💬｜general-chat

[fervent thunder]

is it tho

[verbal osprey]

Hey guys any tips on making comic book style panels in SDXL in one image? Is it possible? For example, "comic book panels, panel 1 a red ball, panel 2 an orange ball"

[static cape]

@echo marsh you could try that with SD3. Not so much with XL. (unless you try it with a lot of patience and regional Prompter)

[low moon]

dman the two weeks still isn;t up

i knew my sense of time is messed up but this si ridiculous

[wet grotto]

What is Omost? Is the same thing then ELLA?

[low moon]

it writes promts

you dont have to write promts even

You just bark at the screen and AI will write your prompt for you

you think of something and omost guesses it and makes an image for you

two papers down the line you wotn even have to think of anything

ai will think for you

[left sparrow]

This is dart

dart is a diesel introduced in the movie Day of the diesels

Dart works with the den at the diesel works

look up a dart the diesel on Google

For something a bit spicier look up dart in the diesel rule 34

[naive thorn]

who got the gmail the weights will release

i got it

[harsh ether]

new SD model stuff

wooo

[naive thorn]

yess

praying my pc will handle it

[harsh ether]

real

it better run on amd if their doing it on a AMD keynote

[visual peak]

Does this thing only make square images?

[naive thorn]

no

it can do anything u want

[spare ether]

What is the self hosting form for SD3 all about? If we already have the subscription then do we need to fill that in ?

[vale steeple]

things is happening

[outer cradle]

mass ping @karmic brook

[quiet delta]

Nice! What do you think the vram requirements will look like

[low moon]

its official now!

[river sparrow]

One day stability ai will fall, and that will be a sad day, because stability has done much for open source

[low moon]

2 weeks !

[opal hedge]

10 days!

[quiet delta]

lmao

[outer cradle]

one day yo moma will fall on my wiener , good job

[rain osprey]

medium?

[vale steeple]

self-hosting is like CivitAI's generation service or Pixart.

We haven't know whether the license is about "commercial model" or "commercial image"

[vale steeple]

2B

[river sparrow]

Dye

[outer cradle]

no ☹️

[hazy cypress]

I bought a 4090 for 2B?

[vale steeple]

hints already told you

[river sparrow]

Me when stability ai isn’t stabile

[opal hedge]

Lil bro out here with the sick burns

[vale steeple]

4090 can run 8B absolutely fine

lol

under unoptimized condition

again

[outer cradle]

thx lil bruh, u violated lowkey but its all love, yo aura diminishing but its all love, shout CCP its all love

[vale steeple]

if you remember the first introduction of SD3 page by SAI

[hazy cypress]

My point was I want the 8B

[vale steeple]

at least it is better than none

that's the point

[spare ether]

What’s the point of stability membership if this isn’t covered then. I need clarity on this

[vale steeple]

Hm, we will see

[median charm]

[vale steeple]

I don't care whether it is commercial-license or not

but I knew some people is going to bombard it like hell

[rare parrot]

Did they rename 2B to medium and then 4B to large and 8B to extra large or something like that?

[royal sentinel]

This is hype

[rare parrot]

Or are we getting 4B first

[vale steeple]

8B have be observed to be "undertrained"

[wet zealot]

the moment SAI falls people will instantly pivot to training pixart and hunyuan

[royal sentinel]

But y not June 21st?

[mystic siren]

YEEEEEEEEEEEEEEEEAAAAA

[vale steeple]

so obviously they probably need more training and time on 8B and 4B

[wet zealot]

cope

you're not getting them

[opal hedge]

800MB Smol, 2B Medium, 4B Extra Medium, 8B THICC

[hazy cypress]

Are you a spokesperson or something?

[mystic siren]

I just read the announcement 🥲

[vale steeple]

cope, you being pessimistic over everything

[sterile raven]

Drop the large.. noone is gonna use medium

[vale steeple]

you sure everyone have 8GB VRAM?

[rare parrot]

And 2B has been observed to be "all you need" by Stability AI all of a sudden last week so I'm not 100% convinced by that

[low moon]

im glad its 2B at least theres hope i can run it

[opal hedge]

Pixart Sigma is really good despite being low param

[topaz portal]

Large models never ever 💀

[opal hedge]

And it uses a similar architecture to SD3

[sterile raven]

My pc i bought in 2017 has 16

[vale steeple]

and T5

[rare parrot]

Ah right I forgot to try it out in the end, what size is it?

[opal hedge]

600M Params

And it's pretty damn good even if a little undertrained

[wet zealot]

hunyuan

[rare parrot]

Hm not bad not bad

Yeah it felt like 3 months

[timid quest]

woahhh

[pearl tapir]

is the new SD3 medium larger than the old SD3 or smaller?

did we have SD3 small or large first?

the new one is more lobotimized?

the old sd3 was never leaked?

[naive bear]

How large of a requirement leap is SD3 from SDXL?

[pearl tapir]

no slowdown is nice

[cedar crow]

With Forge I was able to run SDXL

[opal hedge]

sdxl bigger (3.5B) SD1.5 smaller (~1B)

[pearl tapir]

oh so its in between

[naive bear]

Wait, so SD3 is smaller than XL but a lot better?

[cedar crow]

I hope whatever new models come out of SD3 will wow me.
I will have to wait for months for that though.

[pearl tapir]

i'll take it!

[opal hedge]

Should be yes

[vale steeple]

want to know HunyuanDiT and Pixart Sigma

[pearl tapir]

supposedly it is more neutured though

[naive bear]

Dang that's awesome, XL already runs really well on my hardware.

[pearl tapir]

sdxl is still the most used, not much good art comes from SD3

[naive bear]

Well people always release improved checkpoints anyways

[opal hedge]

There'll be 4 sizes of sd3 this one is the second smallest. There'll be one that's really big and stronk coming out eventually as well

[naive bear]

Sweet

[pearl tapir]

i'd be OK with like 50% slower than SDXL

[steady arrow]

SAI Scamming with this SD3 "medium" lmao

[steady arrow]

The email they sent out said "our most powerful image gen model ever" but its not there's now 2 sizes better?

[rare parrot]

Hopefully

[opal hedge]

Allegedly it'll take 44 seconds to generate a normal sized image for sd3 on a 4090, versus the ~10 seconds it takes now

[pearl tapir]

that sounds very slow

[opal hedge]

It's the same model it just comes in different sizes.

[pearl tapir]

is that the largest size not the 2b?

[quiet delta]

I wonder how much of a percentage stability wants for the commercial model

[naive bear]

As long as it's better than XL or at least enables better checkpoints, I'm here for it.

[cedar crow]

this is no way practical for imo. Most people are running vastly worse cards than 4090.

[steady arrow]

thats some scam shit lol

[opal hedge]

Oh yeah? How much did you pay for sd3?

[vale steeple]

guy

i mean

[steady arrow]

No one ever even heard of any sizes or model family called sd3, they only saw the images that are from the bigger models known as sd3 (XL?)

[pearl tapir]

whats june 11

[vale steeple]

it is not just "same model came in different size"

[opal hedge]

44 seconds for the largest model. The smaller models will be a lot faster

[vale steeple]

by machine learning's term

[pearl tapir]

there are times when i want to generate 100 images over night and i dont mind it taking a minute or two each

[vale steeple]

it is different from middle-of-the-way model to fully trained model or custom dataset size.

[naive bear]

You don't need an image in 44 seconds, I get pretty quick XL outputs with a 3060 Ti, if it takes an extra minute per image, I'm not going to sweat it.

[vale steeple]

it take 40 minutes for me to run XL

and I say it is fine

[pearl tapir]

1 a minute each i coul generate 300 in 5 hours or 600 in 10 hours while im at work or sleeping

[vale steeple]

2GB VRAM for SDXL lmao

yes it is real

[wet zealot]

not sure why people are dickriding SAI

[naive bear]

What kinda hardware can even run it just bad enough to take 40 minutes and not get a rejection message?

[wet zealot]

copium you mean

[opal hedge]

Those are sd1.5 numbers

[vale steeple]

basically why it is so slow is because the VRAM dam have been breakthrough toward the RAM river and large sum of water flowing uncontrollably into the river

lol

[cedar crow]

Maybe I'll rent a GPU instance from one of those sites when the time comes

[vale steeple]

it definitely worked and generate something

[candid trout]

Coincidence? At least give us credit next time for our fun prompt ideas SAI.

#🌠｜show-and-tell message

[naive bear]

DALL-E 3's getting img2img and GPT-4o will likely be able to do a lot of img2img better than DALL-E 3 with tools, so as long as this carries me until then.

[wet zealot]

look toward pixart and hunyuanDIT tunes for the future of local imagegen

SD is dead

[vale steeple]

cope

lol

you have consume too much copium

[visual lagoon]

https://x.com/Teknium1/status/1797467900993036602

[wet zealot]

does he know

[visual lagoon]

what's this "SD3 medium"?

