#✨|sdxl
1 messages · Page 34 of 1
i like this dedication, i will vote for you... even if its the worst image out of all of them. You have secured my vote.
yeah, wanted to shout. bot = dead
no, the ones who submit late studied all the other entries, and then make an fitting image that is contrasting, higher quality, and is more captivating towards the audience
that's what i did to score high usually
Same here
what do you guys do if none of the bot options fit?
i vote to the better looking one eitherway
in a general sense even if both are shit... 1 has to be better for a better direction for the model
but the issue with sdxl is... ITS SPITTING OUT FIRE IMAGES
its hard to pick which is best
true
sd 1.5 walked so SDXL could run
yeah, current bot seems miles ahead of sdxl0.9
Sdxl is out tomorrow?
we can only hope]
Supposedly
SDXL beta was meh cute... current SDXL 1.0 is on illegal steroids
SDXL 0.9 is on legal steroids
What are the benefits of xl?
Better everything
?
still very pleased with this one ❤️ sdxl 0.9
you mean like its super good?
I have 12gb vram, can i use it?
Yes
yeah, with comfy you can i think
it understands you more, it reflects your ideas better to you, its much more vibrant and smart
absolutely
I've heard 12gb vram is the minimum
like getting a good therapist
nice
I need one of those
24GB nice
I have no computer 😦
got a used 3090 for 650
€750 here
how are you on here then
I bought at release to upgrade my 979
confess
970
Well i have a laptop, but it only has integrated graphics
Yeah sd runs on m1 or something
online services, colab, clipdrop
honestly.. i gave up on colab.. Leonardo Ai was good enough
but SDXL > leonardo
its not even close
Well, I'm excited then. I should learn how to edit. Might cull the mutants
Oh I love this pic lmao
sad cat really killed the bot
very nice
||i just wanted to see how the spoiler thing worked||
||i just found out about it lmao||
Oi, don't dig up me memaw like that
||sad cat is taking prompts now||
||
||
||they drop a bomb in oppenhimmer||
Yeah no shit
what you talking about?
I was replying to this

Nevermind
pretty!
(╯°□°)╯︵ ┻━┻
Wow, I like it
double exposure tag is waaaaaay too good to not use 😮
Can you tell me the prompt please
Wanna try it
Xd
got the full comfy + sytan set up?
L: abstract lineart, universe and city
N: low quality, badly drawn, mediocre, average, dull, boring, worst quality, bad, flawed, ugly, deformed, disfigured, misplaced hand, misplaced foot, ambiguous, incoherent, abstract, signature, text, jpeg, scan, sad, sullen, tree, flower, butterfly, angry, bad proportions,low contrast, disfigured, misfigured, missing limbs, missing hand, bad anatomy, alien```
Awww, no, still using sd collab ui xd
Ty, wanna try it later
can someone help me edit
because i don't have idea for where start
is that OG SD 1.5 greatness?!
you never know David Lynch might prefer this model
(sdxl 0.9 local install via A1111)
aight, imma go to bed, else i'll never stop pressing that generate button 😛
couldn't get tensorrt to work with SDXL =[
That one looks fantastic
Ooo neat
yeah, but it would probably look the same but be like times 2 more fast with tensorrt =[
Well when you figure it out can you help me with that as well? A bit more speed never hurts
yeah, but I doubt I'll figure this out. I only managed to get that black magic to work with 1.5 models. ~80% increase with CUDA stuff
What exactly are you working with.. Is it an additional download? an extension? or something else in the python code?
Nvidia's optimization. it's incredibly fast.
Bot not responding, I have to live without sexy Kermit images for how long!?
where would you even put that to apply it to this installation?
that sounds like something a developer would need to figure out.
👋
oh yeah, I did manage to get this to work with 1.5. you put the tensorrt folder into your unet extensions folder, then you convert your model to ONNX, then to TensorRT, so far this only works with 1.5 models with an insane speed increase
other users also reported it to work that way.
I am kinda surprised it works with previews and everything
That sounds really good... Hopefully it will work with SDXL soon.
awesome blender shader workflow
lol
I get some horrifying things typing gibberish into the prompt and setting cfg to low values, one of the images I got once has been burned into my brain for weeks
