#✨|sdxl
1 messages · Page 38 of 1
We have to wait another 30 minutes for the model links? 😢
2 hours and a half sir
grabbing pop corn
It said 2 hours and hour and a half ago
@high skiff join
aslo, thanks to the devs
big respecta mon
the timer goes up not down
we already have 1.0 theyre taking it away from us t
appreciation nation
The real 1.0 was the friends we met along the way.
nothing beats that
also that
the 1.0, the 1.0 IS REAL
Is there supposed to be sound?
prove it
@daring smelt wanna mute this room until #1029055412764422214 is done?
Will we have to wait for 1.5 to get the good one? (Joking..)
stage!!
What's this presentation about?
What's the super stage about? Are they presenting a new model?
We are muting this channel for visibility currently.
Join us over at #1029055412764422214 for our All info and questions will be answered there by Emad and our applied-Team --SDXL 1.0 LIVE Event!
https://discord.gg/nHMkvhsD?event=1130919343941758978
Channel will be unmuted once stage is over! ❤️
OMGOMGOMG, IT'S COMING OUT IN LIKE FIFTEEN MINUTES
welcome back 👀
lets gooo
first
vhttps://stability.ai/blog/stable-diffusion-sdxl-1-announcement
The Stability AI team is proud to release as an open model SDXL 1.0, the next iteration in the evolution of text-to-image generation models. Following the limited, research-only release of SDXL 0.9, the full version of SDXL has been improved to be the world's best open image generation model.
thanks for the stage talk! very informative
nah uh
30 SECONDS
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 push the button
alriiiiiiight !! I'm heating up my GPU in the oven
my f5 button is breaking down!
4040404040404040404040404
🧌
thanks for waiting for me to finish dnd
mood
actually it comes out on the 30th
Post your first SDXL 1.0 image...
now I get to be live with everyone XD
Still getting a 404
404 ✊
15 mins ⏲️
just try again mine is downloading
Stupposed to launch in like 15 mins yall
Hey @visual glade Did you manage to get the Reference Only all sorted to go in ComfyUI?
I have 2 more hours at work
15 minutes
just wait itll take a bit
you gotta believe harder
Question will the base styles be released ? Or can they be released ?
uff ❤️
Models available in 15 mins
Is the stream over?
HuggingFace ❎
FaceHugger ☑️
inb4 huggingface is down
a few minutes of waiting aren't gonna change anything
yes
amazing 💗
Who has the refiner link? Anyone?
WHERE IS IT
Okay thanks
dev said waifu diffusion fifteen times i counted
Weights <t:1690399800:R>
refreshing for 15min gotit
y'all wanna see a teaser?
where did you get that info from
Always
What does reference only mean? for images input?
guys if we all just crash the website maybe we can get it faster
Nice, thanks!
nice work, comfy!
entertain us! we have 15 minutes left to be entertained XD
It was mentioned that Automatic1111 may not be functional - is there an update on this or is it better to just start using comfy instead?
hey you fixed the chart! 😄
So wait more days or?
Nocrypt when? 🤩
I'll put it in a real repo once people actually try it
Please release it
RIP SD 2.1
is automatic1111 gonna be compatible?
No
wer model
I'll have a play with it once the 1.0 weights are released
:(
wait did they not release it?
let the countdown begin
is automatic1111 gonna be compatible?
not yet, it has some quirks but works... use comfyUI
Also no
actually not
lol
so when we getting them
where the model?
The 1.5 version came out today and is compatible, as per the release notes
wen
August 32nd
good question
Waiting for the weights! https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
how can i generate text? what syntax?
thanks!
13 mins

where's the countdown?
Crazy to think we only seeing 25%.
Do you feel pretty confident in the years to come this could get to 50-70%?
Use control net
reference_only if you're talking about ControlNet uses an image as the base, and you can manipulate it.
it can be anything from a pose, to a face, a place, an object, you name it. it's imho a better alternative than lora's in many cases.
@grim brook okie computer restarted
Testing out vlad by preparing it to see how it goes with SDXL, if not good, fallback to comfy :P
HF devops engineers sweating
How did 1.5 get that many votes?
We NEED IT
my f5 will break
is @wicked frigate git link live for install?
how did 2.1?
asked myself the same thing, 2.0 just seemed a bit unstable
the links in my announcement have a timer #📣|announcements message
poor sd 2.1...
anyone know if Auto1111 can use 1.0?
everyone is talking about comfyui but no one is sharing a good workflow for a beginner :/
huggingface DDoS incoming
Release timer is here https://discord.com/channels/1002292111942635562/1002292398703001601
thank you thank you thank you
lol touché
I think maybe just go to bed and wait some days. In bed. No point awake.
<t:1690399800:R>
<t:1690399800:R>
I think it will be with Version 1.5...guess we'll find out. Extensions probably not working yet
Question that can be answered though: What "config" for comfyUI is the most recommended one for SDXL?
They said it couldn't yet but that Auto1111 is working on it.
weeeeeeeeeeeeeeeeeee ❤️ ❤️ ❤️ ❤️ ❤️
This is like edging
countdown party??
this is edging
This is gonna be dependent on what you're making.
what countdown
ya i also just realized that we cant access the channel anymore for the links provided lol
servers at hugging face row are all jet engines rn
Not in slow mode…
So.... You are telling me I have 46GB of VRAM (3090TI 24GB + 4070TI 12GB + 3080 10GB) with Swarm... 
counting down from 10...
