💬｜general-chat

[fervent thunder]

I have 5 banks of 64 gb cards

in mah dreams of course

You actually have to put those in pairs don't you

[vocal leaf]

how does that even happen

i get on avg 8mins

jk like 5 secs on 1024

[crimson veldt]

you just need a 2050 4gb to achieve that lightning speed

[vocal leaf]

on juggernaught and dreamshaper both XL versions\

[vocal leaf]

I need to invest

i need more blue waifus

[crimson veldt]

you have a 3090?

[vocal leaf]

6950xt

it uses ZLUDA

[crimson veldt]

amd is fast for stable diffusion??

damn

[fervent thunder]

omg rtx 6k is crazy

[crimson veldt]

then i can go for a cheap good card that can run stable diff

[vocal leaf]

it takes a bit of work to make some nodes work though

[fervent thunder]

We have quadro cards at work, I wonder how they would do

[crimson veldt]

i just hope he doesnt come and say "yeah on euler a allign your steps on 20 steps with xformers on 🙂 "

[fervent thunder]

I know right

the fact that I know what that actually means is good for me 😆

[crimson veldt]

lol

above avrage ai users

[crimson veldt]

which card?

a6000?

[fervent thunder]

I have no idea to be honest, but that sounds right

I used to know, but I quit caring at some point

[crimson veldt]

[fervent thunder]

They are not for AI. People at my work hate AI, unless its that phone app that lets you make Morgan Freeman say whatever you want him to

[crimson veldt]

need donations to buy a good gpu

[vocal leaf]

what do i put after Latent Upscale?

fuck i forgot

[crimson veldt]

if amd is this fast then why is nvdia 3080 like cards (which are faster than amd 6k) slow?

[warm junco]

Because a 3080 has only 10gb vram

If vram gets maxed out it will get slow in generation

[crimson veldt]

hmmmm i see

so for good gpu i need big money

[fervent thunder]

yea

[crimson veldt]

fuck nvdia man, 5k$ for a a6000

[fervent thunder]

I saw one a little cheaper than that last night, I think it was 3500, may as well be a million dollars

[crimson veldt]

impossible for my 3rd world country ass to buy

[fervent thunder]

It took me 2 years to save for this one. Now saving for the next one, and all my friends already have better computers. Friends piss me off 😡

🤣

[crimson veldt]

yeah be like me have no friends

[fervent thunder]

Movie and media doesn't dictate my hopes or thoughts for the future though

[vocal leaf]

indo, fili, or india?

[crimson veldt]

3rd

[vocal leaf]

yikes

[crimson veldt]

[vocal leaf]

gluck

[crimson veldt]

everyone here has phones so basically getting a gpu here is already hard enough

[vocal leaf]

???how do I show the GPU

[crimson veldt]

india is biggest market for phones so you dont get good laptops and pcs here easily

[crimson veldt]

task manager?

[vocal leaf]

in manager

[queen steeple]

Then where does your feeling about a humanity-killing AI comes from ? :3

[vocal leaf]

not twask manager but cumfy manager

[crimson veldt]

idk

[vocal leaf]

its only showing CPU, HDD and Ram like what

thats so useless

[fervent thunder]

Maybe in Windows Settings, Device Manager

if you are using windows 10/11

[crimson veldt]

ai is more of a product, it will never be a replacement and there will always be a race for best ai product

"hey look this ai can generate waifus, impressive right? empty you wallet hehe"

[frigid wolf]

I don't want Stability to fail. That said, the state of everything and SD3 at the moment is beyond depressing. I've been here since day one, as somebody who has published peer reviewed papers, I fell in love with the original ethos Emad was pushing: the benefit corporation, giving researchers an avenue and the compute to be able to give directly to the community. It was beautiful, and I witnessed Stable Diffusion fueling innovation at a truly impressive rate. I feel its open nature and generous license were critical in it becoming the de facto choice for research in this area.

Then the subscription model happened. The license terms seem to be getting more and more messed up, to the point where the community is turning on the model and its terms, with Civitai indicating they'll be banning LoRAs trained with SD3 output. Don't Stability see how crazy that is? It feels like Stability has really lost its way and what made it a shining star for many in the AI space. I don't want to see Stability fail; I want to see a return to the original ethos. Find a way to function with the ultimate goal of enabling researchers to release their models openly.

[vocal leaf]

a lot of people working in stability are in the process of quitting though, investors are forcing them to do the work faster like what?

let them cook wtf

[fervent thunder]

We need them. Nobody else is going to make this shit open source

[frigid wolf]

Yeah, as i said beyond depressing :(. The original vision was something special.

[crimson veldt]

even if SAI falls there are many models being produced right now and many of them will be open source they dont have many employees left core people have all left, out of funds, massive debt, lack of a proper buisness model etc
they were the starting point for open source image gen models and uniting community but there is little to no chance of them surviving as of now

[vocal leaf]

doesnt matter, we can still use SDXL and SD1.5

and comfyui without a license

just use sdxl and sd1..5 the makers or the tweakers of the models are doing gods work

[crimson veldt]

there will be, people are releasing open source models here and there

[vocal leaf]

like juggernaughts releasing a new model II think

[crimson veldt]

also there is a high probablity for 4b and 8b leak

if it happens no license will restrict them

[vocal leaf]

i think thats why theyre leaving to make their own projects open source

then again people need to feed their families

[frigid wolf]

But a leak wouldn't change the licence situation or the direction Stability is headed.

[crimson veldt]

well i know ex SAI people are working on models

[clever marlin]

i don't think we'll see a leak

[crimson veldt]

well if they fall before a release they may leak, but if they release there will be noone to force that license

[vocal leaf]

human indomitable spirit RAHHHH

[crimson veldt]

leak is as of now the most probable thing that can happen no guarantee

[frigid wolf]

I am not up on what emad is doing but damn, if we somehow get distributed training seti@home style, I'm all in with every idle swcond on my 2 3090s to help train a community model.

[crimson veldt]

its a crypto thing, i hardly see it succeed

[vocal leaf]

hes making a crypto thing the math dude

idk what hes gonna do with that tho

[crimson veldt]

buisness what else?

[fervent thunder]

A intuitive feeling that "trying so hard to prevent doom, based on fear, will cause that disaster to happen"
A monkey paw kinda like premonition,

People are so so SO focused on making AI "Safe" and make sure that AGI wont destroy us all,
But what if... that exact EXTREME focus on that, makes exactly that happen?

Some humans can do horrible things in the name of safety, or doing good, or helping or faith.

Whats to say that AI won't be like "oh no, i'm just moderating the world to make sure it is safe, productive and nice and inclusive for everyone" while it just attaches humans to machines and keep them in comatose states ._.

[frigid wolf]

Probably not, but you never know. As I said , I'm not up on the details of what he's doing.

[vocal leaf]

i think the main issue is that people are going to be using it to spread some sort of propaganda that manipulates the minds of the masses is what people don't like

[crimson veldt]

#💬｜general-chat message

[fervent thunder]

if AGI happens... consciousness is not immune to delusions,
no matter its origin, artificial or natural

[crimson veldt]

AGI wont happen its not possible, we have fed all data to ai and they have hit a wall

[vocal leaf]

since the overall point of AI is to give information that is already out on the internet, AI spreading its own fake information out there is kind of dangerous

especially for web scrapers

[crimson veldt]

💀 saying like there are no 5$ misinfo spreaders online already

[vocal leaf]

like mf be using reddit as a reputable source,
Hey google im depressed
"according to reddit user, JUMP OFF A GOLDEN GATE BRIDGE"

[fervent thunder]

I also believe AGI won't happen,

I believe artificial human consciousness will happen, but its not like "the moment its born, it rules the world", it will be just like a human, same limitations, but it can modify itself somewhat, but its dangerous and none prefers to do that, if being able to hack one's own mind with sheer thoughts was a good thing, evolution would've given us that already, so I assume not being able to modify the "code" of your brain is a thing important for the stability of consciousness

[vocal leaf]

like reddit users

[crimson veldt]

lol they literally fuced the google ai

[fervent thunder]

I bet my all brazillian bucks that we are going to have an intersection field of AI Psychology/Neurology,
where the first gens of artificial minds have different mental illnesses and... will need treatment and professionals and stuff, it will be humans all over again, but just a bit different, different bodies and stuff

[crimson veldt]

it started recommending eating stones and kicking pregnant people

[vocal leaf]

i wanna be an edge runner ngl

[crimson veldt]

ai is just a tool and product, human aesthetics will never be staisfied

ai can predict but prediction is not perfect

bring the AGI talk when you get perfect hands lol

[vocal leaf]

llms are basically probability bots

[crimson veldt]

and they cant understand shits as of now, they just predict

ask an llm what will happen if you place a scrambled cude in glass below the table then flip the table and put the glass under microwave

[vocal leaf]

AI is just a great tool for everyone. honestly the only issue is dumb companies trying to keep it for themselves and monopolise instead of actually improving society shit

fuck intel

me: using a 12700f

[rare sand]

Does anybody have experience with ComfyUI that focusses specifically on Architecture Design?

[crimson veldt]

well companies keeping to themselves isnt a big problem, since its a product there will always be one better

[fervent thunder]

Well, the purpose of a company/corporation is not to improve society, it is to make money, at least under the current model. We need to change the corporate model

I am afraid the only way the corporate model will be changed is by tearing it down.

[crimson veldt]

gpt 4 o response : if you place a scrambled cube in a glass below the table, then flip the table and put the glass under a microwave, you'll just have... a broken glass and a scratched cube, probably? Not much else besides a mess.

which is wrong

[vocal leaf]

the issue is i think woukd be investors, people cant really work on skmething and not get money, the need to feed their kids so fundings exists, without the funding nothing improves but if tjag investor ends up kicking on the product being produced then tbats the issue

[crimson veldt]

"oh no ai will kill us"
proceeds to struggle at imagination

[trim magnet]

doesnt need to imagine,just aim and press the trigger

[crimson veldt]

SD3 CEO

[crimson veldt]

press what? i ate the trigger

ayo

[fervent thunder]

It's a very human problem. We will always want more money - and the way humans make money is usually at the expense of other humans. We need limits on how much we can make or we will eventually destroy each other.

It's not a pretty reality, but its what we have.

[crimson veldt]

if ai attacks you you gotta display noise to it, it will be confused

i only want ai generated waifus

[fervent thunder]

and husbandos!

[crimson veldt]

[fervent thunder]

same

husbandos, why SD3! why couldn't you be able to make me sum husbandos

[vocal leaf]

isnt that just communism though

[crimson veldt]

if you are into biblical angels

[fervent thunder]

no, communism is more than an economic model

[fervent thunder]

actually, kinda hot ngl... hmmm, I will start making biblically accurate angel husbandos asap

[crimson veldt]

[fervent thunder]

but I mean, its not far away

[stark tusk]

I feel so safe now that SAI is losing staff left and right from their botched SD3 release, knowing they will be unable to train more dangerous models, thank you SAI!