I used to do all sorts of node materials in blender
yeah, I know. I honestly don't see why shouldn't it work, TensorRT doesn't effect precision or output quality in any way. it's just incredibly fast. I think this should be implemented by default
and all the unity insanity
Tensort RT limits the resolutions and batch sizes you can generate at
its an optimization, and it takes a long time to run and uses a ton of disk space
you can set the max res to high enough without any issues with 1.5
I used it all the time before I switched to SDXL
We are seconds away from the 26th...
i wish i could take off work tomrrow for sdxl 
all I know is I messed with tensor RT for a few days, and at the time, the most you could do was optimize a single 768x768 image gen
any more than that, and you were above its range compatibility
🎉
I'm just going to stick to this workflow for a while... It is working great...
weird, when I still used my old model with TensorRT it worked with 1024x736 insanely fast, WAY faster than without it
I bet if we would tweak the inference enough it might just work with SDXL, this will make SO many more people use it, because the requirements will be lower.
so glad you like it 
Its awesome... Thank you @high skiff
wasting disk space>wasting VRAM
no problem, just glad to see people actually liking my work
it shouldn't use any less VRAM
ah, so just speed? I never get any CUDA errors so idk
supposed to be just speed from what I remember
it's still way worth it imo. it adds ~80%<= it/s
this potato looks like a mango
Maybe we can work on some more advanced workflows once different things are made for SDXL 1.0... LoRAs, embeddings, etc
thats the goal
I am not planning on spending more time on my workflows until 1.0 comes out
That makes sense
Don't wanna waste a bunch of time for SAI to just decide to change something
yeah, also getting TensorRT to work with SDXL, maybe it doesn't change VRAM, but you could use it with low VRAM settings and it should work at a huge speed still.
TensorRT is dramatic improvement to speed, maybe even VRAM, IDK.
Is a 4090 good enought to train my own model... Or does that have to be done in the cloud with much bigger GPUs?
I believe a 4090 would be cutting it close
but give it some time, there are some new tech's on the horizon that might help
Hey everyone, even if you have SDXL 0.9 locally and are having fun - can I request everyone who can take some time to switch to the bots for a bit and vote? 
Actually....
still looking pretty rough, not sure what thats about
... they may be... sometimes... but that's kinda why we REALLY need you do it
aren't all the release candidates ready to go? what are you guys waiting for?
need us to vote to say its a tie again and then delay for another week 
No no nothing like that 😄
the model they choose to release will be what future finetunes are based off of. chill, these votes/etc are to see what people prefer.
- it's likely the one they'll continue to work off of for future models
most of these images look terrible, so I am not voting for any of them
i vote when there's really good ones
if I see some that don't look terrible, then I will vote
like one really stands out from the other
Any votes is better than no votes
Skip if you wish
but vote on the good ones!
And if there are 2 good ones, open them up and take a look
@dapper current Why do bot results look like they have been run through a computer with the power of a potato clock when I'm at home and get stuff like this???
highly unlikely, even one good image seems to be a rare find right now lol
genuinely what did you guys do... these look dreadful
yeah, if that's what SDXL1.0 looks like, I'll start finetuning 0.9 right now
i will spend my night on the bots then @dapper current !
right, jesus christ these look horrible
been through 6 bot channels, haven't seen a single image thats not worse than 0.9 a month ago
Anyone successfully turn an img2img into a outlined black and white coloring page yet?
I think that's the point
They mentioned it was a blind study
we literally said it
sytan
stop
with
your
stuff
lmao
yeah, it's like dall-e mini💀
its never looked so beautiful
can someone please permaban him?
Really?
❤️
lol

