It was trained to prompt img2img basically.

#

Well not really.

#

Like.... It was trained to understand the difference and then guide the image process.

slate crescent
#

right, but my computer is clearly not running GPT3 DaVinci and I did not previously understand how this model was able to emulate that effect so accurately

orchid wadi
#

It doesn't need the whole model

#

It doesn't need to give a reply.

slate crescent
#

right, it just needs to have trained on the triplets, essentially

orchid wadi
#

Basically it is only half of it. If you read the pix2pix original. The AI was trained to do guide the generation. And then told which was the correct output of the ones it made.

#

In a simple "True" and "false" differnce

#

It got it right or it didn't. And then you just iterate this teaching until it gets it right.

#

The isola paper chapter 4 goes throgh the process.

slate crescent
#

right

orchid wadi
#

I recommend reading the original source.

#

Always

slate crescent
#

most my remaining questions are forward-looking and more speculative
-will pix2pix models on future base models (image or text) be a priority? how much is there to gain, reflective of near-term iterations we expect to see in either?
-tuning pix2pix models requires significant openAI API access, or some equivalent; is there additional value on the table that merits prioritized investment in this by the community?

orchid wadi
slate crescent
#

right, "or some equivalent"

orchid wadi
#

I mean like... Why would we need pix2pix by default?

#

It would be just adding another model of limitations (and opportunities)

#

However it doesn't do anything special beyond img2img. It would be just aswell to train a special img2img model.

#

Then tune it with better captions.

slate crescent
#

maybe I'm focused on my own needs, but for the artists I work with this is a sort of holy grail

orchid wadi
#

Like this is what Stability did with 2.x. They made their own OpenCLIP because google didn't grant access to the CLIP that 1.x used.

slate crescent
#

right

orchid wadi
#

It is more or less a compensation mechanism for bad captioned models.

slate crescent
#

that's an interesting take

orchid wadi
#

Because it understands what you want, and guide accordingly. Regardless of what the ACTUAL img2img prompt might be

#

LIke imagine if some reason.... Getting someone to wear a baeball hat would mean prompting "Wearing diaper on head" this makes absofuckinglutely no sense to use. However the pix2pix AI was trained to derivate the prompts until it got what it was asked then basically put that to memory.

slate crescent
#

all I know is we have had a hell of a time using pure img2img to say, make scenes night, or change the month, change the weather, add or subtract crowds without overly mutating the rest of the editable mask

orchid wadi
#

However pix2pix can't do something that it wasn't trained. It can't turn mämmi to lutefisk if it never encoutered such transformation

#

I just drive it with embeddings...

#

Works 90% of the time.

#

Exactly what I want

slate crescent
#

do you have a comprehensive library of embeddings for context changes such as this?

orchid wadi
#

I mean like I am starting to have.

slate crescent
#

I have seen such for night/day, and they work okay-ish

orchid wadi
#

I got watercolour, oil painting, shirtless, diaper (this was for the donald trump thing I did while ago), blindfolded, mouth covered by x,...

slate crescent
#

but pix2pix promises a generalizable transformation and in my limited testing has lived up to that

candid thistle
#

yo how do i make an art peice?

orchid wadi
#

Those are ones I haven't deleted yet.

slate crescent
#

I do think that libraries of embeddings--private, public, and everywhere in-between--are going to be hugely important in workflows moving forward

orchid wadi
#

I just use embeddings because I am shit at prompting....

#

I am much better at making embeddings for just about anything.

slate crescent
#

yeah, I think that is going to be true for a lot of people, especially those with visual talents

orchid wadi
#

MY issue with prompting is that it is polluted by SEO spam. Like from Amazon/Wish/Alibaba. Clickbait images. And LAION scraped them from google.

slate crescent
#

imo a lot more focus should be put on developing skills/tools and use for embedding and similar training than the "prompt game"

orchid wadi
#

If emad is to be believed this wont be such issue in next version.

slate crescent
orchid wadi
#

As in they know the issue and are working to deal with it.

#

But 2.x is better. Every token is closer to what it is supposed to be in common parlance.

