#๐๏ฝsd3
1 messages ยท Page 13 of 1
Lots of those showing up cheap after DCs brought in H100s
stable diffusion 3 out? we can download locally?
soon
spend 19k on A100
connects it to the PC
PC explodes
It Won't Be Long ... โข
16 gb of vram + 32 gb of cpu ram should be enough for swapping all the parts around but hard to say atm
good call
I see, btw will RTX 4070 able to run it? 12GBvram
โน๏ธ
im on 4070ti, so I hope so
if you can run sdxl right now, you can run sd3
this release, yes
comfy allows for 8bit unet so assuming that works with sd3 it should require a bit less vram, maybe a 3060 (12gb) could work, who knows
Faces are WAY more varied in SD3 , you will get a lot more plain/normal-looking people if you don't prompt for beauty/supermodel people.
I only have a 3070, but Iโll attempt it
Hopefully this will solve sdxl issue where animals always look like they vote for right wing fringe parties
"screenshot from the grand theft auto game set in miami, night time, cyan and pink neon lights"
that will be the first thing fixed
but i dont want plain people i want supermodel babes
2 hours?
not bad
yeah..
add 2 more hours
1 hours now
thats my dad
I mean for whatever minimal parameters they have set, it aint bad. will be better when we can play with the parameters
they're only releasing 2b today either way, that will fit into most potatoes
So when are the others coming out?
who knows
soon (valve time)
you have less than 128gb vram? dont talk to me ever again 
Another long period of dormancy
8b model means the results will be 4x better than the 2b model right?
I'll do a comparison for you via the api
big doubt, probably noticeably better but not like the difference between 2b and 8b llms
its the difference between ps1 and ps2 graphics ๐
Get 2...to be sure.
LLMs drive me insane, they're humongous. and whenever i load them to DRAM it just bluescreens my pc
rip
model: sd3-medium
prompt: "screenshot from grand theft auto set in miami, night time, cyan and pink neon lights"
seed: 12345
aspect ratio: 5:4
model: sd3
prompt: "screenshot from grand theft auto set in miami, night time, cyan and pink neon lights"
seed: 12345
aspect ratio: 5:4
im a biology student so im wondering how good it will be at making animal genitalia for educational purposes?
surrreeeee
seed doesnt really transfer due to image size / different model
maybe check out the pony diffusion models...
Thanks ! ill check them out ๐
Muppets cocks seem bigger in this release
Muppet cocks lmao
Aren't they just peoples arms?
lemme check
They're definitely not using that for biology
never attribute to horniness that which is adeque... eh forget it
yes, im a marine biologist myself 
handjob's razor
lmao
say what cuh?
lol
it is very important, i need it for an exam
i trust you
sd3 medium aint that bad sometimes. its very hit or miss on the api
examination of the file size of the spank bank folder
one hours?
"two hours"
we don't really know what exactly is on the api tbh, better judge it when you have the weights on your drive
Elon said 2 hours , 2 hours ago and still nothing
it really likes them rings
Can anyone tell me the difference between SDXL and SD3
the XL and the 3
the difference is XL is replaced by a 3
we need lada
lmao
sd 1.5 - 1 encoder, sdxl - 2 encoders, sd3 - 3 encoders, sd4....
You pay a lot extra for that many rings on your Audi badge!
more encoders means better text understanding?
Lada
maybe, maybe it's snake oil
nice box
I'll have two please 
Animal Genitalia? Oh yeah, there's a Laura for that, but I think she's in therapy! ๐ฅณ
I can't wait to load t5 and 2 clips with kohya's scripts and get an instant oom as it jumps to 50gb
The Trabant surely? ๐
yeah, its not quite getting it
it gets the general shape, but not the key details
that's understandable.
could img2img it
need a lora for fiat multipla ๐
Was the Multipla the one which looked like Kermit the Frog?!
got that homer simpson look
The Kermit/Homer Fiat Multipla
16 Channel Vae woop woop!
tuktuk rally!
das ist gut
Ha, as a matter of fact, the Multipla was the first Lora I have ever trained (not the civitai one)
A biology student receives a grant, and the other students look confused ๐
could you try without all the racing tags and see if it does better?
i think everyone forgot about sd3 and are just engrossed in genning cars at this point 
ive seen this too many times, please tell me its fake
lmao
same issue.
"a photograph of fiat multipla parked in paris"
blacked
I like the SD3 motion blur.
I hope we see a Fiat Multipla 2024 one day
doug demuro would shit himself
that is a pretty cool rally lambo
This has more of a Renault look ๐
it's getting some of the characteristics right. just not the dove front
honda odyssey rally car
well this is straight to the sd3 api
cant really change any parameters outside of what the api supports
now i know what to try whenever it releases
2 hours
it's the ultimate test. stability.ai will be in shams if it cant make the cut
May seem like a dumb question, but will SD3 be better than cascade?
we shall see
cascade? barely knew her
I don't think anyone will use sd3 because it just takes too long to render an image.
we shall see how it cascades ๐ฅ 
ok pal
sd3 medium does videogame stuff really well
Yeah, I will just ignore the release and use 1.5....
yea sd3 has some cool video game knowledge
What you gonna do if the weights drop and they don't work in Comfy?
I kinda miss Shrek in this image.
not my shrek
where did the info that sd3 is coming out in 2 hours come from?
where does it say that?
it said that 2 hours ago
Elon said it
Yeah, 2 hours
confirmed
you mean you dont trust these nice people in this server? how dare you sir...
no is that ai?
Where did you find this?
I made it with the sd3 api
"a photograph taken in stormwind city" was the prompt
No, there is no sd3 medium
there's a sd3 large
im not falling for that
jokes on you I have a constantly refreshing tab with SAI's hf on it
Please show me the generation where the anatomy will be traced
lol
you find this now
oh shit... finally released ๐ฎ ||https://www.youtube.com/watch?v=dQw4w9WgXcQ||
The official video for โNever Gonna Give You Upโ by Rick Astley.
The new album 'Are We There Yet?' is out now: Download here: https://RickAstley.lnk.to/AreWeThereYetFA/itunes
โNever Gonna Give You Upโ was a global smash on its release in July 1987, topping the charts in 25 countries including Rickโs native UK and the US Billboard Hot 100. It...

