#✨|sdxl
1 messages · Page 36 of 1
I thought a1111 removed the watermark like a year ago
also I thought it was invisible
that "watermark" option in his ui didn't actually do anything fun fact
True
yeah I just read that whole thread
and I thought so too but you're not the only person to have it
ugh, so do I look for diffusers.py?
a1111 1.5
how easy will loras be made again with sdxl? Was it (around)10 pictures or am I remembering wrong
anyone know at what time will sdxl 1.0 come out today?
major vibe kill
when Emad finishes Tears of the Kingdom
could use comfyui for now ig
yeah, suppose I need to learn it, people keep saying it's the one
im assuming when the event happens bout 8 hours from now
if you've ever used blender shader nodes or unreal engine blueprints those are way more complicated than comfyui
I have
comfy's got like maybe a dozen important nodes at most
but there's something uniquely unintuitive about diffusion
there's a few premade sdxl workflows floating around
right on
I will install tomoz
the token limit is annoying me too
so I might as well
Sytan's basic workflow which uses only vanilla nodes https://github.com/SytanSD/Sytan-SDXL-ComfyUI
Mine which consolidates everything into one Principled SDXL node + some extra stuff
https://github.com/Beinsezii/bsz-cui-extras
ty sir
still getting good stuff tho
love the colors on this! What prompt did you use for achieving the atmosphere?
I used a 3 word prompt with my A1111 adjustments
"fantasy lava lamp"?
xD
apparently not
oooh I know..."lava lamp fantasy"!
will you give us a hint? is "lava lamp" two of them?
you know the images contain generation metadata right? Meta for the 3rd recent one
i thought discord strips metadata
aha, so lies
i think you just need to open original file
Images made with ComfyUI can be dragged straight onto the canvas to rebuild the node workflow, too
what? I didn't lie. holdup, I'll screenshot if you guys like digging in my stuff
make of this as you will.
I took the meta from this one
Idk what to tell you. I literally just screenshotted my screen..
also if you guys have magick you can get the a1111 meta from any image quickly using magick identify -format '%[parameters]' image.png
not really useful for comfy since you'll get a 30kb json file
I literally just showed you my settings, why are you still complaining? I did say I used a 3 word prompt with my A1111 adjustments, the screenshot I sent doesn't disregard any of that.
nobody complaining bro, chill
guess you could do magick identify -format '%[prompt]' image.png > workflow.json to extract it for further use
I see you dont have 10 stories worth of prompts clogging up your ui
ever since I made that Principled SDXL node my workflows are tiny
my workflow is smol too. but my notes are... epic
you guys are speaking greek. will I understand this wehn I start using comfy?
14 KB for my principled
ne, milao ellinika
just ignore the fat python node you have to install
eW91IGd1eXMgYXJlIGEgYnVuY2ggb2YgYmFiaWVzLg
actually not that fat tbh. both nodes + the principled workflow is smaller than most people's normal workflows
T.T all my prompts occupy only 5kb. I am both impressed and underwhelmed at the same time
just notes that are fluff on the side
i've got like 2kb of benchmark prompts for 1.5 but they need to be updated for XL
Positive prompt: gaming lava lamp, realistic, desolate, intricately detailed, artistic lighting, particles, beautiful, amazing, highly detailed, digital art, sharp focus, trending on art station
Negative prompt: high contrast, two people, lowres, ((bad anatomy)), bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, (jpeg artifacts), signature, watermark, username, blurry, deformed, (bad eyes), (misaligned eyes), squashed face, pixelated, people
Steps: 67, Sampler: DPM++ 2M Karras, CFG scale: 4.2, Seed: 1177726382, Face restoration: CodeFormer, Size: 1024x736, Model hash: 88320a3dd9, Model: Maybe_a_good_model2, Denoising strength: 0.6, Hires upscale: 1.3, Hires steps: 37, Hires upscaler: 4x-UltraSharp, Version: v1.5.1-RC
you must have some kind of embeddings or something that run automatically and fill the pos/neg with extra quality words
yes, there is, just like I mentioned. there are automatic adjustments I made.
do embeddings like easynegative not work with sdxl?
Ah I didn't see that. At first it sounded like you were confused as to why the meta was so different from the UI
imma sleep so I can wake up to 1.0
or at least the fallout from when it doesn't release
is there a way to update automatic from within automatic or do you have to download from the github page?
wait so not SDXL?
git pull
yes SDXL. see: https://civitai.com/models/114238/tdg8uus-sdxl09-mix
gaddam 13 gb
it's not base SDXL0.9 no. I altered it to make it slightly better.
go to the base install directory and run git pull from an admin/su command prompt
I meant 13 gb is normal
oh yeah, it is
yea it's like 20 gigs for the regular base + refiner
that's why you need 24 or 32 gigs of ram to run them both
I should probably remove it from my 8GB machine then
hmm I've been running them fine with 16
They work with one of the text encoders but not the main large one.
i guess if you unload the base before you refine it could work
bit slower though
Same I'm running it on 16gb RAM and it's fine.
ahh maybe I shouldn't use comfy then. I don't have to do anything in a1111, it just works
my ram usage is typically 22 gigs for my system when using SDXL
It just works in comfy too
If you have more ram it will use more ram.
