#💬｜general-chat

1 messages · Page 143 of 1

slim belfry
#

I was wondering if theres a simple and easy guide to install stable diffusion at my pc and I also have a few extra questions about the engine

flint harbor
loud solar
slim belfry
loud solar
slim belfry
loud solar
slim belfry
loud solar
slim belfry
loud solar
slim belfry
opal hedge
loud solar
teal pagoda
#

I think you can try SD3 for free with 10 images or something on Stability.

#

Check it for yourself

#

I think that MJ still wins

#

in most of the cases

loud solar
teal pagoda
#

really low...

#

they should have doubled the amount

#

like 48 GB VRAM at least

#

so we can play for 2-4 years with AI

#

32 will be good enough for max 2 years from now :))))))

opal hedge
loud solar
#

They want us to buy a new one in less than 2 years 😄

slim belfry
teal pagoda
loud solar
teal pagoda
#

Imagine that they add only 8 GB to 4090

#

and they call it 5090

#

which will be like 2000-3000 euros

opal hedge
loud solar
slim belfry
opal hedge
#

Because Nvidia is making pennies from their consumer gpus and putting more VRAM on them will just take sales away from their corporate hardware

#

Unless AMD steps up and makes a decent graphic card it'll probably just be 24GB on the 5090

loud solar
teal pagoda
slim belfry
loud solar
#

Worst thing that happened here was that I broke my A.I. 🙂

teal pagoda
#

So I won't ever try running SD on an AMD

#

cloud GPUs are the best solution for AMD users

slim belfry
loud solar
#

Who would try to use 2 GPU excessive programms at the same time?

teal pagoda
#

blackscreen gg

teal pagoda
#

but not too excessive

#

like 2 games at the same time

#

like League of Legends very high + lobby in Call of Duty Warzone

loud solar
teal pagoda
#

:)))

#

🤡

loud solar
#

🥳

teal pagoda
#

what do you expect? :))

#

or was*

loud solar
#

To be concentrated on one game 😄

teal pagoda
#

I was.

loud solar
#

Yes ... but by mistake I also started Topaz Video A.I. while running ComfyUI 😄

teal pagoda
#

I was playing League of Legends while I was in the lobby for a Call of Duty game (not in the map)

teal pagoda
loud solar
#

Both programms started to tell me I'll have to wait hours 🙂

teal pagoda
#

so you said you have Topaz

#

how much do you wait to upscale a video of like 1 min to 4K 60FPS?

#

because on AMD is a pain

loud solar
#

To many unknown facts in that question. Point is ... mostly I use it to work to improve SVD Videos ... that's no problem.

#

And I only use FHD an 25 FPS ...

teal pagoda
#

I see

#

Because I thought about upscaling old cartoons, but they would take me weeks, so I said no

loud solar
sage reef
#

🇸 🇩 3️⃣ sadcat

teal pagoda
sage reef
#

🙏 happemad

teal pagoda
sage reef
#

are we really gonna go deep into June and still no release? 😦

teal pagoda
#

really now

#

did anyone test topaz video on commercial GPUs like H100/A100?

#

ever?

#

cuz I never heard about this

#

maybe it doesn't even work on server GPUs

teal pagoda
loud solar
#

There's an A.I. Artist in San Francisco who has his own server ... damn ... I'd like to use it for some days with his help 😄

teal pagoda
#

:)))

loud solar
#

His OK ... I can buy the tickets ... 😄

opal hedge
#

Right, so I looked through the 5090 leaks and all we know so far for sure is that the 5090 will be made from three different pcbs

#

It has a 512 bit memory bus which means it can have up to 36GB vram but no word on that

loud solar
opal hedge
#

There's a million different 5090 rumors. All we know is that Jensen Huang will only put in as little VRAM as he can

#

If AMD doesn't manufacture a competitor we're getting 24GB vram... again

loud solar
#

I'll be a special guest for ASUS at the GamesCon ... maybe I can get some more information ...

opal hedge
#

Would be nice to get some insider knowledge

#

I'm personally hoping for 36GB vram so I can run two cards at once, but I'm not expecting anything so kind-hearted from Nvidia

#

They want people to buy data center cards which are like 20x the price

teal pagoda
#

but hey, zuckerberg bought like 500K H100s

#

:))))))

loud solar
#

Maybe we should start flirting with him 🙂

teal pagoda
#

nope, no homo

loud solar
#

Llama 3 is a thing, too

teal pagoda
#

yea, but sometimes is bad

#

idk, I really think that they just have different versions available and they use/present the best one in the first days of the launch and after that they use inferior versions to cut the prices for energy/GPUs

#

kinda the same with Claude 3 - free version

loud solar
#

Could be a right idea ....

teal pagoda
#

because then how else can the quality decrease so much with the exact same model?

#

there's something shady

#

with all these LLMs

#

hosted online

opal hedge
#

I don't know much about llms, but what if they just do a lower step count to save on compute

#

I know llms don't use steps. Could be quantization

#

🤷‍♀️

teal pagoda
#

but they do something, that's for real

loud solar
#

And I hope there are still a few German boys training it ... so we don't get americanized 100% ... 🙂

teal pagoda
#

:)))

#

You can't even access LLama 3 if you're not from US

opal hedge
#

USAUSA

loud solar
#

I'm Homer Simpsons ... disappearing in a bush ...

opal hedge
#

Hopefully EU steps up its AI game

#

AFAIK you guys only have SAI and Mistral

loud solar
#

EU A.I. not really interesting for E.U. people 😄

#

Sadly we go with mainstream ...

opal hedge
loud solar
#

MP3 was a German thing ... we gave it away for free ... we are not good in selling 😄

opal hedge
#

At least you're selling out them cars

loud solar
#

As long as they are better than Tesla 🙂

#

Cars ... no future ...

opal hedge
#

I don't think being better than Tesla is a high bar

loud solar
#

lol

#

The idea everybody has his own car might work in the USA ... but not in big cities in the EU ...

#

And it even might not work in New York ...

#

Trump closing many streets ... 😄

opal hedge
#

Saving money on car insurance is goated

loud solar
#

I'm not saying that I don't want people to have an own car ... I only think it's not a concept for a future ... so if Germany invests to much money in the car industry ... maybe not a good idea ...

opal hedge
#

I think we'll still need cars for the next decade or two. With AI, who knows what's after that?

#

Personally I don't think any country is prepared for the future

#

It's very likely we'll all be out of work and on our asses

loud solar
#

I think you are right ... cause politic is just reacting these days ...

opal hedge
#

Yup. World's crazy rn

loud solar
#

Welll ... I'm a songwriter now and released my first song ... wtf ... I have been bad in music and art at school 🙂

#

It's damn crazy ... but the positive point is: we are part of it and not just watching.

fierce magnet
opal hedge
#

On the one hand the future looks bright, on the other a lot will have to change to get there

loud solar
lucid cedar
#

hey guys i was gonna get a 3060 but then i found out about nvidia Tesla 24gb, is that a better deal?

#

under $400

opal hedge
#

tesla doesn't run games or other apps, it's a data center card. It's also super slow

loud solar
lucid cedar
#

ah

#

I don't play games on pc, is that the only issue? or do server GPUs just not compatible with regular desktops?

#

I did a litte research, tesla is used for AI only

#

which is what i want

opal hedge
#

That's up to you. Though to be honest, I wouldn't reccomend a data center card to the average person

lucid cedar
#

nvm they're all refurbs on amazon. all the new ones are over $5000

loud solar
lucid cedar
#

alright, 3060 16gb is the only option right now under $400 just making sure

#

or a used 3090

opal hedge
lucid cedar
#

more vram got it

loud solar
#

Both sound pretty fine today ... but nobody really knows about SD3

opal hedge
#

3090 should run sd3. As far as we know, the 8B param model will run on 16GB Vram

#

But the exact requirements are still a mystery

lucid cedar
#

i'm so out of the loop lol

loud solar
opal hedge
#

It's a bit of a rabbit hole. But I do think the 3090 is a great choice for AI shenanigans

lucid cedar
#

this seems like another rabit hole next to the one for mechanical keyboards. why do people keep digging these things.