I recognize the link, I'm not clicking it or the other spoiler
haha
same
https://platform.stability.ai made an account here?
it was lykon but can't find it anymore
ima cry
didn't even bother to find out that SAI's models never have spaces in their name
Whaaat?
day 1 no sd3


Oh, no, there really is._.. and why "coming soon" then?๐
this ^
0_0
ayy 5 hours it is...
what did you expect from sai ๐
just go to sleep
no
probably gonna be an underwhelming release lol
5 hours???
2 hours
2 hour club
its been 2 hours for the past like 4 hours. chill bro, its settled
im jk im actually getting my comfy set up.. im trying to set up stabuility audio but i cant get the stability_audio_tools to install
sry release on hold due to safety concerns, sweet baby inc doesn't approve
SAI in a releasing things on time contest:
release on hold... they discovered traces of booba they forgot to remove
Nooo not the booba
uh oh
everybody laughed but they won't when they see the blurry nipples
Nah it'll totally whelm. Unless you want to make celebrities, booba, waifu, hands, feet or anything copyrighted
Official Website: https://www.yomika.com/home/
Stream: https://lnk.to/L8GZG
Facebook: https://www.facebook.com/mikasounds
Twitter: https://twitter.com/mikasounds
Instagram: https://www.instagram.com/mikainstagram/
Music video by MIKA performing Relax, Take It Easy. (C) 2006 Casablanca Music, LLC
it's still okay because you can make cute cat pics for your family
the entire image gets blurred, not just the censored bits ๐ฆ
same
everyone is waiting for sd 3 hhhhhh
Is possible make a script in python that get the sai hugging face page and notify when sd3 is released ?

you can do anything in python these days
yeah, make sure to set the interval for 1 micro second so you don't miss it
probably will get rate limited or something
and tell all your friends to do the same, I think that had a name... ddos
you think AWS has some sort of outbound ddos detection?
thats what i heard too. in fact they will send an arrest order to the horny police to come horny bonk you if you make anything booba related
I can easily write a JS code that checks every 5 secs if there is a string "Stable_Diffusion_3.0" on their HF
im gonna use my python script to log on SAI servers and release sd3 early 
how you know what its gonna be called?
You can do this with regex ?
I think there was name for that...
Eh, the anatomy is still suffering
We'll have to wait for the big sd3 model
yep
what's the different between the medium/large/etc. models.
lmao
size matters
Just like kebab
I don't, but thankfully I can check for as many strings as I want simultaneously
I hope that this is only a beta type test and when it is released in open source it will get better
real talk, just check for the number of models on the sai page
if it goes up they released something
is that on huggingface?
yeh
oh no, not my booba 
look how they massacred my booba 
already there
hf: pls get out, no ddos any more.
the resource cost of T5 vs the benefit of including it vs CLIP only isn't a great balance, I expect in the long run most people will prefer to run CLIP only vs using T5
aight imma go to bed. see yall in 2 hoursโข
mhm in 2 hours for sure
@raven fern 

