#๐๏ฝsd3
1 messages ยท Page 12 of 1
You underestimate mergers' ability to pump out 500 useless models
buying in means theres no poor people. its classier.
while we get people hopeful that our release is ruined by a leak ๐
I almost exploded with impatience
It's been a while since I've added to this folder, so thank you for your contribution.
lol
i dont hope it's ruined with a leak. i just hope someone at stability is like "oops" and it leaks
a bit out of context maybe, but I'm glad you're having fun ๐คฃ
all these leaks... terrible plumbing work...
The more out of context, the better the inclusion into the folder.
tell me about it, I have been counting days since february
two weeks ๐
mhm
I died a bit every single time tweeted me "you'll just never release it"
because I knew all along but I couldn't say it....
bet
I actually did bet with the Mobius guy
like if the new iphone models showed up in store shelves today and apple was like "oh no it leaked!" kind of leak
Anyone got any estimates for the launch?
oh the guy who is retraining SDXL and SDXL ELLA and an MMDIT from scratch
lmao
a leak would damage us because we have partnerships going on.
Hey it will be almost two weeks from your predicted 30 May ๐
"from scratch"
How many hours, seconds and fractions of a second remain until launch? 
okay. a leak that doesn't hurt partners because it was like a suits episode and y'all did it so crafty like
What even was the deal with him? Kept talking about "de-biased" AI with no proof or demonstration as to what that even means
and like 15B or some other large parameter count
i dont even remember what i predicted before that hahaha, was like may 10 or something
good luck with that on your own
need harvey specter AND don draper on the job for this one boys
he said multiple times SD3 would never be released, I got tired and told him "wanna bet a coffee?" and he accepted
I actually forgot because it was a long time ago
Oh right it could've been 10 lol
its just simply de-biased, don't you get it? 
il get you a doggo plushie
a friend!
โค๏ธ

used to have this dog plushie. it'd say the lines when you squeezed the ear


please stop sending naked pics
kek
he's got afur coat!! but those are also problematic
found a fix for that,ban all fur
diabeetus
sd3's cool and all but the biggest release of the year is calculator on ipad
how I feel
eh, I would like to be one raiding huggingface servers with you all but probably gonna go to sleep
maybe I'll still have time)
tie a string to your finger so you remember there'll be weights in the morning
you got plenty of time if you go to sleep now
"wtf is this string for? OHHH SHIT"
it's the 12th, where are the weights?!
I know this is very early but do you think there will be an SD4? lol
oh yea, but I doubt it can be forgotten ๐
it's not the 12th in ||[insert country that's behind in timezone]||
๐ค
timezones are a bitch
nice, I will use that shortcut
It's day 12 here
amen. got 9 hours of jet lag right now
top secret
good night guys! (or morning)
idea for a meme post. youve heard of elf on a shelf. now thers sd3 in a tree! yeah? yeah i could work and noot be lame, right?
lykon bottom left irl confirmed 
yeah, but...
so clean
not at all, I'm a dog plushie.
awww
How did you get the top right? (separate ones)
is the vae the thing that lets sd3 generate text?
the model did it, in my prompt the text was "SD3 Medium"
it just made variations of it
Top right, with the 3 bottles?
top right is cool
sorry my bad, 2am..
Np, they are all awesome ๐
the vae wouldn't make the model capable of drawing text
similar to this pormpt by the way https://x.com/Lykon4072/status/1799556850284732866
where is the sd3 potion :3
ah yes that one
sooo... i saw the clip_l and clip_g are still around... do we have a clear answer as to what role each plays?
Ty very much ๐
should we just be dumping in the same input to both like pretty much everyone does, or not?
๐คก ๐ฆ ๐ฆ ๐ชฝ
w/stable cascade
smol boi
Takes shorter prompts for me now
I can finally 3D print ai creations? ๐ https://x.com/sedetweiler/status/1793692376524869635
Dusting off my 3d printers....
ah yea Scott is awesome, his comfy videos helped me a lot on some stuff
stl files โค๏ธ
6 min (prob)
https://huggingface.co/spaces/Wuvin/Unique3D check this out
I love how Zavy dropped one more release yesterday before SD3. Hopefully we're all satisfied tomorrow. I'm prepared to see this for myself. My 4090 is ready.
man i cant wait to try all sorts of combinations tomorrow ๐
this is zavy with me playing with the color channels for ancestral sampler noise
Nice dude!