I dont
Switching between then only adds a few seconds to the generation time. It's not an issue. In comfy at least.
assume you can split the steps too? I don't refine most of my generations
adding 2 seconds onto a 9 second generation is a 20% speed loss
better than infinity seconds on my 8gb machien
Depends on the GPU have. It takes longer than 9 seconds on. 3080.
yeah I have a 3080
it switches the one used to the other one in stage 2. so basically in stage one- base is in vram and refiner is in CPU ram. in stage 2 it's the opposite. this allows for a much more optimized experience
They were the best value 30 series card
Sytan has the 10 gig one too
I've got 10gb. Got it at release.
making a $700 gpu or whatever it cost with 10 gigs of vram is criminal
It was £650
that's like not even enough for VR
lol you don't need lots o ram for VR!
I got 4070TI
blade and sorcery maxed out my 1070's 8 gigs, I had to turn shit down to not stutter
I played through HLA on a 2080 super, it has 8gb too
also for extra fun, BeamNG used 20 gigs of VRAM with traffic enabled lol
tp movement all the way
the first VR game I played was Boneworks so I got used to it pretty fast
I can't get used to it 😦
now I hate games that don't have parkour and smooth movement
i think my rift s broke down, it's position keeps shifting, making you instant nauseous
clean the cameras
did that so often already 😦
even thinking about it is making me kinda queasy tbh
I actually go into VR to take a break from programming. Helps the headaches lol
1 minute of vr that way is enough to give you "fun" for the whole rest of the day 😦
I do it to get off my ass. 5 rounds of totf and I need a shower
I did 5 hours of Boneworks once
i hope that didn't give you vr-sickness
what do you even do in that game after the tutorial
good question
at what time is the event?
it's been "in progress " for like years
install Medeival Megapack so you get actually well-made weapons. You can halfsword, mordhau, etc
it should be noted int your local timezone
click the event in the top of the channel bar
Guess the movie....
aco
i wont survive by then
reminds me im supposed to be sleeping so I can wake up to the "where 1.0" fallout
4 am here
20:00 ^^ best time ever
<t:1690394400:R>
7 hours till the world will no longer know whats real and fake anymore
if someone wants to use these as banner arts, you can
I use VLAD A1111 diffusers - no watermark
diffusers has watermark. but nothing stops ui devs from disabling/removing it
those are top tier ❤️
can we make a change.org petition for them to release it in 30min?
So today release???
but how will you wake them, and tell them about the successful petition with 200k signed people
<t:1690394400:R>
and the staff woul sign a petition to ban you
then we make a petition to unban me
#✨|sdxl message
this is what would happen to petition owner XD
we send a raven to them
powerbird !
used to be very difficult to get something like this with 1.5
I'd imagine so
he needs a pirate hat
release the kraken!
Yo whats cookin'?
Khajit is waiting patiently for new SDXL.
and this is why we dont kiss people from krypton with our eyes closed
@boreal bough But baby
A tesla car exiting the conveyor belt of an alien space ship
whats the performance like if you store models on an external SSD drive?
i think that your cfg is extremely high
try to do 60 90 steps, i saw someone that it helped him, or do it with 4.5 cfg
Just will load MUCH faster
Storage for the models is about loading speed, if you have a good interface to that SSD (I guess only USB 2 would be bad) - and it’s a good SSD - it should give a nice boost
Nvme will be much faster still of course
im playing this until 1.0 comes out https://www.youtube.com/watch?v=gzHyGFKnEgk
Provided to YouTube by Universal Music Group
Coke Room · Justin Hurwitz
Babylon
℗ 2022 Paramount Pictures, under exclusive license to Interscope Records
Released on: 2022-12-09
Producer, Conductor, Associated Performer, Orchestra Leader: Justin Hurwitz
Studio Personnel, Mix Engineer: Noah Hubbell
Studio Personnel, Mix Engineer: Nick B...
Hope he s not waiting for TES VI, he will die before
Khajit will have own story and game by the blessings of AI by then. Khajit's adventures will continue.
Nice @timid sonnet are you using some additional checkpoint than SDXL base and Refiner.
Ragin' Cajun is hyped for SDXL's cosmic powers.
RTX on.
Wax On Wax Off
ooh this one came out great
Ya'll give me some credit for the prompt 😉
I did have to improve on it 😄
You did well
A cinematic photo of a tesla car exiting the conveyor belt of an alien spaceship, shiny reflective surfaces, chiaroscuro
ENTER THE PORTAL!
just moving artists around a bit, but this prompt is so creative i love it ❤️
prolly sleeping
Awesome. On an image like this is it possible to display own our text instead of the neon signs?
There is probably some way to do it accurately and reliably using controlnet/inpainting, but it's not reliable with just text2image.
try elon and zuck
generating
a bit more intimate
haha awesome
Looks more like Zuck is up in his face, threatening.