so what, is there no SDXL in any of the bots? I am over 100 images in and have not seen a single one thats not reallyy damn bad
sytan.
leak
is this the sad cat ya'll been talking about
to what?
wanna find out?
what's going on ? 1.0 issues? or nothing to worry about?
SAI literally asked this community to be honest, my opinion is that SDXL0.9 is infinitely better than what the bots use rn. If the bots are using SDXL1.0 I'm starting in this moment to finetune 0.9
^
promote Sytan
I don't understand what the point of voting is right now
I will vote for the images that look good
there we go
it's got non-sdxl mixed in
ok, this one is pretty dope haha
yeah, you can't abuse your power because people are being honest
SD 1.5 and 2.1 are in there.
new color time
Community Archive, hmm
you've been Archived
noo pls noo.. not 2.1
anything but 2.1

5000 votes, and we go back to relatively normal
blame the researchers!
its now 2:30am... my SDXL. 1.0 delievery hasn't arrived yet
the applied team didn't wanna do this to you!
Baseline 0.9 did this? Or you've got a bunch of add-ons?
no amazon prime i assume
I'm just waiting to see an image worth voting for 
wait, we're the product
i thought you were a researcher...
trying to show people just how bad the base 1.5/2.1 was before finetunes?
okay ill find the joe penna papers one sec
lmao
okay so thats why the bots are being funny. i get a decent and i get a bad. this is fun!!
hee-hee funny typoooo
ya'll saw nothin
I SAW EVERYTHING
So 1.0 launch today?
WHO SAID IT WAS A TYPOO
I did a WIP for my altering technique, I made a model slightly better than base 0.9
Pog
Let's hope it's not as bad with faces as the last model 😅
guessing they are just doing matches between 1.5 base, 2.1, SDXL 1.0 candidates, (including 1.5 vs 1.5 for example) just to show how much better the new one is. Though voting on here has some flaws. Wonder how easy someone could just bot it up and make it vote randomly. And how many voters actually look at the prompt and choose the pic is best instead of just voting for what looks the best.
Will 1.0 be compatible with all the extensions that work on 0.9 or will it take a while
xl 1.0 pub on huggingface today?
They shoooould work
what do you use?
interesting... AND FOR SDXL how how... convenient
@hard fractalYou have to be trolling-
I generate one image and it looks amazing, but all I see is trash from everybody else haha
emote on this apperent NSFW pic of potatos cuz why not. 😂
what is happening to the bot lol
NSFW potats
listen, Sytan... we know you're one of the best prompt engineers here – but let's be kind to the others learning ❤️
its called favouritism
its got 1.5, 2.x, sdxl. and more probably
thats not what I mean
I am so confused what is up with the bots
seems like joe has a favourite child .... sytan 👀
I don't wanna be favorites lmao
5000 sadcats coming right up!
got scaled loras working well
What do you mean by scaled LoRA?
Maki, no! You've been there since pre-dreambooth days!
oo, tell us more.
dont remind me of textual inversion
try posting this: <t:1690399800:d>
i think sad cat is a social experiment
<t:1690399800:d>
Is textual inversion working atm? In terms of it being supported by any training programs?
👀👀 👀 👀 👀 👀
<t:1690399800:d>
👀 👀 👀 👀 👀 👀
yep - did a little tweak for y'all
I miss mixing textual inversions so much 😄
i think it worked
Thank you, milk
now please, please please - understand there will be some funk, it's intended. Hopefully you get a bit more good now though
lmao
you can do it in a countdown style: <t:1690394400:R> (use R instead of D)
<t:1690399800:d>
also try posting this @hard fractalbanmeplis
||<t:1690399800:d>||
Doing what I can with 0.9 to get ready to switch my custom ai photo app (lensa-type for photorealism) over to XL. I was a bit concerned with lack of refiner but even 0.9 with the right training has no problems hitting perfect likeness and good realism. I know that girl personally, already the best results I've had with her.
don't help em!
you almost got me...
well now.
..this is garbage?
Think of it as being for me 😉
looks good
I love being a part of this community. It feels like being part of something very important to societal change.
same!
Uh, so we're not even guaranteed that one of the images is SDXL.
ur beautiful marry me
gosh, so hard to choose
were you embarrased?
can we stop the sad cat experiment... i dont think as a abnormal human i can keep with the demands of the people to generate 5000 images for votes
your not red for me so it doesnt count lmao
oo, cool discord skin
"BUT WHAT DOES IT MEAN?!"
he keeps putting it back!
its one of the default ones lol
Will SDXL have a separate inpainting model like SD?
is the refiner going to be slowly phased out after 1.0? or is it sticking around long term.
I assume that depends on them managing to get base up to the same quality or not.
boom, new coloer @hard fractal
not on launch
discord color themes have to be one of my favorite additions in recent
Hopefully it's soon...I use inpainting ALOT
i thought sdxl could do inpainting on its own...
for a bunch of uses, probably fine to ditch it on launch! but people will be making LoRAs and finetunes that will eventually make it obsolete
we wanna make sure an inpainting model it vpred / ztsnr – something that can match full black
Is a seperate inpainting even needed with controlnet?
your gunna make me want nitro premium lmao i only got classic
and full white.
it does seem that base 1.0 does seem to be a lot less fucked than 0.9 base, which is great to see
didnt you say it would require different architecture to get high resolution detail out of base
It is for 1.5....for me at least. I can usually get much better results than with controlnet
@high skiff its official! ANND its gone lol
I did, and it would.
whut lol
Was there even a 1.5 inpainting model and not just for 2.x?
what is gone? lool
Yep
WHAT IS GONE
could i be the inpaint model? i can manage the demand
the orange
probably need a better replacement for the vae
I also saw it
so is everything on track for tomorrow as far as we know?
Halo 3?
what do you mean?
if you are just pointing at me over and over, yeah haha
the release
||"the release of what?"||
what does community archive even mean lol
it's a person abusing his power because you have an honest opinion(I hope it was a joke)
THATS WHAT I AM ASKING
that was a post beta thing at one point
archive of all the prompts and images made in beta i think
yeah, something's off..
is something happening soon
WHY YOU TYPIN
really not a fan of the new weights. any way to get back to what they were before today? they're still the way i liked them on dreamstudio
2% is typing....
IT TOOK YOU AN HOUR TO TYPE THIS?!?!?
usually debating on adding technical details to what joe says but then letting it ride cause its funny to watch the conversations here
new weights on clipdrop i mean

||<t:1690399800:d>||
it does get kinda wild here when joe comes haha
dont tell anyone
sorry if this isn't the channel to ask about that
The bot seems to be running full random now so I guess what we see now isn't representative of final release.
Hi there! If you just came along please note we are doing something a little different right now! Please continue to gen and vote
just imagine that all the time
❤️ @hard fractal
Hey bro what’s wrong with pic generating
1 pic is good
2 pic is very bad
Just check bot1 channel
owo
XDDDDDDDD
owu?
uwu?
do it. I require data 🤖
( ( ) ) \ / \ / ( ( ) )
wonder sai sometimes adds models without rlhf to the bot, kind of curious what the raw model output would look like
so if everybody is hanging out here, that means for the 26th?
whats on the 26th?
or you could be like me and put voting into hardmode #1100170365604483202 message
could sdxl produce a better logo than the new "twitter" X logo?
death lol
I don't know. I heard it from someone else
A kindergartener could do that
Called it
CALLED IT
Doesn't look like the bot it is working for me as I keep getting bad results from stuff that previously worked well.
Dont even lol