https://github.com/ClownsharkBatwing/RES4LYF from the node pack i finally uploaded yesterday
prolly should add a descritpion... lol
some crazy shit in there
wait since when you have a github ๐ฎ
haha
nice
il check it out
and tons for scheduling settings for stuff with samplers
i know you spent a lot of time with schedulers and samplers
RES in particular, which really is the best sampler imo
yeah, this is the culmination of that
pog
got a significant jump in quality vs anything i've seen done with sdxl
you now have a PhD in schedulers and samplers ๐
lol
yeah i was even able to get sdxl to do text with it
no seed spamming, that was the first or second try, i forget
nice
in 15 hours ๐
What is this ?
it's his PhD thesis
congratulations
lol
๐
you're not the clyb of clyb sde are you?
LOL
oh shit you're on here, sweet!
i had no idea haha
still need to do the readme... got too exhausted but yeah, built a ton of crazy shit off the sampler code you put up there
he graduated thanks to you clyb
totally reworked the noise stuff... idk if there's shit i overlooked but i think it's easier to use
RES is my god
res is so nice
i used to main it before i made a 2m modified one (i still need to update the nodes xd) called SENS
the point where i decided i needed to go wild with it is... well
oh shit i wanna see this ๐
@dusky thistle if i remember, you started with res around the time of cascade or was it before that?
i noticed that high ita was bad at the end, high ita was good in the beginning... same kinda shit with momentum etc
yep, that's when i knew it was special
ita is good when its half the other parameter imo
so i made it so you can actually modulate it
0.5 c2 or w/e it is, with 0.25 as ita
half is has high as u can go before you start to see visible burning
you guys know what time sd3 comes out?
dont tell anyone... but around noon EST sshhh
also have a noise sampler (the fractal one) that uses fractal/power/1f noise, whatever you want to call it... and can also have its parameters scheduled, it's a really good one
I've tried doin power noise but without much success 
Interesting to see
you don't even want to know what all that code looked like af ew days ago... my god what a disaster
yeah it's working really well
had to scale it
been doing some weird experiments with fourier transforms to separate the phase/magnitude of latents and recombine them... idk how useful that will be but it's cool lol
was using it to mix in static images into a noise stream
been patching bugs all day... not surprising a lot of shit would come up after all the reworking the last few days
I somewhat got that down by forcing it into the CFG part
https://github.com/Clybius/ComfyUI-Extra-Samplers/blob/main/nodes.py#L563
man clownshark cooked hard
In short, what is its benefit?
oh shit that's right i forgot that stuff was there
been 100% zeroed in on the noise stuff for weeks
large boost in quality, control over the image that you can't get with prompting, etc
managed to reduce artifacts etc quite a bit
oh yeah, Clybius, have you tried setting the noise precision to float64 with RES???
the effect is insane
that's gud, I've been lookin at sde sampler stuff, reversible sde(?) sorta stuff
Yee apparently you can do the sampling in fp64 when the model is quantized to fp8
and get just as good results or somethin
just release SD3 now, what's one day? ๐
iirc RES originally did it in fp64 (or 32?) but I removed that to prevent issues with some other nodes I think I found
mighta just been how i did it, idk
you have to wait for SAI to wake up, brush their teeth, eat their breakfast, read the latest news and then... stretch a bit and release it ๐
this is the difference i saw
i think i changed the model variable to fp64 which caused dtype mismatches 
Might be seed randomness, lemme try
the quality increase held across a lot of images
generated 15k of em in the last three weeks >_>
not all of them look so extreme
but just changing the noise type makes a big difference
yeah yeah I know, just a little sad cause I'll be busy tomorrow :/
i'll drop one in a sec

i've got everything recast to fp64 within the nodes themselves, i think i got everything that mattered at least
the performance hit is pretty much zero
workflow embedded
https://github.com/ClownsharkBatwing/RES4LYF there's the node pack
so we just do a git clone of that in the custom_nodes folder right?
this node changes the default torch precision to 64/32/16 (32 is default)
there's a mysterious... something... that's affected by it that gives a further quality boost
haven't tracked it down definitively yet
Awesome. Thank you
but there's already a huge change by using fp64 latent images/noise instead of fp32
np
@raven fern how many hours for sd now ? 
there's a few noise types i haven't added back in after the rewrite
pyramid shuffle and lanczos
how many hours?
hours
Faster!!
studying biochemistry until sd3 drops
the mad lad
the latent noise hookup always you to override the noise sampler and inject noise of your own
you can even load up an entire folder of images you like and stream them in as your noise sampler
man look at all those numbers ๐ฎ
noodles! those are the real noodles.