Looks like it
sd xl releasing tonight?
we all hope
anyone know if the Center Stage event today will be shown again later? I'll be in meetings at work and won't get to watch it live
Is today the day?
dude, where do you get the prompts from, they too good
uhh... mild modifications on my old prompt... 100% made up stuff lol 😄
i mean, i've tried poems, song lyrics, random bullshit/characters, booru prompts (most boring imho), llm prompting
at the end of the day, it's just trying to figure out what the specific model likes best
and SDXL loves a LOT
but i suppose the big winner of my prompts is : double exposure
that just sends SDXL over the edge in creativity
double exposure shadow puppets
any moment now
i feel it
eface
@floral island Try adding "temporal universe, tesla color gradient" to your p prompt
ooh, good stuff ^^
lets try that next
Tesla color for some reason has good results for me
and using universe out of context as well
Where sdxl
@west breach On today's gameshow of who's really a man?
all of them
hmm, i suppose the rest of the prompt might be a bit much for ur tags, but the results are awesome nonetheless!
I do believe the roll out may be starting..... I see this was updated 17 minutes ago... maybe with 1.0 changes applied perhaps
one of my AI mates said the same too... u guys are f5 abusers!
Are they also launching controlnet tonight?
nope just checked to see what freewilly was about and saw that
hopefully
i think people already are on sdxl 0.9
Should be, we don’t really know
Not old models, but yes
but having a 0.9 lora means nothing -> you'd have to redo for 1.0 again
they will release training tools along with 1.0
They prolonged the launch way too long. I wish they go full berserk with it
nah, i don't mind developers taking long EVER
i'd rather have a good function product
Yes yes definitely
That's what I am saying
From my pov, I have been waiting and waiting
Now that they are putting it out, I just wish the results are good
The crowd gathers...
that is a nice kitty ❤️
@nimble heart Hade you found a way to also include control_after_generate in a custom node?
heterochromia space cat judging your actions!
for all the cool cats i get... sdxl 0.9 fails frogs consistently 😦
Woah! How did you get that flattened look?
i'm pretty sure he prompted it! (lots of ways to get it flatter)
Amazing!
That's the guy I see in my DMT hallucinations that probes my brain
vaporwave, 80s, 90s, retro, nostalgic, glitch art, composite image, glitches, scanlines, 30mm analog film, sunset gradient color pallette, miami art aesthetic culture, simple, flat, minimal shading, surreal, dreamlike
negative: watermark, 3d model, digital art, cartoon, realistic, highly detailed, complex background, intricate details, paint splatters, gradients, texture, messy, scribble
this cat get gets p....
I have been m.i.a. for a while, has anyone figured out a way to run SD online in a notebook without a compute time limit or credits? I used Google Colab for a long time until they implemented the compute time limit for Colab Pro
vaporwave vs synthwave
@west breach
The last images ever taken of these majestic animals.
mmm like I'm sitting in the glow of a warm fire ❤️🔥
🙂
4 hours left
@ionic dragon I got a few in here: https://github.com/twri/sdxl_prompt_styler/blob/main/sdxl_styles.json
gnite. hopefully 1.0 will be out when I awake 👋
24gb
thank you
really curious how good the trainer is gonna be!
Does sdxl get released today?
Does today prompt still train the released model ?
or at least, the announcement
my guess is it's the final model with minor inference parameter tweaks
i asked the most important stable staff member, StableVicuna-13B, how hyped it is:
it's a lil confused but it's got the spirit
you might have overtrained on sonic fanfics but ... 👌
thank goodness Shadow the Hedgehog will be there
also hoping for more stablelm models 😄
was worried he wouldnt make it
i havent tried a local LLM in awhile does StableVicuna have a 7b GGML that could maybe run on my rpi4?

so sdxl releases today?
Where would I go to download it and would it run for stable diffusion webui?
Continuity ... 🤔
when's xl to be released?
close ur eyes
🥲
3h maybe
thats not closed
Of course a thunderstorm takes out my Internet on the supposed sdxl release day 😩
before all this AI stuff when a release date was a few months away it seemed like nothing....now even a week seems like forever
if you take the bottom as nose otherwise LOL
Are we going to be able to train Loras right off the bat? Want to update the Lora of my face as soon as I can.
🤣
doesnt count
In the next 3 hours there’s gunna be so many people asking when sdxl is gunna be released and how to use it😂
I just wonder if 0.9 loras are compatible with 1.0
Yes should be
followed by SD 3.0 wen? and SDXL 2.0 wen?
||psssst i got SDXL 6.9..... who wants it? 👀||
Probably next year
no more layers PLZ
ayeeee
With a1111 already having sdxl support, getting it to run won't be hard, I'll need a tutorial on how to train a Lora for it though :/
you want?
can it generate among us because sdxl fails
kohya_ss is great
thought you were making a joke and was just acknowledging
I got 0.9 working on Auto1111 latest version but it seems like now my 1.5 models don't work as well. On Vlad 1.5 works as usual but on SDXL I only get black images.
it can generate all the us, we's and thems for ya
0.9 works and my 1.5 models work also on auto
I will eventually...I'm just used to AUTO/Vlad
has anyone been able to successfully train sdxl lora's of a 3090?
I'm gonna test some more to see if I messed up my settings accidentally
comfy is so fiddly though, been trying invoke and I cant seem to get sdxl working on there.
I use ComfyUI all the time:)
3090TI
@peak dove what do you do with bad fingers in ComfyUI
Does StableVicuna produce prompts?
