✨｜sdxl

[trim orbit]

big giant ianal disclaimer on this recommendation as it comes with a giant restrictive license attached ot it

[high skiff]

damn, that does look good

[pallid path]

urhmm... I can! 🤓

[rich coral]

lighting on the ceiling is a little whack

[quasi remnant]

and ai does this kinda tiling thing

when u get "too detailed"

[pallid path]

yeah like some upscaler

[high skiff]

your face is a little wack

[pallid path]

omg

[radiant tartan]

comfyui question: so i installed from the portable version, but saying so many things are missing. https://cdn.discordapp.com/attachments/1031343840998731858/1133891677568905387/image.png

[high skiff]

:>

[polar epoch]

closing in now on good looking 2048 upscales

[delicate grotto]

ah yeah, the old school building

[thorn lion]

i'm on a 3070 8gb laptop and it takes about a minute per generation

[rotund cliff]

does comfyui have a "generate forever" button like automatic?

[urban fjord]

Try the tiled vae node or the fp16 vae and see if that helps you run faster with just 8GB VRAM.

[steady bramble]

I’m so amazed that we can now generate text in image

[quasi remnant]

Yes.

[delicate grotto]

what? are you using an asus laptop or something?
it takes me 40 45 seconds to generate through 2 samplers + 1 refiner
3060 6gb vram 120w, 5800h
you should be doing it much faster

[quasi remnant]

It's under extra options

[rotund cliff]

thanks

[rich coral]

as designed

[urban fjord]

Yes, Comfy UI has a checkbox by the queue in the extra options.

[quasi remnant]

it'll keep adding to the queue when the queue goes empty

[clever verge]

Might be a bit better at minecraft style too. Not perfect but hey.

[molten crypt]

Does Automatic1111 support SDXL 1.0?

[urban fjord]

Yes, Auto1111 does support 1.0 base.

[clever verge]

Those are very good.

[fresh path]

I don’t really know what I should download there are multiple vae and different text encoders.. Is there a guide?

[urban fjord]

It does not however support Sytan's workflow as far as I'm aware atm.

[violet ferry]

Yes, but you can't use the refiner model with it which is a major component

[lilac raven]

Do SDXL uses much more RAM (not VRAM) than SD 1.5?

[untold crescent]

refiner works on img2img?

it doesn't error with it anyway

[violet ferry]

Well, technically it can, but it's not a good job

[pallid path]

probably yes cause it has to store 2 models in ram at the same time

[agile quarry]

your missing custom nodes . this will help you install any missing nodes in a workflow https://github.com/ltdrdata/ComfyUI-Manager

[urban fjord]

If you're using base + refiner you probably need 32 GB RAM to make it run smoothly, but with just base then 16 GB RAM is okay.

[violet ferry]

You're supposed to be using the refiner model partway through decoding the latents

[delicate grotto]

yes, but it can also generate images(not as well as sdxl)

[median prism]

You can't use the refiner at the same time on Auto1111? That blows 😦

[polar epoch]

Holy, ok, CFG increase is the way to go. Now to make rest of the face not melty lol

[quasi remnant]

i'll check how much vram it spikes to at each step and report back

[delicate grotto]

16gb can run refiner + sdxl pruned

[molten crypt]

But can you run with the refiner at some slower preset?

[urban fjord]

It won't run smooth with just 16 GB normal RAM and is just dependent on you running almost nothing else.

[lilac raven]

I wasn't able to run SDXL on 16 GB RAM + 12 GB VRAM
Now I'm trying with 48 GB RAM + 4 GB VRAM

[molten crypt]

Like medvram on Automatic1111

[quasi remnant]

16.9 gb was all it was at

[polar epoch]

And even with upscaler. This is generation with 2048 upscale

[quasi remnant]

16.9 GB with no other progs open but this + edge

[pallid path]

what about samplers, what else can we use besides like euler

[urban fjord]

I'm talking about 32 GB normal ram not VRAM.

[frank steeple]

Is there a --listen equivalent for comfyui?

[quasi remnant]

u could get by with 24

i think?

24 is a RAM count some ppl have

[delicate grotto]

you should be fine on 16ra and 6gb vram on a tiled decoder, sdxl pruned+ refiner

[high skiff]

<3

[quasi remnant]

if this is all ur running, 24 would be fine but i think u need more than 16 for sure.

[pallid path]

Sytan do you know samplers that we can use with SDXL that aren't Euler

[rich coral]

architecture feels so mushy without controlnet :/ when control net Mr Guitar Man!?

[lilac raven]

I'll have to test things further, or just accept I'll have to spend more

[clever verge]

--help will show all options.

[quasi remnant]

dpm++ 2m

[high skiff]

its --listen haha

[raw prairie]

dpm++ SDE

[pallid path]

this one?

[quasi remnant]

yes

[violet ferry]

I'm able to almost run it with 16 GB RAM/8GB VRAM in Automatic1111 but it's not stable, I'm just trying out ComfyUI and it seems to be able to handle running both base+refiner models at once somehow

[high skiff]

I use --listen on comfy lol

[cerulean saddle]

I know this has been said about a thousand times before but I have tried Auto and Invoke... 30 minutes per picture. Comfy.... about 2 if that. Anyone with about 8GB VRAM would be wise to use it.

[pallid path]

it looks crappy to me though

[quasi remnant]

it shouldn't if you use it properly.

[pallid path]

do I change the scheduler?

to something like karras?

[quasi remnant]

normal scheduler

[pallid path]

any other setting?

[delicate grotto]

au1111 uses much more ram and vram

use --lowvram, might help
also, use a tiled decoder

[quasi remnant]

example of my settings w/ it

normal scheduler

[brazen patrol]

I'm a trifle confused on the Base + Refiner. I got both models up and running in Swarm, but how do you combine them?

[thorn lion]

I have chrome open so it's using the page swap to load the models.

[pallid path]

thx I'll try that

[delicate grotto]

same here

you should be rendering ~20 30% faster than me

[untold crescent]

I got auto1111 running on 8GB VRAM gpu with --medvram --no-half-vae --xformers

[brazen patrol]

Is the refiner supposed to be loaded somewhere special?

[delicate grotto]

i can link you a vanilla workflow

[cerulean saddle]

Got it running, but it took ages, plus no refiner.

[quasi remnant]

to finish the last 20% of steps give or take

ur supposed to send the base image off still noisy

[untold crescent]

@cerulean saddle it takes forever without that medvram

[pallid path]

it looks bad maybe because I'm at a lower step count

[quasi remnant]

and let refiner finish last 20% of steps

[untold crescent]

like, many minutes per image

[delicate grotto]

@brazen patrol - drag this into the comfyui

[brazen patrol]

Pls do link. I have th HW, and the models, just haven't fiigured out how it is supposed to use both

[untold crescent]

but you can see from auto's gen that it maxed VRAM then

[cerulean saddle]

had all the same setup. its just crumby on my end.

[royal tangle]

mew

[untold crescent]

yeah I did have less issues on comfy myself too

[pallid path]

ok dpm++ 2m with karras looks good on low sample, like euler

[delicate grotto]

wet kitty, small kitty
little ball of fur
happy kitty something kitty purr purr purr

[pallid path]

I just don't know how to really compare it

[untold crescent]

my main issue on comfy was... that you had to manually prompt both base and refiner - which would be fine if you wanted different prompt for them but silly if you do want to reuse the same

you'd think they'd have some text input node

[lilac raven]

It's dying but going on the laptop. I'll have to play around with the options on the desktop

[pallid path]

the same prompt is literally connected into both the base and the refiner lol

[heady forge]

do i need cmdline args for 1080 8gb for sdxl on comfy

have only used a1111 before

[untold crescent]

how?

did they change comfy's nodes recently?

because before you had to wire it so that you couldn't reuse the same prompt node

[delicate grotto]

not really, 6gb is enough for 1440p on sdxl, you're gonna be fine

[pallid path]

try the new example thing

[frank steeple]

Not working for me? Does discord strip the metadata?

[delicate grotto]

try to drag it without zooming in

[pallid path]

drag and drop this into ComfyUI (make sure the save or backup your previous workflow)

[heady forge]

yea i thought so, doesnt seem that slow, just model loading takes forever

gonna try on ssd

[pallid path]

idk if discord destroys the metadata

[lilac raven]

32 Euler a steps in 8 minutes isn't that bad for 1024x1024, I think

[thorn lion]

alright. lemme try again. i'm on windows 11, i have chrome open with somewhere between 50 and 100 tabs open. I have a dual screen running that is also using the gpu. it takes forever and a half to load the models from the hard drive to the ram to the page swap on the hard drive. but once all that is done and it gets dumped to vram, it's probably taking less than 30 seconds. If i upgrade my ram to a faster ram, I'd be faster. if i upgrade my ram to be larger, i'd be faster. but i am quite intimately familiar with how fast my computer is, not just at Stable Diffusion, but things in general. I'm moving around the speeds that I normally get when doing anything else.

so, if i TRULY should be rendering 20-30% faster than you, then if you are available for a private deep dive into my computer settings, that'd be great. outside of that, there isn't really much else for me to try outside maybe a different workflow?