#

Including art periods and art styles.

slate crescent
#

that has been my experience, at least

#

any update on the distilled diffusion hype?

sharp kite
candid thistle
#

how to i use stable diffusion

sonic tundra
candid thistle
#

thank u

cyan flicker
#

Hey, so how come when I generate art, while it's generating, the colors look nice and muted, but as soon as it completes, it always adds harsh, dark, high contrast, super-saturated colors to the image? At 99% complete, it looks lovely, then at 100%, it suddenly turns really vivid and dark. How do I prevent that from happening?

astral goblet
# sharp kite A lot of anti ai artists are circulating this paper which attacks stable diffusi...

they trained their own models, wrote their own vae, and used bruteforce on specially crafted prompts that they intended for recreating images perfectly, and only got results on something like 0.3% of attempts. The data is all heavily soaked in selection bias. If anything the paper's existence shows that it's very transformative and the end user has to go to great lengths in order to intend to pirate copyrighted material it was trained on

astral goblet
cyan flicker
#

What is vae?

astral goblet
#

it stands for variational auto encoder, i guess, but i have no idea what that means. it's a part of the whole chain of systems, and is responsible for taking the data coming out of the neural network, and encoding it as something we know. Like a png image.

#

something about math dimensions. its the basically what outputs the image at the end

cyan flicker
#

Thank you. Appreciate it.

buoyant cargo
#

what's the best way to do it? Multiple batches or one batch with multiple pics?

lilac reef
#

@wise stratus why is it that 2.x broke LoRA and Deltas by demanding bf16 be used to train using them when 1.5 worked fine with fp16?

earnest shadow
#

Good day, I wanted to ask you about your work in the #🏆｜winner-gallery , how did you manage to make girls' faces in this style and how many attempts did it take? And if not a secret, what words(prompts) did you use?

#

@shy pumice

#

your work looks amazing

lethal kindle
#

Were the Dreamer's Canvas winners ever announced, or not yet?

fervent thunder
jagged yew
#

When we would be able to use dream bot?

steep sage
#

Is local text model like chatgpt possible?

empty barn
steep sage
#

Working

#

Good responses

empty barn
#

ChatGPT itself has a model that requires a few hundred thousand dollars in hardware to run. But there is work on smaller models that you can run locally. Take a look at the KoboldAI project.

wide juniper
#

Any announcement about the next numbered release of stable diffusion?

flint fiber
#

what's the best VAE?

#

what's the difference between EMA and MSE VAE?

neat condor
#

VAEs with more training steps create sharper pictures and anime vaes create mode washed out details and colors - that's about it.

#

so for anything cartoon and stuff i'd use anime VAE's and for anything else VAE which has enough sharpness for my taste.

#

look at VAE's as light filters instead of drastically altering something. sure for some ppl like me that light filters also drastically matter but for some they don't.

severe zealot
#

hi my stable diffusion automatic 11 11 isnt pulling from the directory when i edit the file to pull, has it been upgraded lately? do i need to get a new hugging face link? if anyone knows that would be awesome if they linked it here for me 🙂

orchid wadi
#

Example if you want to train someone wearing swimming trunks. It is best to get like 5-6 good clear images where the concept is centred and in focus. Then 1 image without a person wearing it, then a child, adult man, white young man, black man, woman... etc. This is the ensure diversity in things other than the concept. You can imagine it as trying to teach a slice of pizza (or the original cat-toy example) by having it on 5-10 TOTALLY different surfaces.

#

If you use filewords, which are not always needed if the images are good and diverse enough. Then label them with as much diversity as you can. As in derive the word for "male person" as much as you can. Do not have "young man" "old man" "adult man" use instead: "boy", "lad", "guy", "man", "granpa". Because you never know how the AI estimates your subjects.

#

For example teenaged and younger boys commonly get mistaken as "girls" by the AI, so you have to consider that. Then women are ALWAYS nearly universally seen as younger than they actually should be according to human judgement.

#

These are all things you must consider.

#

Also keep in mind that if you dataset has a bad sample (image) remove it. It does more harm having it, than not having it. Good ways to spot it is to keep an eye on the loss count if it regularly keeps jumping really high up, then you got a bad image. You can also interrogate the images and if you get a caption loop as in: "man wearing shit shirt shirt shirt shirt shirt..." there is something wrong with the image.