It's less parameters than SDXL

[vale steeple]

HE DOESN'T KNOW!

[candid trout]

SAI has been ghosting us for like two months on training Juggernaut for SD3 but obviously they follow our twitter account… lol

[naive bear]

I need competent generation more than I need flashy generation.

[shy raptor]

生成3d 的肺气道图

[vale steeple]

seriously

define less powerful

in SD's term

[naive bear]

I don't care as long as it's an improvement. AI's moving fast enough for the company to die without me having to hurt over it, though I did really like them.

[wet zealot]

finding it strange they didnt drop full SD3 when tencent were waving hunyuan in SAI's nose

[vale steeple]

if you can have the same prompt coherency and stability ( not pun inteded ) on all version of SD3

then what is more powerful then?

if all of them have the same finetuning power

[desert dagger]

trust this discord to be complaining

[wet zealot]

its SD2 all over again

spawnkilled by a model that existed prior to it

[opal hedge]

undertrained sd3 on the API is already far better than hunyuan

[visual lagoon]

Likely won't be an improvement. Probably gonna resemble SD cascade

[vale steeple]

but instead of SD2, SD3 is literally a bet or death model

for SAI

[wet zealot]

try *actually* prompting hunyuan

[visual lagoon]

A bit better at text then sucks at everything else compared to SDXL

[pearl tapir]

how so?

[vale steeple]

they already say it is the last image model they will release in short time

[weary heron]

Weren't you the ones who started running SDXL without a commercial license?

[opal hedge]

What does that mean?

[wet zealot]

you can't just 1girl it

[quiet delta]

Couple of bad faith actors trying to ruin the fun for everyone, don't feed the trolls

[opal hedge]

Yeah, no shit

[weary heron]

Like SDXL 0.9 even. The leaked ones.

[vale steeple]

investors are eyeing on SAI bro.

and success of SD3 can bring an inflow of cash to them

even though it is not much

[naive bear]

I mean, these examples are pretty specific compared to what XL can do on it's own. Maybe you don't like the visual styles or render quality as much, but after DALL-E 2 everything's been about coherency over shiny flashy glitter starlight

[candid trout]

We’ve been trying to talk to Colton and the team at SAI for like a year now. And we did everything Colton asked us to do. Waiting on SAI.

[vale steeple]

#🆕｜sd3 message

PonyXL creator is speaking

[wet zealot]

rm license.txt

[weary heron]

Maybe stop being so petty and they could respond to you.

It's a prompt, bro.

[tame jasper]

So is the version of SD3 we are getting going to be bad? What does medium mean? Sorry if dumb question.

[wet zealot]

[weary heron]

"ThAnKs tO RuNDiFfuSiOn fOr tHe pRoMpt" lol, honestly would be embarrassing

[wet zealot]

🤭

[candid trout]

It’s fine. It’s just a prompt. Not trying to be petty.

[candid trout]

Just use a different prompt?

[tame jasper]

Darn 😦

[candid trout]

Honestly it’s cool. All good Joe

[low shoal]

True

[naive bear]

I'm confused, is this or is this not an announcement for a version of SD3 to go open source/weights?

[cinder mural]

hello everyone, is anyone aware of a fix for the img2img generation tab no longer working? i did a git pull update and still no luck, just get a "error connection time out" code

[naive bear]

Are there any example outputs from SD3 Medium to corroborate with the idea that it won't be better than SDXL?

[tame jasper]

How many B is SDXL? I just want to be able to put text on a shirt lol.

[weary heron]

2ish?

If you count the VAE, and the text encs

[cunning aspen]

I.e. we waited so long to end up with a small model that doesn't even reach SDXL in terms of number of parameters? What kind of a scam is that?

[vale steeple]

haven't you see Pixart Sigma

[wet zealot]

yeah

[vale steeple]

big parameter doesn't mean better quality

[wet zealot]

we should be thanking SAI for keeping the 8B

[vale steeple]

it doesn't raise the quality linearly

seriously

[wet zealot]

you people cant handle it

[cunning aspen]

Yeah, it totally sucks.

[vale steeple]

sucks you say

wow

[quiet delta]

5090 --lowvram

[cunning aspen]

How do you know?? XD

[wet zealot]

because its 8 whole B's

you people can barely run 3.5 B's

[cunning aspen]

And what??? My 4090 can't hadle it?? lol

[opal hedge]

There were 5 models advertized. 800M, 2B, 4B, 8B and SD3 Turbo. You're operating on wrong information

[wet zealot]

not at all

[vale steeple]

bro

is this chat informing misinformation in action

[nimble canyon]

Hmmm, maybe we should talk to each other !

[opal hedge]

It can with vram to spare

[wet zealot]

the average user doesnt have the hardware to run 8b

so sai is being responsible and safe

[vale steeple]

4090 is under unoptimized condition

again

[wet zealot]

and releasing a small model they can more easily align and that everyone can run

[vale steeple]

unoptimized condition

read this two word

[wet zealot]

SAI is doing the intelligent thing here

[cunning aspen]

Exactly!

[wet zealot]

lets do some math real quick

8 B's

[rotund forge]

Oversimplification

[wet zealot]

4 bytes (32 bits) per B

[rotund forge]

How many text encoders does SDXL have toughy?

[opal hedge]

I do hope SAI gets their shit together. They really need a community manager to keep us updated and to talk to popular model makers

[candid trout]

Is this Colton? Haha

[wet zealot]

you'd need 32 GB just to load the model

[nimble canyon]

No, this is Patric^W Pony

[wet zealot]

not to mention to handle the processing cache

[candid trout]

Pony XL?

[rotund forge]

Cache me outside

[wet zealot]

with the 2 B's model you just need 8 GBs

[cunning aspen]

Nope. just 8 GB. You can run SDXL on 4 GB. it's 3.5B 😄

[vale steeple]

they already got much better PR manager than some company

lol

[cunning aspen]

12 GB will be enough

yeap

[nimble canyon]

https://tenor.com/view/correct-plankton-gif-14118231

[wet zealot]

its 4 gigabytes per B

[opal hedge]

SAI confirmed that 16GB should be more than enough for sd3 8B

12GB would probably be a little tight

[vale steeple]

that's kinda fair

[cunning aspen]

I think it will be enough for fp16/bf16.

[vale steeple]

SDXL is also around that usage isn't it

[candid trout]

Whatsuuuuup! We should totally talk.

Do I smell a collaboration……

[cunning aspen]

4 GB in Forge.

[opal hedge]

SDXL is 3.5B and it runs with 6GB VRAM at fp16

[vale steeple]

i mean in traditional A1111

[cunning aspen]

6-10GB

[rotund forge]

https://old.reddit.com/r/StableDiffusion/comments/1d0wlct/possible_revenue_models_for_sai/l5q56zl/?context=3

SAI: "We're on track to release the SD3 models* (note the 's', there's multiple - small/1b, medium/2b, large/4b, huge/8b) for free as they get finished."

[vale steeple]

Juggernaut-Pony moment

[cunning aspen]

They must give us at least Large model... But I wanna HUGE one XD

[vale steeple]

that would probably obliterate many finetunes on the street

[candid trout]

We actually did a model with Pony Clip

[opal hedge]

The big one takes the longest to train. More params = more training time

[candid trout]

RunDiffusion Proteus

[cunning aspen]

And what? Let's use 1.5 then XD

[vale steeple]

that's why 2B is all you need

lol

[candid trout]

But it wasn’t Juggernaut

[vale steeple]

i mean

there's an concept in machine learning

[cunning aspen]

I have SDXL with 3.5B, I don't need worst model.

[vale steeple]

where you CAN train only part of the params, instead of all params by 1 epoch

[opal hedge]

I think 2B SD3 will be really good on its own, and 8B will only be a maybe 10% difference

[wet zealot]

8b was too unsafe to release to the general public as weights

[lofty garnet]

yippe sd3

[wet zealot]

just like think about it

[vale steeple]

quality doesn't go up linearly with parameter

think about it

[wet zealot]

how much deepfake you could make with an unrestricted model that big

[lofty garnet]

is there a sd3 large?

[wet zealot]

people's jobs would suffer

[cunning aspen]

Almost a linear relationship.

[vale steeple]

Large and Extra Large

[wet zealot]

artists would lose their commissions to AI

[vale steeple]

who said that lol

[wet zealot]

SAI did the responsible thing

[lofty garnet]

will they relase them or nah

[vale steeple]

we will see

in short

many people being... uhhh

[lofty garnet]

imagine the images it will make

can make

[vale steeple]

well, lets say they expected they are not going to release it due to multiple reasons

[lofty garnet]

i cant wait for anime fine tunes

ah

[cunning aspen]

I said. I know how neural networks work.