Yes 3090’s can do it no problem
been making them on a 4080 all week
do u train lora for the refiner/base?
it wasn't specifically trained for it but it can if you get a good preprompt for it
Can we attach files in SV?
it wasn't trained for image inputs
i train on the base
i'm not even sure if you can make loras for the refiner stage. maybe but i never seen a point
technically possible, though there's very little focus on it
<3 the base
No, can we attach txt files with prompts so we can partial train
StableVicuna is an LLM, it doesn't implicitly have an interface of its own, it's whatever interface you choose. A pre-prompt is a common option that is just a list of example generations so the bot knows how to behave
dad is sdxl out yet
event in 81min
i gave some heavy criticisms about communication i remember. menacing criticisms perhaps. announcing the delay in a secret side channel allowed conspiracy stooges to gain some credibility with their leak.
i'm not going to forget that. the conspiracy side of the community sure won't
I want it to slow down by 5x
I was ragging hard on how shit their results looked, but that's cause they added base 1.5 and 2.1 into the mix, and it took me over 100 images to see one that wasn't unusably bad lol
is the stage thing a video or audio stream?
talk to your government elected officials about slowing it down. they won't listen though. those guys are slow.
ML revolution is here to stay and its barely gotten started. i don't think theres any slowing it down. nobody is regulating it hard enough or making any effort to
EU is passing policy but they've historically been bad at enforcement
Ugh yah not talking about The Man getting involved. Just you know, lets put on some reggae, go enjoy the sun for a few hours then come back to work... i mean, lets embrace this ai making our life better thing. Hacky sack? (Ps, yes i am old)
What is this ancient dialect
Jk
I understood all of it
Has anyone verified that SDXL's VAE is f/4?
old enough to know kick bag too. yeah but i never got good though
regulation is the only means of slowing any of this down. don't count on it slowing down though
I don't think I have ever heard anybody say that
I am too old to post an animated gif aparently
It was animated the first time for me haha
NoooOO Sytan woke up
:C
worked for me. old guy riding.
i was more of a bmx / 4 elements kind of kid
Oh clunky old m1 ipad cant handle it
Radical

What's so bad about me being up? Lol
SDXL 0.9 Episode 1 releases today?
how much longer to the event?
animated gifs old people can hear
If you click on it, it will give you your local time
if we dont get it today...who do we execute?
Well, even if the world froze at 1.5 the tsunami of tools and features would still be full throttle for as long as we go
Oh wait, it's more than 2 hours away, my bad
theyll drop it after the event
they want people to watch it if they release it before nobody will but a few
this isnt an Apple keynote event man
i dont understand why everything has to be some keynote speech these days
me either narp
emad's goal is to be bigger than apple. you gotta fake it till you make it
Check huggingface
To 'control messaging'
I just heard rumors we are getting a 0.95 today
I won't be watching it regardless lol
0.9 episode two episode one
But keynote they could release would have been releasing it on its originally set date lol
directors cut
i want a new sdxlv1.0_report.pdf because i like information presented in figures and graphs and explained in paragraphs. not video / audio streams
i enjoyed the first report
Uncensored
the event should be 30 seconds. just the first 15 seconds of a rickroll. fade to black text pops up "its released, enjoy" would be perfect
But it would be difficult to paste 100 prompts
Unrated
that seems easy but you see, that'd be copyright infringment unless they paid rick astley's label for the rights to that song
memes usually are copyright infringement. people don't really have a good grasp of this
Featuring Latent Diffusion
fair use if its under a certain length I thought
that wouldn't be fair use
context and intent is everything
They could claim it's for educational research purposes and call it fair use
<t:1690399800:R>
What's in 4 hours?
nothing
i use the timestamp website every single time joe
They are a research organization after all 
Oh okay I'm going back to sleep gn
a live stage thing, i dont know why the devs are beating this 'dur whats going on' joike into the ground for 3 days
1.0 on clipdrop!?
Cuz it’s funnyyyy
a business with millions of investment dollars poured into it really shouldn't use a screen door on a submarine legal defence
Oh u guys don't know? That's just so they can tell you the improvements from 0.9 to 1 its not ready to release yet
people are here for infos, help them out
Okay well here’s the info then. Sdxl is getting release during the live stage event. Case closed. They will also go over the improvements and show off there new ui’s.
you seem annoyed
beating a non joke into the ground is annoying to anyone with good humor imo
i was trying to do a build of stablestudio's comfyui branch last night, but i don't know rust tool chains. the commands are like "that command doesn't exist!"
jokes are subjective. your reaction has made their efforts funnier to me.
sdxl won't be released for you in particular >:(
then youve got shitty humor
oh nooo
i'll remember to keep beating a joke into the ground because i know it tickles your funny bone
and like is aid, pseudo got banned for the doing the same thing
unless your a paid comedian humor should be for yourself.
didnt pseudo leave himself
this is funny too thanks.
yeah but he still got a time for repeatedly saying nothing was releasing
I was agreeing with you lol
this is actually hilarious
in 4 hours you'll see why 2023 wont be like "2023"
actually i think 4 hours is too long... lets just make that 30min
30min guys
and then you can generate all the cheese images you want
if i recall, it was more his attitude surrounding the repeated saying it, rather than just saying it. lots of people were in here insisting things were delayed
inb4 stability announces that from now on all generated images will be fingerprinted for tracking purposes, and they will sue this guy if he uses it
i've crossed pseudo's path a couple of times. he's always been so angry with me.
i miss pseudo
they should start the keynote like 30 seconds late just to mess with us
theres nothing funny aobut repeating a non joke as a joke, if you find it funny you must love Maga
bro
well its all about owning the libs and doing things just to try and annoy people, sounds like that dudes comedy stylings served on a platter
this is just getting sulky now
did i hit the mark?
||a non joke as a joke|| << hilarious
Did SDXL 1.0 refiner get trained on bigger res than the .9?