In 12hr and 30 minutes
this is totally fun and i absolutely don't mind being forced to learn it all over again for like the 5th time in my degree
where is that song, "the final countdown"
The workflow you gave does not work with that repo so im trying to figure it out node by node
what's the error
I tried but the clock doesn't want to get fast, sounds like your studying has to be slow
that's odd cuz i just updated everything on github just in case right around when i dropped that image
wtf is that
alien graphs to me, im running
hide under the bed
red alert!
blue: breakdown of glucose
orange: breakdown of lipids
green: breakdown of amino acids
check the console for comfyui... it should, above where all the node load stuff is, show the error
they need to brush their teeth and stretch like lodis said
maybe do the dishes, take the dog for a walk, make coffee then randomly nap before tea with friends
maybe release article writeup and making the example workflow, or no reason at all, idk
either way i am slaying this biochemistry midterm
gootbye for now see you on the other side
My current comfyui had a ton going on. Gonna try it on a fresh one
k
A++!
make sure it has the requirements.txt stuff
everyone that's tested it i think already had those packages, so it could be one is missing
here i'm messing with the color of the noise via the LatentNoiseBatch_gaussian_channels node
screwing with the pattern/sltructure channel is also really interesting
i'll def be going wild with this shit when sd3 drops to see what i find
so there's a bunch of text multi-line boxes. which box should I be editing for the prompt in general?
Does sd3 supports natively image in input or you need clipvision model ?
cute gif :3
whatever you decide to hook up to the reroute node
the other ones you can delete if you want... they're just random prompt generators for testing with the dora
installs python
installs project1 and dependencies
installs project2 and dependencies
project1 now doesnt work because of dependency versions from project2
finally fixes the dependencies, but actually this one library only works between v1.2.0 and 1.2.1
fixes that horseshit and then actually, it doesnt work on your version of python, im sorry k?
gets new python installed... and then im sorry again pal.. doesnt work on your operating system
installs on brand new system, well... try again mate... you cant run it with that GPU...
๐
gotcha, cool, glad it works
that is a thing of beauty
it's my art machine
Hate to bother but mind dropping those graphs to use with it, lol
python https://huggingface.co/ClownsharkBatwing/CSBW_Style/blob/main/csbw_225tkn_r96a48_12epochs_base1-0-fp16_save-fp32_mixed-no_prodigy_decoef08.safetensors latest version
ngl i prefer biochemistry to whatever this is
i sure don't lol
those are the wrong kind of noodles
they should show up when you run the WF
Oh, I thought you made them. It just displays it ok. Thank you.
sd3 marketing is crazy, they slowing discord servers down to increase the hype for sd3 
Aaaa
here's hoping they release the lora training code alongside the weights
suuposedly may be diffusers code from HF
yeah discord's not feeling too good
I tried to upload an image and it timed out
stability ai scheme to get an excuse to delay the release of sd3
yeah it's really bad
as of now, this is the only LoRA under "SD3" in CivitAi
how did he even train it
That's the only one you'll ever need
Ezri the lamancha goat is a real person confirmed
No shirtless pics of Ezri please guys
probably used the api to make images of a goat then trained a lora using something other than SD3
bet it's miscategorized
@dusky thistle Yea good stuff. That cleans the noise up nicely
tons of crazy shit you can do with it
there's some cool iterative ones too that do iterative unsampling
Could anyone explain how this three clip thing works please? Thank you
you see, you type gibberish in each box, and hit queue prompt.

actually i really want to know as well.
SD3 uses 3 encoders, clip L, Clip G and T5xl
put your prompt in at least one of them
Yeah I get the idea, do I need another TE for negative prompt though?
Ill be honest, I havent pulled comfyui yet. I would assume so though
Gotcha
id like to speak to the manager, wheres my SD3. I'm on the textses with my lawnyer right now.
when and where we can download the sd3 weights?
ๅฌ่ฏดๆฏไปๅคฉ๏ผ
Guys it's 3am in the UK, they sleeping
Dont Godzilla me
sd3 should be releasing in about an hour right?
nope
then when?
tommorrow
Depends on where you live.
i live in india
Probably 12 hours or so
Gojira ๐ฎ
probably 6-7 pm for japan time
I envy the fools who waltz in here today and get SD3 while we all got "in two weeks" for months XD
thats 4pm - 5pm for me but not confirmed ig lol
How about China?
7 hours time difference
for sd3 medium. what medium stand for? like memium size model?
release on osrs update time?