were you?
hi cat
calmed down? i mean more and more people are asking when
wait how many inches is that?
aww
tiktok
so using t5 is better..?
lol
is tiktok finally gone?
if you read the research paper, it explains nicely about usage of t5 and clip
could you link ir?
ie the paper says t5 = better text
im stupid
What's the previous text encoder before t5 ?
thanks 
there was some cute graph in there somewhere you could look at and be like 'ah okay'
but not sure which page
T5 is just much more prompt coherent than CLIP
you know, when you have to clarify multiple object in different position
it's just a smart TE pretty much
this one
Can we dowmload the SD3 weights now? 2 hours is up
I look forward to seeing how that graph will change in the current model
2 more hours
man
2 hours mean until it released
can you guy stop gaslighting people
that is OLD sd3 model isn it?
with 2 hours
was it trained further? no idea honestly
@raven fern people love the number 2 
that's what they've been doing since the paper
idk too lol
more likely yes
to be honest
API SD3 is severely undertrained
well that would be nice
what are you gonna generate first? :3
someone that has access to both said that current 2B > old 8B
even more so if it works better without t5 now, the encoder part of t5 is like 10gb in fp16 iirc
that's a pain to load and offload even if quanted
yeah because 8B is also undertrained as fuck
aka need more time to train
Im pretty sure 90% of people won't be using T5 in a month
Its 12th of June
I mean it is an additional feature
it is always nice to have additional feature
Hello, what are the gpu vram/ram requirements for 2B SD3 and what will they be for 8 B?
2h later, pls wait or sleep for 2h
4GB for 2B SD3
you will see sd 3 when you wake up
Most of the additional features are not well-received
assuming you wake up that is 
Are the requirements will be the same as for SDXL?
I think the method would be to either keep T5 in RAM and transformer diffusion model in VRAM, or have them switch when being used
no problem , see nothing or see SD 3
Holy crap thatโs way more efficient than I thought it was going to be. That should be significantly faster to gen than SDXL right?
sir I have a RX 580, can it even run it on day 1

yeah but it's gonna be painfully slow on cpu ram, like +5-10 seconds easily
you can buy a 4060ti 16g
What am I going to generate first is probably photorealistic
did sd3 launch?
maybe try on day 200 ๐
although 99% of the time I just generate anime anyway
4gb for DIT, 1.5gb for clip so all model is 5.5gb?
SD4 im ready 
the sd3 unet is slightly smaller than sdxl iirc so should be even lower
in about 2 hours
i dont think it is supposed to be like that...?
sd3 is not unet
5-10 seconds for cond, then another 5-10 seconds for diffusion. batch sizes are going to be useful for that method
where is SD3
I think there are ppl training clip
wa4-6 hours
everybody asking where sd3, no one asking how is sd3
T5 is finetunable
Google release their own variant of T5 too
....
Awesomee
I dont think ppl will using T5
well no, Google develop T5.
why asking how is SD3.. they have already released the paper and everyone knows the implementation
there you go, a new question :3
we gonna know in about 2 hours
They skipping SD3 and going straight into SD4
in 3 hours everybody will be back to pony
I mean who cares if you need 10s to use T5 in exchange of better prompt coherency
obviously what do you expected
people keep comparing base versus finetunes lol
packaging module error?
ModuleNotFoundError: No module named 'packaging'
they were so smart in school, instead of going to 5th grade, they went to 3rd grade