I mostly use inpainting to add stuff quickly to an image...like adding a guy yo this scene took like 2 seconds.
HEY i was looking at that!
i mean there was a 26%ter in there sooooo i call that a win

not clear enough, better bake another week
this whole launch is a big dragonball fight episodes...
oh it's clear, all right.
Why...
and i recently punched my monitor 🙂
It's not even the right colour.
A photograph of Joe Penna as he winds up to drop a bomb called "DELAYED"
@hard fractal You good? lmao
typically goes that I have something always in the oven to fix the stuff people want but a lot of it is fairly experimental. Maybe we will dive into it some of it tomorrow 😎
when SDXL 1.0 comes out / when SDXL 1.0 does not come out
JOE PENPINIA do be fire farting lol
you dont even understand how... i was about to slap the shit out of this fly on my screen
that looked too real
near success
Nope, I just keep getting results in the bot that is worse than what I got pre 0.9...
i blame the macbook pro screen... shit looks too real
I should try them locally too...
yeah, looks like they are trolling right now lol
Even the ones I can see is from SDXL is a lot worse.
not trolling. just a lot of different models in the bots so they can have a fancy pie chart that says sdxl is better
And many attempts at the same prompt just give me poor results.
trowoll
at this point theres too many models... i just want to draw me a perfect stick man and thats it
aHR0cHM6Ly93d3cueW91dHViZS5jb20vd2F0Y2g/dj1kUXc0dzlXZ1hjUQ==
Nice trick, I can recall the last part of the URL so not falling for that one.
falling for what?
rick astley youtube url? lmao
The double = and random letters should tell you it is base64
hankles
i agree
i disagree
I remember seeing strings of text like that back in the IRC days. Mostly scammers tricking people into clicking them
someone arrest sytan

I really do not get the point of comparing SDXL to non-SDXL models now, they prompt differently so any prompts indended for SDXL will fail at non-SDXL models.
me when I don't get arrested
cat throw up dance? xD
ir does look like that lmao
sdxl looks completely different today from yesterday it's insane
your welcome love it when i see my cat hacking up a hairball at 4 am
In a good way... and do you mean locally or just on the bots?
on clipdrop
MFAOL
This is one of the greatest images I have ever seen lmfaooooo
by the time i wake up SDXL 2.0 better be out
and i mean that in a bad way
Alright to be fair, with lots of attempt I did get one image that surpasses pre-0.9, though still not happy with the cropping.
WHAT HAVE WE BECOME
shit's way too busy and elaborate now
your face is too elaborate
:(
:>
thats actually really good lol
I'm telling ya, it can and will pump out some really great stuff if you try
I don't know if this question has been asked but, does somebody know at what hour SD XL will be released tomorrow?
They are based in london, but london isn't real, so IDK
One example where it was better, I also had a lot of tries where it was worse than pre-0.9.
<t:1690394400:R> maybe..
yeah, that's some harry potter shit
i was in london like few months ago.. looked pretty real tbh
Best guess is it could line up with the event they have listed.
"if aphex twin's music were a place" used to produce some extremely cool stuff and now it's just garbage