2B, they are still training a larger 8B one
sd3 nano*
sd3 128m
it's a cheap way to say "smallest and least but not awful... Just medium"
the sd3 time travel experiment failed catastrophically, and the model was released into the future
we wanted so bad to see if we could, we never stopped to ask if we should
so i just pulled the new update of comfyui
do i need to download t5-xxl?
what is the folder called for the t5?
dont worry. this time its not any different 

Where is it!? "In 2 hours". XD XD XD
i try not to have high expectations when it comes to SD but boy i sure want to try it Soonโข
good observations I do see the ClipTextEncodeSD3 node has a field called t5xxl, I wonder if it's this model:
https://huggingface.co/google/flan-t5-xxl/tree/main
or this model:
https://huggingface.co/google/t5-efficient-xxl
reading the comments I think it's this one: https://huggingface.co/mcmonkey/google_t5-v1_1-xxl_encoderonly/tree/main
last one
oh boy its 10gb
but which folder should i put it in
Most likely models\clip
alr
Not necessarily yet, right?
are we supposed to download all those files or just the safetensors?
Probably going to need the safetensor and the config.json in the same directory
If you're talking about the t5 encoder
yes, this is all fairly new to me. what folder should it go into?
The site says "By removing the memory-intensive 4.7B parameter T5 text encoder for inference, SD3โs memory requirements can be significantly decreased with only small performance loss."
We don't know yet, but probably models/t5/ or models/clip/
ah ok, thank you
Doubt they fine tuned the t5 itself
But I don't know which version they are using. Like pixartsigma uses t5 xxl 1.1
bro just confused us and dropped 3 different sites
Nvm, im a idiot. Did not see mcmonkey said exactly which one they are using here
just to be safe i downloaded that 10gb t5xxl from the 3rd link it seems that's the concensus on which one to get ahead of time, especially if your internet connection is slow get that 10gb file out of the way
the CR VAE Input switch seems to not be compatible with the latest version of comfy, that's going to mess me up ๐ฆ i'll have to find an alternative compatible node
that shouldnt be a problem is its not that big of a deal lol
Is that so?
that moment when you keep refreshing https://huggingface.co/stabilityai for sd3 weights 
hahaha
you can have both vae fed into a reroute, which then feed into another reroute depending on which vae you want to use, then use the vae everywhere

you can also use this node if you have it
will SD3 require more vram than sdxl?
No, less
You'll run the t5 on CPU and regular ram if you don't have a bunch of vram
But you don't have to use t5
What do I need to set up this way?
I see, it have 3 text encoders 
Also, IDK if they have an easy implementation of BNB4 or 8bit for the t5 to run it on GPUs with less vram
=0
you would think they would pin a message that says "SD3 RELEASING MODEL FILE 6/12/2024"
Hopefully they post it in #๐ฃ๏ฝannouncements the moment they release it public
I'm going to guess, they will release it in about 6-13 hours from now (they are in london (almost 6am there now), so i'm guessing noon to 5pm their time)
git good
?
just joking. as if editing makes you better
oh lol ๐
hes still editing the messages 
oh you must have one of those keyboards that the letters themselves have fallen off? my laptop is like that lol
i removed some of my keys, but i've dropped this wireless keyboard over 100 times, easily
I only removed the keys I don't want to press like Caps Lock and Windows key
will TensorRT work with SD3 out of the box?
How does that happen? I mean, I can't remember a single instance where I dropped a keyboard. ๐ค
Rocking a classic 1080TI Toad? Man that GPU holds strong to this day. Incredible card.
EVGA FTW3 version. cost me a lot back then. Bought it when they announced the 2080 was going to be $1,500
shattered my dreams
I dread....to see the price of the 5090.....
Or the power consumption
20 kw
Exactly, the 40 series had some black magic with being able to power limit a card like the 4090 to 75% power with like 1% performance loss.
We'll see what happens now.
I underclock my GPU and CPU so that my room stays nice and cool. If I overclock them, it feels like being inside an oven
how many parameters is the current sdxl
my GPU current temp is 93*f (33.88c)
That's pretty good, I assume idle temp?
Oof, try doing an undervolt
basically, unless you count youtube, etc
they said around 10am EST
I miss EVGA.....