10
Not necessarily
SDxl 0.9, hope it still works on 1 lol :<
I've downloaded it.
9
0.9 works well Auto1111 so yeah I hope for a fix soon.
i don't think it does 1 image across cards
Indeed. I am pretty much as square one when it comes to nodes and whatnot, so any really as i have no idea what to add where 
lol
How did 1.5 get 4% of votes, when literally everyone I know uses it?
GET READY <t:1690399800:R>
Where's the teaser?
i meant seconds !!!!!!
mind you, this is vanilla
THE TIMER KEEPS RESETTING 😭
You'll be able to do 1 image per card but u can't do 1 image spread all 3 cars
Base 1.5 is quite bad, most people using 1.5 is using finetunes.
So it won’t work in A1111 right away yes?
tried restarting. still receiving the same error.
shit outputs
It kinda does? But, like, don't.
there's a hack in the works. a release probably later today if voldy is on it
The only reason I got nice stuff out of base 1.5 was with a ton of style words and even then things were difficult.
I see it will finish on July 18th
the outputs aren't very good
Did they say anything at all about what kinda dataset they trained SDXL on? I joined the event late... Just wondering if they censored it to heck like 2.1...
auto1111 was giving me great output for sdxl lol
they said no.
I expect SD 3 out on July 18th 2024 just for the memes.
not in a1111 yet, but SDNext (a fork of A1111) should work fine.
Will it be released in ten minutes?
Not censored
I feel SD.Next had better outputs imo
SOON tm
nice
lol a spam attack is hitting automatic's github rn
for the a1111 folks,
does it work on comfyui plugin on a1111?
yes, there was an update 1.5 it has sdxl support
he'll be busy all day. don't count on it working
Seems to be working
how i download?
Correct me if I'm wrong, but for 2.1 training they reverted the mistaken filtering for 2.0.
they didn't say that. they said, "2.1 had the wrong training settings, so it filtered EVERY HUMAN OUT instead of only the nsfw stuff."
Does anyone know if a1111 works for SDXL 1.0?
ish
not yet
yes it does they updated today
is there a link for the download yet?
i got nothing out of this live-chat today...
why's ComfyUI in the focus for this release ?
@visual glade Minor thing, the preview on the ksamplers doesn't work correctly when using reference only, it just shows the input image, it doesn't display what it's rendering.
Lol and im over here waiting for NMKD gui 😛
no
They fixed that aspect of 2.0, but it was still missing tons of stuff from 1.5
what ab control net fot sdxl1.0?
then why has everyone been saying no for the last 30min?
Is there any change in the model structure?
it wasnt, in fact there was a new UI that was introduced lol
*nogpu — Today at 12:22 *
Does anyone know if a1111 works for SDXL 1.0?
It does, but people are running into issues.
We recommend comfy or SwarmUI.
Is stable studio updated with the integration of comfy yet?
wtf. Why does your PC ... not have .NET? but it has .NET. But it doesn't. It's weird. Maybe go into system applications and fully uninstall dotnet, then reinstall? sorry it's a really odd error i've never seen before
Any info on animating with SDXL?
auto1111 output
all the wait and it doesnt work in auto?
Where did you get SDXL 1.0?
Should work with auto1111 1.5 released 2 days ago. Stable
YOOO
Yes, I love it
Where can I get the ReferenceOnlySimple node for comfyui?
today we made an achievement. Thanks to @wicked frigate we managed to throttle the microsoft download server to sub 500kb download speeds 🤣
it works, just not perfectly
There's always batch output...lol
its like a ddos XD
lmao oops
thanks.
you'll get better outputs with sdnext
I bet the deforum devs have it updated within a week
do i needa
i dont think they answered how the refiner was trained did they
but comfy is best
what am i missing. i find nothing on SwarmUI.
how so
love it 🤣
@bronze arch same as base but focused on last like 30% of steps
@boreal bough I know you probably don’t know but sense I’m at work is stablestudio updated with the comfy nodes yet?
so... is the refiner just doing an img2img?
what? like whats the training data tho
i see comfy finally buckled and added support for reference, nice
i know.....super odd....i'll fiddle and report back
check the github, downlaod latest master...or update through git...depends on your installation
what does the reference node do?
Was that training setting intentional, or was the AI giving us a sign? Should we be worried? It could be taking first steps towards SkyNet.
yes, in a branch i think
@bronze arch i assume the same i guess
Okay cool
yep, but it was buggy earlier. He did say it will update right after release. so prob will be working by the time you're back?

i dont get it, the output is so different tho lol
Idk, man. I have had bugs opened in Deforum repo for months with no fixes. I'm not sure if anyone is even actively working on it, anymore. There hasn't been an update in a month or so.
Okay sweet thanks caith!
So SwarmUI has a/b/ .../x/y/z support now?
Can’t stop hitting reload on https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
5 minutes™
yep
Comfy for img2img and node base=blahhhhhhhhh
<t:1690399800:R>
Gives an image to inspire the image generation, in theory
damn why did I never prompt for deepdream before, this is kinda sick
oh, the control net reference?