#

What it is? Who the fuck know. Can you fix it? Well adjust it a bit on photoshop and see if it gets out of the loop. If it doesn't then it is inherently bad for some reason

wide juniper
#

I am switching between two models very frequently and needing it to load for a few minutes each time really sucks. Can I duplicate my auto1111 installations so I can run two models concurrently?

orchid wadi
#

However I'd recommend you just turn your models to fp16, since they are half the size and load in half the time. 4 times the benefeit.

wide juniper
#

I have tried having 2 tabs

#

If I generate with different models it will cause it to switch, which is minutes of wait. 24 GBB VRAM btw

orchid wadi
#

You using a HDD? It takes me 10 seconds to swap models.

wide juniper
#

SSD

orchid wadi
#

Model loaded in 10.1s (load weights from disk: 5.6s, create model: 0.3s, apply weights to model: 0.9s, apply half(): 1.1s, move model to device: 0.4s, load textual inversion embeddings: 1.9s). This is me loading a model.

wide juniper
#

pplying xformers cross attention optimization.
Weights loaded in 97.6s (load weights from disk: 15.2s, apply weights to model: 82.4s).
100%|

#

But, on another instance,

Applying xformers cross attention optimization.
Model loaded in 3.0s (create model: 0.5s, apply weights to model: 0.7s, move model to device: 1.3s, load textual inversion embeddings: 0.4s).

orchid wadi
#

Go to Auto's settings and "stable diffusion" you can select how many checkpoints to put in to ram.

#

As you using fb32 of fp16 models?

#

And how custom are there models?

#

My good man there is no point in using 32bit models as ckpt.

wide juniper
#

fp16, but ok i'll keep an eye

#

Checkpoints to cache in RAM is 0 lol

#

Any harm in changing that value to 10? I have 32 GB of ram

gleaming jetty
#

Any idea why my iter number is like 2.5/s ? I'm using a RTX 4090 lol

orchid wadi
orchid wadi
#

If it is still really low, then you might not be using your GPU at all due to some misconfiguration.

wide juniper
#

Sometimes when I push my batch size too far, my SD gets stuck and I'll be forced to restart web-ui.bat. Is it just me or is there no reliable way to interrupt or skip the current generation?

twin mist
#

hello, someone can give me 2 tips on how to improve training results?

orchid wadi
#

For what and on what?

twin mist
#

SD 1.5, on a 3090, i would like to train it only for photos (not paintings), with many hands and feets as it's the main thing the ai does wrong

orchid wadi
#

No I mean like... TI, DB, Lora or HN?

#

I can help with TI a lot, DB somewhat. Lora and HN not at all.

twin mist
#

well im really ignorant about this things

orchid wadi
#

I work exclusively with 2.1 but the system is same. 1.5 is just shittier model.

twin mist
#

i use a custom model

orchid wadi
#

Welll... What are you using to train?

#

Colab? Auto's repo local? Dreambooth extension?

orchid wadi
twin mist
orchid wadi
#

What you mean no NSFW?

twin mist
#

2.0 and 2.1 have a nsfw filter from what i've saw

orchid wadi
#

No it doesn't.

#

2.0 has too strict training set. They fixed it in 2.1.

twin mist
#

ok but i've a particular model focused on miqotes / photorealism lol, idk where to start if i have to train it again

orchid wadi
#

You can get all the naked bodies you want. Just very few genitals. So if you want genitals you need to train it a bit with just abot any training methods since it isn't well represented.

#

First of all tell use what are you training WITH?

twin mist
orchid wadi
#

But trust me... There are dicks to be found in base 2.1 and tinkle caves.

#

Just no surprise penises every time someone has open mouth like in 1.5

#

Open mouth = Alien tongue or a penis. Every fucking time.

twin mist
#

ahahhaha

orchid wadi
#

What is the concept you want to train?

twin mist
#

treat me like an idiot, a super noob

orchid wadi
#

What is the one common thing of the 10 pictures you have chosen you want to be in the embedding.

#

Remember one embedding = one concept. Concept can be style, person wearing a hat. A specific hat. Naked body. Man in a dress.

#

However it can't be man in a specific dress that is 2 concepts.

twin mist
#

my model already does what i want

#

really well

orchid wadi
#

What is it that you want to train in then?