Mine around 19 degrees Celsius. But only because I am at work and it was not used the whole night. 
ripbozo
i am running a game and have stable diffusion loaded
but not a 3d game and put on lowest settings
so barely does anything
time wasting idle game
I'm thinking about loading up the Nolvus modlist for Skyrim for a bit and playing, before I head to bed within the next hour or two. Graphics card destroyer, but worth it.
i tried to do that,doesnt work with the finals but works with overwatch on lowest settings and a 1.5 checkpoint
stable diffusion 3 (t5 on gpu) + cyberpunk 2077 max settings at the same time ๐
and underclock RAM makes my system unstable so i leave it alone
lol
haha
Vader about to get cooked for the second time in that gif.
He added +1mhz over the max OC setting
We gotta try running llama3 + stable diffusion 3 + arma 3 + dark souls 3
Okey time to sleep ๐ Goodnight everyone
Night!
good night sweet prince
Sleeping until the release? That's cheating!!!
real men don't sleep :3
bro's gonna sleep for 2 weeks
Sooooo no midnight release? I've been outside gamestop first in line to cop SD3
anynoe that sleeps should have their downloads delayed by the amount of hours they slept
when people asked "when sd3?", it was "2 weeks", now that we are very close, when they ask, tell them "2 hours" ๐คฃ
release in 9 hours
+- 2 hours
in 2 hours
after that we only need to wait 2 weeks for 8b๐
2 hours, plus tax
Nice reference
you sd3 ceo ๐ฎ
Onetrainer update when?
give money
yea cant wait to train on onetrainer
honestly i'm kinda under the impression we arent't likely to see 8b at all
nah we will, just be patient
Seems like stuff drops around noonish based on the last releases
they were talking on here about the cost of training 8b being so high they'd paused it and weren't sure if they were gonna continue with it
cuz of the cost being so insane
How long do yall think it will take for the custom models to start dropping?
an homage to the waning hours of SDXL
Depends on if you mean decent models or trash.
how long before the first furious anti-SD3 rant on reddit?
i give it till saturday, max
I guess like the juggernaut type
the kind of models that only have a single lora merged prob around 3 hours
I think we passed that point already
cascade remix was cool,, it would be cool to get something like that for sd3 too :3
hoping to find some tricks in sd3 that allow me to do shit like this
otherwise... i'll be training sd3 on sdxl outputs
that's with offset fractal noise sampling
i cant see the car, let me turn on the lights
Hoping it runs smooth on forge that seems to be the only one that works for me
Well... I'm no fan of Juggernaut's: the model cohesion is far less than many others that I've tested. But as for a decent model, I'd give it three-four weeks if things are figured out quickly in regards to what's different than previous generations of model.
Works well anyways
i'm guessing it won't be supported on forge or a1111 for a while yet
if there was ever a time to learn the basics of comfyui, that time is now
a1111 in 2024 ๐คฃ
I don't know what it is but comfy or a1111 take 2 minutes per image
sounds like you need a fresh comfy install
I was using Forge until....well....RIP.
i mean... depends on your pc
but yea sometimes starting fresh helps
What happened with Forge?
Basically going to turn into something completely different from what I have read, but a lot of the enhancements are going to A1111, most of em at least.
@jolly abyss cant wait to try all the CCTV stuff first haha
i might give that a go, havent messed with it in a while
Can't wait for peeps showing how much better 1.5 is against SD3.
Oh yeah.
Yeah, it's gonna come. I'm curious to what finetunes for SD3 will look like in 6 months.
that's exactly what i mean
in six months, we will get research paper for sd4 
i was hoping it would be enough, RTX 4060, i7-13620, 32gb of ram
well that is good
even the one they came out with that was supposed to spit out images as you type didnt work
dont know why it takes so long
forgot what it was called
Cascade 1.1
i feel like it should lol
So regardless of A1111 or Comfy...2 minute gen time? Damn.
i dont think sd4 will ever be made
hahahaha yea they will give us another bastard child before announcing sd4
forge has been the only one that works legit for me 10-15 seconds per image
Half-Life 3 will come before SD4, at this rate.
gta 7
man gta 6 gonna be fun as hell
i really hope they do something like the RP servers that kept V alive
And they are gonna make us all double dip again if we wanna play it either at launch or wait for the PC release like 2-3 years after consoles.
I think it took Red Dead like 2 years, right?
we gonna get a half assed console port on pc like 1 year later,then they will slowly fix it for the next 2 years
only 8 hours and we will see how fucked we are
and then milk gta 6 for 20 years 
I will admit I'm DAMN curious to know what the GTA 6 economy is gonna be like regarding shark cards. Stuff will be even more overpriced.
is red dead 3 gonna happen?