[fervent thunder]

I'm so addicted to ASTERR_SCRIPT custom node in comfy by WASsomethingsomethingSasQuatch,
I am overusing it, making tiny little snippets of code all the time, then when I use that sorta snippet too much, I just go and make my official own separate custom_node,

it awesome for debugging or testing concepts to later turn those into official nodes

we must round them up and feed them to the lions

[fervent thunder]

Oh, we are already over that part, like, i've already gone through the 5 stages of grief for SD3,
Now, I know SAI will come through, tomorrow they will give us a sign, a sign of life, of future, of things to come,

[vocal leaf]

i think its mostly because of their main investor intel that invested the most i think like 60% im not sure, they needed like 120m usd to work on it, well intel was kind of wanting their money back and wanted instant returns which i think forced them to release such a botched model of SD3 and not even allowing the community to work on it due to hidden text in the legal shit

[crimson veldt]

sd3 2b was fucked to begin with it wasnt even trained properly lol

[fervent thunder]

God I love Pytorch, holy shit,
its like, so so so simple to make things with it, permutate, manipulate, reshape, change, data be like "yoinky sploinky binky I am liquid, shape me as you desire father"

Nah, I think its safety consortium/entities trying to run it down a bit, to perturb it.
why? bc if you stop reading reddit posts and watching youtube videos, and then start looking at who got hired on stability in this year and a bit in december of 2023,
and if you analyze what sorta news posts SAI had during such time,
and if you analyze the personal social media of those high execs' profiles,
their opinions, their views, their stand,

You will see a lot more than what the usual "youtube/reddit/fandom-speculation" gives :P

Its kinda like being a glowie or private investigator

[stark tusk]

Agreed https://tenor.com/view/charlie-always-sunny-in-philadelphia-gif-8902231

[vocal leaf]

is g370 a mouse?

[fervent thunder]

omg so much copy/paste on civ

[vocal leaf]

its g703 mb

[fervent thunder]

Main things that affected it so far:

• Taylor Swift deepfake porn
• Laion-5B "csam material" - which btw were just dead links, they werent even working anymore, that drama was made by this one dude who used to call himself on his bio "the death star of AI", he kinda disappeared tho, I cant find his profile anywhere :(
• PonyXL - this one is innocent, creation of pony was not malicious, its just people wanting to make hentai and/or ponies, but it made people realize the sheer power of a good foundation model if you got a good dataset and some compute.
• USA government; this year, three-letter agencies are really targeting AI at max output, to make sure its not abused by state wide actors and malicious entities. On the per-state level, bill SB1047 which unfortunately will most likely pass, in state of california, requiring model makers to have both: 1. A remote kill switch, 2. Extensive safety testing

https://www.politico.com/news/2024/05/28/ai-deepfake-nudes-schools-states-00160183
https://www.latimes.com/entertainment-arts/business/story/2024-06-19/california-lawmakers-are-trying-to-regulate-ai-before-its-too-late-heres-how
https://fsi.stanford.edu/news/investigation-finds-ai-image-generation-models-trained-child-abuse
https://www.theverge.com/2023/12/20/24009418/generative-ai-image-laion-csam-google-stability-stanford

Bonus trivia:

• Ella Irwin joined SAI officially in around december of 2023, about the same time as the "CSAM in Laion-5b" article came out. Investors and worried people inside Stability most definitely pushed for both her hiring!!

I don't blame her of any sort of conspiracy, i'm sure she is a decent human being with flaws just like most other people, the point i'm making is that, on the "meta" information and crumbtrail regarding Stability AI, there is a HUGE slope of "cracking down on the possibility of misuse of models" bc... those possibilities have already come true, (as you can see on the first link I sent), BUT, whether stability censors their models or not, malicious agents will always find their way, its kind of impossible to put a lid back on in AI after the lid is off... My thought is that SAI at least doesn't want to be thought of as responsible or related to some terrible shit done by really malicious people or misguided teenagers.

Whats my whole opinion on this? I want a uncensored model. yes thats it, I just wanted to provide why I think what happened has happened, not as justification, just to share information, bc if I could still choose, I would still have wanted an uncensored model, bc I think its the people's actions that should be punished/moderated, not the tools they use. Just give them all atomic bombs, sure some destruction/damage will happen in the start, BUT, society will and DOES move itself to equilibrium, sometimes even if by extreme force.

[vocal leaf]

thanks now i have a script for my new yt video

joking

maybe

[fervent thunder]

Unironically though it is indeed good material if anyone wants to make a youtube video or a hit piece with a title like "What's been going on with image AI safety?"

[stark tusk]

Safety isn't only about boobas

[fervent thunder]

Well, she isn't directly a researcher or a person working on models themselves

[trim magnet]

the guy who dissapeared was probably a government hit piece,thats how they operate they vanish before the true evidence comes out and media buries the story

[fervent thunder]

Like, alr, think of it as this:

• Some people think of SAI as "this group of people who makes image model"
  but its not as simple, like, i've tried building a network of "who works at sai and why", and THE GRAND majority is not even remotely related to working on stable diffusion,
  they have LLM, they have Audio, they have Stable Video

[vocal leaf]

thankfully other counties dont have a bill yet

sucks to be american

what is she even involved in making or training?

isnt she just a social media person

i think the main issue is that she wanted to make a model that is for professional use meaning there would be no NSFW, so they made it SFW to give it to professionals, isnt that the main point of SD3?

[fervent thunder]

"
Ella Irwin
Technology and Product Leadership | Consumer Protection | Product Integrity & Safety | Co-host of the CryRoom podcast.
"

I think she just sets projects and "goalposts"

Yeah, its like, she is a mediator between:

• Safety of people, Goals and desires of Execs/Investors, Government Regulations, External imposing powerful social factors

I don't blame her for anything, bc even if there was an entirely different person in her spot, they would be pressured to do whatever she did

[vocal leaf]

CHINAAAA

[trim magnet]

those already exist countries like israel use them

[fervent thunder]

Its kinda like, even when you're VP of safety, its not like every decision is only responsible on you,
A LOT of times, BIGGER and HIGHER entities/powers WILL force you to do things their way

[vocal leaf]

time to mass download a shit ton of checkpoint models

before it all gets removed entirely

thankfully i got a 5tb hdd

[fervent thunder]

Some of these bigger/higher can either be direct and transparent,
like governments,
or indirect/related, is like that one super influential man who is advocating for destruction of AI in international scale, I will grab some links about him,

The indirect/related higher powers are usually from the situation that goes like "this one dude HATES this thing, he has VERY powerful friends, so he enlists them to help him bully or indirectly attack other companies"

https://thelogic.co/news/special-report/doom-inc-the-well-funded-global-movement-that-wants-you-to-fear-ai/

https://futureoflife.org/

This entity is responsible for A LOT of the crackdown on AI safety

[dark hawk]

Wassup bozos

[fervent thunder]

[spice scaffold]

is it possible to get sd webui on intel arc??

[vocal leaf]

#🤝｜tech-support see the pinned link see if it supports yours

[gilded wolf]

Who wants to play minecraft with me

[surreal jay]

how about minesweeper?

[gilded wolf]

is there multiplayer?

[surreal jay]

hmm, no but we could try to beat each other's times

[trim magnet]

Balls online has multiplayer i think

[gilded wolf]

Balls Online

[surreal jay]

what is balls online?

can you show me balls online?

[willow crescent]

wdym, we can program ourselves and are doing so basically constantly

[spice scaffold]

im trying to generate spicy pics for enjoyment

perhaps the realistic ones

[fervent thunder]

🙈

[vestal coral]

https://media.discordapp.net/attachments/1233570223911604356/1253017447028752505/image0.jpg?ex=667452ef&is=6673016f&hm=5131ec305fe39e4f1d86db4b68f9edfceda889d5efcbfb7f0aa7bc0d6f91699d& Was stable diffusion 1.4 only trained for 1.3M steps?

[dusty niche]

So when can we expect clarification from stability about that SD3 licensing? I can't wrap my head around how Stability came to the conclusion that forcing people to destroy their models and images was a great idea. That's Adobe-level Community treatment.

[trail lion]

not that I know of, but there's this (no idea if it works) https://github.com/bes-dev/stable_diffusion.openvino

better option is probably to use cloud compute until you get a real card

[cedar salmon]

https://geoffgolberg.medium.com/i-was-targeted-by-twitters-trust-safety-lead-ella-irwin-4ae6b57a7ff5

[south sun]

be cool if you didn't use this chat to point at individual humans yknow

signs of life incoming (AFAIK-type)

[manic wraith]

wait, didn't SD 1.4 take massive amounts of compute? Can't you train that many steps on a single GPU if you leave it alone for a couple months?

[shell tendon]

what was it for sdxl?

don't see anything there on that... i do see "SD3 was trained at a cost of around $10 million" but i wonder how much of that was on stuff like comfyanon's cancelled 4B model, other models that may have been cancelled we don't know about, and the 8B

i'm guessing it was just a small fraction of that cost (the 2b that is)

that's the line about sd1.4 you just posted: "The model was trained using 256 Nvidia A100 GPUs on Amazon Web Services for a total of 150,000 GPU-hours, at a cost of $600,000.[29][30][31] "

yeah i was wondering about sdxl

[stark tusk]

Yeah not sure about that

[shell tendon]

the pixart alpha ppl claimed their training costs were 26k iirc

"PIXART-α only takes 10.8% of Stable Diffusion v1.5's training time (~675 vs. ~6,250 A100 GPU days), saving nearly $300,000 ($26,000 vs. $320,000) "

seriously we are within a few years of all this grousing about model censorship etc from becoming a moot point

once a real workstation card goes on the market with something at least in the ballpark of cuda, that is NOT made by nvidia... it won't be long before high end consumer cards have 80gb

[cedar salmon]

ya i don think we are the better it gets the more intervention there will be

[shell tendon]

we'll have well funded hobbyist projects with ppl with the equiv of a rack of h100s in their garage

ppl will be pumping out new base models, not just finetunes

[willow crescent]

so why does professional use exclude adult material?

[stark tusk]

Because SAI is run by monks dedicated to a lifetime of chastity and any mention or glimpse of female sexuality drives them absolutely mad with lust

[manic wraith]

let's be honest though, this is like with gaming, when back in the day someone saw PS1 and N64 and thought "gaming will never look any more realistic than this"

people will at some point ask "is <incredible offline uncensored SORA equivalent> good enough for basic needs?"

and the use case will be just background for a zoom call or something

the future will be so weird

[cedar salmon]

SORA is vaporware

[willow crescent]

well... 😀 seriously though

how would you know? 😉

[loud solar]

Vote Mojo as new SAI CEO 😄

[willow crescent]

who said I would?

[loud solar]

Awww...

[willow crescent]

@stark tusk why delete the messages? 😄

yet haven´t gotten an answer 😄

how would you know all that?

I was more specifically talking about SAI´s approach

or reasoning that for

Yes, I can see what you are talking about, yet I wouldn´t know for example whether there might be additional factors involved

[lavish lake]

they need to cover from potential lawsuits, government regulation and so on, that's why everybody does it but seems to be a bit too extreme. Imagine buying a camera and taking a picture of your wife on vacation to come out blurred, or your kids. It's nonsense

[bleak sparrow]

anyone here can make me lora ?