Bro brought politics into a non politic conversation.😂😂
i brought another example that was tied to politics
I'm a canadian who wants the NDP to elect a new leader so they can get some standing in parliament again. But i bet you don't actually really care about my personal politics as much as you're pretending to
youre right i dont care about you
you've decided i'm maga already
everyone I disagree with Adolf Hitler!
bold stance
contributing
wat
yeah i guessed. you don't seem to have a lot of actual conviction
someones getting personal smh
somebody better release things quick
O<t:1690399800:R>M<t:1690399800:R>F<t:1690399800:R>G
whats in 3 hours?
still a funny joke even now, 3 min later
i think MJ 5.3 is releasing today too.
you mean 30min
2 hours left! 
fr?
i have no idea whats going on ||https://www.youtube.com/watch?v=Kob0G2hE8IY||
240p for real Men!
Check out http://for10hours.com/he-man-heyeayea...
for more 10 hour videos!
in 3 hours is when Joe finally loses it and bans everyone that's been annoying
387.4 hours left. thats how this works right?
its out anytime i guess and 6.0 is on the horizon tooi
you cant legally ban us since stable diffusion is open source
he should ban himself first
was a great moment tbh
nobodys been more annoying that that dude
Wouldn't suprise me. They updated the last time SAI did a big announcement. Could be a strategy to keep the attention of it's users. Or it's just a very funny coinciedence.
seems like wanting to delete a discord server that one's moderating is a common issue
Wait why 3 hours? Thought it was in 2? O is it like a preparty/ opening ceremony
2 yeah
we pregamin
$2 and you release 1.0 right now? how about it?
drinking game rules. take a shot everytime a newb shows up to talk about mj
okay $1.5
get blackout drunk if it gets delayed again (i cant handle this anymore)
liver failure
youll die by the 5th minute
is this common
nice that you can post localized timestamps
We should ban Chaz for hating on stability staff
Can anyone share how much money they have made through SD
I thought it was 2 hours left. Now it is 3 hours left. It delayed.
also do you also get people that come into a random channel, post the channel name like #✨|sdxl and then a prompt?
probably not but its fun to imagine
0.00. actually negative if you count electricity
still showing 2 hours on the event tab for me
An hour in AI time is a year in regular time
yeah, 2 hours here too
negative a couple thousand dollars cause wow computer hardware is expenseive
who hating. all i said was beating a joke into the ground is like the biggest lack of humor and not helpful to people who are popping in looking for info. but you wanna ban hammer my ass go for it if youre that lily livered
MJ? A lot. 👍
you can swallow now
bro lets stop arguing and just enjoy the fun of the hype
for a while there'd be those poeple trying to use the bots. how to use bot knowledge has mostly been figured out by those crowds though.
you're complaining about them repeating the same thing over and over again by... repeating the same thing over and over again?
sounds good to me
I agree! I’m hype!
I personally believe the joke was funny
a repeated joke is never funny
incredible, they're learning

a repeated joke is never funny
thats why its so funny
tbf he only replied there because someone said to ban him lol
heh
depends, rickrolls are funny sometimes
its the opposite it shows you lack humor because you cant think of anything else
.>
damn i'm funny
what does someone lack when they keep complaining about the same thing?
i also dont think being rude and not helping people who dont spend all day here and dont know whats going on is very funny, esp when pseudo got this big ban over doing it and hes not even staff
"my definition of funny is the only acceptable definition of funny!"
it's always hilarious when i see this attitude actually manifesting in the world
then stop @'n at me about it, you guys are bringing it up not me
May I ask a question? Why is everyone so hyped for the 1.0 release, when it will be still a while before properly trained models will come out of it, so that the regular user might enjoy it?
I get it. 2 hours for the live start and 3 hours for the release
if you talk about how hilarious it all is without @'ing him, he'll still see
it's, uh, the VAE, for SDXL, lol
also, 3rd person allusions towards him will be especially hilarious. he'll think he's a 👻
SD architecture is 3 parts: the text encoders, the latent encoder (VAE), and the unet (the main part). VAE gets separated out often
well for one, people need it released so they can make those models
hilarious requires a laugh, if youre actually laughing at that it doesnt shock me, i mean look at situational comedy, stupid low denominator shit is always the biggest seller
It's the refining step the program applies to the image after diffusion.