Yeah
ah the good old times
can someone prune it
Has there been made any improvements about the refiner for img2img upscaling?
Guys I'm scared HF will just crash lol
Does SwarmUI support prompt editing?
I remember that. I started there in 2018....
Would be tragic if so. I would keep using it with 1.5 if so....but i feel like they'll get it going
@bronze arch i assume cause its trained on the ending steeps, its not designed to output from raw noise at all
Balder Gate 3 in 1 min???!!!!
Only the website will crash, we can probably still clone it
3️⃣
im gonna take off slow-mode temporarily in the last minute 
comfyui is easy. just gotta keep it simple like this
when are we expecting it on HF? sorry if it was mentioned already
then what does it start from?
ok I have normal SD at this time, as a total none PC geek, what is this COmfy and stuff? And is SDXL having an easy update option to go from Standard SD to it or is this a totally different program. I am completely lost at this point...
2️⃣
why dont we elect like 1 person to download from it, then upload a pruned version
This is gonna be epic
slow down
comfyUI is... a UI. user interface
So does automatic1111 UI work with SDXL out of the box?
I just installed ComfyUI today and this made me chuckle
they said its pruned but i assume it can be pruned more?
@hard fractal can we look at your comfy workflow zooned out?
its node based like chainner and lacks alot of features Auto111 has
If I open all my nodes my workflow is way bigger than that lol
not optimally, but yeah
Yeah, they have some great devs, but it would be nice to have a few more. It's been my favorite extension in SD since Jan
SDXL is a new model, that's vastly better than the previous version sd1.5 and sd2.1
it's art - ancient node tree
2️⃣
base comfyui has way more features than base a1111
ya comfy really needs some concepts from good node systems like blueprints in UE etc. Gotta be able to group stuff and 'collapse' it if its gonna take 5 nodes for simple stuff etc
"The Final Countdown" by Europe
Listen to Europe: https://Europe.lnk.to/listenYD
Subscribe to the official Europe YouTube channel: https://Europe.lnk.to/subscribeYD
Watch more Europe videos: https://Europe.lnk.to/listenYD/youtube
Follow Europe:
Facebook: https://Europe.lnk.to/followFI/facebook
Instagram: https://Europe.lnk.to/followII/instagr...
is it out? 
What do you mean by not optimally
comfy is good but complexy
inpainting,image2 image,drawing?
at most it will be a fp32 version on huggingface, fp16 would be half the size but that's not 'pruning'
whats the best ComfyUI workflow? I downloaded a few recommended ones but they are blank
the tree of life
comfy in a nutshell
pretty tree you got there. does it grow a few images every year?

fingers crossed for fp16 on official repo
oh i see
Oh, if the Git is to be believed then SwarmUI doesn't have LoRA support. Alright, I'll stick to Auto1111 until then.
supposedly the optimization causes it to produce worse results, but it should work
0️⃣.5️⃣
Would be nice to be able to run LLM like LAMA2 and others within Swarm have one machine that feeds prompts.
One second.
it's time to party!
1 min
O
M
G
I
T
'
S
C
O
M
I
N
G
SOON ™️
1 MINUTE
60 seconds left.
50 seconds
oh my
heh
3️⃣
45
LFG!!!
its happening 
@wicked frigate I have further questions. Which do I uninstall to ensure it doesn't break anything else that I use?
I can't believe it's nearly the 18th already
StableDream
40 seconds
Ladies and Gentleman it has been a pleasure
SDXL1
Generating Images
Really Super Good
- a haiku by flowwolf
8 days to Balders gate 3!!! whaaaaa
aaaaaaaaaaaaaaaaaaaaaaaaaaaa
30 seconds!
Lets gooo
MUCH EXCITE
its happening
20
20 seconds!
R
E
F
R
E
S
H
R
E
F
R
E
S
H
R
E
F
R
E
S
H
R
E
F
R
E
S
H
R
E
F
R
E
S
H
instant crash epic fail
<t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R>
LETS GOOOO
15 seconds
15
🎉
aaaaaaaaaaaaaaaa
10
I guess the Link will be posted when the update is ready I just hope my PC can handle SDXL
<t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R>
now we talking
inb4 huggingface is down
❤️
gooooooo
5
I'm downloading it! YAY!
AHHHHHHHH
<t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R>
😮
3
let's go
AH
2
AH
MOOMYYYY HOLD MEEEE
1
<t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R>
how to make it f16?
<t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R><t:1690399800:R>
GOOOO
<t:1690399800:R><t:1690399800:R><t:1690399800:R>
0

ITS UP
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0
https://github.com/Stability-AI/StableSwarmUI
🚀
Why is the download speed so slow?
liftoft

It's Live, UNLEASH THE HOUNDS
lulw
YES
💥
IT'S LIVE
i see it
🚀
woooohoooo
Wait, are we DDoSing HF?
GOOPOOO
Its up
Yaaaaaaahh!!!
GOGOGOOGOGOGOGO
ARGGGGGGGHHHHHHHHHHHHHHHHHHHHHHHHHHHHH
First download!
Only 7 GB 👀
YES
It works!