#

jk it's all fun

loud solar
opal hedge
#

On the same token he doesn't get why AI is so exciting for me

lucid cedar
loud solar
#

Girls, guys, whoever (not in a negative way) it's been a nice time with you. I identify as a car ... not older than 25 ... 😄 Have a good day, night or whatever!

radiant frigate
#

When will the bot be available

lucid cedar
radiant frigate
#

Do we have to pay for it?

loud solar
#

Nini

opal hedge
#

Not very expensive but it is what it is

sage reef
astral goblet
# radiant frigate Do we have to pay for it?

yeah. you need to have stability membership to use the sd3 preview. the signup list was a total farce and they decided to charge for access for sd3 instead of offering free preview like sdxl

low moon
#

Does Friday apply to the SD servers?

trail lion
#

it's literally no different than demanding google release a new update to gmail or something

echo marsh
trail lion
#

how does realvis outrank juggernaut

#

that's the problem with crowd polling

#

realvis is one of the least flexible models I've used, so I guess if you only want a portrait, it's amazing

low moon
#

I know there are better models

#

but its a secret which ones...

#

Hidden gems stay hidden.

fervent thunder
#

hey anyone can help me out in tech support

trail lion
#

this isnt tech support

urban dragon
#

how do i run stable diffusion on my pc

fervent thunder
#

thats why i asked to help me there

trail lion
urban dragon
#

i couldnt find the git

#

can u link

astral goblet
fervent thunder
astral goblet
#

would've been even more frustrating if those invites weren't getting sent out

low moon
#

Just use Pinokkio

#

if you are terrfied of git hub!

#

=0

trail lion
urban dragon
#

no no im not terrified of github

#

i just want to run it locally

low moon
#

so then git clone whichever ui you enjoy

#

Fooocus? A1111? Forge

#

or comfyu

fervent thunder
trail lion
#

I imagine the invite thing here, either they underestimated the demand, or simply changed strategy mid-flight to do this api thing

low moon
#

ok

#

Try Forge which is a better fork of A1111

trail lion
#

which is the right business decision honestly

low moon
#

If you just wan tto be a casual user you can also try Fooocus or Defooocus

astral goblet
#

the way they rolled out gmail was meant to make buzz. Google didn't give people invites. They allowed gmail users to invite new gmail users. It was a hype strategy. Which is what Stability's sign up sheet was. But since they never actually sent out invites to those who signed up, and only sent to them to people who they vetted, it was a lot more frustrating and just felt like fake hype

low moon
#

if you have big plans go for Comfui

fervent thunder
fervent thunder
trail lion
#

and users have no recourse because it's not like there is a contract or anything other than a loose verbal commitment

low moon
#

Forge is faster

#

and more streamlined, you egt many features by default with forge

#

like controlnet

#

Well if you want something bigger it sounds like you already have some experience in which case I really recommend Comfyui...

trail lion
#

the way I look at it, I'd rather them do something that keeps them around then keep the software delivery cycle as it was and potentially they close the doors at some point due to bankrupt

fervent thunder
low moon
trail lion
#

so if that's the case, all you're left with is a bunch of whiny users that just want the newest shiny iphone...pffft. they've released so much already

low moon
#

Yes.

#

We whine.

#

but release the weights!

#

😄

#

(Just fix the hands first).

trail lion
#

mark my words if they released weights tomorrow it'd only be a couple weeks before people would start asking about sd4

static cape
echo marsh
static cape
echo marsh
trail lion
#

frankenstein

echo marsh
#

We will be releasing a paper on it soon

echo marsh
low moon
echo marsh
#

It's a new sdxl base model

#

Is the best way to put it

low moon
#

Always looking for the next paper.

trail lion
#

oooh, like pony

#

except without the creep factor

low moon
#

pony sux

#

no offens ebut it does

#

It;s not even perfect at anamoty

#

still has crap hands

echo marsh
low moon
#

I dont mind tities

#

but please fix the hands

echo marsh
#

The title of the paper will be "Constructive Deconstruction: Domain-Agnostic
Debiasing of Diffusion Models"

static cape
#

I'm testing it on imgsys atm - it's better about half the time for me... and it is in almost every A v B, so no wonder it gets so many points.

low moon
#

Poney models have horrendous hands

trail lion
low moon
#

How could that be?

trail lion
#

community contribution is something I can get behind

low moon
#

I thought they were meant to make anatomy great again.

astral goblet
#

Yay papers!!

#

squeeze those papers

low moon
#

Yes

#

Two papers down the line I will live on Venus.

#

In Ted Murtdcoks casinos

echo marsh
#

Here is the main take away from our method and upcoming paper

low moon
#

I will go all in.

echo marsh
#

***The constructive deconstruction method presents a novel and effective approach
to debiasing diffusion models without compromising their performance and versatility. By inducing a controlled noisy state and retraining the model using
advanced techniques, we achieve a new base model, Mobius, that performs well
across various domains and styles. This method’s domain-agnostic nature and
adaptability make it a promising solution for addressing biases in generative
models. Importantly, our method provides a means to create new base models
without having to design them from scratch, leveraging existing models and
enhancing them through a systematic and controlled process.

low moon
#

I am sorry.

#

I am a sad end user who barely makes this work.

#

Let alone udnerstand the magic underneath the hood.

trail lion
astral goblet
low moon
#

Please,

#

AI

static cape
low moon
#

hands have five unfused fingers

#

lol

#

mobius training

#

loop back on that

astral goblet
static cape
#

Btw - on Imgsys I can spot Mobius half the time because it's basically "crisper" or rather has more defined edges in artsy images.

low moon
#

Whats the use of talking about it, no one listens.

astral goblet
#

mobius huh. sounds one sided to me.

low moon
#

Truth hides behind a paradox

echo marsh
# static cape So we could Mobius-retrain existing finetunes?

It enables making entirely new base diffusion models with needed to extensively pretrain a new model from scratch. We can in a controlled way break all the quality and composition associations without damaging the baseline understanding from the preexisting training. Making it so we can completely train those associations without having to start from scratch

#

Basically we just lowered the cost to entry for making new base diffusion models 50 fold

trail lion
#

I have no idea what this mobius is all about, but if it's about creating something like juggernaut or better, all for it

astral goblet
#

barriers to entry. amirite?

low moon
trail lion
#

the vampire?

#

jared leto?

low moon
#

No.

astral goblet
#

thats morb

low moon
astral goblet
#

we don't talk about morb

low moon
#

Any hungrians here?

#

Magyar SD?

#

Bassza meg egyedul vagyok.

#

😢

#

Actually its a spanish movie

trail lion
#

is it nsfw?

static cape
trail lion
#

ditto

#

and that no taboo fetishes are left unexplored by the community

astral goblet
#

i think you underestimate the community

versed yoke
#

what is the diffuion immage model training speed difference between a 3090 vs a 4090 and how much would a (5090 with a 512bit bus and 32gb vram 50%faster then 4090) speed things up ?

astral goblet
#

relies on number of cuda cores more than anything

low moon
#

CUDA hates you.