[crude notch]

sd3 xl is just bad

[rotund forge]

Any finefunes

[crude notch]

the model isnt even out yet

[rotund forge]

SDXL sucked day 1

(no offense Joe 🙂

I mean compared to now....

[opal hedge]

Yeah. In a few months SD3 will be killer

[crude notch]

wdteam is working on a finetune but it'll take a while

[wet zealot]

see what I said:

1. SAI will not be releasing the high power versions of SD3
2. Because they are a security risk with deepfakes and AI propaganda
3. And they would pose a threat to artists jobs using data stolen from them
4. So Stability is being responsible and releasing a more restricted model to the public

its that simple

[crude notch]

no... sai wouldnt have released xl if that was the case

[cunning aspen]

Or not... Just look at 2.1 and Cascade.

[lofty garnet]

ah makes sense

[rotund forge]

Plus it's going to bring back all the good devs who've slowly left due to boredom.....should be a nice spike in new development/activity

[opal hedge]

Lil bro is making SAI fanfiction out here

[crude notch]

3060 can do xl and maybe sd3-med

[rotund forge]

I posted a link to reddit from McMonkey(w/SAI) saying they will release the other models when finished...

https://old.reddit.com/r/StableDiffusion/comments/1d0wlct/possible_revenue_models_for_sai/l5q56zl/?context=3

[cunning aspen]

server with 10xA100 just for $20/hour. EZ.

[vale steeple]

we all knew 8B is severely undertrained

bro

[vale steeple]

you don't even need that much

[cunning aspen]

Exactly XD

[vale steeple]

4 A100 and training SDXL for PonyXL is enough

[lofty garnet]

i wonder will you have to change datasets for training loras, like will i be able to use the same datasets to train on sd3, i know it will prob take like a month before you can train loras

[boreal pier]

License the Disney loog shoes according to the style of the reference drawing

[crude notch]

no, not that much

just re-tag the dataset

[lofty garnet]

why retag?

[crude notch]

it helps (its more natural language prompts than tag based now)

[lofty garnet]

oh makes sense

[opal hedge]

It might be cheaper to use h100s/h200s

[crude notch]

eh

[cunning aspen]

Yeap. something like 14bucks.

[lofty garnet]

idk why but when i see your username i think flint lockwood

[crude notch]

2-4 mi250xs will do the trick

3090 can do it

[lofty garnet]

what param is the autism diffusion based off of pony?

900m?

[opal hedge]

3.6B, same as sdxl

[lofty garnet]

ah makes sense. sorry for the dumb question, i dont really know about image param sizes

[opal hedge]

Based on some of the conversations I've seen on this discord, you're in the top 10% lol

[cunning aspen]

On the contrary, good for you for asking questions instead of pretending you know everything.

[lofty garnet]

yep i like to learn so i can use that info and then use it and share to other people when they ask

[vale steeple]

oh yeah I just look at several website about parameter size

you can think parameter as a hard drive

with 256GB, 512GB or 1TB

[cunning aspen]

Something like that 🙂

[fallow shard]

Gonna take a guess that with SD3, anything even approaching g-re or er-tica, let alone gor-rotica will just default to NaN noise

[vale steeple]

and generally, bigger dataset required much higher capacity of hard drive, in this case it is your photography pics

[fallow shard]

considering i cant even say the words here, i feel even more confident in that statement

[modern girder]

hey guys newbie here, is stable diffusion free to use

[cunning aspen]

yeap.

[modern girder]

then why do people pay for mj

[cunning aspen]

That's why I said I want a big model, not a small one that's even smaller than SDXL....

[vale steeple]

and while parameter and capacity can increase linearly, the quality of the model does not. You have different kind of pics in the drive, and each of the pics can derivate the total time used to move your drive onto your computer

[crude notch]

❌

[modern girder]

ohh that's cool

[crude notch]

it just, doesnt generate that

there is something i want to leak so bad

[vale steeple]

if your pics is like damn high quality, with high camera resolution or something, you obviously get slower speed compared to the low resolution one with only 10MB

[cunning aspen]

With more information stored, the model is capable of greater originality and accuracy in detail than smaller models

[crude notch]

yeah but then i wont be able to do something

[vale steeple]

BUT you can do much high quality photoshop if you have successfully move it to your laptop

the speed itself is regularization, diminshing returns, architecture combined

[keen falcon]

guys why is when my image is generating I can see something a bit blurry but PERFECT and when its finished it gives me A GIGANTIC MONSTER TURD MADE OF HUMAN FLESH ?

[vale steeple]

like well

a lot of math

[fallow shard]

The thing you want to leak, if you leaked it today, would anyone care that it was leaked a year from today?

[vale steeple]

in real case you often have like up-to-millions of pics in your hard drive.

[crude notch]

no, but it would be funny

[cunning aspen]

Highres.fix with denoising - 0.9 😂

[vale steeple]

and it is almost impossible to find low quality or undesired pics to kick out from the dataset

[fallow shard]

on a scale of 1-10 with 1 being a silly cat being a silly billy, and 10 being the titanic claiming another billionaire's submarine, how funny?

[crude notch]

6, maybe 7

[keen falcon]

what about hires step ?

[cunning aspen]

10 tops. After 10 it has almost no impact.

[vale steeple]

at most 15

i dont do highres.fix much

[static cape]

At least it's less than 2 weeks I guess?

[vale steeple]

yes... at least for 2B

or possibly 800M

[cunning aspen]

https://tenor.com/view/movie-one-eternity-later-gif-7900643

[static cape]

Really don't like the extreme corpo speech. Unless the trained the 2B A LOT more it will absolutely not perform as stated in the announcement.

[lofty garnet]

i would hate that. althought it would happen via finetunes

[crude notch]

dont tell anyone but they arent releasing the other models because 2b just beats it

[vale steeple]

https://x.com/Lykon4072/status/1796251820630634965

[crude notch]

2b beats 8b

[vale steeple]

we do have hindsight on SD3 2B

thx to Lykon

after the hints airstrike

[keen falcon]

doesnt work, im out

[fallow shard]

Yeah, from what I've sort of pieced together is that open source is about 1 year behind the cutting edge closed-source stuff, and the spicy models are about 1 year behind the open source stuff

[wet zealot]

you have absolutely no clue

fluffyrock and pony are

uh

yeah

[cunning aspen]

post screenshot 🙂

[lofty garnet]

huh how

[wet zealot]

the goon models are SOTA

for open

[crude notch]

it just does...

[lofty garnet]

weird but good

[fallow shard]

they are?

[vale steeple]

It just works

[opal hedge]

Could be that 8B takes longer to train all the parameters and 2B was more ready for release

[crude notch]

maybe

[static cape]

The announcement states it's the 2B.

[south sun]

nothing is canon abandon your gods appreciate your dinner

[vale steeple]

we knew for sure it is a 2B release of course

[lofty garnet]

they prob send a differnt version to each person that has little trackers, so if you leaked it they would know it was you right away. of course they could have been dumb and not done that

[vale steeple]

just that 800M may come along with it

that's what I meant

[crude notch]

no not model

a screenshot

[lofty garnet]

ah

prob not the same security

[vale steeple]

well

[fallow shard]

i've tried pony, anything beyond a single character or pose with 10 trillion examples on the interwebs and the anatomy goes all eldritch abomination, fluffy rock, never tried

[vale steeple]

I am pretty sure you guy forgot the legal problems if one partner leak the SD3

[crude notch]

hello dont mind me leaking info (this is all public, just not known)

[lofty garnet]

you still could by adding something to change the hash and then when you upload it they could compare it, but a lot less likely since its a screenshot

[wet zealot]

you say this but

[static cape]

And now let's hope the 2B either performs well enough Or that the 8B will get oprnsourced as well (and that they won't taunt the Opensource community with a "demo-Model" that sucks compared to a cash only paid "Pro" Model).

[cunning aspen]

Something tells me SD3 is another fail. I hope I'm wrong.😵

[vale steeple]

"where is your source or evidence" right?

[wet zealot]

no

I know more on this topic than you

[crude notch]

2b is better than 8b fwiw

[wet zealot]

that's what i will say.

[vale steeple]

it's always better to lowering your expectation

on base

so that you can be surprised after someone finetune it

[wet zealot]

SD3 8b becoming available is not impossible

otis knows too

[crude notch]

well yeah its gonna leak eventually

[wet zealot]

😈

[cunning aspen]

I'll try to do that 100%.

[vale steeple]

yeah sure, continue write your essay on how SD3 is not going to change open source community revolutionary

[pseudo coyote]

Hey guys what is stable diffusion 3

[nimble canyon]

https://tenor.com/view/false-wrong-fake-incorrect-untrue-gif-23252753

[static cape]

That'd would be great. Easier to train and to do Inference with.