Can't say, it sucks they abandoned RDO though.
yes eventually,the one that wont happen is Manhunt 3
also what happened to max payne? i hope they remake the first game, was one of my favorites
You forgot to say "Made by a 3rd party company in RPG Maker"
there is a tutorial series on youtube where a guy is making gta 6
its ok they will apologize later and give us a couple of ingame coins as compensation
don't even get me started on the fallout 5 release date
probably won't see that one until 2040 or later
But will it have bugs though?
elder scrolls 6
scary wizard of oz
Love the shadow catching rainbow ๐
more offset fractal noise
yea looks nice
Ironically...I'm playing Skyrim right now and I just hit a corpse with my axe and it exploded. Literally. lol
pretty much the only way to get truly dark images in sdxl
this is way better than just the initial latent offset approach
did you took an arrow to the knee? :3
I'm a civilized being, arrows do not affect me at this time. I am not drunk as of yet.
kek
When are we expecting release?
2 hours

my bet
So when will it be released?
2 hours
in the span of 2hours ~ 2weeks
Hey sd 3 out yet?
Oh 2 hours cant wait
so at launch what are we looking at for inference, comfy ui for sure
i haven't seen any commits for diffusers yet
i mean i know comfy and stable swarm will have it right away, not sure about others
proof?
definitely interested to see how fast sd3-turbo will crank, i'm running sdxl-turbo at like 15-20fps with that streamv2v attention processor
where do you people get the info that its gonna releasr 2 hours
Theyโre mocking you
spoilers: ||#๐๏ฝsd3 message||
our rear end
the best approx we have tho is like 10am-noon EST
but ultimately, who knows :3, prob in 2 hours 
it's out when its out 
it will be out but only for battlepass players ๐
just ask steven seagal, as you can see from his name, he is the sd3 ceo
Wow, I hope he gets good reception up in Antarctica
๐คฃ
Available as a rare SDXL lootbox drop.
The time to the long-awaited "SD3 Drop" is as long as a piece of string!
"How Long Is A Piece Of String?" โข
(Anybody buy Raspberry Pi shares? You'd have made da killing!!!)
sd3 is coming and so am i 
What a long two hours :3
In a few hours this channel will be unreadable. We will need a slow mode.
any time now... :3
yea it's gonna be crazy
some of the stuff we will prob see soon:
how do install sd3?
i have an error
here is a pic i generated
sd3 is bad
sd3 is good
comfy or what?
does it run on my potato pc?
people will immediately begin saying how much it sucks
30 minutes after it is made available for download people will trash it for not having any loras or other tools
i'm hoping for lots fo prompt zombies
also sd3 better than sdxl finetune confirmed 100
people are gunna get a bad image because of some dumb reason for them not setting it up correctly and say it sucks lmao. and also post in here saying why does my image look deepfried
lol
yeah it will get dragged fast
and they're gonna compare to a multimillion step finetune, not sdxl 1.0 base
of course
no booba? this is terrible
this is known
the floodgates are opening
one question
only one?
INTEGRATED GRAPHICS???? ๐ฃ๐ฃ๐ฃ
lol
1 iteration per day
does it run on cpu?
ancient astronaut theorists believe the answer is a resounding yes
get voodoo FX for a smooth SD3 experience
intel arc users 
Trying to allocate 10gb of memory only 4gb available. What am I doing wrong? I am on AMD btw. Classic
it has that uncanny dalle 3 vibe
me about to ignore those doomers
just a friendly reminder: 2 hours left 
That would be at least 14 waifus/s
I can;t try Dalle 3 coz i cant get thru those i am nto a robot nonsense at making a microsoft account
i guess i am a robot
i have no paitence
the fact that i have no patience for that crp prove sim not a robot
this is why i cant stand closed source
what is the conversion formula between waifus/s and booba/s ?
everyone with the bs rules and cpatchas
patience?
idk
huh idk when was that and i already have a Microsoft account lol
rotate the things put thie things do this i have no clue how many or why
fk it
dallee 3 sucks anyway
does have a little vibe to it now cuz of the texture of kim
i hate it so much
i kept seeing those radial shadows
and its much plastic
kims hairdo lol
the noise patterns are horrible
ya that ig
yes from what ive eseen dalle is somewhat lifeless
Does the API use sd3 8b or 2b? Do we know ?
YUPP!!
Where is sd3 ?
2 hours
almost all the mtg colors
sd3 is gonna be the king
and im having flashback of sd1.5, seeing those abyssorangemix vibe
then we go to 2 minutes
2 seconds
kek
2 hours
now kiss
can sd3 do this uncanny valley ahh cgi?