Did they get a SD3 copy already to finetune/train?
... although yes CLIP probably going to remain majority
'they' who?
yeah altho you can fix that by pip install wheel i think
Cooking
calm down guys... you will be able to generate booba in about 2 hours 
The ppl who trained pony or are you saying they will go back to SDXL pony?
I think the weights will fall out in 10-13 hours
or typograpy? if that case you are right
i tried a bunch of things i cant get it to roll
when release tht SD3?
it came out yesterday
yeah that too
After 10-13 hours
Actually, did they say 2024?
I meant that like the community will look at sd3, see that it's not possible to generate nipples and go back to pony, but I also remember that the pony dev didn't get sd3 access x)

i posted pic about that LOL
so fanny
Well guy
We expected 3-5pm ( GMT / UTC+0 )
but in fact, I don't know what's the SD Medium performance
anyone who wants the rough countdown
Ah so it will take a minute for SD3 to get Ponied. SDXL showed us itโs possible to free the nipple.
since they have to working on stuff.
fuck 2 hours I'm gonna play guilty gear
free the nipple ๐คฃ
I'd wager a bit more than a minute but yeah, probably, assuming the pony dev is even interested in training it
??? i'm tired of nipple...with sd i have see all kinds of nipple
oh boy...
They are neets, i got my money on em. Trust.
6pm EST is literally 10pm GMT / UTC+0
is pony just a retrained stable diff?
if you have a lot of RAM and don't mind very long load times it's a bit better. Depends on prompt whether it actually helps much
Alex just needs to put us out of our misery and just release it :3
10pm gmt? omg
Oh you don't want or start that conversation all over again
it's based on sdxl. some loras work on both and you can to some extent mix sdxl with pony models
it is a topic loophole anyway
Yesterday was quite a read here
is it harder to finetuned sd 3 compared with xl?
and it can extend to Pony cultist and whole anti-Pony sentiment
deja vu
Vietnam flashbacks
ptsd
we will see
What!? Really? I thought T5 was what gave SD3 it's superior prompt comprehension?
pony sucks at realism
you can use CLIP or T5 or both on SD3
It's more likely the better captions with CogVLM
which... I didn't expected.
Well, for people that are waiting for it to drop, do yall have the necessary ui to run SD3?
Yo devs can sd3 run on a 12gb vram 3060 and 32 gigs of ddr4?
overkilled
Wonderful
my leader so love pony
for 2B i mean
And what would that be?
I don't think there's really any choice other than comfyui/whatever derivatives of it currently
If Iโm not wrong, comfyui and stableswarmUI currently
@lunar rivet yep went and investigated. They will try SD3 Pony unless SD3 is like a huge flop like SD2 was
that and the generally better MMDiT arch yeah
whats stableswarm offer?
that's good to know, here's hoping it's not a flop
Pony Creator had strong opinions on SD3's license though
Thanks,Iโll stick with stableswarm cause I heard u can use it remotely like automatic1111
so lets' see how the chess goes between all finetuners
Swarm supports SD3 out of the box already
No idea, but thereโs more of an UI than comfy node based
several people are typing
It's happening again
more like A1111-Comfy combination
Isn't swarm a comfy wrapper?
StableSwarm used Comfy as a backend
Yup yup, so currently only two UI, comfy and swarm
I'm excited to see how controlnet works on MMDIT.
that means 10hs later?
Hoping sd3 can give us dall e 3 level stuff if not better
a lot of these other tools will have to change, cause its not unet based anymore
Doubt that stops anything. Ppl always talk about politics and somehow thereโs always a way. ๐
ignore the random timestamps people are making up
I think so, but basically you donโt have to play with node if you are not into comfyui style
i dont hate drama
so is SD3 using t5?
might? or just time for they start servicine
T5 and CLIP, both
T5 is an optional textenc available for SD3 yes
nice
i mean you can use only either one of them
hmmm i should go to sleep right now then...
long story short: t5 is optional, its mainly useful if you really need complex stuff or letters, and it can run on cpu
How much better is sd3 gonna be than sdxl
Is it the recommended one, or are there others?
btw how do CLIP-T5 worked?
lets all take a collective nap
where is this screenshot from?
and dream little ai dreams
I recommend CLIP only for the average person, or both if you have a lot of RAM and want to eek out that little extra bit of smarts
Flan t5 xxl with sd3 is possible ?
that's Swarm
Whatโs a lot of ram? 64gb?
24GB VRAM, 32GB RAM?