But I like Sytan as a person so i'm biased in my choice of workflow

[slender timber]

I've been able to save and then drag

[brazen patrol]

Can someone tell me what I need to do in Comfy (via Swarm) to use both Base and Refiner? Is there some configuration I need to do? I am only shown the option to choose one model or the other.

[proper cape]

anyone has idea how to make SD created a real tileable pattern? it created mismatched tiles. please help... Any prompt to fix this?

[polar epoch]

What can i change/disable to not redo all through ksampler, like different seed, cfg and whatnot, but simply just choose upscaler and denoise? As k sampler kinda messes too much with it. Does comfy have anything similar to highres fix in auto?

[azure oxide]

with somewhere between 50 and 100 tabs open
i think we found your issue here

[clever verge]

alien style works well too.

[delicate grotto]

mmm... the 100 tabs might be it with the dual screen
i used to lower screen res to make the renders faster, try to turn off the 2nd screen and to use firefox or something else for the renders while closing chrome

also, go to your nvidia drivers and make sure you are on full power

+++ make sure than you are not on HYBRID MODE

check with your laptop type if you are on one
it is tremendously slower on hybrid mode

[worthy umbra]

Turn that off in Chrome to stop it from using the GPU at all.

[brazen patrol]

Anyone? Here is all I have. One or the othr. not both

[delicate grotto]

there is an option for that?

@worthy umbra show me where it is

[thorn lion]

yeah that's what i do when speed matters but i enjoy multitasking more than fast generations. my often, i have anime on one screen, a video game on the other, and three or four other things in the background. i really need to fix my second computer

[wary bay]

@quasi remnant I have virtually no difference between using the 0.9 Base VAE you linked (left), and the 1.0 Refiner VAE (right)

[worthy umbra]

chrome://settings/system

[royal tangle]

How do we get NSFW content, will training the model be able to make that happen?

[delicate grotto]

ty ty ty

[thorn lion]

IF i didn't leave chrome on the second screen, which forcibly needs to use the NVIDIA GPU and not the iGPU

[quasi remnant]

I can see the issues by my naked eye.

[jagged ferry]

Any extension recommendations for anime art?

[hasty cipher]

makin some liminal spaces

[unique beacon]

Yes, fine-tuning should do it

[quasi remnant]

it's easier to see in more detailed areas

[wary bay]

Left was 0.9, right was 1.0

[brazen patrol]

Is there a different Discord area where I can get help on this?

[quasi remnant]

when you use 0.9 vae encoder you have zero of those lines

[wary bay]

1.0 looks better to me. Lol

[brazen patrol]

I can see this is not it

[hasty cipher]

we already have aweosme stuff with fintuned sd 1.5, imagine what finetuned sdxl is capable of

[quasi remnant]

To me 1.0 looks waaaaay worse, i can see all the lines

[radiant tartan]

hey, thank you! im finally up and running 😄

[quasi remnant]

like here too

[thorn lion]

sorry i missed you

[trim orbit]

yeah the artifacts your'e getting are apparant. not disputed.

[knotty kettle]

is there an audio recording of the stage event today?

[thorn lion]

actually, that's a lie. i saw your message but i assumed someone with more knowledge would answer

[delicate grotto]

so far it went further than most top tier or head to head with the best specialized... specialized models

[thorn lion]

alas.....

[unique beacon]

I'm working on this now 😎

[brazen patrol]

alas...

[trim orbit]

but you getting them, and me not getting them, suggests something

[pallid path]

if you need a workflow, just download the image (open original) then drag and drop it into COMFYUI

[south igloo]

is there an idiots guide to comfy? it looks so confusing with all the boxes and lines and stuff?

[quasi remnant]

i don't get them; i use 0.9 vae

[lean knot]

It's weird that they didn't see the VAE problem before release

[fleet veldt]

very odd. getting the same with more steps - also tried the model on the install upstairs and it does the same thing too..

[thorn lion]

anywho, use Sytan's workflow. it asked me for the refiner and the base so i assume it's using both

[trim orbit]

you're intentionally getting them. do what works. don't diagnose the entire vae1.0 as busted when youre research is inconclusive

[eternal fog]

It's very subtle, it's messing with the colours, here's a good example. Notice the weird green fringing around the eye.

[brazen patrol]

what is Sytan's workflow?

[thorn lion]

@brazen patrol https://github.com/SytanSD/Sytan-SDXL-ComfyUI/raw/main/Sytan SDXL Workflow v0.5.json

[quasi remnant]

that was someone elses post, lol.

i can point it out whenever anyone posts a pure 1.0 image

[olive flare]

I had this problem before because the VAE used by automatic was the one from SD1.5 in the Settings...

[quasi remnant]

i see the artifacts in 'every single' 1.0 vae image

[fiery plover]

Drag and drop anywhere? or I need to clear the workspace first? new to Comfy and doesnt seem to work

[trim orbit]

that's just trolling. it's likely because people aren't configuring the right launch option.

[timber quiver]

i have same issues with 1.0 vae, no lines with 0.9 vae

[trim orbit]

some people are getting the issue and that's fine. it can be fixed without recommending a restricted vae

[autumn forum]

kitty

[thorn lion]

u too fast. i was coming for u. i got it to load and i got it to make an image when i click generate. so, currently, i'm satisfied until some of the kinks get worked out

[slender timber]

i'll be honest, I'm in the "i can't really tell crew"

[delicate grotto]

yeah, 1.0 vae is borked

[pallid path]

if you aren't afraid of overwriting your current workflow then yes just drag and drop

[brazen patrol]

Ok, and I place this JSOn where?

[thorn lion]

as if i wanna be in the middle of this 0.9 vs 1.0 VAE convo rn.....y'all figure it out and lemme know

[royal tangle]

mew

[pallid path]

if that doesn't work then download from here: https://github.com/comfyanonymous/ComfyUI_examples/blob/master/sdxl/sdxl_simple_example.png

[agile quarry]

I haven't tried StableSwarm but the GitHub page seem to says not implemented yet

[trim orbit]

its not. people are just not launchiung with the right launch switch and a troll brigade is latching onto that issue to misrepresent it

[quasi remnant]

i've used a computer all my life so i'm used to pixels and being able to see discrepancies, my vision is sharp and i see those lines in every 1.0 gen. There's no issue if someone prefers pure 1.0 but it is very visible to me.

[pallid path]

and drag and drop that

[lean knot]

Problem is that people finetuning with 1.0 VAE may encounter problems

[lilac glen]

Any thoughts on whether to use the same seed as the base in the refiner, or a different seed?

[fleet veldt]

what's odd is i've downloaded this 0.9 vae - put it in the VAE fold AND defined it - no change..

[trim orbit]

big fat ianal disclaimer, a restricted license release is not a good fix

[quasi remnant]

i heard it doesn't load in a1111 for some reason if you're using that. if you're using comfy it should work.

[hazy moat]

anyone installed swarmui on linux?

[clever verge]

Graffiti is also amazing. But also a fun 3D-effect on the shoes.

[hazy moat]

I am getting an error that I can't debug

[fleet veldt]

i am infact using the latest 1.5 auto1111111111

[lean knot]

I agree, but it's the one we have now

[heady vale]

my very first XL 1.0 local gen

[untold crescent]

auto1111 loads the vae automatically if you put it on stable-diffusion folder and rename it <model name>.vae.safetensors

[lean knot]

because it seems that staff can't even talk about it

[untold crescent]

kinda hardcoded that way

[fleet veldt]

i'll try that. perhaps it'll work around the bug

[agile quarry]

two prompts or the same prompt to both encoders? what's up with the W & H on clip encode?

[untold crescent]

you need to restart auto1111 as welll

[thorn lion]

uhhh one moment.

[cinder aspen]

Did you guys try the Swarm one?

[fleet veldt]

stand-by

[jagged ferry]

What’s this do?

[fiery plover]

interesting. it works, but i'm not used to comfy. Just overrides everything and whos your workflow & nodes when I drag and drop

[thorn lion]

helps you generate

[fiery plover]

but it worked, so thanks @pallid path

[delicate grotto]

can you please explain?

[jagged ferry]

Ohh okay I’ll give it a try

[thorn lion]

do you know where it goes?

[quasi remnant]

i didn't wanna keep engaging with that, i think they're trolling the chat. lol. the lines are very apparent in any gen, apparently devs can't talk about it though so... people should just use whatever they prefer

[jagged ferry]

Load it up in comfyui I presume

[thorn lion]

just making sure. i wasn't certain how to interpret your initial question

[heady vale]

For those who have a GPU with more than 10gb, like me, does XL still use roughly 30gb of system RAM?