XD
🤪
lets see if huggingface dies
lets go hug huggingface ❤️
yes!!1
ATTACK THE SERVERS
They killed refiner 
It's collapsing!
View from HF servers rn
please don't die servers
Downloads!! 🎉
BABABABABABAAAAAAAA
it works
same config as 0.9?
DOWNLOADS GO BRRRRRRRRR
Whooo!
... okay wtf why is it broken still when you have so many dotnet versions installed?! uh - theoretically just the 7s... or all of them and just note down the versions (3, 5, 6, 7) to reinstall after if things break
BRRRRRRRRRRRRR
what is offset lora model?

everyone STOP DOWNLOADING so i can download
vroooom
at last
damn bro yo internet
YEAAAAAAAAAAAAAHHHHHHHHHHHHHHHHHHHHHHHHHH
Feels like Christmas ngl
Who can beat my record for slowest download? 
100MB/S Yeeehaaa
lol.. you guys killed the throughput
lets gooooo
❤️ THANK YOU!!!!!!!!!!!!!!!!!!!!!!!!
13 minutes, lets goo download
it's a lora that adds offset noise to the model when applied, so you can get darker darks and brighter brights
i know i s lagging today. only at 2.1MB/s now
my internet is sooo slow aargh!
The most beautiful 6.94 Gigabytes ever.
+1. Also seems like they had an update today about 5 hrs ago.....
I have 4 minuted download
should i feel bad
31 mins remain looool
40 sec
🙏🏿
SOMEONE MAKE A TORRENT
envy some of these download speeds
Now, let's test A1111.
where is the refiner?
I am honestly confused, they aren't releasing a 32 bit checkpoint? Not that I care, Im sure it could be made from the other files
Cross fingers!
SDXL 1.0
Base:
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0
Refiner:
https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0
Github:
Offset LoRA:
SDXL 1.0 is also available on DreamStudio, Clipdrop & API:
File Hashes:
Base: 31e35c80fc4829d14f90153f4c74cd59c90b779f6afe05a74cd6120b893f7e5b
Refiner: 7440042bbdc8a24813002c09b6b69b64dc90fded4472613437b7f55f9b7d9c5f
Tensordata Hashes:
Base: 0xd7a9105a900fd52748f20725fe52fe52b507fd36bee4fc107b1550a26e6ee1d7
Refiner: 0x1a77d21bebc4b4de78c474a90cb74dc0d2217caf4061971dbfa75ad406b75d81
fiber bro
WE UP
Can we do the countdown again please, but instead aim for when the download finishes? khtx
we cookin
Nice. It works on A1111.
Yay its out, will it work in A1111?
Thank you.
QUESTION: Is this an update for Stabile Diffusion or something different?
thanks
u dlin slow
Neato burrito
Show examples on auto pls
1.5.0, if anyone is wondering.
Go play with the base, and try the LoRA!
Poor CDN
Nuh uh
I want to see some hot images. Like HuggingFace servers on fire kinda hot
i got the base done
anyone want to explain to me why i would need the refiner?
why is it FP16?
oh yeah. i forgot to account for download time. 1 hour
can someone show us the pipe out at huggingface?
What is the lora for?
to be one of the cool kids than can fix hands in weird positions
Lora's training has started
Time to spam images in here to lower everyone elses download bandwith
I'm happy we haven't crashed Huggingface.
Trying colab hmmm
scam. you can see the fake download buttons in his sleeves
for lora, that one is hers
Any official workflows yet?
i love the weird hands. is it in the same folder on huggingface?
where do i put the refiner lmao? do i put in the checkpoint folder?
looks like its already pruned?
Do we also need to download the VAE for the base and refiner, opr are the VAE already included in the model?
no, it's FP16
oh boy, thanks
They said the VAE is included
downloading as the gods intended it
Downloading base 1.0 AND refiner 1.0 ...
SDXL been out for 5 minutes and still no dreambooth trained models.. I'm disappointed in you people
Ok loaded on my auto colab let’s see
cause majority of people are gonna want to use fp16 i guess
.NET or ASP.NET or both? what about the Microsoft Windows Desktop Runtime?
@hard fractal anyway we could get a LORA config you guys know works well ? The trial and error of that is really tough on gpu time :P
so where do we find the DL for SDXL and is it an update of SD or a different program please?
Got the files!!!
the diffusers version has an FP32 unet at least, so can convert that, but... training/finetuning is going to need FP32 so they should add that
prodigy. lr 1. bf16. rank whatever.
sdxl 0.9 was 13.9gb, sdxl 1.0 is 6.9gb
--network_train_unet_only
It doesn't, I did plenty of training on 0.9 at FP16 and it worked like it should
1.0 is pruned, i am not sure about 0.9
it's pruned, 0.9 was fp32
not a 'bad'thing per say @hard quest , should be possible for more ppl to train it now
y'all using it in comfy or 1111 ?
ayo only 7gb
works in 1111
6.9 GB nice
i remember "pruning" 0.9 from a1111, and one of the models halved its size but the other did't, forgot which
[Info] Creating new admin session for ::1?? ComfiUI
Testing out vlad, but no idea how to use XL with it lol
@proven crane hi 😂
with fp16, it should run on all 12gb vram cards now
i think i pruned and converted it to fp16
@trim orbit not super into loras, so that is tough for me to interp into a config :P
comfy is less finicky an optimised
and nothing
prob lower with optimization
sharex extension for a1111 when?