T5-XXL is a 10 GiB file on its own, plus ~6GiB for the rest of the model and tencs
128GB?!
so ~16GiB free space on top of everything else
oh snap, i'll check it out, so that already has SD3 support? I saw comfy has it too. Any other UI's have already rolled out Sd3 support?
The pony creator is making his own model for non-commercial use only but complains about SD3 requiring a license for commercial use, which will likely cost him $20 per month given the size of his business.
if you use fp8 you can reduce that to ~10GiB
2080ti 22GB is best
well, I'll have to install stableswarm once I get home.
I wonder if t5 will have a 4bit version or if quanting it just destroys the prompt like I remember with the deepfloyd IF encoder 
Whatโs sd3s weight size?

yes Swarm already supports SD3
i cant delete my comfy workflow
Since both comfyui and swarmui are from stable staff, I bet thatโs the only two to support right at launch
oh! i never realised comfy was from SD staff!
stability is pretty dope
quantized T5 works well but the code infra isn't set up to use that rn, somebody will have to do that later on
The one on the screen is in the browser's cache. Just load another one.
if there is a way to use the flan t5, its way less space
anything based on Diffusers will also support SD3 very quickly after launch
Comfy is stable staff 
oh idk that
I mean he doesn't even have to be charged 20$, he didn't have the annual revenue of 1 million USD
glad to hear it might work at least
yeah, ComfyUI already has TRT support for SD3
If you mean Dalle.3 prompt following levels then you are going to be disappointed.
I bet they can look at swarm or comfy code now and actually implement it already
honestly that's a small price to pay, the Pony creator could probably cover that with donations.
It's important that stability find a business model that works for them and keeps them financially afloat while giving the open source community more models
๐ how disappointed
Hy guys๐!!!. When i was entering this discord, it is Free to use! Then suddenly it shuts down ๐! What's happening here!? Did they make now to paid subscription to use!?
you guy are going to surprised how many people will willingly donate money to Pony creator
2000 per day per hour per minute
people talk about my little pony, no one talks about my big pony
surprised by how few or how many?
apart from NSFW peeps...
it's still better than anything else you can run locally, so it's an improvement!
I've not looked at "swarm" at all. How does it compare with Comfy?
I mean you can just see how much LORA and finetunes was trained specifically for Pony in Civitai
SFW or NSFW
No, he would be fine with $20, the issue would be if the $20 subscription doesn't cover it. And it was unclear if it does - random employee comments don't cut it
Hy guys๐!!!. When i was entering this discord, it is Free to use! Then suddenly it shuts down ๐! What's happening here!? Did they make now to paid subscription to use!?so now it's not free๐!??????
Itโs built on top of comfy, but you can use it without the node stuff
Thatโs good to hear,I just hope text creation is gonna be better
Can you create the node workflow and use that in the UI?
I'm pretty sure it IS comfy, just a simple interface using a workflow as a backbone
Hy guys๐!!!. When i was entering this discord, it is Free to use! Then suddenly it shuts down ๐! What's happening here!? Did they make now to paid subscription to use!?so now it's not free๐!?????? Reply anyone!!!!!!!!!!!!
bro what?
other UIs can do that yes, eg auto webui crew is already aware of the code
the old SDXL service was disbanded.
Hy guys๐!!!. When i was entering this discord, it is Free to use! Then suddenly it shuts down ๐! What's happening here!? Did they make now to paid subscription to use!?so now it's not free๐!??????
dont spam
stop
Only in the artisan bot channels.
Calm down
I bet you can use node based workflow as well. (Ps: correct me if Iโm wrong)
Ok
@cobalt moon ok
lol
Then itโs just a matter of waiting. But Iโm just gonna test with swarm instead
ok gn peoples see u in 2 hours

gn
@cobalt moon so, how can now i access it for free!? It will comback๐ฅน!?
Under 1million USD its the 20$ for comercial use, I'm totaly fine with that.
@craggy ridge this place gonna be hell very soon with everyone asking tons of questions ๐คฃ
dont know
Ohh it will be hot yes ๐
I mean until now they ( SAI ) still haven't make any clarification on non-commercial license
I suspect it will came along with the model weight release
Where are you from guys!?
But no one questions how are the questions?