#

HN is a totally different thing for different use.

twin mist
#

hands and feets, the body proportions are right, but it have hard times doing feets and hands with 5 fingers and not multiple ones

orchid wadi
#

Ok so... Issue isn't really something you can solve with Fine tuning. It is a base model issue.

#

You can do what I have done.

#

Which is make embedding for certain hand pose.

twin mist
#

i can merge my model with the 2.1?

orchid wadi
#

No.

twin mist
#

so i would need to start from zero with the 2.1?

orchid wadi
#

They totally different. It would be like merging a horse with a ocean liner.

twin mist
#

i dont have a dataset of pictures for it

orchid wadi
#

What have you used to train the model in the first place?

#

You canuse the same pictures.

#

The systems to train are the same.

#

If it is just textual inversion, then you haven't trained the model at all. Just take your images, load the 2.1 model, and train embedding again.

wide juniper
#

@orchid wadi I'm having trouble training too. Trying to train a face, but the end result is very different. Now I'm wondering what to change - should I add more training photos above my original 14? Or weed out those with heads with unusual orientations? Maybe remove pics with the same person with makeup on because it's confusing stable diffusion?

orchid wadi
#

Embeddings and HP are ALWAYS model exclusive.

wide juniper
#

Doing ti

orchid wadi
#

They rely on the model being the same as when they were trained.

twin mist
#

well it's a merge of wd1.4 / one model focused on bondage and nai, trained on 3250 pics that i've took irl of latex models, as that's my job

orchid wadi
#

With TI.

orchid wadi
twin mist
#

ti

orchid wadi
#

Ok... So 3250 images on TI is like... Absurd but sure.

#

Then just load in the 2.1 model and train the embedding again.

twin mist
#

offf

#

ooof

orchid wadi
#

Nothing changes except the resulution. Train 2.1 at 768 or bigger resolution at the training tab

#

TI is meant to use like 4-20 pictures MAX.

wide juniper
twin mist
#

the ti was with only the word "latex"

wide juniper
#

It's a human face for a model that generates humans

twin mist
#

i've edited the ti templates

orchid wadi
#

Since we need to start from basics.

#

Ill get around to you in few hours.

twin mist
#

ty

#

i install the 2.1 meanwhile

orchid wadi
#

If you want to train a specific face, use dreambooth like Ben's Fast and follow the instructions. You need 10 images.

#

Then you need to fiddle around to get it just right.

#

Specific face needs you to inject that face in to the Unet of the model. If it isn't there already, it can't be found.

wide juniper
#

Ha! When I read about textual inversion, I had suspicions about training phases on ti. Dreambooth it is, then

orchid wadi
#

TI works well if you know the face is in there. Like a specific common celebrity.

wide juniper
#

It didn't make sense how a few kbs is enough

orchid wadi
#

The embedding is literally Just a text file with a vector made of 1024 tensors that point to unique points in the model's Unet.

wide juniper
#

Right. So that's why Kirsten Dunst TI works

#

Or maybe it's placebo lol

orchid wadi
#

If the celeb name is common enough for me to know then it most definitely is in the model.

#

But you can always check haveIbeentrained to make sure

#

Most celeb faces are really overtrained tho. Meaning that you can only get one face and expression of them.

#

To get more, you need DB.

wide juniper
#

Btw how do we know what's coming out next for SD?

orchid wadi
#

All I know from talking to emad here. Is that the next model has more adjusted dataset.

#

When I asked them about the SEO/clickbait spam issue of google results affecting the model quality.

#

The amazon/wish/alibaba keyword spam I mean.

#

And to some extend porn sites keyword spam.

#

I know they are dealing with that. And far as I know releasing some finetuning tools of their own.

wide juniper
#

So they'll move on from Laion5B to something more updated yeah

#

I mean I think stable diffusion would be much better if it uses the better 20% of the text-image pairs

orchid wadi
#

No they don't need to move from LAION which is their own thing.

#

They just need to filter it and make additional datasets.

wide juniper
#

Isn't it the same thing?

#

As updating it

orchid wadi
#

LAION. Has many datasets to begin with.

wide juniper
#

Oh. I thought it's something the few Germanb dudes invented

orchid wadi
#

Problem is that they are all just CHAD easthetic scored datasets. Doesn't actually mean the captions or images are good.