This is what matches the streamlit exactly, but we know it's not the best best:
Team auto
which model should we use in auto ?
base or refiner ?
we are using comfy here
are there any resources on how to use 1.0 models with it ? can't seem to make it work out, getting lots of red text lol
base on text2img and refiner on img2mg
What is this comfyui? Anyone got a link?
😦 100Mb internet isn't fast enough 😢
Thanks buddy, appreciate it
Base to start with. Refiner is an optional second step
hype
up and running
what is that
what does office hours meeting talk about(MJ)?
comfy works best when you
thanks sir, when you say second step please elaborate
says you had an error and the data stream was cancelled
free sdxl 1.0: https://nogpu-webui.com
Experience the power of Stable Diffusion XL with our free SDXL online demo. Try Stable Diffusion Cloud, a reliable alternative to RunDiffusion. Explore the capabilities of SDXL now.
cheers. thanks for the models
you sure you're not playing factorio?
does auto111 work with it or comfy?
Second pass to add extra quality - not needed for normal generations however
already works
factorioXL
when is SD 3.0 releasing?
anyone test a 0.9 lora on it yet?
could you break that workflow down? 
i'm stoked to see what the expansion pack will be

i hope its like a node graph!
Tomorrow
So to install SDXL 1.0 to Automatic1111 you just have to drop in the base model in the stable-diffusion models folder (with the 1.5 pruned checkpoint)? Or is there a token thing to download?
If you run the refiner on the last step of the generation it shouldn't really take much longer if you have 32GB VRAM, so if you want photorealism you will get better results with it.
Understood, Thanks for information.
and thanks for SDXL ❤️
is there a guide to using 1.0 in automatic
Have you updated Auto1111?
When is SDXL 2.0 coming out? Typical stability team with slow releases
ohhh
At least with 0.9 it was just to drop it into the models folder yes. (Just remember to put it in a seperate SDXL folder so you can easily organize your models)
December 1stn't
lfg
oof, running out of memory on a NVIDIA 3080 10GB GPU with SDXL in automatic1111
ye, you want like 12GB for refiner
do u actually need the vae?
indeed, just drop to sd/models, and auto1111 ver 1.5.0 minimum
no refiner
My LoRA still works! 😄
feels bad man
So how does making loras locally work? I previously used kohya and I have 12gb available
Auto colab keeps crashing when selecting xl lol
Use comfy you won't run out of VRAM
😂
whats the best comfy setup?
Download is faster on my phone, I'll have to transfer to my pc 💩
Awesome I'll try it out and see, thank you
yes for my workflow, but u dont u can use the lora from base
I go to the website then have a mental crash as I can not find any link specifically to DL and install the SDXL. Guess I will try later when folks are not so busy...
can someone guide through the entire process of installation of StableSwarm?
I see the safetensors - where do I d/load the diffusers at all?
if im using the refiner do i also need to use vae
you should only need dotnet-sdk-7.0 for it to work, that implicitly includes everything else. This PR raised a similar issue that implied asp might not be including properly? https://github.com/Stability-AI/StableSwarmUI/pull/6
Same VAE in both models I believe
It's not really "censored". Naked bodies are generated fine. Quite surprised, actually. Realistic textures.
see I have no idea what a SWRM or Diffusers are. Totally lost here at the moment..
ohhh ok

on what folder do i put the refiner?
I've been saying this for ages, but the people asking won't care, they want it to make porn
Reminder that NSFW discussions are not encouraged on this server #✍🏼|rules-and-tos
Diffuser d/load link?
How do we get started locally? I can't clone the repository 
Investigating
no need to clone the whole repo
Hoover dam on 512x512 looks like a circle.
seems like there were no major changes to workflow so my workflow still works fine (if you use this change your batch size to be lower unless you have 24GB VRAM)
launching with --medvram worked
@wicked frigate app crashes if you have a space, in the folders leading up to the install. Your starting script didn't use quotations for comfyui
@thorn lion did your reboot work?
Mine works, until I try to add the Lora
SDXL1.0 on AUTO1111
Nope
how
zzzzz
Downloaded. Installed. Running my first generation in A1111 ... aaaaand ... (drum roll) ........ 😐
@primal nimbus just works
automatic1111 doesnt have refiner support right? is vlads version working fine?
🤌
first pic is already nsfw lmaoooo
do you select the base model or the refiner
@z u cant do 512x512
what res?
Probably using wrong VAE
Resolution too low, needs to be 1024x1024
u dont need the refiner
you got rick rolled
that's a 512x512 image, SDXL can't do those
OOOOOH!!!
btw, clip skip 1 right?
0.9 refiner felt a bit animated
my image above includes metadata, you can just drop it in auto1111, enjoy
1.5 is still king
@thorn lion ok I had to delete my old folder then did these steps and it DLed and installed no problem after that.::::::::::: Here is a quick guide to install automatic1111 webui (Stable Diffusion)
You need to install Git 64bit: https://git-scm.com/download/win
and Python 3.10.11, 64bit: https://www.python.org/downloads/release/python-31011/
And check "add python to path" when installing Python.