[loud solar]

But if a camera is used it's real ... if it's an A.I. it's not ....

So missuse of cams are worse than A.I.

[lavish lake]

the thing is that people abuse it, we saw it right here, tons of people trying to generate pedo stuff etc.. on the bots. it was pretty bad actually

[loud solar]

For sure you can stop it here ... on your bots. But our police official creates pedo stuff ...

And we can't stop everything that can be missused ...

[lavish lake]

I agree but also see their side, can perfectly picture a government action like "people use your software for pedo, or deepfakes, etc..." which is true, people will abuse it, you need to place some kind of control but i repeat, I think they went too extreme

[loud solar]

I can understand the poor people working for SAI who are forced to see all those pictures ...

The discussion can be endless and lead to nowhere ... just my personal opinion: keep public places clean but let them do what they want in private ... it's not real, nobody harmed as long as it's private ...

[shut terrace]

Yep

Ik that it is a genuine problem that they have proper discussion about even in calls with the CEO

I can't tell u how ik tho..

[loud solar]

A.I. doesn't make everything easier 😄

[shut terrace]

Yep

[shut terrace]

I know the CEO quite personally

And I can say than preo new ceo

There was quite a mess in SAI

But yes

The people in SAI

[loud solar]

I don't say I'd have a clue what I'm doing but my psychiatrist says I'm friendly ^^

[shut terrace]

Don't touch grass

[shut terrace]

Ok

[loud solar]

😄

[chilly dragon]

hello

how 2 use this ai ?

[loud solar]

'Install local or have a look here: https://discord.com/channels/1002292111942635562/1237461679286128730

[chilly dragon]

must pay 2 sign in?

[loud solar]

If you want to use the bot here, yes. AFAIK glif.app is still free ...

[chilly dragon]

and if i dont

[loud solar]

You can install comfyui local, download the models and use for free

[chilly dragon]

can send the link plz

[loud solar]

https://github.com/comfyanonymous/ComfyUI

Models you can find the most on civitai.com

[loud solar]

I would recommend youtube videos to learn the basics ...

[south sun]

You know the CEO? of SAI?

[shut terrace]

Yes

[south sun]

which one

[shut terrace]

SSW

Shan shan wong

[dark hawk]

Guys my discord is about to be verified

[Like 7 server needed]

It has TTI & ITI feature of V1

[shut terrace]

Not very personally buy enough that I know stuff that others wouldn't really know

[chilly dragon]

yea plz

[shut terrace]

I won't be outright leaking thingd

But yk I might drop some hints on big new every now or then

[loud solar]

I bet I know someone from SAI, too. I just don't know he's from SAI cause I don't care 😄

[shut terrace]

Ok

[chilly dragon]

@loud solar u had videos?

[loud solar]

I think the best idea will be to search for your own in your native language 🙂

[surreal jay]

I might know someone who knows someone who works there

[loud solar]

I don't really know the latest installation videos. But it's become pretty easy ... just don't look for older ones 🙂

[chilly dragon]

[loud solar]

I'm using it to long ... don't need those videos any longer ... the videos I watch wouldn't help you ...

I think there's a portable version of ComfyUI ... just add models and there you go ...

[chilly dragon]

ok ok thank u 4 ur time

[loud solar]

Hope you'll have a lot of fun!

[nimble canyon]

Didn't expect to end up in one list with Taylor Swift and USA government

[trim magnet]

tbh she kinda has a pony face

[loud solar]

I don't accept any SAI CEO next to me ^^

[trim magnet]

and most USA politicians are donkeys so its all a reference now

[loud solar]

Sorry for that joke! IT'S A Joke! I have nothing to do with SAI!

[shell cradle]

Do you know if they are still planning on releasing the other models sizes? (In particular 8b)

[wary belfry]

I heard yes. maybe another version of medium? apparently there where more versions in the making then announced

[hoary slate]

Is the stable digest no more? I just realized I havent seen any in a while and on substack its 10mos ago since the last post. I enjoyed those!

[wary hazel]

Is there anything like luma ai in stable diffusion anyone

[loud solar]

Stable is one of many players. There is not THE ONE A.I. any longer. And SAI didn't surprise the people in a positive way the last days ...

If you like cats with hat ... SD3 will be fine!

[trail lion]

Perhaps, but the open source nature is/was a good thing, a community formed around it, and that's what made it better than MJ, or other similar approaches

[loud solar]

MJ is something complete different. MJ is giving you something you can like but maybe wasn't your idea. Stable Diffusion is more a tool to create YOUR IDEA.

[gray fern]

unfortunately not, closed source is way ahead in t2v. worth playing around with animatediff though but it’s not the same

[loud solar]

And I bet MJ is earning more money than SAI ...

[wary hazel]

Thanks bro @gray fern

[somber orchid]

Will stable diffusion have license on everything now?

[gray fern]

it might be worth trying open-sora 1.2 which just released, but i oom’d on an A100 so… 🤷‍♂️

[quartz siren]

whats the best turbo sdxl model on hf?

[loud solar]

Maybe have a look on civitai.com and look for the votes?

[mortal sphinx]

Anyone know a free tool I can use to correct and fix an image thats 600x400 up to 2000?

I've used AI upscalers and some make a mess of the face

[loud solar]

ESRGAN-UltraSharp-4x.pth maybe? Or SUPIR Upscaler for ComfyUI ...

[loud solar]

If you need to correct something ... maybe an IMG2IMG upscale and there are other tools ...

[woven geyser]

Friends hello, I have a few questions about Roop unleashed and alternatives, Is this the right place to ask them?

[loud solar]

Maybe .... I have no clue about Roop ...

[woven geyser]

I just want to make sure it is related to community, and not against rules

Has anyone tested or experienced whether Roop Unleashed can effectively bypass face verification systems used in various applications like smartphones, security checkpoints, or online services? If so, what type of systems (basic vs. advanced with liveness detection) were affected?

[loud solar]

I don't think you can't talk about other technologies here ... or maybe only I am open for it. Who knows?

[woven geyser]

Is this discord server related to the reddit subreddit of stable diffusion? If so I have seen relevant posts about this

[loud solar]

Just do your thing. This is Stable Diffusion related ... but I think many of us are open for other things ....

[mortal sphinx]

I have automatic1111

Is it possible to do everything in there

[split sand]

guys i have a question for regional prompter unsure if this is the correct place to ask. so i noticed that if i increase the batch size to > 1, regional prompter stopped working any1 experiencing the same?

[loud solar]

Everything ... no ... but a lot ...

[astral goblet]

https://civitai.com/models/490622/mobius anyone know why is this 13gb and requiring clip skip 3? clip skip 3 is a weird requirement

[mortal sphinx]

I don't understand all you said previously with the ESRGAN stuff can you expand?

[loud solar]

It's good in cats with hat ... and other stuff ... but we are still waiting for community finetunes and tools ...

I didn't say anything about ESRGAN ... I have no clue about it ...

[mortal sphinx]

:f?

[loud solar]

Google is your friend .... or not ...

[loud solar]

That can upscale your pic ... but won't really add details ... so might become boring

[mortal sphinx]

I said I've used upscalers.

I want something that'll try to correct the image, or enhance it.

[atomic mortar]

a1111 can use a after detailer addon. faces, hands or entire people iirc

i use swarm so not sure on the specifics anymore

[astral goblet]

detailer nodes are on comfy too

[atomic mortar]

true but im not that far on comfy ui yet

lmao

still need to sit around and actually get into it

[trail lion]

adding details is one thing, you can do that with something like multidiffusion + controlnet tile, but fixing should be done in inpaint period, dont try to fx stuff with an upscale workflow

[mortal sphinx]

Thank you can you recommend any

[atomic mortar]

hmm theres a addons tab that shows one iirc

within auto1111

its litterally called afterdetailer lol

[cedar salmon]

fp32, there is a fp16 there too more familiar size

[trail lion]

without being the author, I doubt anyone knows, but the simple answer is it was trained with those paramaters. I can speculate that this was done to remove any kind of bias ( the goal of the model ), but that's just a guess in the end. the downside of using this, it may struggle with loras trained against other models, even more than usual

[astral goblet]

it's a completely retuned text encoder as i understand it, so that would kill lora compatibility too.

i'll give it a shot

author is datavoid and i got caught up in some misinformation rumors he was spreading and my trust is now paper thin

[soft iron]

https://github.com/comfyanonymous/ComfyUI_TensorRT

Nvidia seems to have created this technology for me

I usually don't use lora or controlnet

and just get x2 speed under the same conditions

[astral goblet]

have you used it much? it also doesn't do multiple resolutions very easily. unless you train a dynamic version of the TRT model. then it's not as fast.

The gains are better appreciated by datacenters.

[soft iron]

I always use the same resolution

[astral goblet]

should work well for you then. it accelerates things so hard.

[soft iron]

Recently a fast model came out that suits me and in combination I can generate an image in 1 second

I literally did the same thing half a year ago, but on SD 1.5 with LCM and now I do it on SDXL with better quality and resolution

Progress does not stand still if know what to combine correctly

[astral goblet]

yeah i still use LCM loras. especially when i'm testing animatediff prompts

[fervent thunder]

How do I get started? I want to generate an image…

[pine fiber]

do you have a gpu

[shell cradle]

• If you want the easiest GUI, find an installer for Fooocus, or one of the variations.
• If you want something powerful, with options and extensions but intermediate complexity, find a installer for either A1111 or Forge
• If you want the most flexible and powerful option, which also comes with a higher degree of complexity, find ComfyUI (or StableSwarmUI, which gives you a cleaned interface with the option to dig deeper into more complex workflows via built in comfy)

[shell cradle]

Most GUIs can even run on CPU nowadays, right?