#

CUDA will deceive and trick you and will take your soul.

astral goblet
#

4090 has 60% more roughly. 5090 will have 50% ish

low moon
#

CUDA will confuse you with red nonsensical messages and will lead you down confusing, frightening and frustrating path. CUDA does not love you.

obtuse marsh
#

is the bot service no longer free? 😦

astral goblet
#

AMD user found

low moon
#

CUDA hates you.

fervent thunder
#

Why does my embedded files don't show up in forge stable webui

urban dagger
#

This place has changed so much since last I popped in aaaaaaa

peak orchid
sage reef
#

mobius? never heard about it

robust sapphire
#

Is anyone available to voice chat? I am really struggling with image generation.

narrow kernel
# echo marsh https://x.com/DataPlusEngine/status/1793803117642854732 muahahahaha!

seems promising, but imgsys just ends up being more of an aesthetic tester than anything. after rating dozens of images, most of the prompts are garbage sd1.5 style tag prompts and 99% of people aren't going to actually read all of the prompt to make sure it's being followed. also, their prompt window doesn't show the entirety of the prompt since it's only one line, so you have to scroll through to see it all. you'll also see a ton of random artists thrown into the prompts and i'd bet money most people don't bother looking them up to see if the generated images match the styles. basically, it's a pretty flawed testing setup with a lot of inherent aesthetic bias. While that is definitely important as well, it just ends up being like 80% of the reason why most users will push left or right.

of the ones that have popped up that your model has been in, it's usually been really good though, but it is still usually a coin toss between the other four models in the top 5. you'd have to then spend a bunch of time actually paying attention to the prompts to truly decide which was better. but aesthetically, it's definitely up there and if the new method of training is as cheap as you're claiming, that's a big win either way. look forward to checking the paper out

#

Like how the hell am I supposed to rate A/B on a prompt that just showed up: "cat seahorse fursona wearing headphones, autistic bisexual graphic designer and musician, attractive androgynous humanoid, coherent detailed character design, weirdcore voidpunk digital art by artgerm, akihiko yoshida, louis wain, wlop, noah bradley, furaffinity, cgsociety, trending on deviantart"

I have no clue wtf I'd be looking for other than "weft pikchur wook pwettyr den wite pikchur"

smoky moss
#

Hello,

crude cosmos
#

hi everyone

smoky moss
#

Hello, @crude cosmos

next hinge
#

hello

loud escarp
#

nice to meet you

#

what are you doing now?

stone dew
#

hi everyone

pearl ocean
alpine plover
#

hiya

fervent thunder
#

Hello

silk badge
teal pagoda
teal pagoda
echo marsh
#

sadly

teal pagoda
# low moon still has crap hands

there's not a single model which figured out to make almost perfect hand, nor MJ or Dall-E 3 (but at least dall-e has decent ones).

teal pagoda
low moon
#

🥺

#

can anyone explain why sigma pixart is so slow compared to cascade but their results ar every similar

azure cargo
#

@narrow kernel Apparently, of all things it was..time? The entire day gens worked fine. About the same time as yesterday I started getting blurs. Also around the same time that chatgpt becomes a blubbering fool. So I suppose it has to do with the traffic.

narrow kernel
buoyant sandal
#

cat

fresh vigil
#

dog

arctic wraith
#

Just now I subscribed on stability.ai website for image generation. But unable to generate images here, what's the drill guys?

eternal ledge
#

(I've only used the API, so I can't actually help on this front)

gilded folio
#

RTX series aims for gamers mostly

#

And they have $40k cards that have huge vrams for datacenters

#

Putting 32gb of vram would hurt h100-a100 sales

#

Datacenters would still get h100's cus there are some other factors to consider

#

But some people and most of the small companies would buy 5090 to train or run their models instead of renting/buying h100's

gilded folio
#

AMD might (and I hope) put that huge amount of memory into their new cards

cursive magnet
#

and HBM2/3 is much better for AI than GDDR

gilded folio
#

amd's mi300x has 192gb vram tho

cursive magnet
#

i know some people buying cheap ass p100 with only 16gb VRAM and telling that they outperform any RTX cards cuz of 4096bit memory. Idk if its true, maybe we have some proud owner of this card on this server who can confirm

spring fern
#

does anybody know if i can stop A1111 from creating folders for dates in the "output folder"? i want it to just get put in the output folder, not for it to make another folder in output and then put that there

cursive magnet
#

there is checbox for subdirectory in settings

spring fern
#

thanks i'll try that

astral goblet
fervent thunder
#

What does hyper lighting and turbo means in model

sage reef
#

faster generation for slight reduction in quality

#

so it's up to you to decide what you value more, speed or quality

trail lion
#

right, but having said that, you can always upscale later, so I can get why people like them

fervent thunder
#

So turbo is fast what about the other two

trail lion
#

because when do you ever get the perfect generation the first time...never

fervent thunder
#

Also anyone know how to stop fooocus from auto downloading checkpoints

trail lion
#

or from accumulating extra o's

fervent thunder
#

Also will there be any problem if I rename the checkpoints and lora

trail lion
#

probably not, definitely not with auto1111, so probabaly not with focus + extra o's

fervent thunder
#

Didn't receive the clear answer on model names lighting turbo and hyper

sage reef
#

i never used foocus but it would be real weird if you cant rename models, you can rename models in all other ones, the name doesnt matter, its all about the "loading part"

dawn mulch
#

is 6bg vram enough to train loras and models?

#

and what are good resoruces for training

trail lion
#

doubtful, might be able to do 1.5 models

dawn mulch
#

i only do 1.5

#

dont mess with sdxl

trail lion
#

sorry to hear that

dawn mulch
#

lol

#

1.5 works well as far as i know

trail lion
#

even with 1.5 you'll probably struggle with OOMs

#

but you can probably get training to work

dawn mulch
#

i hope so

trail lion
#

worst case just use cloud, that's what I do, and I have a 16G card

#

mainly because I dont want to tie up my local compute for training

dawn mulch
#

wish cloud was an option. cant pay for anything ATM. otherwise i would be using cloud.

trail lion
#

ok, well just use every memory saving trick in the book

dawn mulch
#

im still learning on the resource and will give it a try on my 6gb

trail lion
#

like use adamw8bit, gradient checkpointing, keep training resolution at 512, xformers, batch size 1

#

those are the main things for your case

#

assuming you have nvidia

#

otherwise use sdpa

dawn mulch
#

yeah i have a rtx 4050

fervent thunder
dawn mulch
#

if i make a lora for hands and feet, and label images with 5 fingers as 5 fingers, and 5 toes as 5 toes, and anything less or more as less or more for fingers and toes would that address the more than ore less than 5 fingers or toes, or is that not how it works at all?

#

that said i havent been notcing the more then or less then 5 fingers or toes too much lately with stable diffusion models.

fervent thunder
#

Does anyone have any idea

wary belfry
peak tendon
#

what is the best upscaler for 3d renders? Specially vintage nintendo 64 types where you need to keep the vivid colors and the plastic look
Resrgan 4x + anime6B almost gets it right but it's too anime and it kinda removes the 3d aspect

sage reef
#

upscaling Mario 64? :3

peak tendon
#

I would also like to upscale mario 64 renders if I find a good method

dawn mulch
#

if i had more than 6gb vram i would just doing endless test but with just 6gb, i have to be extra cautious. i probably cant even train anything.

wary belfry
dawn mulch
#

sounds like there is hope.

wary belfry
dawn mulch
#

iv tried collab for SD generations, it has a gpu limit issue. but ill give it another look for training smaller loras.

astral goblet
#

^ likely just a business guy with an idea who will want people to work for free until it starts making a level of money he's comfortable enough with sharing. Same old story. If they wanted to pay people they'd be recruiting properly.

pearl ocean
astral goblet
#

What would be the scam? They send you money to get started but need you to send some back? Dudes looking for free work isn't uncommon too. I guess it's also a scam but feels less bottish

peak tendon
mint fiber
#

I guess this is not for me, I am not able to other things.

jaunty sinew
#

Hello everyone, if i purchase credits for stability.ai for api usage does it have any expiration time?

mint fiber
#

I don't think so, Fatih. I had a few creds sitting in my account for quite a long time before SD3 released as an API

jaunty sinew
#

Thanks @mint fiber 🙂

opal hedge
#

sd3 wen

frozen helm
#

+1 month

#

bro added time

opal hedge
#

When did that happen

pearl ocean
teal pagoda
#

And not all of the free work is a scam, some are called "joint ventures". But you really need trustworthy people for these things and maybe even some contracts signed.