- Make a new Folder on your drive (not on Desktop, downloads, documents, Programms, onedrive) and name it Ai for example:
C:\Ai\ - You go into the folder you created in this case Ai, then click in the Explorer bar and type cmd then press enter.
- Then you copy and paste this command:
- git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git
Press enter and after its done you can close the cmd and launch the new webui-user.bat
what is the VAE link again?
Where do you put the refiner model in ComfyUI? The same folder as the base model (The models folder)?
nice, anime became a lot better ❤️
You don't need it, the new VAE is in the model
1 1024y1024 image takes my 5700xt with base model 2 minutes 😄
why...what...i'm install StableSwarmUI
Yes
or 4
stableswarm faster than a1111? can anybody tell
Ah Painful and Lovely at the same time
Yeah, I'm an idiot. Ok. Running it again at 1024x1024 on my 3080 10GB .... aaaaaand ......... (drum roll) ............
@thorn lion ACCKKK sorry wrong one them I have no idea what STable swarm even is totally new at most of this.
base or refiner?
can't get stableswarm to start, dotnet install bugs out, weird as it is
Yeh is vlad working?
base
maybe, but no amd cards supported on swarm
the ui that was released today by @wicked frigate and Stability AI
Has the VAE changed at all since 0.9?
Works on a 3080 10GB and 16GB RAM, with refiner. Seems marginally faster as well.
I just tried putting the SDXL 1.0 in the models folder for A1111 but I'm getting a runtime error, is there any fix for this or potential fix? It just switches back to the default 1.5 pruned model after a few seconds, same happened for 0.9
only in auto1111, right?
did you add this to the webui bat "--no-half-vae"
ok srry
correct
comfyui...?
i did not
@thorn lion so that is not an update to the normal SD??
Please don't share unofficial files. We do not have the resources to verify them right now. No ban thou since you were kind enough to ask 🙂
someone had asked in this chat before. your reasoning makes so much sense that i didn't consider
Hello everyone is SDXL supported in A1111 ?
Can you please throw at me some examples of text generated by SDXL?
add it and you should be good. got it working for me
fek wait which script
any idea where to get the proper workflow for SDXL1.0 on comfyUI?
0.9 vs 1.0 same settings
Ok, sdxl is good, but where is distilled weights? Or its in processing?
yes
prompt?
indubitably my good man
payphones always seem to have a second thing attached for some reason
How do I d/load SDXL 1.0 Diffusers? 🙂
And somewhere you can see how to make it work correctly?
thats way better imo
controlnet?!?!?!
both are awesome but i like 1.0 a lot more
woah 1 is so much better
Yeah I used to use that but I've liked and been using Automatic1111 cause it has directml working so it takes a lot less time
The Lora train with 0.9 works with version 1.0
Its a LoRA of myself along with "painting of a knight with a beard"
just update it, then download the models like every other model
just auto1111
Nice to see more examples were 1 is superior
actually, that's a huge improvement on detail
in that case, is it too much to ask for an official torrent source? or are there too many considerations/complications that make that request more complicated than i imagine?
How do I run this in on M1 ultra?
174 seconds to.load 💀💀
Automatic1111 use: Download model, place in model folder, and use it like every other model.
@thorn lion throw me a friend request if you want and maybe you can explain some of this to me later when it is not so hectic. I usually do voice as I have issues with my hands due to nerve damage at brain stem as well. Spatter
its weird how you can tell this is a real person, vs random generation
Everything working in Comfy
hmmm. downloads failed. retrying picks them up after, but thats sus. i'm gonna have to hash check these guys
@eternal fog can you help me install swarm?
got this error
it really has trouble making payphones, what happened? this looks like the dall-e 2 early days
what workflow is everyone using on comfy
I'm not using Swarm and I haven't looked at it yet.
oh. did we finally crash microsoft? 🤣
several users are having that and I'm not sure why - something's installing wrong with .NET -- you should be able to install manually from https://dotnet.microsoft.com/en-us/download/dotnet/7.0
Automatic1111 support yet?
Anyone know if vlad works with SDXL?
What does no half VAE actually do? I've seen folks using it, but I've never needed to in the past. Does it degrade quality?
microsoft crashed XD
@random bay ya my outputs are all auto1111
just keep your head in the chats and you'll catch up. i'm a little behind on the inference side of things so i have to just the same
perhaps wrong sdk/runtime in script?
Where do you put the Refiner?
whats swarm UI?
do we need a specific vae or is it baked in?
no idea just what i read it needed to run
DEVELOPERS!
Has anyone seen this
why is the first image so small lol
server genuinely didn't survive the influx of users downloading it 🤣
it was already slowed to 500kb
The contrast is weaker
i literally booted up a pure vanilla windows 11 install in a VM and ran the script to verify it worked .-.
just now, you showed me.
sure, make sure to update the repo
@dire locust ya sdxl1.0 wasnt trained with as much offset noise
hmm, failed for me, perhaps having some dotnet variants installed already screws it over? 😮
OH FUCK I FORGOT A PART AAA
I don't know if it's Comfy or SDXL 1.0, but this seems much faster than 0.9
that's my bad hol up
cause ppl didnt know how to adjust it in LORA training so they just wiped it from 1 🤣
Did terminal SNR win?