Perfect
i bet the top elite party member sin NK will try SD3 today
I will try sd3 5 micro seconds after release
i wish i cld generate an image in 5 micro seconds
I bet kim jong un has fast internet
at least 1 mbps
2b is more fast than 8b ?
Will kohya or simular tool be available at release for finetuning?
immediately, yes
launching SD3....
when
2 hours
2 hours
4 real?
yes 
whoa
0
You lose
white
rigged
Maybe
"Will SD3 make me a babe-magnet at all?!" ๐ ๐ฅณ
we both said 2 hours, so these things kinda add up... so 2 + 2 = 4 ok?
whats a good ui for the audio models?
2 + 2 = 22
GRadio
is there like a comfy equivalent?
So i cna just use it in A1111/Forge right?
Anyone know what time itโs coming?
Sorry, don't know ๐ฆ
2hrs,n2minutes and 2 seconds form now
I have a BBC Microcomputer from 1981. It has 32k memory. Will there be a version of SD3 made available for a machine with those specifications?
they're finishing up the hands
they are working on feet now
sorry feet are banned on SAI
rip
I don't even have a static image that will fit those specs!
Is 2B model going to be named Medium? Will Large or 8B models also released?
theyre turning the remaining blood into ketchup
Yes
2B model is called 2B, shocking!
It had a 1k screen-memory (VGA) and all it could show was the old Teletext/Ceefax!!!
I know, I started with a Sinclair ZX81 ๐
The guys who made the BBC Microcomputer have since evolved into ARM
yes. comfyui itself
Yeah, the keys were like touching "dead flesh!!!" LOL
don't do it ๐ฎ
i dont have a choice
That was the ZX Spectrum. The ZX81 had a "touch sensitive" keyboard...but you had to hit it hard ๐
Hello everyone, does anyone know when SD3 will be released approximately?
in 2 hrs
(never gets old)
my kek is top, my lel flows free, you guessed it right, lel kek
isnt SAI in england?
its barely starting the work day
give them so slack
no more 
let them finish the hands
if Elon was ceo he would have released already
If elon was CEO he would have renamed it Y and thrown a sink at it and released everythign
@hallow lion #๐๏ฝsd3 message
You should as well rename yourself Steven 2 Week Gaming 
07:50
Yeah
I wouldn;t be surprised if Seagal actually told everyone he is the ceo of SAI.
its the kind of thing he does.
he is a martial arts expert in every form of martial arts and is a buddist supermonk
sensei is a humble man,he would never do that
6.12 sd3?
maybe
on what universe ๐
in the Andromeda galaxy, sd4 is already out
this one
oh the universe where he is eating toilets
which nodes do you recommend?
Takes cover
depends on what you wanna do. if all you wanna do is put words and get booba then comfy isnt the right ui for you
in 2 hours
confirmed?!

for audio?
booba confirmed
yeah booba
malarkey
Comfyui is not for boobs
actually, what nodes do u guys recoomend for sd3
thats what i said
i dont think this is true
so far theres only a couple
comfyui is for everyone who is serious about making literally anything and having creative freedom
booba or not
tbf probably there are selectors out there that will mask out boobs
probably
allegedly
which node makes boobs out of boobs? like the buildings in inception, but boods
just increase your resolution
booba as in just basic nsfw stuff. you can do the same with a1111, forge, invoke. if all you want to do is just gen then comfy isnt right for you.
what comfy is for actually is customising workflows, images, outputs, inputs, using 20 control nets
booba is a real term not a typo?
in the dictionary 
booba is boobs squared
im a boobamer so this is over my head
i just want to make a zebra drink a milkshake in space man cant a boy wonder
yeah creative freedom. which you also want if you are serious about your booba genning
you are now banned from stable diffusion
it's an honor sir
Two thousand years later. ๐
ITS MA'AM
comfyui was made for serious booba genning
hires 2160p 4k HDR booba genning with inpaints and iterative upscale
one of the first things i'm doing with sd3 is training on my sdxl shit like this
Moles boob
the res checkpoint ๐
when can download the SD3 weights
2 hours
Source : trust me bro
2 hours later?
its been 2 hours for 2 hours now
yes
Released *what * exactly? ๐ฆ ๐
are you sure?