[lofty garnet]

i wish novel ai anime model would leak

[trim magnet]

2b better than SORA too

[vale steeple]

that's like ancient ago

?

[crude notch]

i asked them, they said no....

[wet zealot]

june 11th mark the date

[lofty garnet]

the new one thats really good

[crude notch]

oh my god did lykon actually say that

[nimble canyon]

Ah, fair

[static cape]

Goon?

[lofty garnet]

GOON

[wet zealot]

models for gooners.

[cunning aspen]

February 58th, 2025. 🤣

[wet zealot]

realisticvision, etc.

[lofty garnet]

so tell me why im planing a heist so i hire some goons. they show up and instead of saying whats the plan boss they say when can we start gooning

[fallow shard]

spicy models for private times with a box of tissues near at hand

[lofty garnet]

dont forgor the lotion

chat dead

[trim magnet]

2b better than SD4

[cunning aspen]

cause It's not hardcore stuff.

[lofty garnet]

makes sense

[sage reef]

holy shit.... 300 new messages, no wonder why... so it's finally gonna be released nice,
can't wait to see what the community does with it too, very exciting news!

[lusty beacon]

Guys why only Medium Size Model?

[indigo ermine]

Can I assume my 8gb card won't handle SD3?

[hybrid cloud]

Guys, one of the stability staff has confirmed the plan hasn't changed on reddit. They will still release all the variants. He said other variants still needs more training thus only releasing one variant for now. Once they are satisfied with other variants, they will release them too.

[finite cloak]

i mean yes it was trained a lot more - but of course judge it for yourself on the 12th

[finite cloak]

he was trying to express that an advantage of small models is that training them is more accessible, vs large models require more hardware to train, and words get mixed and the internet did its usual thing of dogpiling hate on someone for imperfect speech

[crude notch]

ah, makes sense

[finite cloak]

Medium is first not only

[vale steeple]

from my understanding you can train part of it right?

at least in theory

[crude notch]

lora ye

it works for all

[lusty beacon]

True

[pulsar shell]

What the native size support and VRAM requirements of SD3 Medium?

[crude notch]

12gb vram is enough afaik

[vale steeple]

probably 4-12GB VRAM

depending how good is the optimization

[crude notch]

prob 8 if i had to put a minimum

[finite cloak]

yes, the minimum training requirements when you do stuff like that are always a bit more than inferencing it, vs inferencing it you can cheat to go below the theoretical requirements (eg offloading)

so eg if you're barely able to fit the 8B in your GPU... you're not gonna do even a LoRA

[lusty beacon]

What means commercial related to SD3?

[finite cloak]

but that same GPU can easily fit the 2B for either inference or training

[vale steeple]

watch #🆕｜sd3

[lusty plank]

will a 4080 be enough to train a lora with the 2B model?

[vale steeple]

PonyXL creator is asking him about that commercial things

[finite cloak]

1024x1024, VRAM usage should... proobably fit under 4GiB? not sure / no promises, but, if you have a modern low-end card like a 3060 or whatever you're more than golden. Anything that can run SDXL is golden.

[vale steeple]

fp8

probably

[pulsar shell]

How about output size support of SD3 Medium? I mean resolution. 1024 or 512?

I don't know at this stage why SAI still not give more details.

[vale steeple]

prob final decision by the upper

i mean commercial license thing

I believe SD3 Medium should handle both of the resolution

[finite cloak]

1024x1024 as i said above

[finite cloak]

?

[rotund forge]

Alex, is SAI giving early access to any of the developers of training tools (Kohya/Nerogar-OT)?

(if you can say) - just a head start on implementation

[finite cloak]

Early access to relevant developers yes, those ones specifically have not been given early access

Kohya at least is based on HF code and HF has early stuff as always so shouldn't be an issue

onetrainer i'm not sure offhand what they got

[rotund forge]

Cool, thx! Super excited (particularly for community fine tuning to begin)

[pulsar shell]

SAI s email doesn't mention support resolution or VRAM requirements. just say 2B . This technically term does not has any meaning for a normal user.

[finite cloak]

Ah yeah sorry that wasn't a very technical publication at all, mostly just meant to publish the date rather than anything specific about the model.

[lusty beacon]

Can I create bigger Images than 1024x1024

[finite cloak]

With SD3 Medium? yes, similar technique as with SDXL (hires fix / ultimate-upscale, using tiling is recommended)

(the mmdit arch gets a bit weird if the res is too high but tiling fixes that fine)

[vale oasis]

im hoping the license will be changed soon

[fluid atlas]

HI GUYS

IS ANYONE HERE?

https://tenor.com/view/gojo-satoru-gojo-satoru-jjk-jujutsu-kaisen-gif-7280593681001594903

[vale oasis]

everyone is on #🆕｜sd3

[fluid atlas]

HI

[static cape]

Oh I will. I'm absolutely looking forward to finally getting my hands on the weights.

Do you have any tips on how we should prepare our datasets?

1024res images seems to be a given.
A combination of tags & natural language captions should work well, I guess?

Are you guys talking about any best use Schedulers for training (like AdamW and the like)?

Any preparation steps we should look out for?

[finite cloak]

1MiP res yeah. (Variable aspect don't lock in to square, same as SDXL)
Natural language recommended, high caption quality is recommended. You might also want to try having CogVLM generate captions for you (this is how half of SD3's training data was captioned)
The rest I'm less sure on but @primal wing can probably give some insight

[keen falcon]

just to be sure guys : the more steps, the better right ? Its just a question of of how much time we want the generation to last ? But if u have a good computer, its fine ?

[finite cloak]

Vaguely, generally, yes.
Practically, eh? Up to ~20 or so, the difference is huge. Up to ~50 it helps. Past that it's pretty redundant but you'll get a tiny bit of quality maybe

There are turbo and half-turbo models that work best at 4/8/12 steps

(Turbo at 4, Half-turbo at 8/12)

[keen falcon]

well im gonna burn my threadripper with this one but lets do it

thx

[static cape]

@finite cloak can you explain somehow how the 2B has a third less data than SDXL and still performs way better?
Quality over quantity?

[static cape]

Kinda ironic since for the Training community these two are the most relevant.

[finite cloak]

AI models generally should be running on a GPU rather than a CPU. If you can afford a threadripper hopefully you can afford an RTX 3060 or something

[errant pulsar]

I'm sorry but does anyone know when the bots will be brought back and if they will ever be brought back?

[static cape]

Btw @finite cloak thanks for your continued community communication. It's much appreciated.

[finite cloak]

Size isn't everything? Mainly.
GPT-3, a 175B model, was beaten out by LLaMA-13B, at under a tenth the size. (the LLM not the chat finetune used as the basis of GPT-3.5)
SD3 is trained with way better data (notably the CogVLM autocaptioning, vs prior models were trained with "whatever nonsense text the internet associated with the image"), has a way better architecture (MM-DiT vs unet), and has a much smarter VAE (the 16-channel VAE in SD3 seems to have figured out a partial feature channel separation, vs the 4-channel VAE in SDXL acts more like a funky color space)

[static cape]

For XL clip upscaling sometimes alleviates that as well. I had success with upscaling the latent & the clip x1.5 before running a second Inference pass.

[finite cloak]

The mostmost relevant is HuggingFace, creators of the Diffusers lib which is the backbone for most training tools, and they got early access.
But yeah it's sad that we didn't manage to seed early access to community developers like kohya this time. It was awesome doing that with XL last year

[finite cloak]

the original bot has been replaced with Artisan now, #artisan-faq

[finite cloak]

the issue with MMDiT large res is a whole new one - the architecture is generally just friendly to unusual resolutions whereas XL was more adaptable. There will probably be solutions to res issues on mmdit published soon in the same way they were for the unet setup, it'll just have to be new solutions for the new problems

[static cape]

Oh I have been using CogVLM for a while. Did you guys prune Cog's mannerisms like "seems to be" "appears to be" etc.? I like I a lot, but also do manual caption refining.

[finite cloak]

there was definitely experiments&refining on the way cog was prompted to generate prompts, don't know the specific details offhand

[static cape]

Ah. I admit I didn't know Huggingface are the devs of Transformers.

[civic crown]

I can't speak for kohya. But OneTrainer is pretty much all custom code. Diffusers is only used to load the models

[finite cloak]

ohhi

[static cape]

Could we get a short devlog on that?

It would help immensely to know how (which prompts, which data curation steps) SD3 was captioned so that our own efforts can align better.

[keen falcon]

RTX 4090 24Go

[finite cloak]

uhh maybe after release we can poke some of the researchers about it, idk

[finite cloak]

oh yeah, burn the 4090 then lol not your cpu

[static cape]

It would be great if you could inquire about that.

TagGUI will also probably support CogVLM2 soon. Looking forward to how it compares to 1.