#

Scroll down my good man.

wide juniper
#

If I want to train poses, like pair skating poses, will I get good results with any method?

#

Method meaning ti/db/lora

#

Or say people being upside down like handstands

haughty terrace
#

Woah there! The owner of Stable Diffusion has requested that Discord block any messages our mostly-accurate robots deem to be explicit, so your message has not been sent.

#

can sum1 help

charred hornet
#

Hi ! How can we make some pictures since the bot is closed ?

trail edge
#

is there any difference in the image output from a gradio online SD and a local SD? (same checkpoints, same settings, same prompts)

bold crater
#

Is there a way / command I can use so the program won't show the "Not enough memory" error when trying to load bigger models. Using 16GB with Nvme as page file and a 1080ti.

warm junco
bold crater
# warm junco memory error or cuda out of memory?

load_tensor(dtype, nbytes, key, _maybe_decode_ascii(location))
File "C:\Automatic111\stable-diffusion-webui\venv\lib\site-packages\torch\serialization.py", line 1079, in load_tensor
storage = zip_file.get_storage_from_record(name, numel, torch.UntypedStorage).storage().untyped()
RuntimeError: [enforce fail at ..\c10\core\impl\alloc_cpu.cpp:72] data. DefaultCPUAllocator: not enough memory: you tried to allocate 9437184 bytes.

warm junco
#

that should work, what model you try to load?, make sure other programms are closed

bold crater
#

Model is the pix-2-pix

warm junco
#

ah okay, do you use xformers?

bold crater
#

I used it earlier today for the first time. Before that I did not use it.

warm junco
#

okay good

#

make sure your webui is uptodate

#

do you get the error with other models too?

bold crater
#

Always downloads latest version when I start SD, also updatet pythorch today. Yes some of the bigger ones.

#

Do you know if xformers is better than the Doggex cross optimization (or whatever it was called)?

warm junco
#

yea xformers is better afaik

#

make also sure you dont use --medvram

bold crater
#

okay, thx. I Will try it with that option enabled again.

warm junco
#

no only use --xformers

#

i said dont for --medvram xD

orchid wadi
#

Also trains TI faster. I can only put it down to the fact my CPU is better in relation to GPU, and RAM is from the quicker end.

warm junco
orchid wadi
#

Not really. If the transfer speed between RAM and VRAM is quick enough.

#

Like with medram xformers. I can do others things with my pc while SD generates in the background

#

Without medram, it clocks up things.

#

Like technically I guess without medvram it is "quicker" but something bottlenecks it.

#

And I suspect that is the GPU's vram.

warm junco
#

hmm okay, that could be it

orchid wadi
#

Also with medvram I can do 768x768 batch size 9. No issue. Still have some computer to even spare for other things.

#

With only xformers I can do 768x768 batch 7-8 9 with the more tamer samplers.

warm junco
#

yea but that shouldnt be faster than without medvram

orchid wadi
#

Well what can I tell you! It is that 0,5-1it/s faster 😄

warm junco
#

okay 😄

orchid wadi
#

It doesn't really matter since I make things in like batch size 9 and XY 10x10 grids.

#

Or 9x100 runs

#

And I'm not home when it does it.

#

Also the passive Vram use between xformers medvram and no medvram is like... 1gig

#

So whatever component is left out of the model, must not be that significant

warm junco
#

i dont think its left out, it just loads lesser files at once into the vram

orchid wadi
#

I suspect it might the the VAE that is put to VRAM. Because it is only called when finishing the images.

#

That is where there is clear delay between no medvram and vram.

#

Like it takes longer for the images to come out. But processing is quicker with medvram.

warm junco
#

very interesting behaviour

#

do you use S.A.M or resizebar?

orchid wadi
#

I have heard other people repoting the same.

#

Nope.

#

Just my stock packet pc wtih 3060TI and autos basic repo. Not sure if the same behavior happens on invoke since I haven't tested that much

#

I don't have xformers on ComfyUI since I can't figure the way to get it installed as standalone

warm junco
#

ah okay i should test it with my 1080 if it behaves the same

orchid wadi
#

Right Reziable Bar is on according to Nvidia settings

#

So no idea whether that does it

#

God they hid that well... Never thought to look for it via the nvidia control panel.