[delicate grotto]

thing is that i know flowwolf knows something...

[hasty cipher]

"devs can't talk about it" i'm calling bluff, this sound nonsensical

[wary bay]

are we not allowed to talk about this anymore lol

[rotund cliff]

is there something like Adetailer for Comfy?

[jagged ferry]

My bad I think I worded my question wrong

[wary bay]

I don't want to be labeled a troll 🙄

[jagged ferry]

Is it better than that .9 refiner json?

[thorn lion]

this is sytan's 0.9 workflow. he's still working on the 1.0

[quasi remnant]

was a posted+deleted message, i'm guessing they really aren't allowed to talk about the topic for some reason.

[delicate grotto]

@trim orbit if you can tell me how it works, id really apreciate it

[agile quarry]

is there documentation about clip_g clip_l ascore and all the other new parameters ?

[thorn lion]

and i don't know who else is out here with a workflow

[jagged ferry]

Okayy thanks

I’ll give it a try

[eternal fog]

I just tested it out on a drawing and it's even more noticeable

[wary bay]

its like chromatic aberration

[quasi remnant]

yup. right is way better

[lean knot]

what? It was in this same channel the comment

[eternal fog]

And that's 0.9 on the right

[outer oxide]

Is there documentation available on what to do with the two text encoders?

Should we be sending the same prompt text to both of them?

[eternal fog]

Yeah someone else said that to me as well

[jagged ferry]

https://github.com/camenduru/sdxl-colab/blob/main/refiner.json I was using this

[clever verge]

Yes

[quasi remnant]

ye, but it got deleted, so i don't think we should try to get any devs involved.

[wary bay]

it's the photography term for that effect lol

[delicate grotto]

yeah, the 2ndary prompt is the style one, my workflow has every style note in it in a large notepad
you should see it here on the left side of the workflow

[thorn lion]

yes to your first question. no to your second. unfortunately, i'm juggling responding to another question but will hunt this answer down cause it keeps getting asked

[delicate grotto]

its like a regular prompt, but shortened into a single word, basically an embedding but with direct prompts

[lean knot]

I'm not saying you should call the devs, just that it's not nonsense

[quasi remnant]

ye

[lean knot]

Probably a watermark for SDXL or they don't want to talk before they have a fix

[quasi remnant]

i'm happy with 0.9 vae results. zero artifacting in any image

[clever verge]

I'm sure they are aware of it by now

[lean knot]

Not a great look either way

[wary bay]

😂

[quasi remnant]

i think they were aware within like 20 mins of launch

cuz told outerside about it back then

[rain compass]

I got that artifact with sd next. But it works fine with comfyui

[delicate grotto]

gn

[compact flax]

If you see these artifacts on the 1.0 releases, appears to be a bug in the VAE. 0.9 VAE doesn't have it. Attached are 1.0 vs 0.9 VAE. Glad the 0.9 VAE is MIT licensed, too.
https://huggingface.co/stabilityai/sdxl-vae/tree/f85c10757eb0bda0222090261de54c5fe17a7f4f

[hard fractal]

It's not a bug.

[quasi remnant]

artifacts on left, none on right

[compact flax]

I'm sorry?

[royal tangle]

feature

[hard fractal]

It's a feature ❤️

[compact flax]

I'll continue to use the unhindered 0.9 VAE

[trim orbit]

--no-half-vae on auto gets rid of it. i dont see it in comfy either

[agile quarry]

Joe is there documentation on clip encoder options?

[quartz sequoia]

the training documents joe 🥺 where are they... i want i want to traaaaain da loras

[hard fractal]

we need sleep

haha

we'll start working on it next week

[quartz sequoia]

sleep

is

for

the

[high skiff]

ping

[spark bear]

its always the same moire colors, some kind of resolution/noise mismatch

[quartz sequoia]

ping

[hard fractal]

"sleep is for the ping"

[nocturne dove]

Start up a discord bot for Joe Penna that just replies "no" to every ping.

[royal tangle]

Host Unreachable

[worthy umbra]

The Dreadnoughts - Sleep is for the Weak: https://www.youtube.com/watch?v=N-M224HdX4M
Goodnight

[fresh path]

emad said only the preview is released today

[quartz sequoia]

sleep is for the peng

[high skiff]

SDXL 1.0 does indeed seem to work well with my 0.9 LoRA's, which is nice

[quartz sequoia]

what preview

[nocturne dove]

Yeah the loras I had seem mostly ok. I'm still retraining them anyway because why not.

[hard fractal]

no.

[nocturne dove]

That's the spirit

[eternal fog]

It's definately in comfy, if you aren't seeing it I'm wondering what's different about your setup compared to everyone elses.

[fair crow]

Did anyone have luck with LORAs for faces yet (non-portrait that is)?

[fresh path]

well he said "Fine tuning API preview launches today"

[quartz sequoia]

ahhh

[trim orbit]

ada arch? ¯_(ツ)_/¯

[dusk carbon]

Wait does Joe Penna work for StabilityAI now? Must have missed that, remember the early Dreambooth discord days trying to figure things out 😂

[high skiff]

gonna be testing soon to see if 1.0 does muscles better, cause that was a big downfall to 0.9

[shell oasis]

honestly i should not have been trusted with this power

[unkempt fractal]

so how does refiner work in A1111

[thorn lion]

i'm cooommmiiinnnng one sec

[molten crypt]

Is xformers automatically enabled on ComfyUI? Also, can something like "--medvram" be used?

[high skiff]

I suggest turning off xformers for comfy, i found its about 10% faster for me without it on

and yes, it does have its own VRAM control commands

don't know them off the top of my head tho

[fresh path]

nice to know hand test still works

[high skiff]

1.0 is out now, so it will all be made on this as a bse

[rich coral]

I don't think it does yet

[trim orbit]

loras made on younger models work on future refinements usually. 0.9 loras all work well too

[high skiff]

yeah

[polar epoch]

With ESRGAN 4x

200 fuckin megabyte

[fleet veldt]

https://tenor.com/view/shaggy-scared-zoinks-surprised-gif-15872371

[gilded plinth]

whaaattt

[high skiff]

WHAT DID YOU DO TO THIS IMAGE

IT SHARPENED THE PIXELS

i am at a loss

[heady forge]

lmfaoooo

[polar epoch]

Aye lol. I testedd 8x, it ate like 20GB video memory and 40GB ram, plus a few on pagefile

[warm hazel]

Hmm. I was hoping to not need negative prompts but it seems like you still get the best results if you follow the positive and negative prompts in the style .csv they posted.

[high skiff]

now thats what I call a "pixel upscaler"

lmao

[trim orbit]

side note i just want to rant about. i've seen some of the ramblings around some conspiracy ideals lately about watermarks on the model and how that should be illegal or how its because of AI regulation and ugh.. it's all so dumb. Think about it for 2 seconds. The same people who complain about it are demanding every shared pic has meta data

[lavish sinew]

https://tenor.com/view/there-it-is-well-there-it-is-gif-8569483

[high skiff]

makes the 1.0 VAE issues even more clear haha

[trim orbit]

conspiracy theories are dummmmb. that is all.

[scarlet pine]

How do I get that text to image page when you click the purple button

[polar epoch]

There, jaypegg quality, but eh, you can see it xD

[fleet veldt]

UPDATE : My auto1111 mess is fixed - had a hard coded vae in the command line args.. works fine

[high skiff]

slaayyyy

[polar epoch]

Fuck sake discord

[cerulean saddle]

i'm having this issue where every image that isn't of some sort of fantasy/strange type has the person coming out like they are a 3d render... even when prompted to be the opposite. anyone else having this issue? eyes are fantastic though.

[warm hazel]

Is the vae not being automatically applied or something?

[candid spindle]

The dot crawling artifacts happen in the 1.0 VAE only right?

[high skiff]

bruh

[polar epoch]

I generated an image larger than discord can show

[high skiff]

[gilded plinth]

lol

[delicate grotto]

There was a dude who put a 16k img in discord, my pc bsod almost, in the end just the driver crashed
Anyone who was in chat that day coulsn't close discord or touch it

[polar epoch]

Using th9is lol

[clever verge]

chromatic aberration as a negative did not work. 😄

[polar epoch]

Couldn't get these upscalers to work properly

[trim orbit]

it wouldn't. it's a precision error

[clever verge]

I'm on 4090 btw, and pytorch nightly

[tropic turret]

[polar epoch]

Keks. You don't wanna see my 63k nvidia ansel image then :P 960MB was it

[azure oxide]

jesus some seeds of dragons are cursed. if its not crazy eyes for nostrils, its xenomorph jaws lmao

[delicate grotto]

At that point you might render the moons surface

[trim orbit]

really seems like theres a brigade operating here of people choosing to zoom in to cropped artifacts. I'd rather see high fidelity megapixel images plz

[polar epoch]

I'd love to tbh

[clever verge]

Yeah, your right about that.