2 hours or your money back

I think Lykon said the target was around 1400utc. No guarantee it will be released then, but I definitely wouldn't expect the weights to be released any earlier than that.
it will come sooner if you keep refreshing this page https://huggingface.co/stabilityai
๐
plenty of time for meme posting
I wonder if huggingface has enough bandwidth to handle everyone dl'ing SD3 at once?
ddos 
large files come from cloudfront now so I think the days of hugging to death huggingface are over
confirmed (100% real)
that's in 2 hours
right. someone should leak it so we dont dl it all on the same time
It's 08.20 here in London UK. SAI Staff are still in Starbucks buying flat-whites and muffins ...
SD3 ไปๅคฉไธๆฏๅๅธๆจกๅๅ๏ผ
2 hours, 2 minutes and 2 seconds
we only have 2 hours left, a leak is supposed to be happen a week before, if its leaked now they would crash servers just by downloading for like 3 days
But it's not entirely fair that we should have to wait longer than a piece of string?!
cant people just reupload elsewhere if it crashes like google drive? or is that against tos
Oh! Didn't they tell you? In the UK, Amazon's Tilbury Depot will be handling the roll-out ... ๐
it should land on civitai too
also just by looking at this chat, dosent seem like theres enough hype to crash the servers
isnt it against tos if used comercially? but i think they would still get angry if someone posted it somewhere else yeah lol
you underestimate the lurkers
redistributing for free isnt commercial use.
also about that, you can just buy a commercial license subscription if you plan on selling images right?
can you test the model they are releasing on the api currently?
sd3 delayed and is releasing next month๐ฑ๐ฑ๐ฑ๐คซ
local or nothing
yes if u make like a million $ a month from those imgs
I just wanna know what we are getting in advance
sigma balls
but is the weights they are releasing also on the api
i know. selling it is
im assuming it would say beta if it isnt?
ok but what if i just take donations ๐
A side question to this but my understanding is the commercial license isn't to do with selling or profiting from images made from SD but from using the SD weights/tech in the back end of a commercial product?
you give those donations to me ill handle everything
the middle man ๐ฎ
confirmed they have the medium and large on api

we do a little tax evasion?
give those donations to the completionist, he will take care of them 
its 20$ a month for creator license
its not tax evasion if you burn the money on a casino in Macau ๐
man, I gotta say this SD3 medium they got hosted on the API sucks
what even is that
ps3 graphics poggers
i fixed the names for them
true^
i've heard the api is just worse than the actual model
why would i play gta5 when i can generate gta5
but dont have too high expectations 
240p ๐ฎ
are those nukes in the air
so I assume I won't be able to download the model to run locally? correct me if I'm wrong
yes you are indeed wrong
you are wrong
so wrong
you are wrong
im 100% gonna refine each image with sdxl anyways untill its community finetuned
8B model soon or we riot
(in 2 hours)
you wont be able to run 8B, true story
Ok, thank god. I was gonna start buggin
oh yeah? well I will when I buy the rtx5090 on launch for $5k
the mad lad
I can run Stable Cascade at 13 GB, It might be a different story though.
(not doing that)
same
still 2 hours?
2 hours and 10 minutes now
2 hours until 2 hours
Better than 2 weeks for the last 2 months ?...
honestly, maybe my previous prompt was too specific or something? some of these arent bad.
i'm going to alt tab and smash all my keys
Bruh release which time? Iโm in CEST
Car crash
2 hours, plus tax
soon^tm
๐
probably noon cest
i thought stability was based 2 hours in the future
why is there a floating fart at the background
its bouyant
i thought stability was based
trying to get missiles launching into the sky in the background. it didnt quite get it
it definitely knows lambos better than mustangs
gta 6 ๐ฎ
(2B model btw)
i thought stability was stable
Bro's making gta6 himself
lets see how well it knows GTA6
Is this a different case or no?
The model just loaded into my DRAM
aaaand it gave me los santos
yeah ask it gta 7
8b unet won't even fit into 13gb at fp16, fp8 might work, not accounting for encoders tho
bruh this is cursed
it will prob give you the leaked pics ๐คฃ
kekw at the pegi 18 attempt
I have 32G
How large could this model possibly be?
you need around 16gb of vram I reckon
I wonder how many samples they are running on API calls
I have 24gb vram and 128 GB Ram
So if it's going to load to DRAM I would only need more than 16g, right?
ok now...
chad
Bro's got a whole server
xd
A.P.I. - Astoundinly Poor Images?
It's not finetuned
thinking of picking this up for SD, thoughts?
thats my girlfriend how did you do that
There is an "SD3 girl-face" I believe?
doin too much
you just need a small loan of a million dollars