[errant pulsar]

@finite cloak Am I to understand that there are no free trials anymore, only the knitting version?

[pearl ocean]

I just have a old vintage 4080

[pearl ocean]

can we get SD3 on local now?

[static cape]

(a bit less than) 2 weeks.

[pearl ocean]

they been saying 2 weeks for 6 weeks now

[finite cloak]

June 12th

[static cape]

But now it's more official. 2B gets open sourced next Wednesday.

[pearl ocean]

😮

[pearl ocean]

can't wait!

[static cape]

Same. 😉 I'm reading all comments I can find at the moment so that I get all the info I can.

[spare ether]

Hey @finite cloak tried to ask via X but not getting a response, is the Stability member (20/mo) covering the ability to self host SD 3 when it releases on June 12th? I am confused as the email and page on Stability site leads to another form. The form is similar to what we filled in when we got the membership but there seems to be no clarity on this. I think it would be good to clear this up.

[static cape]

@finite cloak will Adapters / LoRA's / CN's etc. for the SD3 Model-Zoo be compatible?
For example could at some point someone train a LoRA (or something like that) on 8B and it would be usable on 2B as well?

That would be huge. Otherwise we'll have a 4-way community split down the road, even if most people adapt SD3.

[finite cloak]

that'll be clarified before launch day

[spare ether]

That's ridiculous that it isn't just a yes then.

[finite cloak]

out of the box, no, but the x-adapter project did that before with sd1 and sdxl so they'll probably do it again

[finite cloak]

yeah idk what's up, they got something they need to figure out. I have to presume it's essentially just a yes waiting to happen cause why would it not be, but, idk im just a dev

ftr noncommercial/personal use is entirely free

[static cape]

I really hope that will be the anwser that brings everything together.
When X-adapter released we already had so many LoRA's for XL that it didn't wind up used a lot (Afaik).
For SD3 that could be very different.

[spare ether]

Ok nothing against you and i appreciate you answering. I'm gonna be super disappointed if it isn't in the paid membership as that's something I have been happily paying in anticipation for this release.

[static cape]

I can only guess that they want to avoid another XL which led to dozens of projects / sites monetizing the Model and Finetunes successfully without cutting SAI in.

[low moon]

guess i wont do much next 2 weeks

everythign will be obsolete after this

ella sigma cascade sdxl

no need to jerry rig

[loud solar]

Better plan with 4 weeks 🙂

[low moon]

well either way most of our stuff will be obnsolte soon

just like eveyrhtign before sdxl

[loud solar]

2 weeks till release and 2 more for infrastucture ...

[low moon]

its been worth the weight and now it will blend

[loud solar]

And the old created pictures ... I still love some of my first ones ... they've become oldies now and special in their own way

[low moon]

i cought the tale end of the sd15 era

i barley got used to sdxl

i can never develop a workflow or pipe line

everythign changes all the time

sometime si just pull the slot machine of foocus and give up

[loud solar]

I have started about 2 years ago ... don't really remember the version 😄

[low moon]

hmmm

[forest trout]

SD hasn't even been out for 2 years...

[low moon]

images back then were really bad right?

everything was deformed

and gooey

[loud solar]

Maybe Beta?

[low moon]

first time i heard about these text to images were bad hand memes

[forest trout]

This board started Aug 30 ish 2022.

[low moon]

[forest trout]

Weights dropped that October.

[loud solar]

From todays view it's been kinda dirty ... but that's what some people are missing today ...

[low moon]

Shutterstock watermarks were the key to all of this

[forest trout]

Most likely if you were using it you were using Craiyon, Midjourney or Dall-e 2

[low moon]

and getty

[loud solar]

I've been using all I could get. 😄

[forest trout]

I mean I definitely was using Dalle MIni (Craiyon now) quite a bit.

[loud solar]

Interesting times ...

[low moon]

😏

[next lintel]

😍

[loud solar]

Local installation took me a week ...

[primal wing]

We had some success with zero captions (just style tags), but created some anatomy issues, so we scrapped that.

[vale steeple]

oh hi creator of dreamshaper

[finite cloak]

that is the most cursed input you could have possibly added there thank you

[forest trout]

Good times. First images I ever created with Stable Diffusion.

[primal wing]

It's very likely that style-only captions are very good for pretraining

[loud solar]

Thanks to everyone for sharing information!

[pearl ocean]

I wonder what images ppl make with Stabel Diffusion

[charred mesa]

Hmm can't even imagine

[low moon]

=0 the famous Lykon

[fluid atlas]

hi gays

*guys

[autumn hinge]

who knows how to AI undressing

[loud solar]

Slow ...

[finite rain]

how to get membership of SAI

[low moon]

buy Emadcoin

[rugged mirage]

didn't they originally say we are getting Controlnets with the release? Am I misremebering what someone who wasn't actually staff said?

[low moon]

😮

SAI makes the controlnets for us?

[rugged mirage]

As far as I remember they were planning to this time, but perhaps not anymore since I didnt see it in the latest announcement

[oblique edge]

Good beautiful morning everyone

[vale steeple]

what?

I mean it is Illyasviel do the model thing all the time

[rugged mirage]

historically yes, and this time as far as I remember, they said this time they are doing Controlnets in a new way and they will release them with model release. I'm asking if I'm misremebering and it wasn't SAI who said that

[vale steeple]

include SDXL models

[rugged mirage]

but anyway, I just saw that they haven't been working on them in the end

this is what I remember and was just asking for

[vale steeple]

I dont remember there's a call for that sorry

[soft iron]

this is fake, weights will not be released

[vale steeple]

if not real then it may be better the StabilityAI's Twitter account is fake

[soft iron]

it's all fake, you'll see on June 12

[vale steeple]

sure lets bet

[gusty mist]

how much vram will be needed for inference on sd3 ?

[low moon]

64GB.

[hallow karma]

seriously

[still glacier]

No

[hallow karma]

phew

[still glacier]

iirc, that information comes from the paper and it s the worse case scenario, without any of the community optimisation.

[keen parcel]

in the https://stability.ai/stablediffusion3 form

what is the "discord id" form input for?

is it a string or (discord username) or integer (true discord id) expected?

thanks

[still glacier]

Id say just click your avatar picture at the bottom left of the screen then click "copy user ID" and voila, use whatever it gives you.
As to what it s used for. 🤷‍♂️ But it looks like it s not mandatory.

[teal pagoda]

SD3 will be out

but it's a total joke

They said "SD3 Medium"

So that means there will also be "SD3 Premium/Large" most likely behind a paywall

🤡 these clowns

[still glacier]

Or it could be they ll just release the rest of the models later on like they ve done with their llm....

[keen parcel]

Thanks I wanted some feedback on this

Is StabilityAI in good shape financially?

The Stability HPC cluster looks cool

[iron pollen]

Bruh we gotta pay for the discord bot now???

I’ve been waiting this long for THIS??

I detest the direction this industry is going. What a shame

[teal pagoda]

Good joke. They should have all the models ready by now

10 months since XL launched

so don't tell me they didn't have time to train SD3 in one year

and resources

[teal pagoda]

everything is pay2use/pay2win nowadays

unfortunately

[vale steeple]

buddy

[still glacier]

you can still run SD locally.

[vale steeple]

did you just see the first word and decided it is the announcement of paywall

[vale steeple]

seriously you think they develop a DALL E 3-quality model a year earlier and decided not to announce it

critical thinking at its finest

[low moon]

Such a pessimist.

Stop beign a socialist comunist.

This is AI no room for isms.

[loud solar]

Let's use the time to clean up our harddrives 😄

[low moon]

looool

thats scalty what ive been doing

300 gigs!

[loud solar]

Should be enough for a week after release 😄

[low moon]

😬

[loud solar]

I'll go enjoying the nature. Have a nice day!

[ripe spruce]

From Stable Audio's Terms of Service:

Opt Out. You may reject this Arbitration Agreement, in which case only a court may be used to resolve any Claim. To reject this provision, you must send us an opt-out notice (the "Opt Out") within thirty (30) days after you create a Stability membership account. The Opt Out must be sent to the Stability Notice Address. The Opt Out must include your name, phone number and the email address you used to sign up and use the Services. This is the only way of opting out of this Arbitration Agreement. Opting out will not affect any other aspect of these Terms, and will have no effect on any other or future agreements you may reach to arbitrate with us.
What is the "Stability Notice Address"?
Do I have to send an international, physical mail to opt out of arbitration?

[plush jetty]

Did I read the announcement right that Stable Diffusion 3's release date is June 12th?

[stark hornet]

any news since the last few weeks ?
images kinda still look the same sadly

12th june ? but did they say which year, too ?