[trim orbit]

i know there are resentful side channels on secret servers. please resent to yourselves.

[clever merlin]

Finally got my first image to work. Using automatic1111. 8gb vram, took 38secs, no refiner

[dire rampart]

the pipeline PNG states the size of the incoming image must be 128 before the latent is improved in size and the 1024 size is achieved in refiner, but the image result from 128 is not intelligible with the SDXL 1.0, is there any way to begin with a smaller image or is the 128 option just showing us that it can be of any size

[thorn lion]

Generate tab

[delicate grotto]

Last lowres img of the moon that i saw was i think 12gb?
Maybe 5sqrkm per pixel?

[trim orbit]

is beautiful . reminds me of a trance mix album cover i used to have

[indigo carbon]

still artifacts? I'm blind

[thorn lion]

when i look in a mirror, i see issues. i don't needa zoom in to find problems

[delicate grotto]

Anyway my pc almost died on 16, im afraid to touch the moon

[untold crescent]

comfy seems way faster even when auto1111 runs somewhat fine

[polar epoch]

There we go. 4x ultrasharp
Only 100MB

[lavish sinew]

If you don't find it interesting that there are intentional artifacts built into 1.0 VAE then that's on you, buddy

[clever merlin]

was just a super simple prompt, but turned out good. I got tired of typing a long prompt just to get black images lol

[peak wraith]

Comfy does seem way faster and I can’t figure out how to get the refiner to even load in a1111

[cyan crown]

But refiner in auto111?

[delicate grotto]

Looks good from the phone

[thorn lion]

before anybody asks about the clip encoders again, i'm looking for that answer i saw some time ago. i'll find it

[heady forge]

has anyone got a nice 1.0 workflow with 0.9 vae? ive never used comfy before

[trim orbit]

intentional? i think you took joe's comment too seriously

[slender timber]

yea, outside of deforum I can't see myself going back to a1111

[cyan crown]

Comfy or auto1111

[trim orbit]

yeah it's a brand new day. looking good

[jagged ferry]

What do you use instead of that?

[slender timber]

comfy

I switched over last week

[shy kelp]

comfy is definitely where it's at

[jagged ferry]

I feel like comfy takes forever to generate anything for me😭

[obsidian jackal]

Has someone trained a LoRA using more than 1 GPU at the same time? Using kohya scripts, I'm not sure if I should set the epoch amount to half of what a single GPU training would do

[shy kelp]

getting some pretty sad results with automatic and sdnext

[polar epoch]

Is upscale latent something you do after the seed, denoise and all that? Or before?

[short marsh]

Does architecture very well

[royal tangle]

Let's compare the same prompt on auto1111 vs comfy vs SD-NEXT as close as we can to see how it looks.

[shy kelp]

auto seemed fine to me but idk. but seems a pretty good base model for sure

[slender timber]

Yea, the interior backgrounds on the images are stupid good

[shy kelp]

do you have all three?

[royal tangle]

just two (auto + sd-next)

[trim orbit]

love the grass whiskers

[clever verge]

Like that it has proper whiskers

[shy kelp]

gen something on auto and sdnext and post the prompt

[hushed hornet]

None of my images are coming in even close to prompt, but we're all in this together! Looking forward to seeing more suggestions on prompts + workflows for ComfyUI.

[floral island]

well, i did prompt biopunk... xD

so some organic stuff is bound to happen

[shy kelp]

whose workflow is this? yours?

[floral island]

partly sytan, and then i just started tacking stuff on

[clever verge]

Yeah, saw the prompt. Could have been metallic ones, though. 😄

[outer oxide]

organic stuff? ew. can you get a tonic for that?

[shy kelp]

recognized the saytan bits

getting pretty good gens out of it

[slender timber]

I am finding it to be very hard to steer. Seems like it just wants to do whatever it wants to do

[turbid gulch]

SDXL 1.0 is now live on www.playgroundai.com

[shy kelp]

no one cares

[vital wolf]

Can sdxl create niji quality anime images?

[turbid gulch]

Hundreds of thousands of people disagree with you.

[shy kelp]

i only see three

[floral island]

this better? 😄

[solar tusk]

where to download .9 vae?

[clever verge]

That's the ice version of this.

[vital wolf]

How good is sdxl with anime? Better than niji or comes close to it?

[clever verge]

Damn!

[candid drum]

Anyone having trouble training a Lora on 1.0? Im running a 4090 and it's currently taking over 3 hours for a simple 2800 step Lora. 4.0s/it

[fathom mango]

you made this locally ?

[floral island]

yeah, local generation

[fathom mango]

could you share the prompt and neg ? People on reddit try to get same results as the discord bots, but it seems it wont work

[unreal bobcat]

I create anime like pictures without even using the anime tag. It seems to really easily slip into anime

[fathom mango]

https://www.reddit.com/r/StableDiffusion/comments/15ajpel/discord_bot_vs_local_10/?sort=new

[untold crescent]

I was really wondering about this.. because I don't see any visual difference between the 0.9 and 1.0 VAE here

[floral island]

it's a comfy ui image, so if you drag the image into a comfy ui window, you'll see the whole workflow

[fathom mango]

true

[radiant tartan]

is there a trigger word for the official kaliyuga lora?

[untold crescent]

but I also ran --no-half-vae on auto.. and comfy didn't show difference either

[turbid gulch]

My first splash art image in 1.0

[jagged ferry]

How you get rid of the lines?

[compact flax]

Appeal to Hypocrisy is a falacy. Just because they want or don't shared metadata doesn't mean images should or shouldn't be covertly watermarked.

[lilac raven]

The desktop died again, but at least I got something not that bad

[vital wolf]

Thanks

[fathom mango]

This isnt sdxl ?

[quasi remnant]

the lines are caused by the vae baked into sdxl 1.0

[south igloo]

all you guys making beautiful things and im just making nightmares lol

[trim orbit]

yeah. those who are preaching the "it's busted" aren't considering that their own configuration might be causing the bug. its not a universal fix.

[fathom mango]

or you renamed the sdxl to "maybe a good model? " xD

[floral island]

just slap in some anime terms, usually works ^^

[sweet lily]

so silly me didn't install A1111 correctly, I'm looking to update the current "install" with git pull, but it was just originally an unzipped file. l how do i do this?

[south igloo]

the prompts are certainly smarter at understanding what you want lol

[jagged ferry]

Any way to fix that?

[south igloo]

whats the size limit for comfy? seems to be working at 4096x4096 just 6 seconds per it

[quasi remnant]

Yes; but i'm not sure what the stance is now on explaining the fix lol. it might be illegal

[jagged ferry]

Oh snap😭😭

[fathom mango]

Aynone got same issues ?

[royal tangle]

which is which?

[clever verge]

Not sure if it's manga enough.

[thorn lion]

most people won't bother to respond to your negative self to preserve their sanity. i am not so wise.

[fathom mango]

left discord, right local

[polar epoch]

Think i might favor automatic1111, as dropping image over to comfy uses the image's OG nodes, and not the setup i had lol. Would love if comfy instead added outside the boxes nodes to add to what's missing and auto node to where it needs to be :P

[trim orbit]

the discord bot will use random cfg values , step counts, schedulers, while you have control over all of that locally

[thorn lion]

"it didn't work for me so it mustn't work for anybody. my experience is a universal experience" /s

[trim orbit]

not what i said

i guess that makes you part of team blocked

[hushed hornet]

I have to be doing something drastically wrong, not blaming 1.0. But everything I try (with various workflows) are soooo far off from the prompt and minimal quality compared to what I got with 0.9. Bummed I deleted that VAE already.

[fathom mango]

jup

[quasi remnant]

it's been confirmed "busted" now. there's no point defending it.

[thorn lion]

no that's what THEY are saying. i agree with you. hence the slash s to denote sarcasm

[jagged ferry]

Where do I download workflows for comfyui? They got a discord?

[solemn scroll]

what is wrong with my SDXL? All the outputs are straight up bad. Here are my flags btw :
--opt-split-attention --xformers --medvram

[trim orbit]

brigades from resentful side secret servers are always so dumb

[compact flax]

Then finding a universal fix or at least a checklist should be priority considering it's an image quality issue either way, correct?

[floral island]

wrong size: you need 1024x1024

[south igloo]

using the release candidate? or the older version of it?