the bastard child
hightlighting I think
Oh John Snow yes
Recommend me a free discord bot that can access for free and advanced like midjourney!!!
you forgot one more !
I can recommend you Limewire
Plz provide link
@cobalt moon where are you from!?
Malaysia
Link plz..
is Limewire still a thing these days? lol
not that Limewire lol
One Kick Man ๐ฎ
Hi neighbour
no way
ok hi
uhm... yeah
??????
2 HOURS ALONG??
apparently Limewire have a Discord server
ah I see how it is, they were bought out by some startup
and image generation bot like Midjourney ( except it directed back to the website )
Shouldโve renamed lAim Wire.
the hands are totoally done by now
Who'da thought limewire would turn into an ai generation service
they switch from that to AI
Still as scummy as ever
holy, they even have a shitcoin on their page
It has limitations๐!
limewire used to be cool
i suppose if everyone keeps saying "in 2 hours" eventually somebody will coincidentally say it exactly 2 hours before launch
exactly
forget limewire
There's an hour left!
huh, who are you, why should I tell you that
in 2 hours ๐ฃ๏ธ ๐ฅ ๐ฅ
why does this dude keep asking people where they from ๐คฃ
i wonder if napster does something similar now
If you want a discord bot that can generate images, you might wanna try circles.lab
lmao
I hosted my A1111 as a discord bot
wasnt circles.lab for chatbot
I mean, itโs not as simple as an bot that generate images
by this you're implying it's more than 2 hours from now
they almost went shutdown a few month ago
Thatโs what I was doing previously when I was not using it
NAPSTER IS LEGIT NOW
the swarm discord btw has a discord bot that gens images running off the swarm instance in my living room lol
Yup, itโs back alive again, and there image generation using SDXL or dalle
security camera footage of a masked man stealing pineapple pizza
the crime ๐ฎ
Ban this guy

fam leave me alone
Which platform you used for this!?
to many nacho ๐ฅบ
Link plz..
why is napster trying to force me to download shit
uhm...
Kk
Napster is not under its original creator anymore
ik
it is under multiple shares of at least 6 companies
napster, winmx, limewire, kazaa, man the good ol days 
The odd thing is that it's redirecting me to a download link from buzzhheavier
Does that mean nobody has yet? ๐ค
hhh.9 hours
why would people buy any of these these and run them as new given their reputations?
Just search for circles.lab on google
้ ๏ผ็ไธๆๅ
It just doesn't seem very profitable
wo kao,wo kan bu dong nimen zai shuo shen me
Already did that. But, can't find it. Help me out fam!!!!
speaking the language of the gods right here ๐ฎ
correct
Text creation is amazing in SD3 (better than Dall.e 3 I think) , I just don't use it much.
nice
If I said it in 22mins, would I be wrong?
text is better than any text project i think.
Oh wait, I think they changed to shapes
not just tell us the time?!?! ๐คฃ
Go search for shapes discord server and join the server
yes, he implied it's in more than 2 hours
2 hours, 2 minutes, 2 seconds, 2 milliseconds
in about 5 hours ๐
ไปไน็ไธๆ๏ผ
EN pls...I think, i'm afraid someone speak with russian..
Plz provide link! There are so many servers like that! Confusedddd
SD3 ComfyUi Workflow.
Is the VAE Baked in or seperate ?
the foreign agents are infiltrating
I have
Sent

where download sd3 mod๏ผ๏ผ๏ผ
hf i think
separate
mod? we have mods now? ๐ฎ
SD3 has ComfyUI in the dataset? or did you add this somehow?
Nope, it's more than 2 hrs away still
Will SD3 2B perform better than SD3 8B with the current api?
Sent me the link broh๐!
from google drive
2 hrs
why do we need noise canceling
bro have you even seen the SD steam workshop?
hahaha
why were you have๏ผ
Can anyone ecplain what is sd3zm๐๐น
here is proof if you don't believe me
I dm-ed you brother
0KB ๐คฃ
nice temsor
Someone generated the pic for me , he used this service : https://www.seaart.ai
;p
it must be really efficient at 0kb
what ๏ผplease tell me how to download
Try running that! ๐
My dream
only two more hours guise
its real guys
I believe in you , and you guy even do it without PS
two more hours to go
(:
Time to bed, no more SD3 will come out today
dont cut yourself :3
I asked internally before and they said I'm not allowed to tell anyone a precise time, so... 
I think your download failed, should probably enable lfs
oh are they going to do the license announcement too
we all know its coming in like 5 hours :3
@lofty canopy my bad i generated this with stable cascade
Then hint it 
if you ever asked them
i envy the fools born today who will grow up and will prompt up perfectly awesome blockbuster movies in 10 years
But they didn't say you couldn't tell anyone a rough time?
in fact u wotn even have to promt
@cerulean cove gotcha!but no use ๐!!!there are no option to generate images.!!
we need is an SD3 model that can run independently from the cloud server and on the local computer.
you just think of somethign and ai will deliver it
I believe it's going to be indicated in the release announcement today
I know it, the exact time is in less than 24 hours!
Today lol
Brother, if you want a quick and easy to do it, then shapes isnโt for you
What is sd3 can anyone explain!?? ๐น
loooooool
Be less rough! ๐คฃ
oh prob yea, but you downloaded a bit too much
Thats what we are getting today...
like stable diffusion 2 but 3
correct
you need slow mode in here emad