Some people say that it is right or wrong to use VAE when the VAE of large models is destroyed
Of course, that's understandable. But shouldn't it less of a "Can't wait for SDXL 1.0 to be out in 2-3 hours!" and more of a "Can't wait for SDXL 1.0 to be usable in the next weeks or months!"
chaz I gotta say you are a great comedian, this conversation is very funny
Vital Anatomy Engine
Wait what site/where is sdxl 1.0 going to be posted?
layer 8 issue
thanks sponge, that means a lot coming from someone like yourself
usual place is hugging face
in simple terms the VAE is a bit like the refiner
i mean even without finetunes it looks pretty good, even if its worse than other options rn its fun to mess with
I appreciate that
thatsa more confusing way to put it. the vae takes the latent dimensions of the data and renders it as actual pixels
Well, it can only be usable AFTER it's released
Honestly I'm just being provocative, it's exciting as hell. :D
For people training on top of it, it will be usable in 2 hours 🤔
so it's less about being excited about using SDXL and more about it now being possible to use it
and knowing if it's worth it, in a way
lot of people will be wanting to use it right away after it's public. Broader public adoption will take a bit of time, naturally
yes
Aigth thanks
wont be worth it probably, i think the SD community should permanently stick to using SD1.5
Some people might train a few models in the first day, it would be a good way to test if a model in SDXL is considerably better than one in SD1.5
anyone got tips on how to build the comfyui frontend version of stablestudio? the release is the first version
SD1.5 has worse prompt understanding than SDXL, atleast from what i've seen
does it need to be built on linux?
woulda been funner if you didnt include the /s
yeah, but but nsfw or something
I think we should all stick to dall e mini
I think SDXL is "censored" in a way that it wasn't trained on NSFW
not that it's uncapable of doing it
0.1.0 release tag is up to date
That wont matter m8. It will be trained in by other people later. 👍
why would it take time to get it installed? isn't 1.0 using the same architecture as 0.9?
it is in the way that NSFW images just look... wrong
but it still works, so it's just a case of fine tuning it
oh weird. i must've done something wrong and wasted so much time trying to build a version.
☠️
a semi-official comfy workflow with an example picture to reproduce is all I ask
one that you'd consider a good result
@visual glade
https://github.com/Stability-AI/StableStudio/releases/tag/stablestudio-v0.1.0 this has some magic to it
is now a front end for comfy from what i understand
there's a pinned one on this channel
stable studio is a front end. no optimizations going on, just an interface.
I was talking about the sd1.0 release, unless nothing's changing on your side?
I said it'd take time for the broader public to adopt it ie time for human people to decide they want to use it in large numbers, which will likely only fully happen after a lot of finetunes are out and a lot of tools work on it fully
Wasnt there a tweet or osmething where they said it would release with custom models and everything
when's that?
with comfyui as the backend
Cause they had been working on it
july 18th wasn't it?
heh this joke again
Whenever it is is fine by me, just get the multimeters working, Joe
does anyone know vram req for training lora?
ah I see. will it be possible to use TensorRT on SDXL1.0?
8-24gb vram
if we change the windows date to june 18th can we get it now?
o ok. let's delay release then
Worth for the multimeters.
when the project is finished, devs get unruly
yeah give them something to work on
i work remote i have no idea where the server room is
for only 50 cents a day, you can make sure devs have projects to keep them busy
the cloud obviously
i wish i could just walk into a room full of all our GPUs... i might borrow a few accidentally permanently
listen for the whooshing sound of heavy AC units and fans. cant miss it
all it takes is 50 cents a day to keep 1 dev from using discord
probably better for their mental health tbf
for anyone's
i mean if we really want to improve everyones mental health we should just delete discord
since some devs are here currently, will TensorRT work on SDXL?
I'm curious how much of an improvement there will be from 0.9 to 1.0.
the rest of the internet still exists. what do now?
delete it all, revert to monke
instructions unclear. i'm stuck in the internet now
11.1(1)% improvement 
just wait a few years and you can be a virtual monke with perfect immersion
purely from a naming standpoint... 10%? I say we forget the 1.0 release and go have fun with 0.9
that's what I'm asking.. I tried to use it on SDXL0.9 and it didn't work.
when I tried it it worked fine
it gave me an error about SDXL's shapes
then you didn't do something correctly
square peg round hole problem.
Can SDXL be as widespread as SD1.5? This is something I've been worried about
itll get there
you cant really predict how prolific it will be until the community gets it
going from 512x512 to 1024x1024 is huge
how long does an 1024x1024 take on a 3090 ti
The demands on computer hardware are getting higher and higher to the point that I have to go to a cloud server to run
that old addage has been flipped right upsdie on it's head recently. apparantly the people who say it haven't really ever tried playing the game. https://www.youtube.com/watch?v=6pDH66X3ClA
2021 Meme where an OCD girl cant take the wrong shape going into the wrong hole.
Original video can be found here:
https://www.youtube.com/watch?v=9nSQs0Gr9FA
featuring Alison Burke
https://vt.tiktok.com/ZScybDAB/
and brock1137
https://vt.tiktok.com/ZScyGwsu/
just dont expect it to immediatelly make all other models obsolete
bro beat me to it, i couldnt find it in the gif finder
jinx you owe me a coke
AssertionError: must specify y if and only if the model is class-conditional any idea what this is?