[slender timber]

lol, he was agreeing with you @trim orbit

[thorn lion]

no did i just lose flowwolf because i was misinterpreted and they didn't realize i was agree fuck imma be sad

[lean knot]

Too many people with fresh installs for SDXL are having the same problem

[unreal bobcat]

Try changing vae to automatic. Sdxl doesnt like old vaes in my experience

[lean knot]

and staff/devs don't or can't talk about it

[thorn lion]

they're one of the good ones......sigh

[quasi remnant]

they can't

[trim orbit]

its an issue for some. undeniably. but there are those who are hyping the wrong idea around the problem. let it sort out in time.

use what works. if 0.9 works for you, use that. i just don't see the point in going aroudn shitting on everyone's generations

[solemn scroll]

I'm using the hugging face 1.0 release

[compact flax]

When you're so blinded by hate and ignorance you hate people who agree with you

[south igloo]

of automatic i mean

[lean knot]

It's a little silly to blame users for their setup

[warm hazel]

I mean I really want the staff to come out and say "the recommended settings are this:"

[torn canyon]

Yeah I tried every possible setting in comfyui to fix it, and only reverting the VAE helped to solve the issue

[thorn lion]

imma give them the benefit of the doubt. not everyone knows about the /s sarcasm tag

[solemn scroll]

I just git pulled a few hours ago, I'm on the main branch I believe

[torn canyon]

So there's something weird going on for sure, whatever it is

[compact flax]

True, I see

[trim orbit]

just stop. really.

[quasi remnant]

It's literally stated as #✨｜sdxl message it doesn't get much clearer than that. this is directly quoting an image with the noise which means that it's acknowledged & intentional to be messed up like that.

[floral island]

do my images have it too?

[azure oxide]

not silly but after helping out people on reddit and here for months, it's definitely to be expected =]

[unreal bobcat]

Also that image size is probably too small to work. 512x512 is probably the lowest you can go

[trim orbit]

let it go. y'all are being so obvious

[floral island]

no, 512x512 will also fail

[compact flax]

Elaborate?

[quasi remnant]

don't bother; floww is taking everyone off-topic & trolling. just ignore/block him.

[hushed hornet]

Ok I take it all back this one came out great.

#jesusinthestreetsspaghettiinthesheets

[azure oxide]

there was a dev quote from the stage chat earlier saying "Please don't generate 512x512 on sdxl :)" or something like that. I don't think it makes any sense to do so.

[south igloo]

now im hungry

[solemn scroll]

Just did. @floral island helped with his suggestion actually, 1024x1024 runs nicely on an RTX 2060, I'm suprised

[trim orbit]

this makes it sound like i'm the only person ever who doesn't have this issue. seroiusly stop mis representing things

[warm hazel]

Horrifying. I love it.

[slender timber]

the sai-comic book style seems to do pretty well

[floral island]

i need to rinse my eyes after that lol

[quasi remnant]

you haven't posted any examples with a workflow that works that doesn't produce that issue with the pure 1.0 settings, until you do there's no reason to believe.

[thorn lion]

someone tell flowwolf that I was parodying the "those who are preaching the 'it's busted'" by attempting to satirically portray their perspective of their individual configuration and subsequent fixes being equivalent to a universal experience. I am well aware that this is not a universal experience and, tbh, i don't even have an opinion about the VAE or whatever else is broken.

[supple latch]

Is there anything additional you’re using to generate that image accurately? Is that sdxl by itself? Teach me because that piece is awesome.

[wary bay]

[shy kelp]

ayo?

[trim orbit]

yeah. i don't need to satiate you though.

[quasi remnant]

now that's a clean image.

but that's 0.9, yeah?

[floral island]

that's just sdxl 1.0 stock on comfyui

[quasi remnant]

there's no clean examples of 1.0

[polar epoch]

This. It's a 1024 model :P It's like 512 models, doesn't do too low good.

[dull cairn]

is anyone else getting really noisy images with auto1111 1.5? i havent changed any settings from the default but all of my images look like they have the equivalent of heavy jpeg artifacting, this is with SDXL

[floral island]

the workflow is actually in the image: if you drag it into a comfyui window, you'll get the exact setup i used to get it

[compact flax]

That's not the gotcha you think it is

[lilac raven]

I'm positively surprised, the first picture I made took 8 minutes, the second took 5 and now it's consistently 5. I was scared I would no longer be able to use the laptop for SD

[untold crescent]

is there a way to made these fit on comfyui?

[stray lake]

any kind soul want to walk me through adding a mask to comfyui workflow? Im using vae encode inpaint and it's taking forever

[wary bay]

never going back

[supple latch]

I never knew that

[west breach]

omg you guys! 1.0 release day! 🎉

[turbid gulch]

My first text attempt in 1.0. Here is the prompt:

text saying "SDXL" written in beer and suds with a black background, realistic, splash art, shiny reflective surfaces, liquid, ray tracing,

[floral island]

so you can use them in multiple parts of the generation?

[trim orbit]

when there's a brigade attempting their best to school yard brigade me, it's a "try harder" that i think it is

[torn canyon]

I guess it's likely that the VAE issue only shows up on certain hardware or versions of pytorch or similar, so that's why some people don't see it

who knows

[quasi remnant]

same... means we can't use the images commercially though lol

[azure oxide]

curous, where was the confirmation? i mustve missed something

[thorn lion]

i haven't even given any opinions on anybody's generations....what happened here and how do i repair it?

[wary bay]

I have never sold anything ai beforeee

[quasi remnant]

look at my post after that, i linked the confirmation

[azure oxide]

ah kk

[wary bay]

I'm just here to have fun 😂

[quasi remnant]

same, just the typical commercial restriction stuff applies cuz this is a really restrictive license lol

[smoky badge]

SDXL alpha (dreamstudio) vs SDXL 1.0, a big improvement

[wary bay]

yeah...

[timber quiver]

imo vae 1.0 is optimized for jpg, no artifacts there, vae 0.9 is for png format
check vae 1.0 in jpg and png:

[compact flax]

If you have settings that work, the quickest way to shut them up is probably to prove them wrong. Even if they don't stop, you've helped people troubleshoot a legitimate issue and your credibility remains in tact.

[floral island]

is that mathilda from leon?

[trim orbit]

thats my suspicions too. but people insist on childish brigades while they try to dunk on the team that did all the hard work. its the way they're misrepresenting thigns that is just.. well... toxic

[nocturne dove]

Sort of. Not the best likeness.

[thorn lion]

flow what did i do please

[quasi remnant]

left is 1.0 but jpg?

[supple latch]

I’m new to sdxl and all these additional add-ons so I appreciate your insight. Thank you.

[timber quiver]

yes, SwarmUI saves images as jpg by default

[thorn lion]

someone please tell flowwolf that I was parodying the "those who are preaching the 'it's busted'" by attempting to satirically portray their perspective of their individual configuration and subsequent fixes being equivalent to a universal experience. I am well aware that this is not a universal experience and, tbh, i don't even have an opinion about the VAE or whatever else is broken.

[timber quiver]

imo thats the feature

[azure oxide]

any ideas on saving comfyui outputs as jpg? don't see an setting anywhere

[vale quartz]

Base vs refiner in comfy. Im impressed.

[quasi remnant]

yea i'm not sure, i'm trying something rn but i don't think it'll work

[floral island]

i'm using the WAS extensions, so i can use a single number (for example, the number of steps, or when to change from base to refiner in a single box

i feel that makes it a bit easier

[south igloo]

how does it look if you do a full body shot? im getting super crazy long limbs lol

[lavish sinew]

Nobody is trying to "dunk on" the team. We're asking questions. I understand that you've decided to be a stalwart defender, but it's goofy.

[azure oxide]

all i can find is this https://www.reddit.com/r/comfyui/comments/13uw3w2/is_there_any_way_to_save_output_as_webmjpeg/
but if you read the warning from the dev, even he isnt very confident in himself lmao

[south igloo]

it does portraits well but the hands and arms and legs seem crazy still

[compact flax]

Probably won't be of any use, mans seems too blinded by his own hatred to see that, otherwise they would've gotten it from the get go.

[nocturne dove]

Are you using 1:1 aspect ratio or did you change it. If you change it...

[south igloo]

1024x1024

[quasi remnant]

yea, and my trick didn't work either, guess that's for whatever swarm does. Not sure why it's such a big difference?

[untold crescent]

is anyone actually having 1.0 VAE issues with comfy?

[vale quartz]

I had the same problem. Had some really aweful hands aswell lol

[compact flax]

Yes, I had it on Comfy

[slender timber]

Yea, lots of weird limbs. Adding back in some of the 1.5 negative text embeddings helped a little, but I'm sure there are issues with that

[torn canyon]

Yes, I have issues

[untold crescent]

can you paste here comfy image that has it?

[warm hazel]

prompt:geode inspired spaceship, flying through solar system, nebula, black hole, best quality steps:50 width:1024 height:1024 guidance_scale:7.0 sampler:DPM++ 2M Karras seed:1649116019 data_model:SDXL

[nocturne dove]

Hmm. Usually I get the "long limb/torso problem" when changing aspect ratios.

[untold crescent]

so I could try to see the difference on my end

[quasi remnant]

Hmmm. But this also shows no real difference btween the two images, ergo, if 0.9 is for png format that's better imo since i like pngs. Interesting discovery tho...

[untold crescent]

so far I don't see any difference on 0.9 and 1.0 VAE here with comfy

[thorn lion]

i would prefer for flowwolf to make that decision once given the information. if my word for word explanation can be given to them, that would be appreciated. i would appreciate an advocate. anybody. to just copy paste it.