it was generated with SDXL^2 pro max
Here is pre-0.9, a lot better cropped.
ultra
in real time?
Advanced AI simulation
yessir
no way
rendered on a Jupiter Brain
they did it the distilled model its about damn time
dGhpcyBjaGF0IGlzIHRvbyBkdW1iIGF0IHRoZSBtb21lbnQsIEknbSBsZWF2aW5nIHVudGlsIFNEWEwgcmVsZWFzZXMu
what you doon
@urban fjord i had a dream about tricorn lizard guys like that when i was a kid. insane seeing someone else post them on here
are you having an aneurism?
by the way i dont know what aneurism means
he's having a magnetic moment it seems
||base64||
ok im going to sleep
sleep, peasant
wasnt london destroyed during the filming of gi joe: retaliation?
so this is the kind of thing i would get yesterday from sdxl on clipdrop
and now it does this shit
for "if aphex twin's music were a place"
do you see why i have a problem with this after blowing ten dollars on it that i could have spent on gas
tbf, quality will significantly go up in about 18 hours
i don't want it to go "up" i want a way to generate something like this again
is there any fucking way to make sdxl work the way it did just yesterday
i mean look at this fucking garbage for the same prompt now
not a clipdrop user. You're probably better off asking in either #🤝|tech-support or #1010934719455707218
thank you
time to listen to the Selected Ambient Works again 😉
Wtf is a super stage
why was he so upset about that.. i mean if he could read they literally said the bots are crazy rn because of them gathering data on other models. its not like he needs it or he'll die lmao
clipdrop as well though?
im not sure how clipdrop works.
yeah i'm not using the bots i just now joined the server here because it was the only place where i thought people would be able to answer my questions
just spent like 2 hours wondering why my Principled uber-node had worse results and it's cause I forgot to use the XL-specific clip encoders...
mod support is official Stability AI Support - so you'll definitely get a good answer there, though it may take a few hours
gotcha i didnt know you were using clipdrop my bad. im not sure how it works 😬
i was under the impression that clipdrop was using whatever the latest builds of sdxl 0.9 were and any changes were a result of that
if they're fucking around with their instance on their own maybe i need to ask them about this instead
but the official page for sdxl specifically linked to clipdrop as the way to use sdxl online
so i would assume there's a fairly close relationship
@high skiff what's the official term for the refiner -1 step detail hack so I can make it an input field
Fractional Misalignment?
no one's going to know what that means
Discover the power of SDXL 0.9!
Unleash your creativity and enhance your images with this cutting-edge AI technology.
#StableDiffusion #StabilityAI #Clipdrop
Try it for free on @clipdropapp today! https://t.co/FP57AeLqAN
117
I have been told by so many people that its very clear what that means lol
refiner_short_sample = 1 or something
i'll shut up now sorry
you misalign the latents... by a fraction lol
oh wait, i think you are talking about the less ideal verion, full step misalignment
Add a tooltip to it that explain what it does.
where the refiner ksampler has 20 total steps and the base has 21 total?
oh, no, thats fractional
desu
the input will have to be an integer
What is the benefit of doing that?
so idk "alignment fraction" is a weird name for a whole number
refiner_misalign_step maybe
Seems like that would be very nearly the same as skipping a full step, is it really that noticable of a difference?
from what im getting in the bots that i know for sure are 1.0. im seeing a decent quality improvement i am very happy with. cant wait to see more tomrrow. (hopefully)
fat fucking node. has partial diffusion, misalign step, hires fix, and img2img all in one.
holy hell XD
principled, love it 😄
does the refiner use clip L or cip G
i thought karras is better sampler
Where can I get this node???
Im making it still
it'll be on my github here in a bit
Only real downside is I don't know how to cache latents.
So if you change a refiner setting it re-renders the base again
Just created a reddit post with some of my SDXL experiments including prompts:
https://www.reddit.com/r/StableDiffusion/comments/159rio2/some_of_my_sdxl_experiments_with_prompts/
0 votes and 1 comment so far on Reddit
"Is here" he lied as naturally as he breathed
These eyes seem ok? no?
a bit of a lazy eye but thats okay. lazy eyes can be cute
I would inpaint the eyes
the right eye isreally good
#1100484581037195384 message vote for my picture so i can get into pantheon xD
Combining high-res fix and prompt-editing with SDXL gives some cursed results.
How do I get an image into showdown? does it have to be created with a bot?
Yeah, images that gets the most votes in the bot channels goes to the showdown, as far as I'm aware unless Joe is typing to correct me.
what was the default pos/neg ascores again?
Mental prep
Just overheard the funniest thing: "You realize they proved that they are better than midjourney, right? You realize this is the best model ever made? If you don't like it, it's your fault. You suck at prompting."