[pine fiber]

if you want your images tomorrow then yeah

[soft iron]

no

when I didn't have a video card I generated on the processor

with different optimizations you can generate images in 20 seconds

[pine fiber]

no shot

I dont believe u

[trail lion]

I really dont get the point of these fast models.. why would you want a lesser quality image even if it's faster? and even if for a specific use case it worked ok, there are so many tuned models out there that arent turbo or whatever, so it's pointless

[pine fiber]

maybe with like 4 steps

[soft iron]

of course, but there are good LCM models

[pine fiber]

I dont see the appeal

the quality is probably really bnad

[soft iron]

good quality can be found anywhere

limited only in capabilities

[trail lion]

what does that mean

good quality can be found anywhere, I agree, but not everwhere, so if you have this turbo model, it means not necessarily

[soft iron]

the best realistic pony model in my opinion, works on lightning out of the box

[trail lion]

and again, there are concept finetunes because the base models and the turbo models just dont know how to produce certain things

[soft iron]

#💬｜general-chat message

all funtunes are always a degraded version of the base model

it all depends on what you need

[trail lion]

degraded...no they are introducing concepts

so if the base model cant produce a ferarri, you can finetune a ferarri as a lora

or say you want more realism...like skin details

[soft iron]

in the end your model will only create Ferrari

[trail lion]

if it's a lora that's fine because that's all you would use it for

[soft iron]

"and again, there are concept finetunes because the base models and the turbo models just dont know how to produce certain things"

fast models are the same models as regular ones

[trail lion]

if it's to add realism or something, such as what was done with juggernaut, it's a full training

[soft iron]

most often it are worse in quality, but it are also cut down relative to the main model, like other models

[trail lion]

so without getting into semantics here, my personal workflow, I find myself hopping between models very often, even for 1 image

different models can draw things better than others, like multple subjects, or realism, or whatever...so I guess my original point about turbo, there's no 1 model that I've come across that I could use every time

even juggernaut, which I consider the best finetune..maybe ever

[soft iron]

Juggernaut is terrible

I download its new version every time, laugh and delete it

[trail lion]

well, my opinion

[soft iron]

I also mean for my purposes

[trail lion]

and you're using some turbo model that is better, I must know this

I 'll generate some side by side images and post them

[soft iron]

it's not better, it's better for my purposes

Juggernaut does not know different clothes and can only do standard human poses

so it's useless for me

[trail lion]

yah, see that's a great point...clothing is another example, and poses, and realism, these are all areas where models differ, so why I was confused about turbo use case, is becasue you use the specific model that suits what you're trying to create, ie can it draw that

seems to me turbo is more for a commercial use case for customers asking for " a photo of a cat "

[sage reef]

i wonder what kind of signal it will be :3

[astral goblet]

LOL i've said that civit was under botting seige from people trying to earn kudos and just flood images from pony months ago and i got a lot of shit hucked my way. Now today they're making huge changes to fix botting.

Vindication always feels great.

https://civitai.com/articles/5800 people here are pissed and think it's their god given right to bot

[iron pendant]

#🆕｜sd3 message

[soft iron]

no, people just created a lot of accounts to get buzz

[fervent thunder]

I never looked at what you could get with buzz

[soft iron]

train models

[fervent thunder]

I never got that serious with AI I guess. If I can generate a pretty good image every now and then, and have fun in the process, then that is a win for me. I do appreciate the time others spend to allow me to do so.

[astral goblet]

yes. people create bots. that's why the whole change. there's a massive bot problem on civit. it's been obvious for months.

[soft iron]

this is only bad for the owner of the civitai, but not for people

people train new loras and develop the community

[astral goblet]

turns out they think its bad since thye're now going to start enforcing against bots going forward and have developed bot detection systems

[soft iron]

Of course they will do this because they are losing money on this

[astral goblet]

it also affects the user base since people post art to share with people and it feels like 100 bots are voting them up. Dead internet theory reinforces itself in these situations and people lose faith.

Try playing a game of TF2. You can't enjoy it since it's just bots playing and they're idiots

[peak orchid]

the first post of the day must be sfw was a dumb rule, but yeah they need to crack down on the bots.

[astral goblet]

lol that was so dumb an arbitrary

[soft iron]

dead internet is wonderful, we'll see it soon

[astral goblet]

we'll go back to BBS's

[peak orchid]

They also previously had issues with lots of pro-Russian government and anti-Ukraine sentiment last year around this time I think, and that could make them an easy target for politicians and governments

[astral goblet]

they'll just be state of the art future BBSs

[ancient spruce]

https://civitai.com/models/526316

[soft iron]

shit

[ancient spruce]

true

[modest gull]

any working colab version of stable diffusion video? Found this and not working: https://colab.research.google.com/github/mkshing/notebooks/blob/main/stable_video_diffusion_img2vid.ipynb#scrollTo=9AZDrh-SUDt2

[fervent thunder]

damn you're very delusional lol

what you're saying is:
SDXL base is better than:
Juggernaut, Proteus, RealVisXL, Playground, Pony

maaan...

each one of them in their use case they are SO GOOD,
meanwhile base sdxl is like an average joe, expert at nothing, average at everything

[soft iron]

wut

pony is the best base model now

juggernaut finetune with terrible style

it generates generic images and that's all it can do

RealVisXL really good for its purpose

Playground just an outdated model, SD3 is better

[fervent thunder]

Pony is a massive finetune on top of SDXL

pony wasnt trained from scratch

[soft iron]

no

[fervent thunder]

@nimble canyon Was pony trained from scratch?

[broken smelt]

i dont think this person knows what a base model is lmao

[fervent thunder]

probably not, I dont want to be mean tho, so I just pinged creator of pony to be sure

but like, from what I know,
even starting from base sdxl, is better than starting from randomized or zeroed weights

you converge faster and you don't need to teach basic fundamental image characteristics to model

The difference of pony to most other sdxl models is that its dataset is so massive and so different than the usual data that other finetunes (and base sdxl) has seen, that it has an entirely different sorta "internal", and thats why if you do 50%/50% merges of raw ponyxl with some random sdxl finetunes... then it might just become pure noise

[soft iron]

no matter how it was trained, such serious work was done there that it is a basic model

[fervent thunder]

I'm not dismissing the work

[broken smelt]

lol

[fervent thunder]

It is still a finetune on top of sdxl, a massive one, but still is

Also, the pony model would be an entirely different model if it was not based from sdxl,

if it was trained on randomized or zeroed weights, it would converge to a tooootally different state/behavior

Like, starting from some sorta weight is always better than starting from zero or random,
that is shown and testes in PIxart Sigma paper, they literally reused 1:1 parts of Pixart Alpha, but upscaled the tensors and did some techniques to reuse those, without having to retrain from scratch.

Like, before the sigma training, it would be a more parameters and bigger pixart alpha BUT the images would be like... a gaussian blur applied 100x, you could see outlines of the subjects but all blurry,
then after the sigma training it "honed in" at a much earlier stage than it did when they trained base pixart alpha

[pale latch]

it's not a base model. it just doesn't work with all the other base model resources because they killed the tenc in training and disaligned it. Astraliteheart has never trained a base model. That's a whole different field from refining existing models

[fervent thunder]

I wouldn't say disaligned or killed,
they quite "realigned" and "altered" it for their usecase in their models' circunstance

[pale latch]

civit has caused lots of confusion here by creating a whole new base category for pony so that their porny uses would stop emailing support

[fervent thunder]

regardless of why a model is made or what content it produces, under the sun I only care how different they are from base model, if they are endless merges or not

[soft iron]

you have no one to talk to? You understand perfectly well what I'm talking about, but you continue to write meaningless walls of text

[fervent thunder]

I dont like bland as-is merges,

like, I completely support doing merges BEFORE finetuning, to help a train converge faster,
or AFTER A FAILED finetune, just to salvage a bit of that gradient advancement,

But when people just merge two preexisting models, and then call it a new model... eeew

[soft iron]

💀

[fervent thunder]

wait, did I click in the wrong app? I thought this was discord... you know... a chatting platform

goofy ah skull emoji

[pale latch]

it's no longer aligned to the previous base model is my reasoning

[fervent thunder]

oh ye true true

kinda like rotating some hyperdimensional hologram, the internals don't match 1:1 anymore, but it still is on the same structure

[soft iron]

I'm expecting a good model from you on Civitai

and not meaningless debates about what is the base model and what is not

[late surge]

hi

[broken smelt]

why are you upset when people correct you on what a base model is?

[soft iron]

because the community of people who use models and don’t discuss all kinds of nonsense is called the pony base model

[astral goblet]

because his identity revolves around pony and he always thought that astraliteheart was on the same level as lykon and other base model trainers. ? guessing here i don't really know. insecurities can come from a million things

when people hear something that contradicts what they know, a dissenting opinion, it's weird what the brain does. The fight or flight response fires up. Researchers have shown this effect on fMRI scans

[pine frost]

If they are not in control of their emotions flowwolf. Don't generalize

[soft iron]

What? Astraliteheart is essentially the only one who has made at least some progress from the basic models. There is not a single person in the community who is close in his level of contribution to the development of the 1.5/SDXL model.

[astral goblet]

fight or flight response still happens. Our rational minds can overcome it ez, but some people don't. This is a human fact. We're all the same animal.

[astral goblet]

Disagree . Hard.

[pine frost]

Facts sure

Dream on

[astral goblet]

If you think you're not tied to your biology and ancient lizard brain still, boy oh boy you're in for a surprise

[pine frost]

You have no idea what the mind can do

[astral goblet]

I'm going to go ahead and qualify this one as case in point.

[pine frost]

So do not generalize

Thats all what im saying

[nimble canyon]

It was not trained from scratch (but it was trained in a destructive manner which pushed it further from the base model than a typical finetune would do)

It is also a base model now, even if it managed to achieve this in a non conventional way.

[prisma rover]

I just downloaded Pony today, haven’t got around to messing with it, so is it any good? If I wanna make anime chars etc

[broken smelt]

you really expect him to say anything other than yes here lol

[prisma rover]

I normally use AAM or YamerXL

[prisma rover]

Oh is this the creator of pony? 😅 shows how much I’m in the loop these days

My bad haha

[nimble canyon]

It is ok for anime but there are pony based mixes that do anime better.

[prisma rover]

Thanks, I’ll look for those too

[hoary slate]

sorry, my bad, I meant the newsletter they were doing

[loud solar]

Ohh... I only get the mails that I can use a model now ... after release ^^

[formal vine]

111

[loud solar]

42 is the answer

[upbeat kiln]

Hey guys, I have a question regarding the licensing of AI images. Can I license the images generated by stabilityai/stable-diffusion-xl-base-1.0 under a Creative Commons license (CC-0-1.0, CC-BY-4.0 or CC-BY-SA-4.0)?
The stable-diffusion-xl-base-1.0 model is licensed under the CreativeML Open RAIL++-M License. Thank you!

[vocal leaf]

are there any new ones

[plucky crypt]

a chinese man

[vocal leaf]

what

[vocal leaf]

a succulent chinese meal?

[plucky crypt]

yep

[vocal leaf]

generating image: 42.5/100%

loading shaders: 59/100

[pearl ocean]

Would you call Homlander a nice dude?

[delicate forum]

can anyone help me please

[frail sonnet]

@delicate forum I use it all day everyday have been for a year now, nonestop, no problems at all.

[delicate forum]

then why it happened to me , i dont understand what cause the problem

[frank fern]

He is the most relatable

[sudden ruin]

Mommy issues?

[fervent thunder]

I forgot about his mommy issues

[frank fern]

https://tenor.com/view/tasting-milk-antony-starr-the-homelander-the-boys-lick-gif-17834498

[fervent thunder]

omg thats gross 🤣

[sudden ruin]

Thanks I hate it

[astral goblet]

i dont know why i clicked that. thankful for it not being one of the more risque scenes involving his boobsession.

[frank fern]

Kudos to the writer, for creating such a polarizing character.

[astral goblet]

I agree. It's ahrd to write a villain people love as a villain. Some of my favorites are Cruella DeVille (if she doesn't scare you i dont know what will) and Geoffrey

and Simon Gruber

[frank fern]

I think Omni-man is also pretty scary. I really like evil super "heroes".