astral goblet
#

free work is pretty scammy

versed yoke
#

and they can just also increse the vram of the datacenter gpus

desert dagger
sterile raven
# opal hedge sd3 wen

Probably when its too late and gpt4o drops their image modality. Then they'll panick drop but itll be too late cuz everyone is using the better product. The ecosystem will never develop as well as SD1.5 cause of their fault

low moon
#

So since refined models just get better with time is it safe to assume that just more refining and cheaper better GPUS eventually will lead to perfect images? Nothing really seems to beat a good model. no amount of sigma and llama and T5...

low moon
clear oyster
#

when sd3??

sage reef
#

by asking that, you now added another month 🙂

clear oyster
#

this is firsttime asking this month lolol

clear oyster
sage reef
#

the truth is we have no idea, as far as we know, it could drop tomorrow, it can drop in a week, or in a month idk

opal hedge
clear oyster
sterile raven
opal hedge
#

Lil bro is projecting

peak orchid
#

What none of the Stability AI staff are going to tell you is that SD3's training is going to take until August potentially. They laid off their best employees who could figure out how to do it faster, and can't afford a larger GPU cluster either

wide heart
#

I producing like 20 images and I've already reached my credit limit

peak orchid
#

The other elephant in the room is that the for profit company masquerading as a charity known as Thorn seems to exert some level of influence over Stability AI. Thorn recently failed in their supervillain style plot to take over the world until the guise of "stopping csam", and their efforts were only stopped by the European Court of Human Rights (Chat Control). Thorn had the head of the EU, many of its ministers, the head of Interpol and others doing whatever Thorn wanted (including breaking the law, which they did to try and help Thorn), and I seriously doubt their corrupt cronies are suddenly free from their leashes

wide heart
#

Any ideas why I can't produce my than 20 images on the free trial?

peak orchid
sterile raven
trail lion
#

so the woke thought police are here ,thank goodness

dawn mulch
#

Who was it that said stay woke, was it Nanci Pelosi.

#

If its off topic than disregard

#

By the way, are fake artists basically artists who are focused on money (i know i might get a lot of negativity from this), are they still having a crisis with AI generated images or have they all sort of secretly started to use AI themselves. Or are they not there yet. I haven't been seeing anyone complaining about AI art much so i assume its because many of them have secretly started to embrace it.

opal hedge
sterile raven
dawn mulch
#

doesnt microsofts copilot image generator use the same tech as openAI but suppoesdly less censored.

opal sierra
#

hello

#

nice to meet you

nova zodiac
narrow kernel
#

Within a five years, there will be sweeping changes to the internet you know. Microsoft and Intel have been hinting at it for years with an almost DRM style verification system for everything. Wouldn't be surprised to see GPUs and ISPs getting on board as well. Basically, everything would be fingerprinted by the CPU down at the IME level and/or they'll implement something similar on GPUs, that can't be bypassed at the OS level. ISP servers and other servers will have to authenticate requests to do certain things like using CUDA features on a GPU. This would cut down drastically on things like bad actors and botting.

dawn mulch
#

So they dropped the idea of just forcing everyone to use cloud services where no one has a computer but instead an interface and everything is processed and done through the US government.

#

Are there any SD models that focus on making political art or images? This way normal people can start making their own propaganda.

narrow kernel
dawn mulch
#

[Attached to my previous post]

pine fiber
#

any news on sd3?

pine fiber
nimble locust
#

stable diffusion

covert herald
#

It is possible to buy credits as company and dont pay tax ?

teal pagoda
fresh ruin
#

can anyone guess what checkpoint this IG profile uses? @silva_florina

twin roost
#

What steps can we take to ensure that the images generated by Stable Diffusion resemble realistic human appearances?

trail lion
#

Avoid terms like animal, creature, cartoon, livestock

fervent thunder
#

Is there any way to get a token to just have a really minimal effect in a prompt? For example, "a (red:0.01) cat" isn't much different from "a red cat", especially compared to the contrast with say "a (red:2) cat"

trail lion
oblique elk
fervent thunder
#

Yeah, I'm aware comfyui doesn't normalise the weights by default

fervent thunder
trail lion
#

Sometimes you have to play around, oruse like cnet reference with a photo

#

Using words to describe something specific can be hard, like if you wanted a red stripe

fervent thunder
trail lion
#

And assuming the model knows how to draw that

fervent thunder
#

Yeah, I'm aware it's kind of a crapshoot with specifics

trail lion
#

I've had some luck with inpaint sketch also

#

You can insert colors you want and rerender that

fervent thunder
#

That's for sure useful, I'm just working with a limited toolset in this instance

trail lion
#

Yeah if you hang out in the prompt help channel for a while you'll inevitably see people trying to make something very specific with just prompting, which has limitations that sometimes can't be overcome

fervent thunder
#

I'll give the substitution a go though, I think now that you mention it, one of the really bizarre and powerful generators on the perchance site is the fusion generator, which tends to construct prompts like this:
"[many photography abstractart black topaz champagne sheer striped abstractart with warm hyper maroon Naturallightsources incadescent on grape dimly-lemon mildy chartreuse Dreamy:giant enemy crab:0.35]"

fervent thunder
fervent thunder
hexed pilot
#

Yo, I'm new to making LoRAs and am about to finish my first one on forked tongues (body modification). I have a variety of pictures ready. I heard it's good to include similar but not identical images for better learning through captions, so I've added some pictures of normal tongues too.

I'm unsure about naming and captioning. I've named the LoRA 'fork3d_tongu3' and want it to trigger with 'fork3d tongu3' or 'split tongu3' or something in that manner. Do I need to put these tags first to make them the main tags, or is it the name that determines this?

Also, how do I teach the LoRA to distinguish between a normal tongue and a forked one? I've included normal tongues in the dataset but am unsure how to tag them.

Any advice would be appreciated HACKERMANS

astral goblet
static cape
peak orchid
# astral goblet Lol this guy is actually mad about an organization that fights CSAM? holy fuck....

No, I'm mad that Thorn wastes their time and resources on trying to maximize profits from CSAM at the expense of human rights. They don't actually care about stopping the issue, and they have a very disturbing 'ends justify the means' worldview. They're a tech company that has EU politicians in their pocket: https://balkaninsight.com/2023/09/25/who-benefits-inside-the-eus-fight-over-scanning-for-child-sex-content/

astral goblet
#

you can look anywhere theres money and find corruption. even in the most altruistic seeming places.

that doesn't disqualify the entire organization's efforts though. thats so ultimative and feels like ulterior motives

peak orchid
# astral goblet you can look anywhere theres money and find corruption. even in the most altruis...

There are better organizations in the field, and why they may be doing some good, the harm they are trying to cause certainly outweighs that. Thorn was was behind the anti-encryption complain in the US against Apple. That 'Heat Initiative' group behind the anti-encryption campaign is part a web of fake charities and advocacy groups to promote their harmful worldviews, and they had government ministers break the very laws that they were meant to enforce by using illegal microtargeting in the EU.

astral goblet
#

I don't believe thorn was ever "anti encryption" . That was just qanon hyperbole nonsense.

End to end encyrption has a LOT of implications. It's a giant buster claymore sized double edged sword. The concern is about E2E encryption on services that children use

#

"I should be able to close the door with the child and me alone in the room!" is the argument being made against their concerns

peak orchid
#

They want encryption backdoors mandated on all communication services, and they don't care about the consequences. Its just about getting rich for them as their AI mass surveillance tools would basically be a legal requirement to use

astral goblet
#

They're registered as a non profit. I'm not sure how that's a get rich quick scheme. There are a lot of other ways of getting rich that are much easier.

peak orchid
#

If you've followed the news on the subject, its been playing out for a couple of years, and Thorn had plenty of time to correct the record but choose not even when confronted by journalists from reputable sources

astral goblet
#

Open a casino. EZ

peak orchid
#

Lots of groups are listed as "non profit" even though they really aren't, and nothing stops them from changing after they get their way (see openai)

#

The point is there are better groups for companies like Stability AI to partner with, ones that aren't trying to undermine human rights

astral goblet
#

I think outlandish accusations shouldn't need to be corrected all the time. Same way Keven Hart was like "no. i will not apologize a third time in order to host the oscars" and instead didn't do what critics demanded of him. Why jump when told to?

peak orchid
#

It refused to disclose Cordua’s emailed response to Johansson’s May 2022 letter or a ‘policy one pager’ Thorn had shared with her cabinet, citing Thorn’s position that “the disclosure of the information contained therein would undermine the organisation’s commercial interest”.

astral goblet
#

Yeah and? Even non profits have expenditures

#

They need a business paln

#

if it were all a matter of money they woudlve just got into arms trading instead

peak orchid
#

Experts and journalists are saying that is not normal, especially in the EU. I trust them more than some random person

astral goblet
#

theres 10000 other ways to make money more easily. they have a different mission

#

They seem to be experts in their own field. They've been effective in the past

peak orchid
astral goblet
#

Remember earlier how i mentioned e2e encryption specifically? and not all encryption?