Is the local model open for download?
it came as a bundle with free ram
Smol Sus? where is Smol Boi?
What's it broh!? ๐๐น
greater than 1 hour and less than 20 hours
yes, in the dream
No, Sometime between now and midnight, they will not tell us.
it's just the most recent stable diffusion
they all download in their dream
ugly
Ok war
i havent watched freiren yet, is it good?
You sure?
Can SD3 change the world
๐ซ 
it's the best anime of all time
yeah it's aoty easily
Boom! done, the planet is now lego
Solution : wait until weekend
almost 6am here, i didnt even sleep yet
SD3 will bring the messiah.
Yeah it is pretty wonky
you're wonkier
ill go touch grass till it releases
What the fuck
oh you sweet summer child
probably stable swarm or something
Please do not trust any download link until it is officially released.๐
exactly
I only use auto1111,so this ui and workflow lines scare me
we have been here for a while, so we know things
He's new here I see
Come on guys let's talk!! Come to voice channel!
how to download
Lykon prob half-joking
You forgot to add the prompt, otherwise it's good ๐
Need sd3 more vram then sdxl? I have a 4070 TI.
you are set
no that's comfy
The minimum is 4GB VRAM
swarm is the really easy friendly UI
Tired of typing?Everyone come on guys let's talk in voice channel!!
Way more than enough
Is stable swarm something different
comfy is the deep noodle ui
where are you from
yes
I thought you could use it on comfy
Can you add your own workflows to Swarm?
I haven't used Stable Swarm so i have not a damn clue
Swarm includes comfy on a tab, but also has a friendlier UI - https://github.com/Stability-AI/StableSwarmUI
I use comfy, I know
you can use it in comfy.... in 2 hours 
ok
Yes, you can (A) use the workflow tab or (B) import to generate tab to use power features like grid or (C) create a Simple tab from your workflow. See also https://github.com/Stability-AI/StableSwarmUI/blob/master/docs/Why Use Swarm.md#i-am-a-comfyui-user
Foocus is very nice, best Speed and quality for sdxl.๐
@viral plaza how do you reply to the bad comments about swarm in here? lol https://new.reddit.com/r/StableDiffusion/comments/1ddqn5c/stableswarmui_now_supports_sd3_you_can_ditch/
So fucking long 2 hours
Any cute girls in the voice channel?
there's no public release time, so people are just repeating 2 hours over and over until it eventually comes true through luck-by-volume
this is the stable diffusion discord server, absolutely not.
Stop spamming!!
nah, I'm good.
@raven fern things are going downhill
Sreehari over is buggin
time to bring the ban hammer ๐
Oh soory!!!
the mere thought of sd3 releasing is driving them to madness
it's fine
is 12 GB of vram good for sd3?
Overkilled ( 2B )
aight excellent
the best 2b
I mean do you remember how long is taken for A1111 to merge dev to main for SDXL conpatibility
I think 2 months
Does it support Cascade yet? ๐
Will comfyUI be faster?
the dev branch got a massive boost recently, 15 PRs all at once
im sorry i only talk with people that have a 5090
:3
And this action make Forge to transform into extremely experimental repo
forge or foocus ๐
sounds reasonable
Forge, do you see the announcement
rip forge
I mean
It is a sooner or later thing
For memory optimization
Since Illyasviel ( the creator of both Forge and Fooocus ), dont wanted the Forge to be a "technical" clone of A1111
Illyasviel, one of the gods ๐ฎ
understandable why people are upset with the title of that post there, that's a bit fight-provoky. I replied to some of the comments in the thread
i didn't know that thread even existed til you linked it here
didnt see yr comments
or should j refresh again
now i see it
I'm off duty. Is this two hours away?
so wait is sd3 a new VERSION of sd or just a new base model?
Same same but different
whole new thing
ok, i kept seeing the word "model" on stability.ai and was starting to wonder
can you have this and a previous version installed?
Silly, people dont even have any idea what is transformers
Bro ๐ญ
yep. some people are even going to instantly try what happens if you use them both at the same time, one input into the other ๐
how tf would you even do that?
Fed
the model and architecture is all new, you can use one UI for many different models of different architectures
you can currently use SDv1 and SDXL and Cascade and etc all in one UI happily
oh sick
or unhappily
man i thought i knew a lot about this shit