[torn canyon]

Look carefully at any high constrast edge

[thorn lion]

someone pretty please tell flowwolf that I was parodying the "those who are preaching the 'it's busted'" by attempting to satirically portray their perspective of their individual configuration and subsequent fixes being equivalent to a universal experience. I am well aware that this is not a universal experience and, tbh, i don't even have an opinion about the VAE or whatever else is broken.

[torn canyon]

zoomed in

[untold crescent]

no I mean, full image generated with comfy.. so I can get the metadata

so I can try to regen it here

[compact flax]

One second

Does Discord wipe metadata?

[agile quarry]

I'm not sure what you're asking but maybe this is what you want https://github.com/ssitu/ComfyUI_NestedNodeBuilder

[azure oxide]

not the ones we care about

[untold crescent]

doesn't appear to

[vital wolf]

How are full body shots coming along in sdxl?

[warm hazel]

It used to, but I think it doesn't now.

[compact flax]

Thanks

[urban fjord]

It should work.

[south igloo]

lol

full body shots 😛

maybe i needed more negatives like long limbs etc

[untold crescent]

I mean look at this, both lines have overlapping text.. if I could resize the node wider it would work.. but resize option doesn't seem to do anything

[urban fjord]

Full body shots works fine for me.

[compact flax]

Try this one

[quasi remnant]

No, Discord keeps metadata however you have to open the image in a new tab

[vital wolf]

Try it

[timber quiver]

i pref png too, png has loseless compression, jpg not, even at 100% quality

[quasi remnant]

it's nice to finally have confirmation that there's no difference between the two though when they're fully denoised

[vital wolf]

Can you do realistic one?

[agile quarry]

do a1111 imgs load prompts+ in ComfyUI?

[wary bay]

0.9 vae vs 1.0 vae, zooming is pretty vital for comparing for sure

thats my best example of any issues lol

[quasi remnant]

i don't need to zoom in xD i can just see it with my naked eye every time nearly

[slender timber]

I found the fix

[royal tangle]

I need to get 0.9 vae

[slender timber]

don't zoom in

[floral island]

full body shot? XD

[quasi remnant]

that's why it was bugging me so much

[south igloo]

thats really terrible quality

[quasi remnant]

but i'm glad it's not a big issue

[south igloo]

wonder what happened?

[untold crescent]

opens

[wary bay]

Lol well some people do need to see it like this

[compact flax]

Wow

[azure oxide]

im literally required to save as jpg's since i prefer to generate infinitely and cherrypick from a thousand or two of images, PNGs take up lotsa space fast with my methods lmao

[quasi remnant]

Yoooooo. Looks good

[trim orbit]

there's an over fixation and its due to one cause. brigade.

[civic sigil]

Maybe they should just release the 0.9 vae with the new license

[lean knot]

Some people refuse to live in the same reality

[quasi remnant]

doubt they can, not usually that simple.

[compact flax]

0.9 VAE is MIT licensed

[floral island]

"less bitching, more generating" 😄

[untold crescent]

hmmm, I've only tried VAE from refiner so far

[civic sigil]

Oh then why is everyone complaining

[untold crescent]

I'll do some testing

[compact flax]

https://huggingface.co/stabilityai/sdxl-vae/tree/f85c10757eb0bda0222090261de54c5fe17a7f4f

Here you go

License is listed as MIT

[south igloo]

still crazy long haha

[vital wolf]

How to get good full body shots then?

[south igloo]

no idea

[slender timber]

passing time between runs of course

[civic sigil]

Huh I thought that one was the 1.0 vae just released early

[gloomy sage]

@urban fjord like how ? Do we use the img2img pipeline? Would you share the workflow

[compact flax]

Nope, this is the 0.9 VAE. Either way, it doesn't have the artifacting for me.

[civic sigil]

Wouldn't it be possible for someone to create a ckpt with the 0.9 vae though? Not sure how that works

[urban fjord]

I don't get your question, in both ComfyUI and Auto1111 you can set the resolution to 1024x1024.

[south igloo]

artifacting is a deal breaker if you are going for high quality large images =\

[compact flax]

It doesn't have it, which is a good thing

[trim orbit]

#✨｜sdxl message

[compact flax]

SDXL 1.0 VAE does have it

[quasi remnant]

yup; and since i like my quality images i shall never return to the new vae lol

[compact flax]

From my testing, at least

[trim orbit]

no ckpts

[floral island]

this is our actual universe, we just don't know it yet!

[south igloo]

yeah and when people start blowing up the images to 4k and larger its going to be really obvious

[compact flax]

Good thing it's safetensors

[left coral]

I'm not familiar with Comfy, I'm a SDN user. Tried to load and got this

[indigo carbon]

maybe it's all just a huge safetensor 😉

[compact flax]

Even if it were CKPT, it's on the StabilityAI Huggingface account, you can verify the hash

[quasi remnant]

If you want a workflow w/ no custom nodes you could use mine (you HAVE to reduce batch_size though!!!)

[left coral]

Why? I have a pretty beefy pc

[quasi remnant]

24GB VRAM?

[left coral]

yea

4090

[quasi remnant]

then u can use my batch size lol

[left coral]

❤️

THANK YOUU

[compact flax]

I can run batch size of 4 on a 3060 Ti, what's your batch size?

[left coral]

This UI is WILD bro. Looks cool but feels over my head 😅

[agile quarry]

missing custom nodes this will help https://github.com/ltdrdata/ComfyUI-Manager

[left coral]

Just wanted to compare and see if SDN's implementation of SDXL has issues or not

[quasi remnant]

just prompt in green/yellow and then negative in red.

[untold crescent]

yeah I can verify the 1.0 vs 0.9 VAE difference with that image @compact flax

[quasi remnant]

more or less

[left coral]

you tha best, thank you!!

[untold crescent]

0.9 VAE from civitai and sdxl-vae repo seems fine

[quasi remnant]

you need the 0.9 vae though from

[untold crescent]

you don't need to sign any 0.9 agreements

I dunno why it has different hash tho

[quasi remnant]

from here: https://www.huggingface.co/stabilityai/sdxl-vae/resolve/e6e6fee4ec9c36d1f8e534a0874f7a773fb44bf2/sdxl_vae.safetensors to plug into the vae loader @left coral

[untold crescent]

1.0 VAE doesn't seem to glitch on all images

[left coral]

Oh wow you're the best, was already on it. But love the support in here!

[slender timber]

dreamshaper has an alpha out there that seems to work pretty good

[left coral]

Can I config the model/vae/whatever folder location?

[slender timber]

I'm playing with it now

[quasi remnant]

Not sure, i don't mess with much outside of the prog, but there's a place for vaes, models/vae in comfyui

[compact flax]

Diffusers VAE also works but has a different hash

[quasi remnant]

yea i'm using the diffusers vae

and it works great

[wet nacelle]

I don't understand.

[left coral]

true, I just have a ton of stuff already in my other SD frontend folders

[quasi remnant]

....and i'm not on diffusers.

xD

[inner ruin]

is this a LoRA?

[wet nacelle]

This is from the new UI

[gloomy sage]

I mean you said there is no need for inpainting model with sdxl, I am curious how to pass the mask ? AFAIK the unet model needs to be modified to have 9 input channels( latents from img, masked image and mask)

[molten crypt]

Does SDXL even begin generating in Automatic1111? It doesn't for me.

[urban fjord]

Just use the inpainting tab as normal in Auto1111

[compact flax]

🤯

[untold crescent]

well this explains it

[outer junco]

If you are on 1.5.0 it will but its somewhat buggy

[turbid gulch]

For text just use "text saying "INSERT HERE"

[untold crescent]

this hash is diffusors file here https://www.huggingface.co/stabilityai/sdxl-vae/blob/95f3fcb9e4f59982c462d0120529a5e7aa793453/diffusion_pytorch_model.safetensors

[outer junco]

Can someone please explain to me what the Refiner model is for?

[untold crescent]

again, no need for the 0.9 repo

[quasi remnant]

Is there actually any difference? lol. should i actually be using the regular one? the diffusers one weirdly worked fine

[agile quarry]

comfyui can use models from other setups https://github.com/comfyanonymous/ComfyUI/blob/master/extra_model_paths.yaml.example

[azure oxide]

the devs decided that the way forward with this model is to first generate a picture normally with the base model, then use a few steps with the refiner model to top it off. its mainly for details

[untold crescent]

see the announcement blog post https://stability.ai/blog/stable-diffusion-sdxl-1-announcement

[magic dawn]

What should be the training learning rate? Any official resources on this for SDXL?