#

As if anyone actually uses it

upbeat tangle
#

hello guys. Is there a site where we can learn artificial intelligence developments and stable diffusion tactics?

warm junco
orchid wadi
#

Yes... I know what it is and does.

warm junco
#

ah ok i misread it then xD

orchid wadi
#

I just don't know whether Python can call it it.

#

Or this system can call it.

warm junco
#

yea thats the question if its usable for SD

orchid wadi
#

Not every game can even use it

wide juniper
#

Is the auto1111 dreambooth still broken?

orchid wadi
#

I recommend asking at the extensions github page. They know the best

#

Well don't go above 50 steps or below 20.

#

That is about it.

#

It is best for long and complex prompts

#

But it is rather primitive in it's approach

#

I use all samplers, since they all can do things well. And you never know which will be the good one.

verbal vale
#

Hi all, I’m working on an application, and currently struggling on achieving somewhat consistent faces of characters. Are there any guides for best practices in this area? Should I be utilizing a separate trained Lora model for each character. The primary use case is portraits avatars. So I don’t think in painting would be necessary.

#

Im currently separating scenic images from character images. So I don’t have to worry about full body consistency. Just shoulders up mainly.

orchid wadi
#

It depends. If you have a model with the characters in it. The Textual INversion is enough

#

If you don't then you must use Lora, DB or HN to get them in individually

verbal vale
#

The app will routinely be releasing new characters so I’m trying to think of a strategy that allows for that.

orchid wadi
#

Which of these is best depends interily on how and what kind they are.

verbal vale
#

User customization isn’t a feature yet. But perhaps they characters might react to the stories they are placed in.

orchid wadi
#

Well in that case I'd say HyperNetworks

#

Heavy to train, but they are most versatile. Example: The NovelAI image system is built on HyperNetworks

verbal vale
#

Okay I’ll research that.

orchid wadi
#

To put it simply. HN allows you to do a concept within a concept.

#

As in if your base model only has "male" with HN you can have a network starting from "male" that breaks down to "Young male" "Adult Male", and "young male" to "Boy, Lad, Teenaged boy, young boy..." so on and so forth

#

And these networks can be stored individually, they do not affect the base model, however work only on that base model.

verbal vale
#

Interesting. Would you be okay if I friend requested you so I can give you more precise details on what I’m trying to achieve. And see if i am I understanding this correctly.

tawdry summit
#

Does anyone know an online group centered around deepfaked voices?

verbal vale
#

No but I wish someone would release an AI that does like a dozen voices really well, so I could integrate with it for character voices. I know the big fear is that voices would be deep faked, like with V-ALLE. But instead of allowing any voice to be generated, just allow a few good ones that aren't the voice of some celebrity.

glass bear
#

leão ouro ¯_(ツ)_/¯

wide juniper
#

All of my dreambooth training classification images look like black and white photographs. Am I doing something wrong?

#

Also, it's giving me images of people of different races. Not sure if that's what's supposed to happen

nova tusk
maiden crystal
#

In your opinion what are the best diffusers out there now?
Specially for logo design, abstract, pop art, graffiti and advertising styles?

nova tusk
#

Someone help me, in my discord the "dreambot" does not appear

fervent thunder
#

@warm junco yo whats up check general-with-images I've uhh changed style a bit lol

nova tusk
#

or where can i find pls?

rigid chasm
#

is the dreambot in this server free to use?

ripe eagle
#

Just saying hi. I'm knew.

fervent thunder
#

how can i use this?

fervent thunder
orchid wadi
#

Well that is how it actually does work. Imagine it as a tree. One tree can only have so many branches until the prompt - latent interrogation ≈ 0 condition is met.
The more trees you have, the less significant one tree is. How ever with more trees, you can find more things overall.

#

However there is logic to the prompting you should try to follow. The first token starts it all.