that's just the nature of increasing the resolution tbh, the multidiffusion process where you use low amounts of vram to create a large picture in small chunks at a time should still work
ill give it to SDXL, just ask SDXL for a coke and u will get it
should be a matter of seconds, 10 or fewer, depends on specific settings n wotnot
yours popped up the second i hit enter. i was like "wtf"
Perfect!!
we haven't got there yet for general usage - if you have a 8 GiB graphics card you're good for SDXL, if you have 4 or 6 you can probably run it just a bit slowly
yeah im praying it works reasonably with my 2060 (honestly it amazes me how much can run on this thing)
if you're personally on a very small card and can't upgrade and don't want to wait for ultraslow gen on tiny cards... well we're also doing our best to ensure cloud services are pretty good too
In China, we rent a 4090 for only 2.5RMB for an hour
ComfyUI will gen 1024x1024 with 4GB in 3 mins or so, nothing too complex controlnet etc dog it down
it means you need to read the code a bit more to see how it works
I tested running SDXL on free tier google colab yesterday. idk how much google will allow that but it works lol
WHAT
it's not something I can just say "do this to fix it"
oh, I did. when I changed the code it gave me a CUDA memory error
oo, 35 cents US (google conversion), that ain't bad pricing at all
yet another string ban incoming 
So we are getting close to the H hour 
Stability AI unveils Stable Diffusion XL 1.0, now available on Amazon Bedrock for developers worldwide #StabilityAI #AI #AIcommunity #TextToImage #OpenSource #AWS #SDXL
@visual glade , if you really did manage to convert SDXL into TensorRT, can you make a tutorial on how you did it? I'm sure most people here would love to know.
it'll be an honor to have my new UI personally banned by google
sdxl 1.0 github link plz! :D
oop
oh hey lmao
Plzz theres too many links and read me's I can never find anything

apparently i got granted access to sdxl, no idea what these people are thinking i applied by saying "i dont work for a company i just want to make cool shit" and they actually gave it to me
where could I download the model?
no huggingface link?
what's the 1.0 model name to specify in the API?
people will get around it as per usual
1.0 isn't released yet.
there was just an announcement
you can apply for access for it
yeah bit confusing
ooo i walk off for one second
It's available on github but I don't see the model there?
im confused
I don't find tensorrt to be useful for local inference so probably not
how to download github
Where does one apply
give them a while to put it out on github
is 0.9 1 or wat now
??
no weights lol
thats 0.9
SDXL 1.0, now available via Github
so community gets 0.9, and api gets 1.0, lol
dont spread misinformation
its not misinformation i was just wrong
is it not this?
https://huggingface.co/stabilityai/sdxl-vae
When 1.0 is out it’s just out thought right? No application
If there's no weights then it's not an open model.
why not? it's about twice as fast
404
thats a vae
yo where the link at
how to load github.safetensor
Yes, 1.0 is out!
For everyone. Not just the API
where lol
aah sry
where?
how to specify in the API?
am I dumb or blind but why can I not find the 1.0 thing the github lmao
Struggling to locate this :D
where
Can auto run xl without any config anymore?
404?
Keep in mind all of you are probably destroying the servers trying to get into it ASAP lol!
Stay with us while we try and work on any technical issues ongoing
model download isn't public yet
where is the model I CANT FIND IT
isn't tensorrt limited to 2^31-1 elements which would make the bigger res gens which sdxl wants not doable?
I CANT SNIFF IT
model isnt public
wait like 5 minutes guys
SOME ONE GIVE IT TO ME
worked fine when I tried it
hmm
they changed the 404 icon
github readme still only has links for SDXL 0.9
is 1.0 on huggingface? i cannot find it
API is 1.0, community gets 0.9
you can use it on clipdrop
thank you this is what we want lol
@wicked frigate what about finetuning insturment? will it release also today

breath people.
how do i get 1.0 up and running
Testing my setup in comfyUI on a 1080ti... 95 seconds for base+refiner.
is it not released yet
This is incorrect. Please give us time to work on any issues people are having. Technical issues right at release are not out of malice.
is tensorrt stable?
Can auto run xl without configuration
tensorrt works well but the weights can't be replaces so for local inference it's pretty useless
Will it be located at https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0 when its out?
it makes SD twice as fast? why do you think it shouldn't be done on local installs
my question too
thanks
if they ever add functionality to replace the weights it will be useful
it's super nice for non local inference though
what even is tensorRT
what is the engine name to use in the API? 
isnt this vae the same vae for 0.9 aswell?
https://huggingface.co/stabilityai/sdxl-vae
I just wanna know how to download and add it to automatic1111 🙂
@upper sage all good, but it cant be that hard to make sure something is uploaded before announcing to thousands its out :P
how much difference is there between 0.9 and 1
oh, then it doesnt make sense as of now
0.1
🧌
I can certainly see why someone would feel that way ❤️
about 5 meters
you just switch to the .trt file, right?
bruh
its not out...