#

and called it a double edged sword, figure of speech typically implying one effect is good and the other is not as good

#

The corrupt people with ulterior motives are more likely to be the ones flinging hyperbole. I wouldn't trust those journalists.

#

C-SAM and abuse is global problem that's worse today than it ever has been before. I don't think it's the time to lighten up on it. From Canada, i'm cheering on the efforts of this US based company. Hurray.

peak orchid
#

but the evidence seems to point to it being more than a few bad apples, if you trust the journalists of course. Thorn's products also seem to be regraded by customers as overly expensive and not as reliable as competitor products (as of a few years ago). Thorn also claims to have over 99% accuracy with their AI model, but it doesn't take an expert to know that such high accuracy is a straight up lie or the result of messing up training/testing

#

I did a deep dive into this industry because I wanted to see if there were tools that people here can use on their scraped data to help eliminate and report such content. And there was basically nothing available the open source community can use. I also would appreciate it if you amend your first reply to me so that its not making baseless accusations

rigid harbor
#

guys please i need comfyui help. how do you increase batch size for img2img workflow??? empty latent image node which has batch size doesnt work for img2img

peak orchid
# pine fiber how do you know?

The model is obviously still under trained and the devs have been saying that. Rumors have been circulating that it will take at least a couple months to fully train it, especially if they decided to change some of the model's architecture (like the devs mention on reddit). This is compounded by the financial issues that the company currently faces

rigid harbor
peak orchid
astral goblet
#

if their product isn't as effective as others, that's something wholly different

static cape
peak orchid
#

at the end of the day, things usually come down to money and people seeking power who shouldn't be allowed any power

autumn tiger
#

what a big chat

astral goblet
peak orchid
astral goblet
#

End of the day, i have no reason to retract anything i've said

astral goblet
peak orchid
astral goblet
#

It's just quick and easy money. Ethics get in the way though

#

since most people have them

teal pagoda
#

it's not like in gta

#

you have to invest thousands of dollars just to do this trick to somewhat work (not even 100%)

#

you have to look for a good team, you need a good hacker who can take out the cameras, you need a good driver, you need motorcycles/cars for swapping everything, you need an insider and the list goes on

#

so you have to invest tens of thousands of dollars into this and you're not even 100% sure that it'll "work"

astral goblet
#

armed robbery happens multiple times a day in nyc. and if everyone was doing it, the cops would have no hope of keeping up.

this isn't even considering the OBVIOUS analogy. If it were just about money they'd exploit children instead of protecting them

#

motives beyond money do exist

teal pagoda
#

yea, even worse ones

#

it's pure logic

#

no one from this entire planet which has power is a 100% good person

#

because you can't even achieve that power by doing the right things

astral goblet
#

good and evil are entirely subjective concepts

teal pagoda
#

presidents, politicians, CEOs

teal pagoda
astral goblet
#

morality has always been a hairy topic

teal pagoda
#

so let's say that because of some substances from a factory, some people from a village unexpectedly died. The CEO knew very well that those substances can affect the other people's health who doesn't have protection equipment, yet he just chose to do nothing, but to exploit them

#

can you ever say that this is good?

#

if yes, you're legit out of your mind

#

The CEOs have their heads full of "killings" and innocent lives taken

#

but they seem to not care at all

astral goblet
#

I typically don't like loaded questions. Of course i wouldn't and i dont know what i said to prompt that inquiry.

peak orchid
trail lion
#

Why all the focus on the CEO, they may not have founded the company, or be on the board, it, and depending on the size even know about certain decisions

astral goblet
#

Just titles and context matters a lot

#

often the CEO is just a patsy meant to take all the heat. Like say Youtube

gilded folio
astral goblet
#

Canada got those too

gilded folio
#

Cant they just shoot these robbers

astral goblet
#

the CEOs? please

gilded folio
#

wdym CEOs

#

oh analogy

#

btw I kinda remember seeing video of a robbery of an apple store while no one reacting to it as if it were a daily thing

#

it was in california iirc

astral goblet
#

san fran. shit is wild there. people are pissed because of the skyrocketing costs of living caused by the tech industry and the city does nothing for the people growing up there to make things affordable

low moon
astral goblet
gilded folio
low moon
#

yeah my dog died

astral goblet
#

america the sequel

gilded folio
#

well I mean you're technically an American

astral goblet
#

same way a brazilian is

gilded folio
#

yup

astral goblet
#

actually, i change that. peruvian. cause of lake titicaca.

#

same was as a peruvian final answer

gilded folio
#

One can be a lot of "x-ian" at the same time

#

you could be [Name of the hospital you were born]ian

teal pagoda
#

imagine that in my country, half of the salary is going to taxes (pure taxes, not food, insurance or something else)

#

50% from everyone's income

#

:))))

#

how can you ever become rich as a worker? you cannot

gilded folio
teal pagoda
#

unfortunately

astral goblet
#

i love roads an shit, but yeh

#

taxes. am i right?

teal pagoda
#

there are only bad roads here

#

they steal all the money from the taxes

#

and the judges do nothing

gilded folio
#

at least we're not paying 27 grand for a frickin childbirth

teal pagoda
#

they are all in

low moon
#

most of it goes for weapons and bunkers for the illuminaters who want to kill most of us and ride out the nucular holocaust there

#

allegedly

#

some goes for roads too

#

:))))

teal pagoda
#

where?

gilded folio
#

outside of the US

astral goblet
#

uninsured ameircans

teal pagoda
#

unlucky for real

low moon
#

to think that at some point the usa fought the brits to refuse to pay taxes...

astral goblet
#

the shitty thing about US healthcare costs is that the insurance companies don't have to pay the same costs. they negotiate a 10th of the price generally

low moon
#

how ironic theyre worse now

astral goblet
#

only uninsured poeple get the full price

teal pagoda
#

so if you get sick in the US, you better k1ll yours3lf than paying these sums

#

I'm joking, but still...

#

27k for a birth

low moon
#

thats nuts

#

no wodner birthrates ar eplummeting

#

cant afford not only to raise a brat

#

even to give birth is alreadsy too much!

#

:)))

#

wars are messy why dont we make it too expensive for them to exist that way theyll die out quietly

#

hahaha lolrothchild illuminater laughing in the bohemian grove

#

allegedly

teal pagoda
#

saw a really interesting reel the other day which showed why the 2000s generation can't really afford too much on their own nowadays (I'm talking about the normal people, not about of apes, people doing something illegal and so on). He compared the prices from 1990 with the prices from today and everything is 75% more expensive today. That's why you can't afford to buy even an apartment with all the cash down (you have to take credits from banks) and most of the people are living paycheck to paycheck nowadays

#

a really bad situation

low moon
#

Well AI will difinietly fix the economy.

teal pagoda
#

how?

low moon
#

sarcasm

teal pagoda
#

imagine that now it's even harder for all the manual/AI artists to make money from art

low moon
#

AI will pprobably be the thang that will likely send this zombie of a monetary system down the pit

teal pagoda
#

because everything is oversaturated

astral goblet
#

iw as about to be like "AI IS TRTAINED BY THE HUMANS HWO HAVE BIASES AN WONT FIX SHIT" but i didn't catch that /s

cerulean merlin
#

Does anyone know for loading clip vision with an sdxl model what model to use?

low moon
#

i just want nice hands so i can render myself some heaven while the storm rages

#

please fix the hands

#

theyre attrocious

#

even in these sd3 images

teal pagoda
#

and as you can see, the ones trying to do something "good" (open-source) are going bankrupt now

#

stability ai

cerulean merlin
# 

Error(s) in loading state_dict for Resampler:
size mismatch for proj_in.weight: copying a param with shape torch.Size([1280, 1280]) from checkpoint, the shape in current model is torch.Size([1280, 1664]).