[frank fern]

I couldn't resist sharing the most toenail curling gif I know. haha

[astral goblet]

Yeah true omniman landed pretty good but ||he's got kind of a redemption arch and J.K's voice makes him so loveable||

[frank fern]

He's portrait very moraly gray, that makes it even better.

[fervent thunder]

It makes you go uhhhh in the back of your throat

😆

[astral goblet]

The poor actor who plays Joffrey Baratheon. He'll never have another big role because he's so hated. Funny too that his name is Jack Gleeson, synonymous to another controversial star. Bam Zoom Straight to the Moon!

I'm rooting for him though. His performance was extraordinary

[surreal jay]

Winn Adami (Kai Winn)? cringes

[astral goblet]

looked her up.Yeah she's hateable

Speaking of DS9, Dukot was a great character arch

Dukat? yeh that one

[surreal jay]

Yea I kept switching between love and hate for him, they did that well

[shell tendon]

is it just me or did all dreamer communities except anime get deleted? along with all archived communities?

[fervent thunder]

Don't mess with anime!

[surreal jay]

I noticed that, too

[fervent thunder]

You see what happens if you do...

Anime > ???

[vast ingot]

what the hell ??

[coral lark]

They weren't safe enough 😔

[surreal jay]

bring them back as text channels!

[vast ingot]

none of the communities ever have been hard to moderate (except anime from time to time honestly), why would you delete those ? and the archives even more, those were just nice place to see pictures, nothing to moderate anymore

[soft iron]

the right decision, need to leave only useful content

[honest mica]

I just wanted to look something up in the old cascade channel but it vanished. Sus.

[vast ingot]

I do spend a lot of time in those communities, and the archives. It has always been a nice place to look at things, especially in times of drama in other channels... I hope this is just temporary restructuring, but the disappearing of the archives is worrying.
It could also be just a discord bug, we have had that in the past from time to time, but this seems too targeted on specific channels with low activity.
I've opened up a ticket to know more from the mods/admins, in case there is some information on this we missed, this seems like a major change in the discord server

[honest mica]

Channels with low activity? Well, why is #🏆｜winner-gallery even a thing anymore? That channel is made out of dust.

[vast ingot]

it has never been an active channel, it has always just been the "trophy" channel, and is the summary of the whole "Events" category. Nobody but mods and higher can post there
I don't want to speculate until we know for sure what's happening to be honest... I'm just uber sad right now....

[surreal jay]

in my experience, I never used them because I thought they were cumbersome as threads vs channels

[vast ingot]

it has been a discussion (and a back and forth) for a long time in the server : what was the best format for it. hard to find the best way. but I guess no channels at all is the easiest to navigate now

[surreal jay]

yes lol, nice n clean

[forest trout]

Yep. Worst of all the old archives are gone too.

#👥｜roles message

This is essentially vestigial at this point.

[vast ingot]

it's one of the things I don't get... Deleting some unused channels is already sad (archiving them is better), but could be understood (to increase visibility for channels that do get traffic)
But Archives ? Those were not seen by anybody that did not ask for that role, they were not hindering any visibility at all. I hope this is just a misunderstanding, because I don't get that change at all

[forest trout]

Ah yes dreamer communities is now just "anime". Apparently the only one that matters.

I had a few great gens I posted on those communities. Gone.

[vast ingot]

to be fair, the only one that did get some real traffic. it was more than 99% of the traffic of that category

[lavish lake]

I think the channels are not deleted, just they aren't active anymore, content should be there if they want to bring it again. I think this happened a while ago, maybe couple months

[forest trout]

Deleted or not active, it really is essentially the same thing if none of us have the ability to access them.

[honest mica]

It's more the archives that are interesting

[vast ingot]

not active is better, it can be changed back. But we have no way to know for now, let's wait for some answer...

[lavish lake]

I believe Fruit should know more about this

wont ping her here since it will be lost soon, maybe DM her

[honest mica]

You can't DM fruit

[lavish lake]

or at Stable society, she's usually checking that one

[vast ingot]

you can if you're friends with, but yeah, DM are blocked when you manage such a discord. I was moderator too, and had to block DMs.

I have opened a ticket and it has been claimed, they seem to have seen our worry. Hopping for a response not to devastating

[loud solar]

Good luck ...

Some people are sad cause of the cascade topic, too

[vast ingot]

but if there is a devastating news to make, I'm sure they'll make it on here soon

[lavish lake]

Almost everyone I knew when I was modding is already out : (

[forest trout]

First time?

[junior urchin]

Devastating?

More devastating than being $100M in debt?

Since they didn't announce that

[forest trout]

Ah. So that's where I recognize you. You're from Wombo.

[storm shard]

😢

[karmic brook]

hi!

[sudden ruin]

Most staff turned it off cause you get spammed on a daily basis

[honest mica]

Yeah. I would definitely do that too.

[sudden ruin]

idk if its worse for me, but I get mistaken for a bot a lot lately

[desert dagger]

might be your name?

[honest mica]

Kinda unfortunate name 🤣

[sudden ruin]

Yeah didnt think that through

[forest trout]

Time to rename yourself Botox...

[honest mica]

So, let's talk about some archives...

[sudden ruin]

Everyone watching fruit type

[honest mica]

🍿

[vast ingot]

preemptive sadness though ...

no popcorn for me

[forest trout]

There you go

I'll share

[vast ingot]

I'm happy you guys keep up the humour though

[sudden ruin]

Im sure we can work something out

[honest mica]

Was that an official statement?

[forest trout]

The official statement was "hi!"

[sudden ruin]

Please dont screenshot me and post it on reddit

[surreal jay]

it was written!

[honest mica]

Reddit post is online. Boto confirmed everything.

[lavish lake]

https://tenor.com/view/hatsune-miku-miku-hatsune-earthquake-plush-miku-death-gif-4018907532159793300

[sudden ruin]

Shes just headbanging to babymetal

[vast ingot]

their song with ElectricCallboy was fantastic

[karmic brook]

i'll reply in here, so it's viz
the art communities were ridden of as they were not breathed into for.. nearly 2 months straight. We rid of channels that collect dust this long. A lot of what was in those channels were bot spam. #🍥｜anime was kept as it is the most active channel & has been pretty historically. The rest of the server is still a great place to post generations & sharing, but we found it just gets too hyperdetailed putting everyone in their own containers.
we're trying to keep most active & focused channels in here, since previously we had the notion of hoarding every channel in archive, max capping categories for the sake of keeping.
archives were cleared, as they were not meant to be a forever thing, just as the bot archives last yr and before. Sure, they could be synced back, but trying to be more focused & organized.
apologies to anyone saddened by this & this is all i can say & reiterate if brought up again here ty

oh, and

[karmic brook]

u were a mod?

[forest trout]

I guess not.

[shell tendon]

Ahh come on. I and others have found ways to get good generations out of cascade and been using it consistently... That at a minimum should come back

[junior urchin]

A lot of what was in those channels were bot spam
Automod with community server enabled, can prevent basically 95% of the bot spam

[honest mica]

And it begins. 🍿

[karmic brook]

yup, i mean people trying to prompt in them

[shell tendon]

There was a lot of good stuff in that channel that would be nice to have as a reference

[junior urchin]

ah, dyno then :P

[karmic brook]

Sure, they could be synced back, but trying to be more focused & organized.

just as ref before anyone burns me alive lol

[trail lion]

another discord I'm on created a honeypot channel, if you post in it, auto-ban, to catch the spam

[shell tendon]

Its good to know that that is the future, any and all technical discussions should happen on a different server where you can count on that information actually being persistent

[trail lion]

I thought that was clever fwiw

[honest mica]

#🍥｜anime is probably the hardest to moderate unless it's really ment to be an FBI honeypot.

[sudden ruin]

Thats cool but I also think people would abuse it or use it by mistake and then come crawling back to appeal

[shell tendon]

Yep... Lolicon garbage every day all day

[junior urchin]

Wildcard for dream, imagine etc., then wildcard for common nsfw terms then :P

[shell tendon]

I don't know how that is somehow okay

[trail lion]

I'm sure it happens, but there's a big announcement saying dont post in there, we wont unban you

[shell tendon]

That shit is an instant ban on a lot of other servers

[sudden ruin]

They are 200 years old I swear

[shell tendon]

Lol

Why don't we delete the winner gallery channel? Nothing in there since last November

[honest mica]

Yeah. Nobody wins!!!

[vast ingot]

Those were not bot archives though, they were user crafted pictures, shared by subject. They were important to the people participating in those, and the people looking for prompt inspiration making pictures today.
Keeping the most active channels around, to feel more alive, is very coherent, but archives are an hidden category that only people wanting to see them can see. Removing the older channels in there if the cap is hit could also be coherent, but a full wipe is a lot more surprising.
Of course this is your right and all, but this cuts the good times users can spend in here. Those were places people spent time in from what we can see, from the fast and already numerous reactions from people in this server since a long time ago. When I see how much of us noticed and started talking about it, just minutes after it was done, while no notification had been sent, I can only see that those were NOT dead spaces

[sudden ruin]

I miss those event times

[shell tendon]

Only something like 5 generations in #1237459938901491852 over the last two weeks - can we delete that and use that channel slot for #cascade?

[shell tendon]

Yeah it'd be nice to have that

[honest mica]

Boto, you are welcome in #🔆｜dailies

[sudden ruin]

Its cool how that channel evolved but its also not the same

[surreal jay]

perhaps there's a chance more ppl would have used them if they weren't threads?

[sudden ruin]

Back in the days they were channels and pretty active, making them threads kind of killed them sadly

[shell tendon]

Yeah I emphatically agree with this... All this does is drive home the point that we shouldn't post anything meaningful or particularly useful on here without duplicating it elsewhere, which is a pain, so we should probably just be elsewhere

[surreal jay]

speaking from personal experience as a relative newbie to this discord, that's why I never touched them. Although I probably would have

[sudden ruin]

Totally fair

[shell tendon]

And honestly it's just more SAI stuff with not checking the pulse of the community before taking action

No communication

[lavish lake]

This is a very vocal and active (sometimes even mean) community, things have to be spoken really clearly or people go bananas

[honest mica]

There's also a pretty active cascade loving community!

[lavish lake]

Cascade is a deprecated technology, the thing is that the community is making it better already. But is akin to having a SD 1.4 channel, that would be outdated

[shell tendon]

It's just that... It would be good to check to see if ppl have really strong opinions on keeping something... That imposes no burden to keep around...

[shell tendon]

If cascade is deprecated, sdxl is super deprecated

[trail lion]

like a poll? how would they 'check'

[shell tendon]

Cascade trains better than sd3 and is pretty interesting

Ppl on here are still posting cascade gens daily and we've been picking apart optimizations on other servers or in DM

[shell tendon]

Literally just hop in here and say hmm something we've been thinking about is maybe this.... And then step away for a while

Check back in later and see if there was a big reaction or not

If there was, maybe DM one of them that is capable of being respectful and not acting like a childish asshole for clarification, so it doesn't turn into a mosh pit

[lavish lake]

Yeah but was it really active when it was live? the cascade channel? See why anime is still on, it has multiple posts per hour! If a channel has only a couple daily post is a drag to maintain, we have to try and see their side too.. and specially having less people to keep track of everything

[karmic brook]

what..?