The 1.0 weights will be out on huggingface later right?
guys just wait
I'll make sure to share as many links as I can as soon as possible! Stay tuned! 
you need to build one for every single different checkpoint
github page is lacking any up to date info
is it available through the API?
and forget loras
links of?
the announcement says its out on github, I'm confused what it means to be out on github but not be a publicaly downloadable model?
last news on it is from july 4th and no metnion of 1.0 support
yeah
theyre working on it

Switches are still getting flipped. Please stand by and we'll update asap!
does this situation legally allow me to complain on discord
You guys are so impatient just let them figure it out and fix all issues
o ye flip that switch
Calm down everyone lol, it will be here when it's here
They just released SDXL 1.0?
it's still incredibly useful. no one minds to wait half an hour for all their gens to be god speed.
wait, there are still issue?
just asking simple question: is it available via the API? yes or no?
just wait, its not 100% out yet
Kinda there's some issues
only in ur dreams
lol i know. just seems like poor coordination for the announcement. the annoucnemnt pings people who are waiting on it
i dont think so
yep, let's all relax, devs are doing fantastic work and release it for free, let's be thankful
repo isn't public yet
I want it all and I want it now!
🚢
how to download the pci-e plug
The announcement probably got posted a little too soon and they struggling to catch up
where?
titanic?
It was posted at an odd time afterall
if someone manages to download it. lets get a torrent up?
Stay tuned! Everything is rolling out as we speak! I'll pop the links in here as soon as I can!
yup not listed on Hugging face yet either
I hope we don't crash huggingface the moment it gets available 😄
why would you need a github torrent
@daring smelt how does it feel sending out a message of that magnitude to so many people? Like clicking that discord message do you want to send this to 250,000 people? Has to feel like something..,
why isn't anybody using tensorrt on SD1.x then?
RELAX guys 😂
Nothing wrong with being premature 👀
there is a consumer use case which is using multidiffusion to do biiiig upscales, which tensort helps with a tad
🧊
if the site is overloaded then a torrent is a better way of distributing it

I am 
is it available through the API or no? "we’ve finally reached our winning crowned-candidate together for the release of SDXL 1.0, now available via Github, DreamStudio, API, Clipdrop, and AmazonSagemaker!"
could always put it on Civitai for the lolz and watch them melt
immortalize it or something
Whilst everyones here am I correct in assuming SDXL (any version) is essentially the current most recent progress in stable diffusion?
announcement really fell flat. everythign it links to should be ready. i've only been PM on a few smaller projects but that's a big key to communications
speaking of, will there be a new paper for 1.0?
ICE BERGE RIGHT AHEAD
yes they are. I am atleast, when I used 1.5
time to crash hf
if one of the stable staff doesnt come and personally install SDXL on my computer this very moment I am prepared to complain on discord!
thank you for your amazing work
everyone has 1.0 but the community that RLed it 😆
if its gonna be released on github I dont see any chance of it being overloaded, but sure
someone did that but they got banned i think
Like is just switching over to img2img gen (as SDXL comes in two models) just better at making images or something?
i feel like i got enough experience to say this announcement was done incorrectly and all the links it provide should be in line with it
I think it's also fair to mention that Stability AI is a unique company for creating this enviorment :), without @marble dew 's and @hard fractal 's leadership it wouldn't have been possible
I know the 0.9 was leaked there
It happens. Yeah, it kind of sucks, but they are clearly working to get stuff up
absolutely!
i'll come back for the event. maybe things will actually be released then
Let's just breathe and wait for them to update in thread
what I'm looking at to speed up local inference is AITemplate, it's a bit more flexible than TensorRT
does it work with amd cards 
can somebody merge 0.9 + 2.1 to get 3.0
um.. yeah.. the event is to talk about the release which i would assume would coincide with the event.. so why is everyone worked up?
#📣|announcements is the thread people are waiting on
doing gods work 🙏
cheers for the open source-ness of it all, thanks to sd1.5 I have generated countless booba
what's that? never heard of it
I tried asking if the arxiv article will be /updated/ and I got the non-answer of "our team publishes regularly"
Pfft yall are waiting on 1.0.... Iv literally had 1.4 and 1.5 for months now
yeah, but the safetensors is not available yet
these are already in the code, so i presume they're either getting uploaded or are private for pre-release
SD is the blender of AI
i wil follow your career with great interest
lol i remember 1.4, seems like years ago
you only have 1.4? 🥱 i already have 2.1 🖐️
whoa
someone made some AITemplate nodes for comfyui but they keep breaking because it's something that needs to be properly integrated in the core software itself and I will do that when I have some time
now we just gotta wait for SD 3.0
YO
fr tho it genuinely feels like 1.5 and all that came out years ago
Just to clarify, SDXL will be released specifically as a .safetensors
Please do NOT download any files from anyone sending checkpoint files or other unofficial sources!
Stay safe everyone. We're getting things live ASAP!
the github readme says only 0.9 support
whaddabout SD 6.9
how many of these reporters are in the discord. there's already like ten news articles that sdxl 1.0 has been released
how much speed increase are we talking about?
Wow, there are a 1.7 GB Jenna Ortega LORA in civitai , are xl loras use to be that big?
ok but im waiting for the SDXXXL
im selling 1.0 $1 a pop
Thanks for looking out for us!
please pin it
doesnt a 'custom node' system kinda make no sense if its not deved in a way that allows the custom content to do what it needs lol