@thorn lion Back in my day we had DOS 3 to 7.0 then C++ after that it has simply moved faster than I can keep up so I have become more of a end user than anything and it still boggels my mind at times.
where's my "i"? 😄
it does seem a lot faster for me
when do I use the refiner? using auto1111
IIRC it basically disables one way of compressing weights (simplifying), because some VAEs don't work with it, and it corrupts their weights if you try to use it
probs comfy, they made a lot of adjustments, the cuda-malloc thing was huge
Do I run SDXL1.0 using ComfyUI on my Mac M1 Ultra? will it be able to utilize GPU?
0.9 vs 1.0 Longer prompt: "Deep within the vast, starry expanse of the cosmos, an intrepid space traveler floats alone, his visage transformed into a macabre, skeletal mask. The skull's empty eye sockets seem to stare out into the abyss, while the jagged ridges of the cranium appear to radiate a faint, otherworldly glow. In the distance, a trail of glittering debris streaks across the velvet blackness of space, a haunting reminder of the unforgiving nature of the cosmos"
I'm running on M2 Max 38 core
It's time...
although i like sd0.9 more here, i suppose 1.0 adhereed to the prompt better
Previously it took like 40 seconds for one SDXL 0.9 image
thats deadly
yeah I was taught my educational principles by people from "back in the day" so I go for a level of understanding and mastery that most don't aim for and....it's really tough to keep up. I have my youth to help me keep up but it's tough for sure
i feel like there's a little artifacting but idk if it was there in 0.9 too like everything i gen is having some lines of artifacting like that
comfy?
@smoky patrol i think its better to add a channel #sdxl0.9vs1.0 channel so people can share their comparisions
first few simple prompts. loving it.
can see it like here too
wtf light mode
Yeah this is so much faster, it's like 60 seconds to do a whole high-res fix chain, it was about 160 seconds before.
If you're thinking of training on SDXL, first try prompting, it might just be there already, this is how hyped they are about SDXL 1.0. "SDXL’s improved CLIP model understands text so effectively that concepts like “The Red Square” are understood to be different from ‘a red square’. This accuracy allows much more to be done to get the perfect image directly from text"
Chat is this true?
@toxic vortex that is a hot sonic bro lol
can i use SDXL in automatic 1111?
i too am old enough to have been forged by the terminal. My first being very BASIC (hehehehh)
@high elk Hey I still use command prompt and dos at times in windows 10 to get around some issues. LOL
':\ComfyUI_windows_portable\python_embeded\lib\site-packages\torchsde_brownian\brownian_interval.py:594: UserWarning: Should have tb<=t1 but got tb=14.614644050598145 and t1=14.614644.' does anyone know what this means? I had it for 0.9 too
can i use comfyUI with amd ?
You guys use any negative prompt?
So sd 3.0 doesn't exist, but SDXL is essentially 3.0?
Hi, I'd have two questions:
Will this model run on 6GB VRAM and 16GB RAM? Do I need a refiner and what is it please?
Confirmed. Won't run on my 3080 10Gb in A1111. OOM Error. Tried with Xformers and No Half Vae
yes, i don't know for sure but it seems like comfy (the author) also uses amd
SDXL is based on a slightly diff arch
Was the event recorded ?
well I'm gonna give a pretty confident estimate that my face is not in there already and, if it is, i'm a lil scared
Use comfy
but i get error no nvidia card found xD
I had on rtx 2070 and it ran with medram
medvran or lowvram, try both
Comfy said it's just a dev warning and can be ignored.
NoVideo
@trail bay they said it was and would be available later
Launch with —medvram
any tuts on using with A1111?
no.
I remember seeing that SDXL uses both RAM and VRAM, but can I make it use just VRAM? I have 64gb vram
only comfy fans here.
you need to add --lowvram or --medvram
womp womp
Ooh okay. So why use sdxl over normal sd? Not a rhetorical I just don't know
do we need to set up certain vae etc?
If you are using Comfy you can use the --gpu-only launch param
you just put the model in the models folder like any 1.5 and it works for auto1111
Where are these links located?????
uhhhh it'll be 20more minnnnn
btw i use a1111 what is better about comfy?
How were the votes in #1100484581037195384 used to improve the model?
yep how to use it with a111
better
Here is my last 0.9 vs 1.0 I dont wanna spam
download speeds are never fast enough

Any idea which branch of kohya i should use to train 1.0? also any documents?