if that is your want
but i gues i have a lot to learn still
@junior dune how dare you accuse me of being a fed
he is not a fed, he is a crook :3
I'm evil ๐ ๐ ๐ ๐ ๐
just before i go to sleep, nice
sd3 available now?
in 4 hours
well, cooool
3.75 hours now
Good Night everyone
I mistakes u for the vc guy lol,Iโm on a dark screen and I thought u also pinged me
ah, that's ok
Here's q question. If sdxl is a 6.6b parameter model, why should we expect SD3 medium to have better "knowledge" and prompt adherence? Are we expecting any performance falloff?
Size doesn't always matter
that's what she said
Most don't, it's just that 2b will be able to run at home and for now 8b is in api only and undertrained (that it is undertrained might make it worse for some definition of worse, we'll see in a few hrs)
how undertrained is it?
so in essence, the 2B model will fall between 1.5 and SDXL, but be able to take what it has and perform at a level higher than SDXL
or its overtrained... we don't know
Iโm most hyped for the text generation in sd3 tbh
Nah should be better than sdxl
ya think?
will it be better than sigma+SDXL tho
Like literally I wonโt be disappointed w anything else unless prompt following and text in images is gonna be ass
will it instantaneously be better than everything else ever tho
I doubt it. But it will be a good base to supercede anything before it
will it be better than SORA?
I use MS Word for text generation, it's faster then SD3 or when I am lazy LLAMA3
Does sai work on other models or is 100% focused on sd3
sigma this 
even if it snot better than sigma+sdxl at leats when controlnets and ipadapters and stuff come out for SD3 theose will surely put it ahead of sigma
Release when?
2 hours
Swarm UI is ready to run SD3?
I hope Stability will release controlnets themselves rather than leave it up to chance
๐
question, is the medium model available on artisan undertrained?
today's launch won't have controlnets. Might in the future, but either way the way to make controlnets is very straightforward on the mmdit arch
Medium isn't on artisan yet
wait
only the old 8B and the Ultra workflow
im kinda confused
When will it release in CEST
Yes
i think people generatef some of the images on 2B here
Will todays launch be sick
Stability works on a huge amount of stuff, we published Stable Audio Open last week, and some LLM and 3D stuff recently too - check https://stability.ai/news
I hope that Xinsir guy will make controlnets for SD3 then
SDXL has a ~2.6B parameter unet, SD3-Medium is only a bit smaller than it. But SD3 has a much better architecture and much better training (the cogvlm generated captions do a ton to increase model's understanding)
They are uploading it now apparently, so whenever the upload is finished, depends on thier internet speed..
I'm guessing 12:00 BST in 31 mins.
We need to calm down. Too much hype cna onyl lead to dissappeainment.
Pretty much just guesses
I'm downloading it rn
Ill upload this when the hand sar edone hea means ๐
me waiting
Google needs to get their act together
No cus it is not there yet
I never thought we'd ever see a well captioned dataset, especially when cascade came out, but then SD3 came by surprise.
Ha that's cool to see stable audio is open now. Can i test with diffusers lib ?
that's the joke
Is stable audio local now!?
I hope it will come soon so I can play in about 3 hours
In this case I don't blame google entirely, but the act of getting info from google and grabbing the top info is messy - that's a "correct number" but one that needs context: the 6.6B count is counting the text encoders, and counts both the SDXL Base and the SDXL Refiner together.
2.6B is the size of only the unet, which is the best equivalent to the 2B size mmdit for SD3-Medium.
If you count all textencoders and etc you could say SD3-Medium is like 20B or something but that'd be very silly
yep
20B SD3 confirmed.
not sure what libs support Audio or not. I know comfy is working on support, and there's a hacky comfy custom node for it already too
Can I run it on my 4090
20B ๐ค
I love spreading misinfo, its so much fun
๐ฉ๐ฉ
no 2B will require 308923712398654 Terrabytes of VRAM

oh ๐๐
i mean technically yes it is there at that hf link lol you just can't see it
nah its around 4 GB iirc without T5
SD3-Medium? yes easily
Okay cool ๐
oh nice a longer version
you mean this one ||did-I-get-you-with-this :D||?
wtf
I alt tab from my game to see this