[untold crescent]

basically this is similar workflow to what auto1111 does for highres fix but instead of swapping the resolution mid gen, it swaps to different "refiner" model for final steps

you can also swap the resolution for refiner I suppose

[winged tree]

probably better to just do that after in another pass

[gloomy sage]

Ok will try that out, I wasn't considering auto1111/comfyui, I'm more of interested in using the ckpts in my own inference pipeline

[unreal pecan]

so if we use a vae loader node with sdxl vae it will get rid of the artifacts?

[chrome furnace]

where do the windows "load vae" & "dualcliploader" come from?

[untold crescent]

I just drag the wire from VAE pin and you get VAEloader there

as one option

one could probably just swap the embedded VAE from 1.0 files for simpler use

[shy kelp]

hi should i use comfyui or a1111 what do u guys reccomend?

[untold crescent]

not sure if auto1111 repo can do this

[torn canyon]

Interesting, images generated on dreamstudio have exactly the same color fringing issues

[outer junco]

Now I just gotta figure out how to get it working in the latest Auto1111

[untold crescent]

comfy seems to work WAY better with SDXL

it's a lot faster and actually works proper with refiner

[royal tangle]

Goodnight guys excited to see how things progress tomorrow!

[outer junco]

ComfyUI will work better but Auto1111 can be easier to understand for beginners

[timber quiver]

hey, are images from dreamstudio saved in png?

[urban fjord]

If you intend to use the refiner then ComfyUI, if not Auto1111

[shy kelp]

ah cool ok, also wondering, how would we train something? i got a 3090 dunno if thats enough

[torn canyon]

yes, but that seems reasonable

[shy kelp]

like a meinamix for anime style

[torn canyon]

I'd wager the reason jpeg looks better is because the compression reduces the artifacts, while png is lossless

[shy kelp]

👌

[timber quiver]

i have no issues with 1.0 vae when saving image as jpg, but with png is awful

[polar ravine]

Should I try training a lora on xl base?

[lament vessel]

Just did my 1st generation using SDXL 1.0 on auto1111... which took 9mins using only the base checkpoint. Really hope this is NOT normal? Used to take ~8s per generation using a standard sd1.5 checkpoint. Any pointers?

[nocturne dove]

jpg is generally chroma subsampled. Without knowing the exact details I couldn't say whether the green/red lines would be guaranteed to average out in a jpg or not. But at least they would be reduced.

[outer junco]

A 3090 should be plenty up to 512px images, after that can be hit or miss. I have used both the Dreambooth Extension and https://github.com/bmaltais/kohya_ss/tree/master though neither are officially compatible with sdxl yet

[azure oxide]

someone other than me needs to use this training document by the controlnet dev and train the usual controlnet models on sdxl ASAP https://github.com/lllyasviel/ControlNet/blob/main/docs/train.md

[elfin flare]

Seems there's currently no method of running SDXL on intel arc

From what I can tell

[outer junco]

due to arc's low adoption rate id give it a week or two

[elfin flare]

That's alrighty

[untold crescent]

there really isn't fix for the overlapping text on comfy?

[hard fractal]

Make the node bigger

[untold crescent]

how?

[chrome furnace]

can someone point me to something to read about comfy. Im used to auto1111, but cant get sdxl running there. comfy is generating images now for me. would like to read into stuff. how do is use refiner with comfy for example. whats a refiner? 😛

[hard fractal]

Drag the bottom right

[untold crescent]

ohhhh

[peak wraith]

change image size to 1024 x 1024

[untold crescent]

I tried dragging it everywhere else 😄

and resize option only made it smaller 😄

thanks!

[lilac raven]

Is it possible to generate an image using only the refiner model (still downloading the refiner)?

[frank steeple]

was able to go 1536 x 1024 on an 8GB card

[urban fjord]

It is possible but you don't want to.

[high skiff]

😳

Can i maybe have this prompt?

[floral island]

sure, just steal away ^^

[civic sigil]

Yes it's basically like a normal model

[high skiff]

is it in the image itself?

[floral island]

yeah, is comfy image, so workflow is included

[high skiff]

awesome, thank you, I love tigers

[floral island]

is based on your workflow, with addition of WAS (easier doing stuff with the numbers)

[high skiff]

oh nice!

I love WAS, just can't use it cause of the wiki guidelines haha

[fathom mango]

What happen here ?

[thorn lion]

I have communication issues due to my /disability/ and that's where this question is coming from. How would you ensure that someone knows that you are being sarcastic? I was taught that this method was the general convention. I don't want to make communication mistakes

[untold crescent]

this seems to do the trick on auto1111 for swapping the old VAE in permanently

[urban fjord]

Here is a corgi I made using just the refiner, it tends to struggle with shapes a ton if you don't run steps with the base first.

[elfin flare]

I wonder.

[floral island]

ah, that's the WAS extension 😄

[elfin flare]

Does the current automatic1111 repo support SDXL?

[fathom mango]

rip 😄

So have to install that

[urban fjord]

Yes Auto1111 support SDXL

[outer junco]

Someone have a good sdxl prompt for 512x512? Gonna do an XY of Samplers to CFG

[elfin flare]

If so

Then it's possible I might be able to do it through DML

[floral island]

https://github.com/WASasquatch/was-node-suite-comfyui

[civic sigil]

Why 512

[urban fjord]

SDXL does not support 512x512!

[floral island]

that's correct, the new default is 1024x1024

[outer junco]

wait you cant do any gens below 1024?

[floral island]

correct

[urban fjord]

Stick to one megapixel images or downscale afterwards.

[jagged ferry]

Would it be possible to take an image and use sdxl to generate new images of the same character?

[outer junco]

hmmm surprised there isnt a huge jump in vram

[high skiff]

its very efficient in that way

[azure oxide]

ya you can use img2img or if ur in comfy, unclip for that

[outer junco]

nvm 22gb with hires fix

[lilac raven]

I'll fund Micron next month, get the 64 Gibes

[jagged ferry]

I tried it and it made something completely new😭

[azure oxide]

might have to play with the denose values

[outer junco]

so would the new equivalent of 512x768 be 1024x1536?

[high skiff]

no

[quartz sequoia]

@visual glade is it possible to automatically make a node connect by just dragging it on a connected like between two node? cuz currently its only manually... basically something like how blender node system works

[jagged ferry]

I forgot about those. I’m tryna set up a workflow for img2img atm

[high skiff]

you want the final image res to be the same as 1024x1024

[elfin flare]

Wait.

[high skiff]

so you want it to equal about 1.04m pixels IIRC

[elfin flare]

Does SDXL work on Vladmantic's repo then?

[urban fjord]

There is a list of intended aspects ratios if someone wants to report it

[polar ravine]

What happens if I merge a previously made lora to a sdxl based model?

[azure oxide]

is this true even with hi-res fix?

[urban fjord]

1.5 LoRA cannot be merged into SDXL

[high skiff]

^

it will just give errors

just like 1.5 through to 2.1

[civic sigil]

0.9 loras might work but you might as well remake them

[lilac raven]

How many steps should be base and refiner? Is it a proportion?

[azure oxide]

just do the ollobrains approach and merge it anyways

[polar ravine]

I'm not sure I can train a new one on 12gb of ram (4070 ti)

[high skiff]

reworking my research on that

it changed since 0.9

[urban fjord]

Just test your 0.9 LoRA, they might work well, they might not work well, your millage may vary.

[civic sigil]

You could train an 8 rank lora with the right settings

[outer junco]

Still someone toss me a quality testing prompt

[fading kayak]

whose got some juicy non-basic workflows for comfy SDXL 1.0?

[azure oxide]

wait i wonder if he merged sdxl into megamodel yet lmfaoooo 😂😭

[high skiff]

my full mixed diffusionw orkflow for 1.0 should hopefully be out within a couple days

[slender coral]

Am I doing smething wrong, my prompts are not as good as 0.9

[winged tree]

refiner is first?

[urban fjord]

Add more style into to your prompt.

[winged tree]

[urban fjord]

Do you want photo, drawing, clay etc..

[upbeat summit]

use the noise offset lora

[azure oxide]

nah its last here

[winged tree]

it goes bottom up

[wet nacelle]

@azure oxideDo you think you can help me with getting the Stability UI working?

I've been trying for the past four hours.

[azure oxide]

yep just follow the nodes, man i love node based gui's

yeah whats your issue?

[winged tree]

yeah, but refiner is loaded first and the sampler uses the refiner and an empty latent

[azure oxide]

could also try posting in #🤝｜tech-support , theres some really intelligent folk who lurk in there

[civic sigil]

Can't wait for controlnet and reference_only, animations are gonna be so good

[wet nacelle]

I was the one trying to get the new Stability Co UI to work.

You said you had the same issue and just installed the .net sdk 7.

I've been doing that for hours and it just keeps opening and closing with less than second.

[high skiff]

@hard fractal Speaks truth, 0.9 LoRA's run better on 1.0 by a considerable amount

This is my Na'vi LoRA I trained on 0.9 (left), and it looks better in 1.0 (right)

[wet nacelle]

They said they couldn't help

[trim orbit]

#✨｜sdxl message

[winged tree]

You get it fixed?