Did you make this image last week?
I just made it now 😅
Testing my soon to be obsolete LoRAs
Oh. Maybe you mean that you'll ALSO have it next week
🤷♂️🤷♂️🤷♂️
Just save your training config and you can redo them for 1.0 if they won't work.
Not complaining at all, I got a decent sense of what to use for 1.0 🙂
Is that a text-LoRA btw?
woohoo made it to pantheon for the first time thanks for those who voted for me 😛 #🏅|pantheon message
The 🤷♂️🤷♂️ is questioning whether or not you are going to need to retrain…
Maybe you won't
w-why does this sound like a delay? 
next week is sdxxl, they'll release some experimental 1280 finetunes
any chance on getting the old chonker CNs just for fun before ControlXL
@hard fractal JOE ANSWER ME
WHAT DOES COMMUNITY ARCHIVE MEAN LMAO
You keep trolling me lol
We don't want to put our stamp of approval on that kind of architecture, not for SDXL.
Even better 😮
From anger? Embarrassment? Maybe you're just hot
cant those files be made public without a stamp of approval?🥺
In my opinion, maybe via API, but certainly not open source
like not related to SAI, unless sai resources were used in the making
any release targets for ControlNext? we need rage fuel when deadlines are missed.
Otherwise, it's going to split the community, just like how 1.5 and 2.1 were split -- and LDM versus diffusers
but its not a base model, anyone can swap the new CN into workflows the moment it comes out
I have a target date of right now to stop giving out target dates
where is the fun in that
these channels will get boring, do you really want that
it creates engagement, you know you want to
Then target date is July 18
wtf where's CNEXT was supposed to be last week, im going back to MJ this is a conspiracy
YOU SEE
But what if I start putting out the target date?
@high skiff mild quirk with your workflow I found out making my super node. The end step being equal to the total steps means theres noise still left over I guess. You just don't see it cause refiner is set to not return noise.
If you set the refiner end step to like 10000 or something, it finishes the denoise at the cost of an extra sample which improves quality slightly, making your actual runtime total_steps + 1.
Whether you account for it or not is up to you but it might be worth at least knowing about.
forgive me for bothering you mr. penna but do you know anything about how clipdrop handles its instance of sdxl? the results i'm getting today have drastically changed from yesterday and for the worse.
My fractional offset workflow is not intended for all use cases, that's why it's not being merged into the main branch
It can be very detrimental to certain styles, specifically non-photographic ones
that's not fractional offset
that's your regular workflow
partial diffusion
if you don't have anything to do with that then i understand but i'm really desperate to get some help with this
Oh, that's just a downfall of the refiner to begin with
The 0.9 refiner would always degrade the finer details in stylized images
For digital art and painting, I recommend just removing the refiner as a whole for 0.9
Alright, yeah, I'm sorry
My mind is all over the place
I'll have to take a look again later, sorry
this results in there still being leftover noise, surprisingly
if you adjust it, it'll refine out that noise instead of just removing it at the cost of an extra step
so you'll run 13 base 8 refiner
i have no clue why it's like that, comfy would probably have to explain it to ya
Interesting
I'll have to mess with it
What value is going into steps, 20?
yes
Seems like an issue with comfyui then, thats a little counter intuitive
I only noticed it cause I decided to just leave last_step unset internally in my principled node, then all of my gens suddenly ran for an extra step. Setting force_full_denoise=False when calling common_ksampler showed that it did indeed have leftover noise when last_step was the same as steps
so then I got curious if we were doing the refiner 1 step too long, and set the refiner's start to base end + 1, so 0/13 and 14/999 but that causes there to be excess noise when starting the refiner which results in a really extreme version of the "fractional misalignment" phenomena
I think it's because for start_at_step and end_at_step to be the same it would have to not do that step twice, it would make more sense imo to do start_at_step = end_at_step + 1
For the second pass
that no work.
base 0-20, refiner 20-999 on left
base 0-20, refiner 21-999 on right
clearly "misaligned noise"
lemme try with basic bitch euler to make sure it's not a ddim quirk
Oh
interesting
so it's the ddim_uniform scheduler that causes that
using the normal scheduler causes base 0-20 refiner 20-999 to run for the proper 30 steps instead of 31
today xl release?
@high skiff the mystery deepens
doesn't ddim start at a weird step?
¯_(ツ)_/¯
yeah, 999 by default i think.
ddim with the normal scheduler works as well
just ddim with uniform scheduler
imma just have my principled node default to Euler cause I'm a coward
yea. I guess that extra 999 timestep means you need to have an extra sample for it on the final render?
lol wrong VAE oops
nono, perfect vae, it's just magically enchanted
Yeah I was gonna say depends on your definition of enchanted 😄
I think I ate the wrong type of mushroom...
🍄 
https://www.reddit.com/r/StableDiffusion/comments/159oy2v/my_friend_and_i_just_released_sdxl_on_our_free_ai/ caught my attention, I wonder, are such sites in violation of the license?
where is SDXL 2? it's 26th already 
1.0
I heard SDXL release date gets pushed back a day everytime asks where it is
Look at the stage event for likely release.
yeah later in the day, I was being sarcastic 💀
We are stuck in SDX-hell
my pain is immeasurable and my day is ruined
🤣
So no controlnet?
got mr. Scott to comment so nvm this

Im sure joe will drop some hard dates on controlnext release soon..two dukes will be locked in the office with an IV drip of red bull, Adderall, and crack to make sure the deadline is hit.
what time is it now
bruh. get a clock
it's 04:11 where I'm
there is no european time on the clock my friend
Delayed
try turning up your resolution
I wonder as well!
what should i turn it to? its already on native res 1024x1024
faces that are at this depth in a composition will mostly be distorted one way or another. sometimes the refiner can help with that. but usually subjects at that distance in an image are hard to do. try to get a closer shot 😉
will do! I am using the refiner as well 😛
sdxl was trained by amateur art restoration experts
When encountering this situation in sd1.5, I usually perform Hires.fix or try to get a closer shot
I wonder if afterdetailer works for sdxl on auto1111
after detailer is a great 'turn on and forget' tool for this exact scenario
basically performing an automatic inpaint after detecting a face
Does anyone ever use dpmpp_2m_sde and what scheduler should it go with and how many steps are recommended?
yeah, I was going to try that out
Use the defualt in sytan workflow
I do all the time... I just wanted to try different things
Idk any others u can try also make sure to keep the bae refiner and upscaler all the same sampler and scheduler when u mess with it thoigh
Oh. If a text entry is single-line on the main node, it stays single-line on any primitives connected to it...
face looks a little better, still pretty messed up 😦
especially the eyes
ill keep playing around with it
isn't it weird that the bots are still running the 0.9 A/B tests though? don't they need time to actually refine the RLHF results?
I dont think so
the polling isn't closed. i don't think we're getting arelease today
Not currently anyways
I guess that is a definition thing, but a release candidate in software would be usually something like SDXL 1.0 RC1 + variant etc
looks like it's all set up for a/b voting to me currently