So a good way to start prompting something is: theme, subject, action, object, image properties, style As in "Steam punk boy holding a potato, warm colours, drawing"

#

The equivalent embeddings go to equivalent place. Then once you find something that works, you can further refine by adjusting prompt.

velvet verge
#

hello! does anyone know how to create midjourney like results with stable diffusion? midjourney looks so much better to me

fervent thunder
velvet verge
fervent thunder
#

that is checkpoint training data using a UI like automatic1111 you can do img2img also

#

and since its 1.5x based you can use the 1.5x inpainting checpoint data for that

astral goblet
#

i don't think you'll find an img2img quite like midjourney can do

#

its more than just a model they're using for their image gens

#

the legal battle hasn't slowed them down. there's no court orders at all. just a filing

#

emad hasn't said anything to the contrary. deepfloyd testers have regularly been putting out new examples

fervent thunder
#

I fail to see how sd with embeddings and good prompt isnt already good

astral goblet
#

midjourney is just a very popular and known product is all. It puts out a lot of high quality images that make their way around the net and get scene. It's easy to understand where people are coming from when they ask about that

fervent thunder
#

ye for sure

astral goblet
#

i'd be using midjourney's models if they released them seriously, that shit is good at what it does. Total shame that it's closed off SAAS behind their terms though

fervent thunder
#

it would be interesting to use dreamlike and make a custom embedding based off midjourneys outputs ;')

long marten
#

hello peepo

#

I have question: is it possible to get SD to influence only certain frequency bands of an image?

#

Like for example, only influence low-frequency image features, but preserve high-frequency features

astral goblet
#

i also don't think DF is going to run on people's home machines. They'll liekly be a website offering services instead. Likely has memory requirements closer to 40gb

long marten
#

For example in the case of changing a lighting or atmosphere / color palette but retaining all the subject matter

astral goblet
long marten
brazen fox
#

analyzes your image using MiDAS and generates a depth map, and uses that depth map as a guideline to make a new image- so it generally retains composition

long marten
#

as it is right now with image - image I have to repaint all the "goobified" faces every time I want to try out an edit 😂

brazen fox
#

or using an extension you can supply your own depth map- also works with logos

fervent thunder
brazen fox
#

extension in the a1111 extensions tab

fervent thunder
#

also something else I find helpful is doing an image in 512x768 then using outpainting mk2 to extend the background

long marten
#

oh yeah! I've tried those things, but if you can imagine, you land on a face you like, or you just paint a face that's good. Now you want to change the lighting conditions in the scene, currently with img-img low CFG value will only affect details, high CFG will change the structure, I think the depth-preserve model might be what I'm looking for in terms of preserving structure but changing low-frequency information like lighting without moving objects around

#

but thanks for the references though, I'll give them a read for sure

fervent thunder
#

entirely depending on the checkpoint data and tags

astral goblet
#

it estimates the depth map based on training, and then the model is trained to understand that depth information as it generates a scene. It mathematically treats the mask in a way that allows it to generate new details while keeping overall composure.

also, i'm not entirely sure how it works. i've only groked a few paragraphs of the paper and i'm still sort of unsure i groked that

fervent thunder
royal mica
#

Hello, I'm using anything diffusion and I was wondering how much is ideal for stemps? and do {} influence anything like in NovelAI?

warm junco
astral goblet
#

in what where now

#

wait, i dont wanna know

#

looks like another over merged situation that looks as "general purpose" as protogen or any other.

#

merged models get really great specialized results, usually for pinup girls

royal mica
#

what is the best sampling method?

astral goblet
#

the one that gets you the results you like the best

#

experiment and find what you like, x/y grids help a lot

wild steppe
placid wing
#

Hey folks, anyone know a good checkpoint (2.X) or embedding for buildings and architecture?

lethal kindle
astral goblet
#

i think the fix to all my stability issues was as simple as using an up to date version of python 🤦‍♂️

fervent thunder
#

the potential of AI is scary

twilit willow
#

@narrow jackal I think it's a very complex topic

narrow jackal
#

I think that AI learns through them

fervent thunder
#

I believe that if we also check the very "pedantic" details, then the AI can't steal your art as stealing is defined as something which you do not have anymore because someone... well, took it from you :P
I believe that's the legal definition... why am I sweating? I-I'm sure this is kinda correct! :P

narrow jackal
slate crescent
#

I mean, AI should not be conflated with piracy or IP theft (regardless of viewpoints on those things), since it, well, simply isn't

#

there exists a luddite narrative that AI training/use is blatant IP theft only permissible through pedantic technicalities or legal oversights, and it's best to give this... worldview as little oxygen as possible

twilit willow
slate crescent
#

but, it's possible?