[plush jetty]

2024

[stark hornet]

can you link that announcement ?

that -would- be some news 🙂

[plush jetty]

I have this: https://www.reddit.com/r/StableDiffusion/comments/1d6szfg/sd3_medium_release_on_june_12th/

[trim magnet]

just check #📣｜announcements

[stark hornet]

ah, they didnt state the year, just 12th june 🙂

[soft iron]

most likely the larger version simply will not run on a regular PC

[stark hornet]

well, -some- people might be able to run it nonetheless ?

[soft iron]

those people who can launch already have money for paid generation

few people buy just for fun 5 teslas to run large models

[stark hornet]

may be a personal preference, but i like to run stuff locally

[soft iron]

Sure, but would you buy 5 Tesla cards just to run a large model?

[frank plover]

Hi

[stark hornet]

i wonder if the sd3.0 will still generate images that can immediately recognized as being done by stable diffusion

i m not really sure what you mean with tesla card

[soft iron]

Enthusiasts buy several cheap Tesla cards to run large text models

[soft iron]

But the pony can already handle it

[stark hornet]

i didnt understand either sentence (what pony?)

i got a 4090 here, which probably aint enough soon anymore, but maybe a 2nd 4090 could solve the problem already

[trim magnet]

every model out there (1.5 and sdxl) can generate either generic looking ai imgs or very realistic ones that look like they were taken by cctv or by webcam

[soft iron]

model trained for anime/cartoon porn

it is so well trained that it makes pictures indistinguishable from art

[stark hornet]

ah, i m not into anime/cartoon porn at all sadly 🙂

i -am- very much interested in art though

[soft iron]

its modifications can do this

[stark hornet]

i got a few nice pictures done today even

can you link me the pony model ? and why can cartoon porn model make art ?

[soft iron]

most often it is very easy to recognize due to the insufficient color palette

and in the case of realism, because of the same faces

[trim magnet]

yea but u can make them better with loras

[stark hornet]

i put three of my latest works into the #1072236442463518882-with-images channel

maybe you like them, too 🙂

[soft iron]

Depends on your goals. The model is trained to make characters. It can't do anything else.

[soft iron]

you don't need characters then pony are useless to you

[stark hornet]

my goal is to make ai art that doesnt scream "this image was done by stable diffusion"

[soft iron]

ok

[soft iron]

it's very easy to do

[stark hornet]

how?

[trim magnet]

just go to civitai look at all the img posts there and copy the settings/prompts from whatever img you like then try it in your webui

[soft iron]

correct model, prompt, loras, comfy nodes

[stark hornet]

i feared it would be a generic answer, you got some nice pictures to show by chance ?

[soft iron]

I don't do that kind of thing. I'm interested in following ai and being able to work with it, but not creating. I don't feel it as creativity.

[trim magnet]

like these ones?https://civitai.com/models/162136/pablo-picasso-infinity-sd-xl-10

[stark hornet]

well thats another thing i dont want to do, i dont want to do "image like picasso did", but create something new

(and yes, i know that the model itself cant really make up new things, but what it for example could do is mix things in new ways)

[soft iron]

the person works exactly the same

[trim magnet]

yea with sd u can combine several artists styles or loras at different strength to create a unique look

[stark hornet]

have you done this successfully yet ?

[soft iron]

https://civitai.com/images/7491833

I don't think a person is capable of creating something like this

so it’s strange to treat it like "image like picasso did"

[stark hornet]

looks like minecraft meets old school runescape 🙂

[trim magnet]

yes i have i combined 5 different loras with block weights at diff settings to create a unique look/style

[still glacier]

no it would be pointless there s no way SD workflow that can unify memory across two gpus. You would be able to run two instances tho, one on each gpu. But that s about it

[stark hornet]

thats at least some kinda new ideas, even though its still recognizable styles (like minecraft and old school runescape)

[winged jay]

Hey guys can you tell me which channel should I go to for help in using stable diffusion in local

[soft iron]

youtube.com

[stark hornet]

how about this one ?

[winged jay]

Can I ask my doubt's here

[stark hornet]

sure

[winged jay]

So I installed stable diffusion following the instructions from the website https://stable-diffusion-art.com/install-windows/ but my photo's are coming out bad

[stark hornet]

most people are using comfyUI these days, i think, not a1111 (but it may be a personal preference)

[winged jay]

does the images come out good in that?

[stark hornet]

can you show a screenshot of your settings and the result you got ?

[winged jay]

I can't seem to upload it here can I dm them to you

[stark hornet]

put in the #1072236442463518882-with-images channel

[winged jay]

Done

[wet grotto]

guys where is the better wayto have multi prompt in A11? masks with in paint upload?

[silk gulch]

I’m confused wasn’t the biggest model 8b?

Are they not releasing that

[still glacier]

#🤝｜tech-support

[foggy halo]

sd3 when ? June 12 yayy

finally

[arctic sedge]

I have offical infomation that the 2B will require 15gb of Vram to run. (With the T5 encoder) I gather this can be decreased with the use of Xformers, and the FP16 format? But does anyone have a idea by how much that will improve vram usage?

[charred mesa]

T5 encoder with bnb4bit is going to make it way less than 15GB in total to run

or just use it on cpu, with the bf16 of fp16 weights it loads okay and doesn't use as much ram as the full fp32 weights

I just wonder if a ggml implementation of encoder models would help with this

where did you get that official information?

[arctic sedge]

Yeah. I'm wondering by how much we could get this down by. Not so sure. SD3 is a real heavy beast.

[silk gulch]

Yes but I thought their biggest model was 8b?

[tropic frost]

sooooooo

anyone here uses pony?

[soft iron]

yes

all uses pony

[tropic frost]

I want to give it a try, so i want to know a little bit about it.

[tepid burrow]

Is there any chance a 3090 will sell after 5090 comes out? I want to buy a 4090 when the price crashes.

[soft iron]

The standard model only works well with lora styles

Modified models work well initially, but are very limited in what can do

Score_% replacing quality modifiers from other models

[glass cloud]

Any AMD users here?

[pastel lily]

hey anybody here who needs an video editor if yes then pls dm me?

[true canopy]

i wonder why they decidede to call it sd3 medium instead of 2B, maybe it sounds a bit fancier

[tropic frost]

Are the any compatibility issues with SD Zluda?

[true canopy]

btw i called second week on june for sd3, i was spot on

[soft iron]

don't know, I have nvidia

[tropic frost]

yeahh... i use amd

[trim magnet]

my condolences

[glossy zinc]

do we know what SD3 API versions correspond to in terms of model size? sd3 vs turbo etc?

[dull field]

So the new SD3 is 2B model (at least the version they are releasing) but requires (I know it's optional tbf) 11B LLM to condition it?

[arctic sedge]

Yeah. I was hoping maybe we could some how squeeze it to 10GB with the encoder.

[astral goblet]

Omost is pretty spiffy. really ups the prompting game

[manic wraith]

Start with AutisMix and prompt it right (it has some weird tags that are absolutely necessary). It can do anime style better than any other, even AnimagineXL.

[tropic frost]

honestly, thats what im looking for, im trying to get anime screenshot like results

[glass cloud]

Do you have stable operational, this is my first venture and am struggling a bit to get it working.

[tropic frost]

works decent i guess...

[manic wraith]

Just make sure to put the source_furry, source_pony etc in the negatives, also those score tags, just copy a good prompt off civitai and you'll be good to go

[manic wraith]

Doesn't the A1111 installer set it all up for you these days? Or ComfyUI for the nerdier experience

[glass cloud]

I have zero idea, what you're talking about. I'm new to pc, and stable diffusion. I'm a console gamer, and have been using generator apps, like dreamerland.

[fervent thunder]

can I tell from a lora metadata whether regularisation images were used during training?

[astral goblet]

nope. only if the author put that into the meta data.

[fervent thunder]

nvm, I am stupid:
"ss_num_reg_images": "1010",

[cobalt turret]

As far as I know SD3 large is able to run on a rtx 3090 (24gigs of vram) judging by the research paper. Also if I can run SDXL on 4gigs of vram, and SD3 Medium is smaller then SDXL in terms of parameter count, I think it will be possible to run Sd3 Large (maybe with some vram optimizations, making it slower) on a decent graphics card. (or cpu if you want to wait 1 day for one image 😅)

[rugged mirage]

pretty sure the math has been done a bunch of times and you can run the largest sd3 on 24gb and even 16gb

[cobalt turret]

Btw, I have another question: does somebody know what exactly the non-commercial license for Sd3 medium means? Does it mean I can still create a bot with my self-hosted Sd3 fine-tune, if it's simply giving it for free to my friends (100 people) who I invited to my discord server. Or would this already fall under a commercial use case?

[atomic mortar]

i think it boils down to, do you or they earn money of it

[cobalt turret]

What if none of us earned money with it? (for my little discord bot)

[rugged mirage]

the it's non-commercial and you are fine

[manic wraith]

well then you have a lot to learn, but it'll be fun! mostly. you have the power to do anything, even the power to make it harder for yourself.
just google automatic1111 and go from there.