SDXL may be out, but Marge Simpson as Mona Lisa is still... well, a bit off.
i had horrible results with amd and 16gb vram on linux with 0.9 using highvram flags, but medvram or default made it possible to use just fine, and it was kinda quick (for 1024x1024). with 1.5 i can't even generate 1024x1024 without running into oom
Does the model fit on the free tier of Google Colab?
how big is .9?
my 0.9 LoRa works on 1.0, and i think it looks better on 1.0
I guess SwarmStableUI is just broken
so how does this thing work? Can't put text in there and it only says undefined
Maybe it's b/c I have a learning disability, but for some reason Comfy makes my eyes cross and my brain melt. I can't wrap my head around it. A1111 was simple. I have very specific workflows and no clue how to replicate them in Comfy
well could be the seed, but also the noise offset now being reduced
gud shit tho ❤️
what is the freewilly thing ?
literally everyone except 3090 users crying rn
lol are you saying that an RTX 3080 is not enough to run the model, and I have to make it run on my poor RTX 2060 lmao? This thing is about to burst in flames
@hard fractal were there plans to release what you all learned from training/rlhf? I think one of you mentioned a big writeup on practices and such?
llm
Custom LORA trained on 0.9 (style)
LLM refined version of Llama
a large language model by stable
look up a guide on youtube
dumb question time, trying to update a1111 and it gives me error: Your local changes to the following files would be overwritten by merge: requirements_versions.txt ... how do I fix that
i think that's what StableSwarmUI is supposed to help with but I wouldn't be able to tell you till I get it installed
@deep dagger ok as you seem to have it working a basic ignoramus question here: Do I update my current SD with this or is it a totally new DL and install then I get rid of my old SD folders?
Does sdxl work in automatic1111?
👋
if you have a motherboard that supports resizable bar it can overflow into ram but it will be slow as fuck
hey mate 🙂
FreeWilly is a local LM that beats ChatGPT
Comfy is great
are you using the offset LoRA in there?
Nice to see that the the base model size has shrunk.
great. Is there a civitai link?
@hard fractal
fp32 version for finetune releasing later, or finetunes only via 0.9 + special contracts with sai?
also RTX 3080 CAN run the model what you talkin bout 💀, you using AUTO1111?
So never available to average 2060 users?
its a completely diff arch?
@hard fractal the lora was trained on 0.9 with the same offset as 0.9 was trained on ya
The base model works at least, but I have no idea if it's automatically using the refine model too
well i just downloaded the safetensors file of the sdxl 1.0 base and placed it in the model folder of a1111. tried to select it as a model and it doesn't load the model and just kicks me back to the previous model i was using.
thanks
Where is the refiner model ?
Any mac users here? comfy is better for running it? and how do you use m1 gpu?
@hard fractal Thanks for your work and releasing it, it's very sexy.
update it
1.0 worked well for me in Auto1111 and gives better results than 0.9 so that's a win.
git checkout requirements_versions.txt if you want to revert your changes
else something like `git commit -m 'whatever' -a; git pull --rebase'
well in the guide he just clicks it and another window appears, not for me
you guys gotta check out the pin tabbed
Auto1111 apparently uses more VRAM or is less efficient, with ComfyUI you can run it with only 6 GB (ampere, idk bout 20 series)
Refine model?
Thank you Stability Team
Good work, I can't judge 1.0 based on one prompt but it looks like an improvement despite my experiences with the bot.
fixed in new commit
Any output in the server logs? That happened for me as well and it turned out to run out of VRAM.. had to change launch flags to include --medvram
Okay, I'm cautiously removing the timeout from @high skiff
totally agreed
hires
anyone else getting size mismatch for model.diffusion_model.output_blocks.8.0.skip_connection.bias: copying a param with shape torch.Size([320]) from checkpoint, the shape in current model is torch.Size([640]).
?
@me if he acts up again.
i run a 7gb sd1.5 model with 4 controlnets on a 12gb consumer GPU with Auto1111, seems fine
on the topic of the offset lora, the name irks me. sd_xl_offset... ? i changed it to sdxl_offset for huge wins. now it goes faster
1.0 is better than 0.9 on automatic1111
@wicked frigate wtf.......okay i'm gonna try uninstall all versions of .NET, ASP.NET, and the Microsoft Windows Desktop Runtime and then rerun install-windows.bat to see if that works
Oh, awesome! Pulling
I am back now
your outputs are having the same issues i've been seeing crop up in mine, there's like lines of noise still left
how do you get it to work im getting errors when loading it
getting 34s/it on auto1111 3080 and that is not a typo
so just add git checkout requirements_versions.txt to the bat?
testing 1.0 at the moment, very very very interesting results
@rough harbor yep I go there but can find no actual DL link to use and half of the stuff on the page is like me trying to read chinese and I do not speak or read chinese. Sorry I am simply a old back country DF trying to just stay in pace with the world these days. Any assistance would be greatfully helpful.
welcome back 🎊
@high skiff ping it came out
Isn't the whole point of sdxl that it combines two different model? Something like starting out with a lower res image using the base model, and then upscales it with the refine model..
can't figure out what's causing that
It fixed! thank you sm
i updatedf and now have ERROR in red letters above the model select
bruh use ComfyUI, I literally got better speeds on 6 GB lmaoo
wherever you're using git pull
lets check my rev animated lora :P
how to use it on automatic? my collab crashes everytime when i start it
base resolution is 1024 correct?
1.0 listens way better to things like "cinematic"
but I have a question for @hard fractal
A genuine question
Comfy is so much faster, its crazy
What's problems ? maybe you can DM to me or go to other channel
@hard fractal i think its better to add a channel #sdxl0.9vs1.0 channel so people can share their comparisions
Wildy stupid question here, but... can Swarm utilize multiple GPU's that are in different PC's - somehow?
For example, my wife and I both have 3080's on the same network