File "E:\ComfyUI_windows_portable\ComfyUI\execution.py", line 151, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI_windows_portable\ComfyUI\execution.py", line 81, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-0246\utils.py", line 381, in new_func
res_value = old_func(*final_args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI_windows_portable\ComfyUI\execution.py", line 74, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\IPAdapterPlus.py", line 711, in apply_ipadapter
work_model, face_image = ipadapter_execute(work_model, ipadapter_model, clip_vision, **ipa_args)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\IPAdapterPlus.py", line 340, in ipadapter_execute
ipa = IPAdapter(
^^^^^^^^^^
File "E:\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_IPAdapter_plus\IPAdapterPlus.py", line 70, in __init__
self.image_proj_model.load_state_dict(ipadapter_model["image_proj"])
File "E:\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\nn\modules\module.py", line 2189, in load_state_dict
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(``` Does anyone know the fix?
low moon
#

yeah but i mean im talking about a more fundamental change

teal pagoda
low moon
#

kinda like a transition from slavery to non slavery

#

just like slavery had to end at some point maybe wage slavery gotta go now

teal pagoda
#

:))

low moon
#

kind aobvious the way things are going theres no future in this

teal pagoda
#

their plan is to make 15 mins cities

low moon
#

ive lived thru the fall of commies

#

so now i get to see this global whatever go down

#

once u live thru the collapse of a regime its not so hard to imagine how it can happen

#

and it happens fats too

#

like booom

teal pagoda
#

imagine that microsoft will launch an AI that takes screenshots with everything you do on windows and will send them to microsoft

#

recall it's the name of that tool

#

it'll be on windows 11 out of the box

low moon
#

i guess we have to stay on windows 10 forever

#

and disable pdates

#

😄

#

updates

teal pagoda
#

that's the way

floral umbra
#

Is C++ build tools gone for anyone else in visual studio? As i'm testing out a image to 3d model comfyui workload, but suddenly build tools is just gone

teal pagoda
#

someone said that people better try to keep themselves healthy, because in the next 10 years something revolutionizing will appear and the life can be longer with some implants/medicine/interventions.

#

:))

low moon
#

you cant be happy in time, the fera of tomorrow and meories of the past ruin any chance for real joy

#

only outside o time can there be joy

#

so in reality no one is happy

#

not 100%

teal pagoda
#

kinda

#

would be really cool if the souls will go into that astral world after death

astral goblet
#

if there is a soul then we can engineer it

low moon
#

😆

#

AI soul

#

man you better give those souls better hands

teal pagoda
#

:))

#

you really have something with those hands

#

they are onto your nerves

#

it seems

low moon
#

theyre by far the worst in AI image gen right now

#

As fra as the body goes anyway

#

electronics mechanicla parts, writing uhm also ar enot stellar

#

but hands are worse

trail lion
#

trying my first LLM captioning tool, I seem to lag like 6 months behind with most new tech. I get stuck in my ways

#

it's called internlm, we'll see how this goes

astral goblet
#

the new mircosoft vision model looks pretty nice. light weiht

wet zealot
#

stability cant catch a break can they

trail lion
#

?

#

sure they lost 100M but they generated 5M, I'm not seeing the problem

low moon
#

Maybe Emad will lead the revolution.

#

We need like the Americna civil War or the French Revolutnion. but globally. at this point.

#

Money has become a hindrance rather than an incentive.

#

A lot of bitter people out there and that means only one thing: wars.

#

I have a feeling the system will collapse before I finish my game.

#

thast's my luck.

#

but it was worth it

sage reef
fallow veldt
low moon
#

All hail Emad.

#

His heart is in the right place but the Illuminaters got his balls.

solid whale
#

I want to do some face-swap stuff with Stable Diffusion. I have a video of someone dancing and want to face-swap faces onto this video. However there is a catch: I want a new face to swap onto the dance video every couple of seconds and most importantly I want a smooth transition to appear when swapping from one face to another, just like can be done in AnimateDiff. Does anyone happen to know if there is any way to do this? 🙂

low moon
#

reactor?

keen falcon
#

hello people, whats the correct channel for help please ?

trail lion
#

I guess it depends on what you need help with, if it's technical support there's a techsupport channel, if it's psychological, doubful anyone is qualified

#

if it's moving a couch, I guess it depends on your location

keen falcon
#

nah, its with my command prompt, I got an error, and Chat GPT arrived at a point where he keep repeating me the only solution it think there is

#

(spoiler alert : it doesnt work and it keeps giving me the same procedure)

trail lion
#

sounds like you need the tech-support channel. and as a tip, maybe describe what you're doing...you installing something?

keen falcon
#

well, something-something "torch" apparently

#

but I did installed it, and I did also "force reinstall" as it call it

trail lion
#

torch, ok yah that's pretty much the glue that makes all this work

keen falcon
#

it gives me this : File "C:\Users\swatb\AppData\Roaming\Python\Python312\site-packages\torch_init_.py", line 141, in <module>
raise err
OSError: [WinError 126] Le module spécifié est introuvable. Error loading "C:\Users\swatb\AppData\Roaming\Python\Python312\site-packages\torch\lib\shm.dll" or one of its dependencies.

#

ive checked the PATH thingy, and its normally the correct one, so I dont know whats the problem by this point

#

well if anyone got an idea to help me u can ping me

peak orchid
#

He cleared a path for everyone

solid whale
pine fiber
#

seems pretty good

#

does anyone here know much about the fusion mixed modality models? Like 4o, I was wondering if they could dethrone diffusers for image gen

soft iron
#

there is no problem to run it, the problem is the speed, it will be incredibly small

dawn mulch
#

anyone tried the krita AI with comfyui? i tried the animation option it has and will it dorsnt seem to be able to do auto animation on top of an animation but it can do frame by frame where you have to generate one for each frame...

wraith patio
#

Canción aserradero

tight jewel
#

人物IP、潮玩、三维、潮流、街舞、牛仔，c4d、

rapid ore
#

anyone know of a frontend for stable diffusion that can work on 10gb ram and use things like real vision I remember being able to run the realistic models on my machine with medvram now it just instantly crashes

fallow veldt
rapid ore
fallow veldt
#

maybe you should troubleshoot, check if you need another flag for your system, or reinstall

fallow veldt
lean pilot
#

has anyone already tried sd3?

loud solar
lean pilot
#

can you suggest some of web ones?

loud solar
lusty beacon
#

Guys SD3 release in May?

lean pilot
peak tendon
#

hello my frends, is there a way to keep the transparent background when you upscale with Stable Diffusion?

#

I use automatic111

soft iron
soft iron
tame gulch
#

hi people

rapid ore
#

anyone know of a good model to take an illustration or anime and output it to a photo like image with good accuracy for keeping the same facial features

oblique elk
worldly heron
#

Hi guys, just saw that I am in r/SD and SD discord server. Is there one more active or one that is archived. or both are relevent?

#

if I may ask my lil noobie question here also :

I guys, I have SD installed with ultraspiceXLTURBO_v15.safetensors [069737f0a7].

I am wondering what video or website should I used to get a better grasp of all the configs please? Thank you

halcyon granite
#

any game Devs here using SD AI imagery in their game? I'm really paranoid about using it a I want to release my game on steam.

I'm using a custom civit ai model

oblique elk
# halcyon granite any game Devs here using SD AI imagery in their game? I'm really paranoid about ...