#

like it's all emotional, all a matter of branding and framing at that point

twilit willow
#

anyone have any thoughts on Niko's (from Corridor Crew) take on SD basically being a visual search engine? It just doesn't sit right with me. He suggests that all of the data is there, all the results are predetermined. So it's akin to looking for a book in a really large library, it's super hard to find but it's already there

slate crescent
#

I mean that's just, not how it works

twilit willow
#

no, on the flip side, it doesn't exist until you find it

slate crescent
#

if I can be frank, and I'm speaking as someone who thinks that anxiety towards AI should be met with empathy (and those gloating about people "losing jobs" should be thrown out the airlock), all these ridiculously misinformed takes on how AI does or doesn't work should mostly be ignored and not discussed

#

spreading around bad stuff is just Toxoplasmosis, regardless of whether or not you think you are fighting it

narrow jackal
slate crescent
#

labor is scarce

twilit willow
#

I'm obviously team AI btw KEKW

#

but, it's always good to play devils advocate

slate crescent
#

the potential energy to move a ball up a cliff, or to alphabetize a shelf of books, or to organize drops of ink into a story, or to put 1s and 0s into a pattern where they make a picture, can only come from a pool of limited resources and opportunity costs

narrow jackal
slate crescent
#

it is vital that society permit these activities to be rewarded

#

like all human interaction, it's a big Ally vs Betray prisoner's dilemma

#

"stealing" the book, movie, or drug is the Betray option

#

if everyone picks it, the means of its creation is imperiled

narrow jackal
slate crescent
#

society exists as a series of contracts that enable people to choose Ally without being made chumps

narrow jackal
slate crescent
#

you are hung up on a semantic distinction that is completely irrelevant to everything I'm saying

narrow jackal
slate crescent
#

no one cares what word you use, I just would like to get paid and not starve

twilit willow
#

was supposed to be implied

narrow jackal
twilit willow
narrow jackal
twilit willow
#

I don't think any of this is theft

#

tell you what though, as I'd predicted (and well, lets be honest, we probably all saw it coming) the anti-ai group was all like... only train dead artists! until that was also disrespectful

narrow jackal
slate crescent
#

well, as a case study, compare the US and China with regards to IP law

narrow jackal
slate crescent
#

China is, for all intents and purposes, a free-for-all with no real IP law (or rather, IP law that is only selectively enforced for a few favored native corporations)

#

intellectual property, the topic we are discussing

#

there have been loads of cases of Chinese "asset flips" of western games

#

you take a year make a game, they flip it in 3 days, uploading their copy to the storefronts

narrow jackal
slate crescent
#

yes, China is the opposite of anyone's libertarian paradise, but the fact remains that in terms of IP law they might as well be Mogadishu

#

no--Chinese asset flips aren't unjust because they are mean or not nice, they are unjust because the inevitable outcome (the person who did the work going out of business while randos in another country profit) is logically opposed to the intended values of our systems of law and commerce

#

the central tenet of Libertarianism is the non-aggression pact, the ability for independent actors to assume reasonable stipulations about how they choose to be interacted with

#

the ability to say "I made this thing, please pay me $X if you wish to enjoy it, otherwise please do not use it" is very much in line with the NAP

#

the option of "no, **** you, I will not pay you and enjoy this thing you made, purely because you can't stop me" is very much anti-NAP

narrow jackal
#

*lose

upbeat tangle
#

RuntimeError: Not enough memory, use lower resolution (max approx. 320x320). Need: 0.0GB free, Have:0.0GB free i got this error

#

-_-

slate crescent
#

that has nothing to do with the example given--if the Chinese in question simply made their own game (bigger and better), then there is no claim against them

#

in that case, the system is working as intended

#

(the people who made the best thing got paid)

narrow jackal
slate crescent
#

yes, in the purest nature, Betray is the only strategy

#

ergo we construct civilization, a form of mechanisms by which Ally becomes possible

#

a rule like "don't attack your neighbor"

#

"don't throw your garbage in his yard"

#

"don't copy a movie that he made but asked you for money to see"

#

these are just rules we make to achieve a desired purpose