right now they had (or still do) 3 different SDXL model variants running on the bots
i'll be pleasently surprised if it releases. not expecting it
I hope they delay so they can launch it with controlnet weights ready day 1
Okay I think the Principled SDXL node is in a usable state. It's located here along with an example workflow and the rest of my stuff:
https://github.com/Beinsezii/bsz-cui-extras
ping @spring fulcrum since you asked earlier
from all the hints that we got. it will be releasing during the event. also the bots are just for the pie chart they want to create basically saying that sdxl 1.0 is prefered over all the models. even old ones like 1.5, 2.1, etc. the bots are using multiple models.
uhhh - that sounds interesting 🙂 nice
it's actually not a lot of python code. just does a few checks then calls the built-in nodes
definitely looks scary though lol
well that sounds even better 🙂 thanks for sharing it. I will definitely check it out
a power node 🦸♂️
Whatever happened to the comfyui presentation layer
theres comfybox. you mean that project?
No theres supposed to be something new
The A1111 killer
first i would've heard of it
it sure is a shame that no clean UI over comfy is dropping the same time sdxl is
as usual, anythign calling itself a "popular choice killer" is going to fail
oooo nice.
Any sneak peak images mcmonkey?
for example, the 100s of games that called themselves "The WoW killer"
here's a screenshot of copilot making fun of me when i made a typo while trying to fill in legal stuff, does this count?
lmao no that doesnt count!
that is definately copywrong
Not really.
there is no A1111 killer, everything is a tool. Sometimes there are better tools this doesnt mean one tool suddenly useless. I am excited about this UI comfy backed thing, maybe that will be a comfy killer
i'm going to disregard any project being referred to as "another project killer"
If you replace a tool completely it is a killer of that tool.
not a comfy killer, or anything else killer no
in the case of comfy it's explicitly an empowererer
that's just a bad mission statement right from the get go
take comfy and make it more powerful
no one said it was a killer of anyuthing lmao
Does it have checkboxes for booleans
and by that i mean it's definitely not that because it doesn't exist nothing exists, i don't even exist, you're imagining me
HAHAHAHAASfjasfjoasf listen
I would love to see it happen but I'm kind of doubtful that it is really getting up to par with Auto1111.
i'm just going to block you. it was literally said.
don't be so highschool. fetch isn't going to happen
Comfy has to be trolling with that lol
I actually had the issue of not being able to identify the booleans properly because they're not booleans they're strings
i need to talk to him about that lol
please do im tired of your negativity around sdxl and stuff XD
Yeah, get on that ☠️
ehem
@visual glade
do you think we can get the boolean inputs swapped to boolean instead of "enable" / "disable"?
17 hours to go?
expecting sdxl tommorrow to not release isn't negativity. it's setting realistic expectations.
taking delay speculation personally.. hmmm
maybe be less emotionally invested. good for mental health
And checkboxes for bool inputs
I think that was the standard for software before I was born 🤣
Thanks you
form UI's are still the standard really
new experimental prompt editing node - could get interesting. testing now: https://github.com/asagi4/comfyui-prompt-control
Not needing Lora nodes could be convenient for sure
that allows lora's in the text like embeddings!? YEUS!
i hate wiring in loras
it's the destroyer of workflow
Well well well... we're done!
and sliders for the resolution inputs instead of the awkward numeric draggy thingy. Maybe have it like, idk, snap to common values but let you hold shift to move it more precisely? Maybe even be really cool and have it use power-of-two scaling so it feels really natural to position it exactly how you want, vs a linear scale which is awkward. Yknow something like this:
that's just a mockup though, not real, no such system exists or is possible
she lost her finger in the war
best crop ever
checkpoints or it didn't happen
Oh, did I crop?
now it makes sense!
poor 2.1
😦
Let me plug that into uncrop...
oh hold on guys i got you
what a tease
oops
lol caq that was my idea too
so not fetch