Well here you need some luck and faith. Even if the custom model is fine for commercial use of the results, it might have merged model which are limited to personal/research use.
In addition when an artist decide that the artwork of him was mainly used to create that specific style,...
On the otherhand i am pretty sure there will be ways ($$$) to settle such a thing and it will be a hell of work to get real prove for a copyright law suit

#

But all depends on your local laws etc.

pastel turtle
unkempt hatch
#

I don't understand, even if they go bankrupt, what's to stop them just releasing it?

#

just release the 8b model. Even if it's not fully cooked, the community with fine tune it

vale steeple
#

you know that base is extremely important for finetunes

#

like how you choose potatoes to make baked potato

#

if you decided to use a half-grown potato to bake, will it good. Yes, but not certainly as delicious as full-grown potato

#

sure in this past few weeks people have been dreading on the unable to release SD3. And well the SAI needed to find the balance between time taken and completion

#

seriously even some of them went through stages of grief

unkempt hatch
#

but like, Pony was basically a completely new fine-tune, right? There exist folk who can do some extreme fine-tuning?

#

and they've been using the API online, so it can't be that undertrained, right?

#

the online API is working ok?

vale steeple
#

Even if SAI gone bankrupt and dead, with scientist worker run themselves to elsewhere. There are always open competitor against closed source, like welp HunyuanDiT made by Tencent team.

#

It just get replaced, that's pretty simple.

unkempt hatch
#

if SAI dies, couldn't the workers create a new open source co-op together?

vale steeple
unkempt hatch
#

they should just release the 8b model then, and stop worrying about all the others. Why make like, 4 models that are unfinished when they could just finish 1

vale steeple
vale steeple
#

except those who have 4090 / 24 GB VRAM

unkempt hatch
unkempt hatch
#

better that nothing

vale steeple
unkempt hatch
#

and folks without enough vram can still use online services or SD1.5

vale steeple
#

Lightning though, well. It just needed time

unkempt hatch
#

ok, but a distilled 8B might be comparible to the 800million, right?

#

and we already have SD1.5 for people with weak GPUs who don't want to use online

vale steeple
unkempt hatch
#

releasing a bunch of new models just divides the community even more... and SAI doesn't have the resources to finish one

#

like, again, 1 finished model is better than 4 that never get released

vale steeple
#

wlel actually that one is subjective

unkempt hatch
#

ok, but 1 is still better than 0

vale steeple
#

I know some folks in #🆕｜sd3 would like the 2B version

unkempt hatch
#

I'd rather 1 than 4. But everyone would rather 1 than 0

vale steeple
#

I know one personal project to make a model out from scratch ( or from the paper )

#

Simo Ryu ( the person who introduced LORA into diffusion space ) 's LavenderFlow

#

he use 1% of SD1's budget to get LavenderFlow, as he actually went out to FAL to borrow 2 H100

#

the result of the model though... will remind you the days of 2018 AI Art

#

but it is deem to be possible to actually train as individual or group.

unkempt hatch
#

SAI could also partner with some kind of stock images site, get free already-captioned images to train on, make money through AI use on their site

#

because the stock image sites are using AI already, but not very well

vale steeple
#

There's a goal to make Qualcomm / Mediatek phone running SDXL-quality AI Art models.

#

with current scale of NPU in their processor

fervent thunder
low moon
#

Where are the "in two weeks" jokes?

#

Wait there is less than two weeks left of May.

#

could this mean...

#

that aliens exist?

low moon
low moon
#

the models we have now sdxl sd15 midjourney etc they ar elike the early huge clunky computers that only could be built and ran by huge companies and very expenive

#

now you ahve more power in your pocket

#

not too far down the line you will train sd3 in 3 days on your wristwatch

#

with quantum computers in minutes

#

Qubit SD3

#

and even the hands will be good

ashen sun
teal pagoda
teal pagoda
#

it's legit impossible

#

so they can't sue you for anything, because they don't really know what model you used to generate

teal pagoda
teal pagoda
teal pagoda
timid gust
eternal violet
#

@sudden ruin can you remove this grifter?

arctic sedge
#

When do you guys think HunyuanDiT will be on Civit?
How long has the weights been available? 2 days? I'm not sure.

teal pagoda
teal pagoda
#

run the cell and open the gradio link when it'll show

#

tried it with an advanced prompt and it's not really good at that

#

and it launched on 14th of May or something

#

so not only 2 days old

oblique elk
fallow veldt
narrow kernel
# teal pagoda it's legit impossible

Depends, in theory they could watermark it in latent space to where it shifts some pixels and clusters of pixels around the image like an invisible QR code. There is also invisible watermarking that you can use with SD, but that's something you intentionally do yourself. Both methods can survive certain amounts of post processing people might try to remove them. Like technically, modern QR codes can still be readable even if like 70% of the code is missing.

#

So no, it's definitely not impossible. But they'd have to train the model around it and do some tinkering to pull it off

devout prism
#

Is it okay if i just read stuff, I am just starting with Stable Diffusion on a very, very elementary level; so don't have much to participate with this at this time?

meager yarrow
#

Help? Create a prompt?

frail dew
#

hi

plain raptor
#

stablilty Ai partnered with Tripo Ai

#

so..why is there nu 3D channel

#

<v>

plain raptor
#

i mean there is but

arctic sedge
tired creek
#

I guess emad jumped ship because he saw the model management issues coming, sd3 nowhere in sight

trail lion
plain raptor
#

for example, say i want to make a kitty, i would say, fluffy kitty, white fur, blue eyes, laying on pillow, sunset, sunlight through window

mellow junco
#

Hi people ! How can I try to generate an image in discord?

plain raptor
#

and anything u dunt want in the picture, type that in same form on te negative prompt

trail lion
plain raptor
astral goblet
plain raptor
#

too bad u cant download Tripo Ai like u can stable diff

#

kinda sucks tbh

#

but ig its fine, Tripo is their own thiing, and only partnered with stablity Ai

trail lion
astral goblet
plain raptor
#

i find it kinda cheeky tbh to make ppl pay to use stable diff, when stable is free, its like the ppl who make u pay are taking advanagte of those who dunno how to do that

astral goblet
trail lion
#

they arent, you can still use all their released models free, it's just the new one that you can pay for the privilege to use in lieu of a free release (assuming that ever happens)

astral goblet
#

such is the nature of open source. Google doesn't pay to use the linux kernel and samsung doesn't pay to use android

plain raptor
#

xD

#

i dunno all the hand shakes tho that were prolly made

trail lion
#

and honestly, it's hard to fault them when they are struggling financially

plain raptor
#

but, seems like they worked together

astral goblet
#

That was probably just marketing speak. maybe they just emailed a question to one of the staff members

plain raptor
#

XD

#

prolly

#

bc, i mean, i imagine it would be free

#

like stable

#

not proprietary

astral goblet
#

i've never touched their website for instance.

#

but i'e used that model alot

trail lion
#

I have zero interest in the api, or any of their base models. I just want to see what the community does with the weights, so if they never release it, meh

astral goblet
astral goblet
plain raptor
#

this is wut sum1 said on the reviews on github

#

"They are never going to give away their latest SOTA model as it is a commercial product, if they will update the model in TripoSR with something a little better that depends on the competition, at the moment TripoSR is the best open source model, unless a competitor releases a better open source model for a different project, then they have no incentive to release another model for TripoSR. Even so, they would just release a model a little better than their competitor had and nothing close to their commercial model."