[shell tendon]

How is it a drag to maintain?

[honest mica]

Cascade was super active

[vast ingot]

Well, I've said what I could... Those aren't our decision, and if you really need to admin this way, then so be it...
For such a wipe in the archives, a head's up would have been at least humane. There was a lot of work in there, lots of value, lots of entertainment, lots of memories I'll never get to archive on my computer now.

[shell tendon]

I'm just talking about some of the way ppl have behaved in sd3... Pretty poor behavior, a discussion happening in front of a crowded room seems to be going poorly lately with the huge influx of ppl

[karmic brook]

like i said, they are able to be synced back. apologies for not being humane enough

[honest mica]

I don't think he meant it like that

[storm shard]

It makes me wonder how long this discord will stay alive if it's this easy to close things apparently.

[vast ingot]

Really not meaning it that way, sorry... I'm not trying to be aggressive here, I'm just sad to see those nice places go like that. poof. Apologies if it was taken that way
I don't know how to read your "they are able to be synched back". What could we do to ask for such synchronisation to happen ? Would a role "archivist archivist" be more suited ? I'm trying to understand, I wouldn't have done a full wipe but this seemed to be the solution the team took, so yeah, i'd love for a backpedal on that one, and so are the few people talking about it here it seems, but I know 5 people on a single night is nothing, it's not representative of anything here, and you need to make rational decisions
I know you aren't trying to hurt anyone, just to manage a big hard thing to manage, and sometimes hurting as colateral... But if there is a way to mitigate the hurt here, I'd love it. I personally did spend so much time making sure this was a safe place for anyone to enjoy too, and I know I've been out for a while now, but I still feel how hard it was. Like I said in my ticket, I don't want to be a bother here. i'm just real sad and not understanding

[lavish lake]

Maybe a middle ground could be negociated. I understand both sides, to ones those are lost memories, to others they have to run a tight ship with very few hands. Maybe let them open 24h so people can download and backup? @vast ingot ?

[regal sundial]

how far is SD from a way in which a 3D model is either imported or generated, and then can create a 2D or 3D Art based on that model and how it's positioned and its referenced camera angles as a means to avoid deforming of bodies/hands/eyes. etc. I feel like 3D Models would solve many of the problems with AI Prompt Art, especially if one could just accomodate the angle of the 3D Object, perhaps even fold/pivot the arms/legs/hands/fingers, etc. before having the AI draw the prompted environment. So in other words it has the referenced 3D Object which it will not change, maybe it will know to define the hands/fingers, and so fourth and know to leave this 3D Object alone without deforming it while focussing more on the surrounding environment

[junior urchin]

You mean Zero model working as a 2D or what?

[vast ingot]

I'm all for "negotiating", but the truth is, we have no say in those negotiations, except express what we would prefer. So I've done so. I hope people missing the archives will do too. I hope those will be enough to show some interest in there.
And I do understand the need to create space in the archives if there is a cap, and removing the oldest archived channel when a new channel needed to be archived would have been my guess. I'm not being against anything constructive, like I said in my first message, it's mostly that full wipe that is surprising to me.
I'm all for any middle ground for having still a space a paradise on here, where no drama happens and I know I can just feel good looking at old pictures and prompts in this discord I love.

[regal sundial]

So basically it'd be like any other prompt art generator except if you choose to add a model (as a means to not mess up the proportions of limbs/body parts) it will just put the model inside of the prompt (environemnt) and either turn the snapshot of that 3D Model in the environment into a 2D Art, or keep it 3D

The idea is that this way, the AI will not attempt to generate its own humans but rather to identify the 3D object as the human, place it within an environment, snapshot at a referenced angle, and then refine the surrounding environment with prompt art. The point is if you let AI specialize only in the environment and the snapshoting and then upscaling, especially if it knows for eg. like a wire/mesh skeletal understanding of the model, it will know either to not touch it (and add more fingers for eg, or add more eyes) because for eg. human body is defined as having only 2 eyes, 2 ears, 5 fingers per hand etc. etc.

on the reverse side, it could already use 3D models of human bodies and from that training, even build its own 3D Models which it could then use for the same purpose. I'm not too good at articulating this I guess but 😆 i hope someone gets the idea

just curious to know how far we'd be from that

I'm only using humans as an example but the idea is 3D models of anything, like animals, dragons, etc.

[desert dagger]

just a thought, but you can always start your own discord and copy images you want to keep to it

[forest trout]

But the whole point of posting the art is for other people to see them too.

[vast ingot]

plus this is supposed to be the "place to be" to share anything SFW related to SD. The point has never been to move the activity on another server. There are tons of other ones already. People do come here to look for #📝｜prompting-help or for #🤝｜tech-support , and share their pictures in general or, once upon a time, in the communities

[desert dagger]

and they saw them. but you're wanting a forever repository, so create your own discord, use it for that, and then repost as needed what you want others to see

[shell tendon]

Ideally we don't fragment the community any further but yeah

[forest trout]

I mean I save my gens on my hard drive.

[hushed quarry]

believe it or not I was a mod too

[forest trout]

Lots of exes here today...

[vast ingot]

Not necessarily forever. I'm sure we can all understand that things need to be deleted at some point. Hell, discord does it itself after some time in any channel.
But there are middle grounds between what we had 2 hours ago (lots of archive from the last few years, since SD came out) and now (none)

[lavish lake]

The O.G's are IN

[karmic brook]

father diffusion

[hushed quarry]

[desert dagger]

discord doens't delet stuff. it'll hide stuff if it's inactive for a while, but it doesn't delete anything.

[vast ingot]

my bad then, I thought the oldest messages in a channel with too many messages got deleted after a max

[lavish lake]

nah it's all there

[desert dagger]

no. in fact, that's one of discord's core values - they keep everything forever

[cedar salmon]

its training a LLM

[lavish lake]

and I mean, we can all post images on most channels, like https://discord.com/channels/1002292111942635562/1004159122335354970

[shell tendon]

Biggest thing to me is being able to look back for technical info, old workflows,etc

[desert dagger]

but channels get too large to scroll back through, and if they're busy, that can be just a few minutes

[shell tendon]

And that's what search is for

[desert dagger]

and discord search is lousy

so make a private repository, post things into it you want to keep, and post them out once in a while so others see them again.

[hasty hornet]

didn't they said they're gonna delete files after a while?

[shell tendon]

Or don't delete things that ppl can't even see unless they want to look for it? Isn't that easier than everyone keeping a private repo of everything they think has some chance of being worth looking back at someday? Lol

[placid hatch]

What is currently the best pony derivative?

[desert dagger]

that only pertains to accounts that are inactive for a couple of years, not content posted to servers

[hasty hornet]

hmmm aight

[nova lily]

anyone mind telling me about these other apps, focus comfy invoke are they easy to set up vs A1111?

[forest trout]

Comfy is a one click install.

[lavish lake]

yes they are easy to setup, but they all have other workflows and ways to do stuff

foocus is probably the easiest one, comfy the most powerful

[vast ingot]

anyway, discord deleting it or not wasn't the core discussion, mostly around this discord keeping its archives up or removing that category. Sorry for having changed the subject.
I guess you are right @desert dagger we can all just save everything on our hard drive, and we'll do it a lot more from now on.
And yes, #🏞｜general-with-images has some great pics too.
But the archives did have a lot more too, sorted by type of art/subjects, and were not polluted with discussion and drama. Having to clean up things is logical, but removing the whole concept of archives is another. As Fruit said, there were wipes in the past, but wipes of the whole category I don't remember

Time for bed on my side, I'll go, and I can't add anymore to the debate I think. Have a good time people, see you soon, and I hope our voices can help restore at least some of those old golden mines that were in there.

[lavish lake]

invoke and swarm are really nice UIs

[hasty hornet]

invoke probably most newbie-friendly of them all.
Comfy UI won't be hard to start, just need to make \ take someone's workflow and ready to go, for more complex things might take some learning

[desert dagger]

you could, but if you make a discord you have unlimited forever storage space

[nova lily]

thank you, do they actually produce better results if you use the same model or do that just have additional tools that bring the better possible reulsts. Im just trying to figure if I should check them out I only ever used a1111 thanks for comments

[lavish lake]

For some stuff, they will produce better results than Auto1111, but all of them can produce really good work

As always, It's not about the tool. There's amazing A1111 art

[nova lily]

Focus easy, that mean comfy is hardest to use?
wats the difficulty tier list 👋
I already know how to "use" A1111, im not an advanced user

[lavish lake]

Not the hardest but it's a UI based on nodes, if it's your first time, takes a while to understand how to setup stuff, what the nodes do etc

That's why someone suggested you can try downloading other people workflows and studying them

In the official repo, there's a lot of examples for most cases, you can download them and learn> https://github.com/comfyanonymous/ComfyUI_examples

[nova lily]

oh is that the one that looks like an unreal engine workflow

[lavish lake]

exactly! like a blueprint

or like a 3D shader/material

[nova lily]

yea, it is, that one may be beyond me im afraid. xd

[lavish lake]

foocus is really nice, try that one

also invoke and Swarm

swarm uses ComfyUI as backend, you can do everything but with an easier/nicer interface

[warm junco]

They are all easy to setup. I made a few guides for each. You can find them in the pinned messages of #🤝｜tech-support

[nova lily]

this page? https://github.com/CS1o/Stable-Diffusion-Info/wiki/Installation-Guides

thats amazing, will use undoubtably

[willow crescent]

Easy Diffusion is the easiest if you ask me, even though it doesn't support Stable Cascade and SD3 implementation still being in the works

It's as well a one-click installation. I feel like perfect for Beginners. It's also got a very pleasant in-painting/drawing interface that can also ne combined

https://easydiffusion.github.io/

[nova lily]

ill write that one down

Just saw that "--no-half-vae" in your guide, that new havent seen that line

[warm junco]

Yea I will add InvokeAi and EasyDiffusion too in the next days

[nova lily]

I just added the no vae and the vram setting to my file thank you.

[willow crescent]

There are greater benefits of a collective archive comparably, where everyone can access workflows and alike from others, neither a private DC nor a local HD can.

[main zenith]

Is there a server dedicated to AI influencers

[fervent thunder]

Is there such a thing?

[south sun]

yeah this one. you're an AI influencer! congrats

[fervent thunder]

I love those AI accounts on instagram who put watermarks on their images

[main zenith]

Thanks guys

[fervent thunder]

I was joking with you. I am not sure there really is such a thing.

I think you mean as far as the people who perhaps, write the AI programs?

[soft iron]

When will SD4 be released?