[astral goblet]

nice its there!

[glass cloud]

Okaly dokaly

[verbal osprey]

#release the 8b cut

[arctic sedge]

@finite cloak I'm slightly concerned about the licensing. Might the current direction of licensing for SD3 2B hamper community development, and finetune creations?
I hope i'm wrong though.

[finite cloak]

noncommercial/personal usage is entirely free, commercial usage they're figuring out and will clarify when they can

so community development should be fine

[rugged mirage]

is commercial likely to include e.g. making a youtube video that has ads?

[arctic sedge]

Thx for everything you and the SAI staff have done for AI image gen btw. You're all doing great. 🤗

[finite cloak]

(I am not a lawyer don't quote me this paragraph is just guesswork) I think you should be fine on that? Youtube stuff is usually pretty lax in general with most things. And like the point of the $20 membership fee is that it's to be paid only by people that are making enough more than $20 that the cost is trivial to em, so if you're barely making cents a day or even a few dollars a day on YT it shouldn't matter.

The usage of outputs generated by AI is a weird legal mess area so if you have non-trivial-income commercial plans you should consult a professional lawyer before making decisions.

For a precise/accurate answer on Stability's definitions of things check https://stability.ai/professional-membership-agreement or https://stability.ai/membership and/or contact the memberships email there to ask the actual professionals about it

[rugged mirage]

okay thanks

[wise ginkgo]

Why? j/k which gpu?

[wise ginkgo]

you're dreaming? why would the price for 4090s crash? IF anything, they'll go up slightly briefly or stay the same - 4090s are already low stock aren't they? then they'll be very few when the 50 series is released

[rugged mirage]

it'll likely go down

3090s still went down after 4090

[trail lion]

3090s are still inflated, there's still demand for them...so if 4090 goes down, it wont be by much if the 3090 is any indication

pretty much anything that says nvidia is inflated tbh

[mellow kraken]

Guys when is the Stable AI is coming bacK?

[trail lion]

the bot? check out artisan on their site

there's no free one

[manic wraith]

I suspect the used 4090s might drop to almost used 3090 prices, and used 3090s would drop by $50-$100 at the most, because as you say there's still demand

AMD and Intel keep dropping the ball on every aspect, but if they merely released a 48GB "consumer" card, that would push the envelope and people would probably do the free labor of making AI work well on AMD/Intel cards

what I expect of AMD: same VRAM as Nvidia, slightly lower performance, slightly later/earlier, slightly cheaper, supports some of the things Nvidia supports

[trail lion]

yah, AMD's highest performing card right now performs worse than a 2080 TI in terms of a benchmark specific to SD. https://www.tomshardware.com/pc-components/gpus/stable-diffusion-benchmarks

and that's not taking VRAM into account even

[visual peak]

how do you change the size of the image or dimensions, what command?

[opal hedge]

Did they leave somewhere?

[mellow kraken]

Well this server used to have a free ai named stable AI and they vanished

[trail lion]

free as in making them bankrupt free, yup, they did

[pale latch]

the cost of running the bot probably wasn't what hurt their bank accounts

[trail lion]

I was being a little silly there, dont need to take me so literal, but the point is if they are hurting and having problems paying cloud providers, giving that away for free makes zero sense

[loud solar]

And an early version of SD3 isn't the best marketing ..

[molten yarrow]

I have just installed stable diffusion locally but the first tests are not at all conclusive, any idea how to improve the rendering qualitatively ?

[loud solar]

Have you tried community modells? You'll find at civitai.com

[molten yarrow]

will take a look, thank you

[loud solar]

And model Version can also be a thing ... but many other options, too ...

[loud solar]

If the problem stays you can send a screenshot with your settings here https://discord.com/channels/1002292111942635562/1004159122335354970

But usually you can read the most important settings in the model description ...

[teal pagoda]

I'm just realist. :)))

[loud solar]

A pessimist telling he's a realist ... that's depressing ^^

[forest turret]

hey, did anyone have problems signing up to use stable artisan?

im trying to subscribe but after logging in it sends me to the stability assistant chat bot

[visual lagoon]

have they shown any examples of what "medium" level model does

[nova jacinth]

Hi, is it possible to use Stable Diffusion for Architectural project, to make previews of how an interior would look, starting from a drawing?

[low moon]

It's not so hot with anythign involvign straight lines and mechanics

[fervent thunder]

DATASET CURATION

Hello viewers, would someone be willing to look at my dataset for training a celebrity LoRA? ...and please give me tips on what to add and remove from it?

[low moon]

if you have a pretty detailed drawing and u use img2img u cld theoretically get somethign usable

[south sun]

Should be fine if you just go ahead with Assistant. If problem you can blame me

[south sun]

Try the sketch and structure control APIs first to see what you could do with a neat workflow. On platform.stability.ai

[gilded wolf]

hey

[fervent thunder]

I have a 3060 too

12 gigs of VRAM

I’m having VRAM storage issues

So I’m almost in the same boat hahah

i think i have 12 anyway

yeah, 12 gb

[low moon]

so is 12 good for SD3 (medium)

?

will it produce 1024/1024 image sin less than a minute?

(without loras contorlnets whatever)

20-30 steps on average...

[wispy vessel]

so,Is it a 2b size model?

[balmy rune]

Finally

[balmy rune]

Yes 2B

[wispy vessel]

I hope my gpt can run it safely.hhh

[forest turret]

what do you guys think of SD3?

i haven't been very impressed so far, i just tried a few prompts

Dall-E is giving me better results. it's frustrating that SD can't do hands properly

[blazing mica]

Give it time to local release and before long custom models will have it entirely solved

[trail lion]

mmm, custom models

[forest turret]

ya im hoping that the local release will be better

when is that happening anyways? june 12?

[blazing mica]

Yeah I believe so

[trail lion]

and then several weeks more till there's something worthwhile

[forest turret]

what version of SD are you currently using for your image generations?

[trail lion]

and maybe even longer until the degenerates get their due

[forest turret]

im trying to generate sprites for some games that im making, but i'm not having very promising results

[blazing mica]

1.5/xl

[wispy vessel]

1.5

[barren dove]

怎么画画

[wispy vessel]

部署webui或者comfyui

[willow crescent]

I beg your pardon? 🙂

[willow crescent]

remove every celebrity 😄

[willow crescent]

@low moon no, seriously, wdym?

[low moon]

it sucks at mechanical things and straight lines

its too gooey

someone was asking if he can use SD for rendering architecture stuff

[trail lion]

conceptually it shouldnt matter to SD. a bed, a table, a house, a building. so ultimately someone will need to create a really good architecture finetuning to get it to understand how to draw those things. there has to be a desire for it first.

there are some on civit, I just have no clue what their quality is

[blissful slate]

Anyone know of a reliable tool to check for AI deformities? I've been building out a full automated pipeline with the SD3 API, and while the outputs are usually pretty good I occasionally get images with random non-sensical objects, weird hands, detached hair, etc (the usual AI nonsense).

Tried using GPT vision, and a couple other tools, but haven't had much luck yet.

[trail lion]

reliable tool, your eyes 😉

if you're someone trying to mass produce for commerical purposes, good luck

[willow crescent]

#🏞｜general-with-images message

[blissful slate]

Yeah, that's my goal in this case lol

Nothing crazy, but I'd like to avoid needing to manually check if possible

[trail lion]

anyone else notice the flood of influencer loras on civit...like who are these ppl

"when is a celebrity not a celebrity?"

[blissful slate]

It no doubt stems from Aitana, everyone saw her making bank and decided to try on their own

[narrow kernel]

Because parasocial stalker incels be doin parasocial stalker incel things

[valid beacon]

who is aitana

[blissful slate]

https://www.instagram.com/fit_aitana/

she was all over the news for a while

[valid beacon]

oh she looking fine 😏

[cedar osprey]

hi how are you all

[fallow veldt]

so SD 1.5 looking

[verbal osprey]

I hope the rtx 5090 has 50gb of video ram. AI let's goooooooooooooo

[willow crescent]

naaah 😉 😄

[blissful slate]

They did a decent job of face consistency, but if you look at the body it's pretty obvious

There was another called Emily Pellegrini that also managed to do videos (probably through face swapping). She deleted all her posts though

there's "fan" profiles you can search if you're curious though

[willow crescent]

319k followers, what bs 😄 As well AI-generated?

[valid beacon]

🤑 💰

[fallow veldt]

I would say the opposite, the body was ok but the face it is very AI looking. Good face consistency, but on a really AI generic face. Also it suffers with so many pics with the same lighting, and in most a soul-less look (!?)