#

and that made me come to the conclusion, thats is all property of Tripo

#

which sucks u-u

#

the model that is there, is, not..really all the developed

#

and likely prolly wont be based on that, and, the way the Ui is designed, doesnt really allow u to, intuitively install another model

trail lion
#

there are other things out there, but I'm not seeing a bunch of open sourced weights

#

elon released his grok or whatever, that was cool

#

I just hate the idea of closed source because ultimately you'll just be spoon-fed someone else's agenda, that's the end game

plain raptor
#

its up to us, the community, to rise above the robber barens, and make like a linux community of open Ai, for both txt to image and txt to 3D models. When i say linux like community, i mean, open source and free

#

:3

astral goblet
#

microsoft just released Phi-3

#

it's funny to me that these days, MS is actually quite the champion of FOSS

#

Github has remained pure since the acquisition. They've given so much

trail lion
#

this is the most I've ever seen goog threatened, they have the doj monopoly thing with their play store, they have the threat of ai assistants such as chatgpt that threaten their search business, they have all the news publishers lobbying to get link laws like cali and canada and australia

astral goblet
#

And Meta just released Llama 3

#

Grok models are a pile of crap that was never meant for release. He only did it because not doing it really weakened his "open ai needs to open their tech" arguments. I don't think anyone has managed to rig up their own deployement

trail lion
#

yah, I have no opinion on the quality of grok, but even if it's garbage, maybe someone can do something with it

astral goblet
#

doubt. llama makes more sense for community hacking. grok is something like 80gb

#

oh wait no. the 770 files that make up grok weights are 296GB

#

Phi-3 models are 4-8gb each, stand alone, and have 128k context length vs 8k. grok is just not the cheese

vale steeple
#

the response given is decent but LLaMa 3 can do the same. If I won't wrong.

#

And Grok aren't even supposed to run at home hardware.

astral goblet
#

Llama 2 is more effective than Grok on all the common metrics if i'm not mistaken.

Even as a datacenter level model, its far less efficient than all other options available. it's just far behind the research curve and is more of a historical record of what grok was than something that will help future development.

#

like if microsoft open sourced directx 7 today. Even that might be more interesting though.

#

ok more like directx 3 when things were really bad and they hadn't found good solutions yet

peak orchid
#

All these LLMs and VLMs are basically using the same datasets, and the only differences are the size of the model and small differences in architecture

#

If you want better open source models, we need better datasets. Its that simple

unkempt hatch
#

would be cool if an LLM were trained largely on biographies / true story narratives as well as high quality fiction

rapid ore
nimble canyon
lucid cedar
#

how do you calculate it/s that people are posting? is it just the it/s from a single image that's listed in the comfyui terminal output, or do i have to run a specific script to get an average?

sharp rapids
#

【图片内容】一只疲惫的老虎蜷缩在草地上，眼神无神，身体疲惫，毛发凌乱，周围环境昏暗，呈现出一种疲惫和无助的氛围。

【正向提示】A weary tiger curled up on the grass, with a dull gaze, tired body, disheveled fur, and a dim environment. This high-quality image captures the vulnerability and exhaustion of the tiger, evoking a sense of fatigue and helplessness. Perfect for projects that aim to showcase the emotional depth and fragility of wild animals.

【反向提示】blurry, low quality, pixelated, (unattractive), (small size), (angry), (vicious), (unnatural colors), (out of focus), (unclear), (dirty), ((extra tail)), ((extra paw)), ((extra ear)), ((extra eye)), (out of frame), (bad composition), (too bright), (too dark), ((extra stripes)), ((extra teeth)), (poor lighting), (bad color grading), (red-eyed), (morphed face), (unnatural posture), (awkward pose), (frozen animation), (low-res), (bad framing), (insipid)

【参数】Sampling method: DPM2; Sampling steps: 20; CFG Scale: 7; Seed: 3856492; 最优长宽比: 3:2

teal pagoda
#

but not the model itself is good (like SDXL), but the fine-tuned version by the community

#

the base sdxl model is a joke

#

like the base 1.5

teal pagoda
teal pagoda
unkempt hatch
#

if you've got an image with someone upside down in your data, is SD able to use that info to better draw upside right people?

keen falcon
still glacier
#

so no, you didn t install the correct python

#

install any python from 3.10.6 to 3.10.11

keen falcon
#

oh ok, so is there a problem with recent version ?

keen falcon
#

I unistalled it and installed it at least 50 times each, im getting tired of this sh*t

arctic sedge
keen falcon
#

Ok can anyone help me on this please

#

Traceback (most recent call last):
File "C:\Users\swatb\Documents\essaiNSFWredhead.py", line 1, in <module>
from diffusers import StableDiffusionPipeline
File "C:\Users\swatb\AppData\Roaming\Python\Python310\site-packages\diffusers_init_.py", line 5, in <module>
from .utils import (
File "C:\Users\swatb\AppData\Roaming\Python\Python310\site-packages\diffusers\utils_init_.py", line 97, in <module>
from .peft_utils import (
File "C:\Users\swatb\AppData\Roaming\Python\Python310\site-packages\diffusers\utils\peft_utils.py", line 28, in <module>
import torch
File "C:\Users\swatb\AppData\Roaming\Python\Python310\site-packages\torch_init_.py", line 141, in <module>
raise err
OSError: [WinError 126] Le module spécifié est introuvable. Error loading "C:\Users\swatb\AppData\Roaming\Python\Python310\site-packages\torch\lib\shm.dll" or one of its dependencies.

#

There is a problem in there apparently I cant find. Ive tried unistalling/installing torch, Pytorch, check PATH, now try with a different version of Python, and nothing changed

#

Ive checked the shm.dll file which is apparently weird out, but it gives nothing to me

#

I dont know where to look anymore nor what to do

still glacier
daring nest
keen falcon
still glacier
#

we don t see the command, we don t see what it tells before that, etc

rare parrot
#

And once you get the same speed on the last two images that's your it/s

#

Tho on it's own it/s is meaningless if you don't also say what size the images were

lucid cedar
#

also, can i copy the entire comfyui folder to back it up before installing custom plugins that might brake things? I'm on linux. I understand it's a 'portable' app, so it should be able to be moved to a new computer if i want, right?

rare parrot
#

ComfyUI stores all plugins in a single folder

#

If you remove the custom_nodes folder it removes all the extensions fully

lucid cedar
#

nice

rare parrot
#

Also all your generation parameters are saved in the output images

#

So you can just drag and drop your old generations into Comfy to get the whole pipeline

#

Anyways, get comfy manager plugin, makes downloading other plugins way easier

lucid cedar
rare parrot
#

And updating and such

rare parrot
lucid cedar
#

metadata, i should've known that lol

rare parrot
#

Encyption is when you want a message/file to be only readable if someone has a password

#

Or similar

lucid cedar
#

yeah i wasn't thinking right

rare parrot
#

Nothing wrong with confusing the terms

lucid cedar
#

I have the manager installed, would it help me fix missing nodes errors? a few of them are conflicting...

rare parrot
#

Trying to use some very large pipelines or something?

lucid cedar
#

it's not that big of a workflow

#

I'm missing these: Text box
SDXLPromptStyler
ACN_AdvancedControlNetApply
ControlNetLoaderAdvanced
ScaledSoftControlNetWeights

velvet pilot
#

hey, im looking for a good 2d checkpoint, not like anime style but more of a cartoon

#

any sugestions?

loud solar
velvet pilot
#

yea ik ik, but there's so many that I dont know whitch one to pick hahah

loud solar
#

Yes ... but who knows better than you wich preview pics fit best to you?

eternal ledge
rare parrot
#
  1. ComfyUI Manager has a button to install all missing nodes
#

So that should help in theory

empty star
#

It would be real swell if in the future the open source community will be able to mass train models (like a folding @ home type of distributed setup) that we all benefit from. Instead of watching the big corpos pass us by

rare parrot
#

Nobody says we can't

#

Other than several million dollars of compute cost for making a single base model

#

So go on, make one

eternal ledge
# rare parrot Other than several million dollars of compute cost for making a single base mode...

I think the idea is that the community could share the compute cost by offering unused time on their own GPUs, I still agree on the "go make it" point though.

@empty star , it seems strictly possible that one might be able to package the "help train community models" as a custom comfy node or something if you're worried about distribution of a new standalone thing. You're free to use this idea (and the next) how ever you want to 😛

Or, you could start a new blockchain that somehow uses training community models as its mining function 😂

pale latch
#

shouldn't

charred mesa
#

lol

#

Thank god I didn't get excited

pine fiber
soft iron
#

yes

#

sd3 release

astral goblet
#

new stable assistant version. still not any weights released

wet zealot