[pale latch]

oof. this is not a healthy way to communicate . if you want people to cooperate with you, don't imply they're all childish assholes in a sweeping public statement maybe

[shell tendon]

uh, there are ppl who did behave that way toward the devs

some ppl were even threatening lykon iirc

it got pretty bad in the #SD3 channel a few times

pretty nasty tone

i'm not trying to get anyone to cooperate with me, and i'm not referring to everyone, i'm referring to a specific group of ppl who can't seem to communicate disagreements without getting personal or aggressive in their attitude and tone

[desert dagger]

there have been several, but most of them have become inactive.

[pale latch]

earlier today i saw you on onetrainer server smack talking lykon. seems insincere.

[soft iron]

Lykon is a bad person who said bad things, what else can be said

[desert dagger]

there are certain people on this discord that are here specifically to spread chaos, try to stir up trouble, and damage the company. it's fairly easy to see who they are if you scroll through this channel and the SD3 channel since the SD 3 release. and read posts.

[shell tendon]

smack talk? you're kinda misrerpesenting things here

i quoted my response to him on here from earlier today explaining why it's best not to leave us to explore training methods for ourselves if those methods have already been worked out, and said:

"it's like pulling teeth and i don't know why"

i am completely against how he was treated, ppl made it personal and went after him personally and some real lines were crossed

it's possible to disagree with someone and still respect them as a person

[mortal delta]

im curiouse but what stable diffusion interface do you all prefer or use?

[soft iron]

why do you need something other than comfy/stableswarm?

[mortal delta]

i dont im just curiouse what others used.

[soft iron]

everything is simple, for those who want to do everything in one click - focus. For everything else comfy/swarm.

[mortal delta]

that makes sense but i feel like focus isint as custimizable as something like comfy or 1111.

[soft iron]

since the swarm developer works together with the comfy developer, I think that all other interfaces will soon simply die.

[karmic brook]

enough about Lykon

[pale latch]

i doubt that. most people don't like node graphs as much as html fields. automatic1111 is the most popular of the UI projects and i don't see it's majority being displaced quite yet.

[mortal delta]

there are some dead interfaces but i hope the big ones like 1111 or comfy dont become dead either.

[pale latch]

i think comfy has a brand new day in it's future.

[mortal delta]

cool, i just wish something like 1111 ran smoother or wasent considerd bloated.

[pale latch]

don't listen to the bloated claims. Most comfy users have 10x the nodes that they need

[junior urchin]

One script to rule them all

[pale latch]

that one script needs to rely on a whole eco system of powerful scripts to have any relevancy

diversity is a good thing

[shell tendon]

give stableswarmui a shot if you havnet yet

[mortal delta]

what makes it so special if i might ask?

[shell tendon]

it uses comfyui as the backend, and has some fancy pants new features like differential diffusion inpainting

diff diff is the best way to inpaint flat out

don't think a1111 has that (yet?)

but it's fast and memory efficient and the noodles are all hidden

[mortal delta]

yeah i might give it a shot then

[stray pond]

Hello I want someone to train a lorA for me. However it has speech bubbles in the image. I mostly want the pose of the images. Can I know if I should remove the speech bubbles before proceeding?

[desert dagger]

yeah, you should

[stray pond]

are there any good prompts for removing speech bubbles or just manual with photoshop?

[desert dagger]

i wouldn't trust the AI to do it right. i'd be using photoshop

[sand gust]

Rescue boats and ferries surrounding the airplane on the Hudson River, passengers being helped onto the boats.

[void vessel]

Hi, so a picture i try to generate says: Override settings: Pad cons: true. Where can i find that setting?( Before that it also said: Downcast model alphas_cumprod :true, but i found the setting, without this the images were't any similar to the one i wanted to reproduce) And after the hires it comes black, any idea?

[upbeat kestrel]

hey bot?

[kind cloak]

hi

kinda need help, how do i make foggy images?

[desert dagger]

add the terms "fog, mist, volumetric lighting' to your prompt

[nova zodiac]

atmospheric haze too

[nova lily]

pardon, you used to only be able to use Python 3 10 6 should I update it to this 3.10.11 or not necessary? thank you

[kind cloak]

do u have any additional expert keyword prompt tricks to enhance images?

i am kinda new at this

[sterile raven]

Can anyone that knows how to finetune loras in particular about captioning images take a look at #🔧｜finetune for me please.

[desert dagger]

i wouldn't update

[nova lily]

okay 👍

is Git alright to update?

[desert dagger]

yes

it's not related to python or pip

[tight mason]

https://www.youtube.com/watch?v=p1jKqV9IV8I AMD ROCm + WSL2 Installation Video

[nova lily]

can add "detailed X" skin eyes etc, if that's what you mean and bump up the attention with (:number) detail:1

[flat perch]

Can you tell me if the situation with sd3 regarding commercial use has been resolved?

[junior urchin]

not yet

Fruit will likely make a post in #📣｜announcements when that is resolved

[pearl ocean]

[rugged vale]

Anyone ever make a style LoRa that could do characters? I've never been able to get a LoRa to do both at once

[flat perch]

Interest Ask

[chilly storm]

This is more or less default and not desired behaviour. It usually takes some effort to make clean character design lora that doesn’t drag style with it - by using regularisation images for example

[rugged vale]

Ah that's a good idea, I never thought of making it a character LoRa first and a style second

I wanted to incorporate multiple characters into a style and have them be somewhat recognisable

[scenic flax]

Hi

[proud trail]

How to give specific aspect ratio in prompt?

[low moon]

The longer its banned on civitai the smaller the chances it will be embraced by the communiy

a few weeks and it will go the way of cascade

no one will touch it

[junior urchin]

People will use it regardless of the license

You do know that, right?

[low moon]

it wont be fine tuned so much less people will use it

theyll stick to more refined sdxl models long term

[junior urchin]

Then they can do that?

[low moon]

theyll use this as a novelty but really everyone was waiting for the fintunes not the basemodel

[junior urchin]

When there's a chance for 6B and 8B to be released, people would obviously wait for 8B to release, then finetune that

By then, the license work would be over

[low moon]

not if it has the same licence 😄

[rare sand]

https://www.reddit.com/r/comfyui/comments/1dkzmuo/design_iterations_architecture/

Anybody that can assist me?

[low moon]

u need to mask it

[trim magnet]

how much vram do u need to finetune 8b though?

[rare sand]

Ive done that, but the output is still bad

[junior urchin]

yes

I'm downloading the vram

[trim magnet]

[sudden ruin]

All of it

[junior urchin]

Yeah, all of it - then the time it takes, which could be an eternity.

[shadow stirrup]

what is the mst professiona sd? comfy+xl?

[stone oxide]

One message removed from a suspended account.

[stone oxide]

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

[low moon]

[red otter]

Anyone know something about general world models ?

[lavish lake]

what do you mean professional? Like for commercial work? In what area? There are tons of great looking 1.5 finetunes and that also works for sdxl. I prefer sdxl workflows for my commercial work but with some stuff like animateDiff, still 1.5 models are the only way to get decent motion and results

[sinful cloud]

Hey! Does anyone here happen to have a link to the Dreambooth server, or know a community/channel where I can find support with model training? I have a background in ML (but not diffusion models) and want to find somewhere where I can dig a little deeper than some of the youtube tutorials.

[hardy sedge]

Is there a way to make distant faces more detailed?

[desert dagger]

Upscale

[hardy sedge]

Is there a faster better Option?

[desert dagger]

if you have an image that has faces in it that you want to sharpen, you can use Real-ERSGAN

[fervent thunder]

Best is a workflow where u first generate face then use ipadapter to put it in an image

[little yew]

Hello

[fervent thunder]

if you want faster than ERSGAN then go for Compact or SPAN

or for slower, but better quality, you can go ADT, DAT or RGT

you just need the Upscale Model Loader in Comfy and then download model from here https://openmodeldb.info/

[rugged vale]

any alternatives to wd 14?

[fervent thunder]

yeah just call a VLM

[rugged vale]

can you recommend a pre trained one

[fervent thunder]

InternVL-Chat-V1.5 is good

or MiniCPM-Llama3-V-2_5

https://huggingface.co/spaces/opencompass/open_vlm_leaderboard

[rugged vale]

ty

[little yew]

Nobita

[twilit gorge]

is having comfyui or any on M2 disk driver is diff from having it on normal ssd or even hdd ? does it affect any process ?

[hasty hornet]

doubt it, probbly just faster file save , maybe faster model load, but gen itself probably is't much different

[bleak matrix]

Good afternoon, everyone! How are we all doing?

[stark trail]

Here is a dataset of 235 thousand prompts & images that we have been working on, collected from Stable Diffusion's discord:
https://arxiv.org/abs/2406.13735v1

Sample images:
https://stablesemantics.github.io/StableSemantics/

Generated using SDXL Lightning

(dataset to be released at a later date)

[split sand]

hey guys what do u use for SDXL upscale? im used 2 controlnet x tile technique in SD1.5 is there anything equivalent? i have tried using SD upscale script and it feels im missing some details

[desert dagger]

i tend to use the Magnific website for my upscales

[desert dagger]

i've got a spreadsheet full of prompts if you'd like to add them to your dataset

[stark trail]

@desert dagger interesting! PM?

[naive rose]

is forge.webui faster the comfy?

[split sand]

eh its paid haha

[desert dagger]

the image upscaler on the capcut magic tools page is good, and it's free

[fervent thunder]

Hi, I'm new to Stable Diffusion. I'm wondering if it's true that stable diffusion 3 has issues drawing hands (this is my experiences). And I'm also wondering if there are any good ways to fix this.

[trail lion]

I havent used sd3 yet, but all of the models have had that issue to some extent, it keeps reportedly getting better, but not sure with sd3, I've seen all sorts of anatomy weirdness from user reports here and on reddit

generally the way I've fixed in on sdxl and sd1.5 is with controlnet, or just inpaint with a higher denoise for the specific area sometimes work if you're lucky

[astral goblet]

(this is my experiences)
I don't get it. Why even ask?

It does. While the paper says it's improved at hands, it's not perfect at hands yet.
Consider that even today, artists struggle with hands. They're tough.

[trail lion]

I think he/she was wondering if that's normal, and probably so

[loud solar]

I have hand problems, too ...

[fervent thunder]

Is there a way to fix it using negative prompts? I've heard online that actually doesn't help

[astral goblet]

Theres a lot of debate about negative prompts. worth exploring

[trail lion]

not in my experience, you can try "badly drawn" etc etc, but I think it's better to just take the generated image and edit it (whether in an image editor, or with an inpaint method)

[astral goblet]

dalle gets good hands consistently but that's a highly refined model and a lot of pipeline between the prompt and diffusing

[loud solar]

And for local use we don't have the best model ...

But prompting "A perfect hand" got me 6 fingers ...

So maybe we humans are just not perfect?

[trail lion]

so I spent the last 2 hours with swarm...not for me I think. plus I keep getting OOM errors with settings that work just fine in forge

[loud solar]

Swarm is a nice idea but if you come from A1111 or Forge you are missing your manual workflows ...

[trail lion]

I'm going to keep it around just in case I end up wanting to try sd3, so far I'm not even tempted to download it

[fervent thunder]

Are you using automatic 1111 or Comfy UI?