[valid beacon]

she can finally compete with all the soul-less looking celebs on IG like the Kardashians

[blissful slate]

Yeah they definitely aren't amazing. It does look like the same person though. Definitely enough to fool the average person (especially based on the comments)

[fallow veldt]

lol, yes, I've seem so many AI generation I can spot them on the fly, mostly if they are made with SD 1.5 xD

[blissful slate]

Lol same, I've been poisoned to the point of excessive perfectionism

[fallow veldt]

Like from the pose, the lightning, the expression, the perfect overall averaged face, even with the same hairstyle

[fallow veldt]

The first one fool me, but it is a photobash or face swap. Others are kind of bad photobashes, or full generations. But there at least th face looks real, in a way, not SD 1.5 looking https://www.instagram.com/tucumandigital/p/C1ndOWvrhXY/?img_index=1

[blissful slate]

Yeah the videos she had were decent, I think that's what the first image is from. They were definitely all faceswaps though

[small prism]

hello everyone, can anyone suggest me a model from civit.ai where i can drape garments on models.

[ashen lark]

hi

[bright wagon]

I am a novice (intermediate?) SD user. I have been using Automatic1111 now for six months or longer (installed locally). I was using the DreamShaper model and I had almost mastered prompts for the type of generation I was doing.

The prompts all followed a format something like this: "Photograph of skinny (21 year old Triqui woman from Newvo Leon, Mexico:1.5)(with face of Lun Bawang woman from Brunei:1.5)"
I have a nearly-complete list of ethnicities from every country and region in the world and this method was creating mixed-races incredibly well.

The images I was generating were also extremely high quality.

I decided to add some more extensions and found out about "wildcards". I added about five extensions. After adding these extensions, the image quality of SD suddenly went down to very poor quality. The images were extremely grainy and the faces had those ugly color-blotches (what causes those??? I get those often with various models...Is it something I'm doing? Adetailer won't clean them up). Faces were also coming out distorted.

I re-installed Stable Diffusion, but the images that Dreamshaper was generating were not nearly as good. Many looked much more artificial and perhaps doll-like (not quite the right description, but you get the idea). The image compositions were totally different. Epyptian women would be wearing headdresses, for example, and set against i.e. pyramids. This never happened with the previous installation. All were in normal settings. There was also a lot of writing AND almost every image was a split image featuring two different women.

If I just gave it a single-word prompt of "woman", it generated a "fair" (but not great) image of a woman. But as I added many more details, the images became "weird".

I tried many different models and none of them even close to decently.

I reinstalled wildcards extension only, and immediately the image generation became horrible (or rather more horrible) again.

Stubbornly, I tried it one more time and wildcards screwed up the image generation again. I'm about to re-install SD again, not using wildcards, but the problem is that I won't get the "nearly-perfect" images I was getting in my first installation.

Does anyone have any suggestions? I've scoured the internet for any clues, but found absolutely nothing.

I am not sure what "Dreamshaper" model I was using originally, but I wouldn't think that it would matter THAT much.

[mental crescent]

Did anyone encounter this weird things? The default checkpoint using is A. When I change it from A to B, it will first load the B and showing a loading UI. Then after the B is loaded, it immediately loading back to A. Therefore, I couldn't change the checkpoint at all.
I am using runpod.io to run the stable diffusion web ui with the latest version

[loud solar]

For sure the more/longer you prompt the more the A.I. looses focus on what might be most important to you.

[honest bloom]

hello

[heady warren]

GMGM

[little lava]

gmgm

[bright wagon]

Yes, but it's not a problem of losing focus. The actual generated image becomes lower quality. It's sort of like putting a ton of parentheses around various parts of the prompt. I mean...if you simply write a very long prompt, yes it won't include most of the stuff you have in the prompt, but the image quality itself should not degrade. But that's what happens if I start putting in even a moderate amount of text now. Certainly, if I use a full prompt that I previously used, the image becomes distorted and grainy and just generally much uglier.

I frequently get color blotches on faces, sometimes covering half their face.

[queen steeple]

Hello everyone, is somebody using regional prompting with Forge ? I've heard it was a bit buggy but couldn't manage to find if it was solved.

[wise kernel]

怎么开始使用？

[vale steeple]

https://www.youtube.com/watch?v=lNGEeUCL8NE
你可以看这里

哦不过那个是一键生成视频（需要额外的步骤），不过大致安装流程都是差不多的

[undone fern]

👋 SD noob here. Any recommendations for a bulk (batch) creative image upscaling, similar to Krea / Magnific but for multiple images?

[sudden ruin]

Please keep it english in here!

[willow crescent]

so have you tested a clean install + no wildcard yet?

[bright wagon]

Yes, three times

I even uninstalled python as well as python remnants

[willow crescent]

so Wildcard is deinstalled and then you did a clean SD installation ?

[bright wagon]

I did not deinstall wildcard, but figured it was all contained within the SD folder that I deleted (?)

So it would automatically uninstall?

[fervent thunder]

How to use SD on Discord?

[willow crescent]

don"t know, could be, not sure, just guessiing

[bright wagon]

I posted an image generated by my reinstall of SD in tech support...whereas before it was a high-quality photo image every generation.

[willow crescent]

https://youtube.com/watch?v=JOiK8z-Ffp4

Upscaler software: https://github.com/JoeyBallentine/chaiNNer

Model Database: https://upscale.wiki/wiki/Model_Database

though you can load whatsoever upscaler from wheresoever

[willow crescent]

this is the batchsetup: #🏞｜general-with-images message

[late surge]

hh

[undone fern]

thank you. I will look into that. But are you aware of some general upscaler (style independent) that can add details , not just sharpen and increase resolution? Maybe I should be looking for some good img2img or controlnet rather than upscaler

[lusty plaza]

怎么使用

[keen shard]

Hi all
health to you and your loved ones!

[balmy abyss]

cool group, everyone is so kind)

[bleak matrix]

Good morning, everyone! How are we all today?

[honest jackal]

https://tenor.com/view/excited-excited-man-excited-funny-rubbing-hands-plotting-gif-27652478

[manic wraith]

SUPIR

[pearl thunder]

Good morning,everybody!

[lethal jetty]

Hi everyone. Excited to share our latest work (on VLMs and compositionality, and also importantly related to Stable Diffusion and other T2I models). Have a look, would be happy to chat about anything related.
ArXiv: https://arxiv.org/pdf/2403.20312.pdf
GitHub: https://github.com/jaisidhsingh/CoN-CLIP

[tropic frost]

out of curiosity, its normal for pony stuff to take longer to generate than SD 1.5 right?

[still glacier]

Yes as pony is based on SDXL.

different architecture from sd1.5

Also 1.5 is for 512x512 stuff whereas sdxl is for 1024x1024 (with some margins and before any upscaling)

[upper stag]

Is there a chat where I can ask questions on the stability api ?

[tropic frost]

I usually go for small resolutions to test waters out and get a rought estimate of what i want. I only got for higher resolutions once i get a good baseline and use it in image to image.
Pony is taking its time even on small resolutions unfortunatly.

[still glacier]

Yes it s gonna take longer no matter what. But even worse you can t use it at 512x512. I mean.... you can but the outputs will be garbage as it s designed for 1024x1024. for sdxl stuff, the low res limit is probably around 768x768

[still glacier]

it depends, you can try asking in #🤝｜tech-support but keep in mind that it s a community driven discord. So if you re having trouble with your account it s probably not the place to ask about it. We re not staff, and if you require help from the staff it s best to go through the official routes (website, mail, etc)

[fervent thunder]

Hi, anyone know what the current OP method of making AI art is locally? I remember using Coomfyui with some great results, I currently have Fooocus but for some reason it is really really bad at the prompts im trying to do

[wet grotto]

hey guys

sd.next already have google colab or not yet? I wanna try it 🙂

looks like not yet =/

[haughty patio]

what are your favorite assistant/tool/trick to make full prompt from an idea or a thought in your mind that you are unable to fully express in words.

[wet grotto]

pick the correct model, controlnet

[untold moss]

hey - where is the old stable diffusion UI where I could enter prompts?

[fervent thunder]

so... what is a good source for real regularisation images?

nvm found one
so... should I crop/resize them?

[dry shoal]

what webui are you using?.. i just installed the latest version today and have a prompt input

[fervent thunder]

https://huggingface.co/laion/CLIP-ViT-g-14-laion2B-s34B-b88K

how to use this in kohya?

[gilded wolf]

Hey

[whole minnow]

can i pay someone to help me train my own models? having lots of errors and feeling lost

[fervent thunder]

do reg images need to be captioned?

[peak musk]

does anyone know what happened to deepfloyd ai they released one model and then completely dissapeared i know its been some time since then but i wonder what happened there?

[trail lion]

no idea, and that will plague me till the day I die...not really honestly, I dont remember being all that impressed

[jade sentinel]

Hello, everyone? how are you?