[high skiff]

oop lol

[azure oxide]

unlucky. but ya, you installed net from the link that mcmonkey linked right?

[civic sigil]

That doesn't work with img2img, comfy said he will support i2i in the published version of the node

[slender coral]

ffs thank you

[wet nacelle]

I'm not sure.

From the git

[winged tree]

💖

[wet nacelle]

The git links the list of downloads not just the one I should download.

[azure oxide]

alright try downloading this

[azure oxide]

once you install it, it should work. if not, someone else said that this step only worked after they uninstalled all other versions of net

[hoary saddle]

prompt weights still work right? doesnt seem to make much difference

[wet nacelle]

That's the thing. Which?

[proud ruin]

Hi, stupid question maybe but how to install SDXL using Google Colab?

[high skiff]

yeah, 1.0 works even better when using my special LoRA post processor

Left is 0.9, right is 1.0

[azure oxide]

honestly i had the same question 😅 so i clicked a link that said "not sure which one to DL??" and it brought me to here https://dotnet.microsoft.com/en-us/download

i ended up downloading this one

from there, i installed it and it worked

[tiny walrus]

just dipping in to say that yes, indeed 0.9 LoRAs should work on 1.0, and all the ones I've tried look like 10x better on 1.0.

[outer junco]

I love this it knows who laika is \

[floral island]

lovely

[wet nacelle]

I am going to fucking scream.

I installed it. Should restart my computer?

[trim orbit]

its the best

[worn quartz]

i take it the team knows about the 1.0 vae issue?

[azure oxide]

hahahah. nah i didnt need to restart

[spark bear]

contrast looks too high on 0.9, probably why they backed off offset noise

[lament vessel]

anyone tried sdxl1.0 on auto1111 using the --no-half command? this take my generation time from ~12 sec to ~10min

[azure oxide]

joe acknowledged it and promptly left afterwards

[wet nacelle]

I'll uninstall the UI files as well and reinstall them.

[trim orbit]

they were told to not talk about it and joe mentioned not to ping devs about it

[floral island]

if this wasn't the SDXL channel, i'd ask which movie did you download

[outer junco]

it is interesting nolonger being able to do a bunch of commas and instead needing to write sentences

[tiny walrus]

it's still worth retraining on 1.0 in some cases--the models I've gotten out of it are amazing even compared to the 0.9 ones working better now

[lean dawn]

hi

[trim orbit]

i definately will retrain some of my datasets. its nice to see all the ones i made on 0.9 working SO much better too

[agile quarry]

any sorta activation token on your LoRa kali

[lean dawn]

any help?

[quasi remnant]

Yes, you're not using a vae loader + do 1024x1024 not 1080x1080

[outer junco]

is it stuck?

[hushed hornet]

Lokos like you can indeed get to Flavortown with SDXL with a bit of work!

[lean dawn]

yes

[outer junco]

for how long?

[lean dawn]

for tooooo long ._.

[wet nacelle]

@azure oxideWait. How long did it take you to install the UI from the bat file?

Each time I did mine it just took less than a second and I had my file list.

Now it's taking more than that.

Should it be taking longer than less than a second?

[lean dawn]

maybe 30 mins

[slender coral]

Could I ask why in this case, would like to understand. Other than pow of 2

[novel birch]

where to download SDXL model file to use with comfyui?

[civic sigil]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

[novel birch]

thanks 🙂

[lean dawn]

._.

[tiny walrus]

the first time I tried I gasped 🙂

[azure oxide]

yeah it only took a few seconds

[lean dawn]

any one know what to do?

[quasi remnant]

well, just 'cause the images are trained on 1024x1024, 1080x1080 still works fine i've used it.

[azure oxide]

come to think of it youre right, it was about a second mabey less

[wet nacelle]

It might actually be installing the proper things now.

I'll just let it run

[civic sigil]

Wtf 1.0 looks way different than 0.9, I'm kind of worried noww

[winged tree]

yeah

why wouldn't it?

[untold crescent]

can you append text node with another text node in comfyui? to get your style presets done?

or you need custom nodes for this?

[lean dawn]

?

30 mins

[civic sigil]

I mean worse, although it could just be because I'm using a 0.9 lora. Gotta try retraining it on 1.0 to be fair

[lean dawn]

is it fine?

[azure oxide]

you want this combine node

[hoary saddle]

SWARM worth checking out?

[untold crescent]

awesome

[outer junco]

may need to restart the install unless you have really bad internet, I think gfpgan is a few gb

[untold crescent]

I tried searching add and append

[west breach]

results from the 'grumpy style' lora I trained this morning

[outer junco]

Do I increase it....

[winged tree]

So you mean your LoRa looks worse on 1.0

[trim orbit]

those are really disgruntled people

[lean dawn]

oh how many 10?

[floral island]

yeah, ur just screenshotting dude

very nice 😄

[wet nacelle]

@azure oxide It still didn't work.

How did you get it running. What within the file list did you run?

[upbeat summit]

thank you!

[outer junco]

hmmm here actually

[lean dawn]

thanks for your help ❤️

[azure oxide]

after running the install, did you launch the shortcut on your desktop?

it closes immediately still?

[outer junco]

stick this in the gfpgan model folder

[wet nacelle]

That just closes as soon as it opens

[lean dawn]

332? very small

[azure oxide]

hmmm youre running on windows right? the link that i sent above (not the one that mcmonkey sent) was for the windows version of net 7

[lean dawn]

it must take 2-3 mins not 30 ._.

[wet nacelle]

correct

[timber quiver]

@wet nacelle FYI I had same problem with .NET and SwarmUI, had to install required dependencies with Visual Studio by opening StableSwarmUI.sln project and VS installer.

[slender coral]

This is going to be a dumb question, but do we know what sdxl was trained on?

[urban fjord]

You shouldn't share unsourced models in a non-safetensor format.

[wet nacelle]

ASIYDGVASDYLUIOF

[outer junco]

its straight from my model folder

[tender timber]

I have had no luck with using git to install ComfyUI into Windows. I make a folder and use

git clone https://github.com/comfyanonymous/ComfyUI.git

And I end up with what's in the screen grab.

Then... what? I found the run_nvidia_gpu.bat file in C:\Applications\StableDiffusion\ComfyUI\.ci\windows_base_files but when I run it, I get:

**C:\Applications\StableDiffusion\ComfyUI.ci\windows_base_files>.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build
The system cannot find the path specified

C:\Applications\StableDiffusion\ComfyUI.ci\windows_base_files>pause
Press any key to continue **

I'm sure it's something stupidly simple I am missing . I can download the standalone, but when I do that, the update batch file does nothing but happily tells me it's done when when there have been update within the last two hours

[lilac raven]

Some really amazing images today

[wet nacelle]

Okay. How do install the dependencies?

[lean dawn]

where is that

[kind pewter]

NSFW content sucks with SDXL :(

[floral island]

give it a week, nsfw models for days!

[outer junco]

auto1111 or comfyui?

[outer junco]

they were worried about it generating photorealistic CSAM

[uncut zephyr]

SD.Next has sdxl 1.0 working fantastically!

[eternal fog]

I do wonder what people are trying to have it make. Because it can do nudity fine.

[slender coral]

ya'll need to get laid, AI is not a porn generator 😄

[kind pewter]

like if it's gonna be stopping somebody to train it if they wanted

censorship is always uneffective

[outer junco]

yeah but they are nolonger legally responsible if that happens

[winged tree]

it definitely is harder

[kind pewter]

and messes more with normal people than actual bad guys

[fading kayak]

were the style prompts released with 1.0?

[azure oxide]

id ask this in #🤝｜tech-support , the folks there should be able to guide you through it

[eternal fog]

Dunno seems pretty easy, I can't post it though for obvious reasons

[kind pewter]

since when gun manufacturers are responsible for what you do with it? 🤔

[lean dawn]

what?
can you send me the folder path

[winged tree]

tbf it was 0.9 i was messing with before 5 minutes ago

[leaden fable]

Left is discord bot, right is local with comfy, same prompt same seed. does the discord bot use some sort of invisible default negative/positive prompts ?

[outer junco]

Im using webui so stable-diffusion-webui\models\GFPGAN

[civic sigil]

Probably yes

[urban fjord]

Probably different, sampler, scheduler or CFG.

I've seen no signs in the past that there is any prompt medling in the bot

[kind pewter]

yes it is.
@eternal fog just a vulva or female genitals are very hard to get

[azure oxide]

actually remington paid out millions to the families of one of the shootings a few years back

[eternal fog]

It's not a porn creation model

[civic sigil]

You're the one that said it's good at nudity

[kind pewter]

Art can have genitals as a part of the piece yknow

[eternal fog]

Yes, general nudity, it can do the top half fine, the bottom half is more difficult. But you aren't getting close-ups

[kind pewter]

and you told it was easy

[urban fjord]

This is not the place for any of it...