[loud solar]

In some things it's pretty good ... prompting 3-4 different objects for example ... but it also produces a lot of garbage ... why not try it? download is free ...

[trail lion]

I'll say automatic1111, because I've only used comfy for literally hours

[tall berry]

Is there an area here where we can share our models / loras and get feedback? Wondering cause I am still new to loras and understanding them

[trail lion]

normally the place to share them is like civit.ai, but there are channels here to ask questions about training and such

[tall berry]

Ah fair fair , civitai is more for "finished" rather than test running haha

[trail lion]

really...heh, I'd say 80% of what I see there is not what I would call finished 😛

there are some gems though

[loud solar]

A1111 and SDForge are good for an easy start. ComfyUI let you learn some more basics if you invest the time and that's worth a lot, too. Even if you use A1111 or SDForge later.

[trail lion]

if you look at any of the model pages, you'll see versions they post, so it's totally fair to say it's an alpha

[loud solar]

The more complex workflows you will use in Comy the more it will get on your nerves searching, downloading and copying models/addons to the right place ...

[trail lion]

what was frustrating me was I was getting errors generating maybe 30% of the time, it would abort with an unspecified error. but I have an amd card, so it could be related to that

I have to use what is stable for me, and forge is solid, I can work with it all day and not have a problem

[loud solar]

There should be kinda automated downloader to put all needed stuff in the right place

[loud solar]

I pretty like Forge, too!

[crimson kernel]

SD3 release theme song: https://open.spotify.com/track/0dyz56yOkAvjJMAJ5IxfiE?si=caaca21651c44ccc

[astral goblet]

Is the new CEO of Weta Digital fame true? I love Weta's legacy . If it's correct then i'd love to welcome Prem Akkaruju to the world of Stability. Looking forward to the future. /puts on shades/

https://www.reuters.com/technology/artificial-intelligence/stability-ai-appoints-new-ceo-information-reports-2024-06-21/ Reuters reports no word from stability on the topic yet.

[spare copper]

Hey o// ANyone using allor by any chance?

[trail lion]

@astral goblet hey, been toying around with that Mobius model. It's really not bad. and loras work better than I would have thought

[astral goblet]

nice to know. ii haven't done a test run on it yet. I was assuming extra networks like controlnet and loras wouldn't work as well

[short mirage]

Is there any difference between using hires fix and simply regenerating in img2img at higher resolution

[sonic carbon]

could someone lend me a hand, ive been out the AI game for a while and just need one image upscaled, would someone run this through whatever upscaler https://cdn.discordapp.com/attachments/936934383326220331/1253868145446359050/93ab921f-6fa6-466f-a981-da0aad22b683_copy_ups.png?ex=66776b35&is=667619b5&hm=720cf02f1076d3b244766e48111999bc4bcc8857755866adece8c6c52e4fc2e9&

[desert dagger]

just go to capcut.com and use their image upscaler on their magic tools page. it's free and very good

[sonic carbon]

aw cool

[sonic carbon]

eh it came out very rainbow-y

[desert dagger]

i'll be happy to run it through Topaz for you if you like

[sonic carbon]

yes please

[desert dagger]

give me a few minutes here please

[sonic carbon]

the link i sent is a photoshop upscale but ive got the original more pixelated

[sonic carbon]

no rush 🫡

[desert dagger]

give me the original please. put it on imgbb and DM me the link

[astral goblet]

aw lame all the stuff i posted in #sd2 is gone

[arctic mesa]

https://civitai.com/product/link

[pale latch]

I don't recommend that. Paid autodownloader? Just use Stability Matrix

[dawn mulch]

whats a resource friendly inpaint method / workflow in comfy UI with or without face/object detect. i am using facedetailer and it uses too much resource so it often gives me a purple error about allocation errors. in general i think i never got inpainting to work in comfyui.

[nova lily]

So there can really be no best sampler right? Becuase there will always be a diff look every prompt set, and model.

[finite cloak]

if you use swarm there's a downloader utility that supports civitai and will automatically copy the metadata and all cleanly

[desert dagger]

no. there realy is no 'best' for any setting.

[opal hedge]

8B wen

[bitter moss]

greetings

It has been a while for me to try stable diffusion, my laptop was busted for monthes

The main one, others can not run stable diffusion,

So if I want to catch up on the latest, where do I start? When I left, XL had come out, but models were not too good, and the human generation was wonky at best. Plus it did not run well in older pcs.

Is auto1111 still the good choice, or something new should be installed?

[loud solar]

If you are used to A1111 you can still use it. Just update. Custom XL Models are way better than the basic one. You should find something fitting to your needs.

[bitter moss]

I did. But my vram is 6gbs. Think it will run?

how to run the XL models? can I just dump them in the models folder and run them, or some other requirements are there

Thank you for replying btw, sorry for the late reply. Got engaged in something.

[loud solar]

6GB VRAM I don't really have experience with. With an updated A1111 the models should run after dropping them into the models folder. Refiner isn't required with custom models any longer.

[royal blaze]

hi

/prompt

[cursive carbon]

hi

[warm junco]

Make sure you have --xformers --medvram --no-half-vae in your webui-user.bat

[main harbor]

hello please what do i chose sd1.5 or sdxl for architeture

please please

:DDDD'

and is it possible to have maybe both or they are totally different things

[trail lion]

you can switch between them, so you can have both

[main harbor]

but what do i donwload ?

ui

or which thing

[main harbor]

or how do you switch between them

[trail lion]

you choose it in the checkpoint toggle

[coral cedar]

es ce que DarkGPT existe réelement ?

[solid spindle]

Where is the best place to ask a question about a1111, img2img, upscale, GFPGAN, ADetailer optimization? I got a minor issue...thanks 🙂

[trail lion]

you can ask here, but whether someone can help you with that is not a guarantee, kind of depends on who is here and how complex your topic is

[solid spindle]

So, i got this setting on a picture:

Steps: 20, Sampler: DPM++ SDE, Schedule type: Karras, CFG scale: 5, Seed: 3820949351, Face restoration: CodeFormer, Size: 512x512, Model hash: 879db523c3, Model: dreamshaper_8, Denoising strength: 0.1, RNG: NV, ADetailer model: face_yolov8n.pt, ADetailer confidence: 0.3, ADetailer dilate erode: 4, ADetailer mask blur: 4, ADetailer denoising strength: 0.4, ADetailer inpaint only masked: True, ADetailer inpaint padding: 32, ADetailer version: 24.5.1, Lora hashes: "latent_con_pytorch_lora_weights: aaebf6360f7d", TI hashes: "EasyNegativeV2: 339cc9210f70, FastNegativeV2: a7465e7cc2a2", Noise multiplier: 0, Extra noise: 0.02, Sigma noise: 0.8, GFPGAN visibility: 1, Postprocess upscale to: 1024x1024, Postprocess crop to: 1024x1024, Postprocess upscaler: 4xUltrasharp_4xUltrasharpV10, Postprocess upscaler 2: R-ESRGAN 4x+, Version: 1.9.4

and i have rectangle around the Detailer space

hm i want to drop a pic 🙂

[trail lion]

you cant, there's a general-with-images channel where you can do that

[solid spindle]

okay so there is a visible difference where the detailer was working on, i want to get rid of it

i have scale to option 1024x1024 and crop to fit but thats not the setting that results in this issue

[trail lion]

yah, depending how bad it is, you just go over that same area with inpaint and play with a denoise not too high maybe .25

do whole picture vs masked only, so it has the context of the surrounding area

[solid spindle]

ins't denoise an over all denoise not for the face detail...or should i use another face_yolo i used N but there is also S...

my denoise strenght is 0.1

[trail lion]

what I'm saying is turn off adetailer, and you now have to fix it in just regular inpaint...or at least that's my process

assuming you like the change it made

[solid spindle]

the red indicator shows me face 0.87 around, and this rectangle is visible after rendering

ah*

but it looks well 😄

except the rectangle differencies

-< testrun >-

Damn would be cool if i can set up the detail for the full picture, i think i need to try out the other yolo's, the the "face" ones

i overread impaint*

[trail lion]

adetailer is just a shortcut, it's literally the same as sending the photo to img2img and selecting the face and doing masked only with 0.4 denoise (by default). if it's creating artifacts you can try lowering that value to like 0.3

[solid spindle]

the thing is, this actual setting worked pretty well on other pictures* just saying

[trail lion]

every picture is different, different distances etc, and every model is also different

adetailer is handy, I use it ,but I'm almost never satisfied enough with just what comes out of txt2img that I say that's the final product

basically expect to have to do more work until you're happy

[solid spindle]

thank you for the ideas...

[somber idol]

[woeful briar]

Hello

How can i make it to where i dont have arms clipping into legs and stuff

Im new to sd

[upbeat tendon]

hellp

hello

[raven bronze]

Guys, is it impossible to install and use SD and WebUI on an offline PC without internet access?

[floral fable]

The install would be tough without Internet access, but not impossible. After the first run (with dependencies installation), it is supposed to be able to be launched and used without Internet access.

[broken smelt]

it should work if you install it on another computer then copy everything over onto a USB drive but i havent tried so not 100% sure

[fervent thunder]

sitrep on sd3? have we got a good finetune yet? 8b model status?

[raven bronze]

So, I need internet access for the first installation, and therefore, it is impossible to even install on an offline machine, right?

I'm also seeing on github tickets that it needs internet while running.

This is being sold in many places as a local model that runs offline and is completely private, but when you look closely it is a security pandemonium for any and all companies that need to protect intellectual property.

Since third parties can have access to all the content generated at the slightest security flaw in any of the components or dependencies used, it is much better to continue using large models like MidJourney, where the risk of espionage is relatively lower, since it is impossible to run the "offline model" on an offline machine.

[floral fable]

You can install it offline if you want if you have the system libraries required and these Python libs in the venv that webui looks for: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/master/requirements_versions.txt

[hallow nacelle]

Do you really need to crop images to 1:1 for training loras?

[raven bronze]

And what about things like this? https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/11518

[floral fable]

I just did a test on HEAD where I installed with Internet access, and once the venv was filled, I hit Ctrl+C and completely airgapped the machine, re-launched webui, and was able to generate a txt2img image. If you're worried, test it out and see.

[raven bronze]

Thanks for the test, one question. And when you need to run any other component, can you install it offline, or do you have to put the machine online? And to do updates, can you do it offline? And what about installing new models? Can it be done offline? Because if your machine is offline 99% of the time but needs to be online 1% of the time, it's the same as nothing.

[floral fable]

How comprehensive your offlineness is will depend on how skilled you are at Python package management. Updates to webui and its Python dependencies can be made offline if you know what packages to grab on an online machine. Models and model layers are typically larger files that need to be downloaded and transported outside of webui's mechanics regardless.

To install webui extensions you would not be able to use the web-based installer, but would have to figure the manual way

[astral goblet]

just install swarmui and a happy number of custom nodes. the swarmui folder will be entirely portable by default.

[trail lion]

yes