i don't believe that's a mockup :phillipjfry:
Anyone nervous about having the general public try and improve the 1.0 model? The public are generally idiots
i concur
We don't RLHF directly for that reason. Otherwise, you end up with images of pretty ladies when you prompt for potato
no. i'll laugh at the idiotic shit poeple do, but i'm not concerned
oh yeah haha it would be insane to take vote data directly into the training. people vote stupidly.
some people vote with malice
potato, seed: 1, SDXL_1.0-If-We-Did-RLHF
@hard fractal Why doesn't your username appear on the list of users on the side of discord?.... Why is your existence so secretive???
joe penna, much like me and any UIs i may or may not be developing, is a figment of your imagination
Bro is a director?
irish waifus
@hard fractal how long do you think the super stage will last? ill be at work D:
Is the UI like external from comfy? Will it be easy to try in colab?
So if red is 3.42, and the next colour is 4.63...
||@hard fractal will you finally tell me what community archive means? 🥹 ||
it goes until there's only one dev left standing on the super stage royale
what ui? we know nothing of no ui
'tis an unconfirmed rumor!
i've never heard of any ui either
Heyyyy uuuuuuu iiiiiiii's
if i was aware of any UI i'd definitely go check if colab can run it though
what is that?
@high skiff @visual glade Do you know if there is a way to lock certain nodes to keep some settings the same but change other things when you drag an image into the UI?
See that wasn't so bad was it everyone? 😄
hehe
Yes, yes it was

ORANGE AGAIN
your name is orange again
orange name 
Sytan got archived 
WHAT DOES IT MEAN
And now we have our Uncroppable evals.
Can I get a cool role color... like for channel clown or something?
I hope those don't melt
ill swap you if you do. i want black
@high skiff It's a role you can apply, it unlocks the ability to see archived communities at the bottom of the server list.
Aaahhhhhhh


I wanted to make combos with only two choices automatically become booleans in the frontend but never had time to implement it
Joe keeps making me orng!
comfy, did ya release those Reference Only nodes?
is my name orange now 👉 👈
Nooooo
it was for a second lol
It only goes orange if the primary "Member" role is taken off you
I would be interested in this personally
ohhhhhhh
I guess I'll clean them up release them then
I totally missed that mine is different now... I am in the Stable Society role and my name is black... I like it
YOU ARE SO SMART OMG
Awesome! Let's release it at the same time that SDXL drops!
Which, who knows when that is...


Is the even gonna be a blast?
This is where he accidentally breaks comfy for the SDXL release
Oooo, me me, I know, I know Joe!
august 22, isnt that when sd 1.4 dropped? gotta stay consistent
ORNG AGAIN
Musician/Gamer/Artist roles sound more fun than member XD
We don't randomly break things... Naaah.... Bot has never been down unexpectedly.... Perfect uptime 😇
it would, but wouldn't make any sense
If only you looked in the roles channel
RVC?
if something doesn't break at random times I don't trust it
it won't break anything because it doesn't require any core changes
lol. you do have the musician role though
and comfy is freaking amazing at what he does, jus sayin
I dowo?
I know, I was joking 🙂 You do like pushing to the Main Branch after all lol
Comfy is probs state-of-the-art inference expert.
comfy's a magician
I DO WOO
nothing can convince me otherwise
nobody has complained about me pushing directly to the main branch yet so that's good
🧙♂️ 🪄
We're always like,
"Hey comfy, can you--"
"Done."
is there a way to mark a field that it should be a node input by default instead of a widget
@hard fractal even the inpainting and controlnet would be supported as soon as 1.0 is out?
any news on release time? someone on reddit said tomorrow?
Seriously. The amount of times we've heard "Oh I just did that" or something along those lines...
no and no 😦
not at launch.
updating comfy is like unpacking a present. except you get to do it on a 2 day basis rather than a few times a year 🤣
Could be months, years, who tf knows 🤷♂️
so tomorrow isnt a commonly known fact gottcha
Wait how do u download sdxl 0.9?
theres a event tomorrow where it could possibly be released then but who knows
when is inpainting gonna get supported and when might you release controlnet models?
If this was anything like me growing up as a child with my father the more times you ask the longer it takes
are we thereeee yeeet
unless you wanna delay SDXL base model launch?
Vote 🅰️ to delay
Vote 🅱️ to release ||you know wen||
bruh
^^ me until the thing happens
Delay for controlnet support on release 
That's it, call it - the community wants a delay
<t:1690394400:R>
the mob has spoken
NOO
We must do what the people want
I was threatened with riots in the touhou ai discord
It's what the people want ™️
How
Anyone down to play chess?
heyy yall cant vote on your own poll
SDXL Chess where every time you move your piece is generated into something completely new?
😉
only AI generated chess
the cookie monster is wearing a shirt that says Still SDXL, still from the muppet show and than playing the seed lottery! 🎲
delayed until 1pmish central
it would probably give better results with a real inpainting variant but inpainting does work with SDXL
Lul
@hard fractal how good is the new controlnet at qr codes?
I wonder if I could make a chessboard lora and make chatgpt play against whatever position sdxl spits out
sounds doable with enough overfit
It might be good, but I'm BAD at it.
That would be interesting or when stable diffusion gets so good it can defeat the best chess engine just off an image
someone teach me.
SDXL is really something
Havent played around with it too much since no controlnet
Its so impressive
is it better compared to v11?
lot of those came from a custom trained QR code controlnet model that some chinese company made and never released publicly
This is my favorite one so far:
yeah it's very very good 🙂
oh! TIL
yea all the ones that link to that QRcode pretty-ification website are from that private model
Hm, we could probs do that, eh @sour obsidian?


