✨｜sdxl

[rigid laurel]

You need to install that in comfyui, then restart it. Then it should run no worry no fuss.

[woeful patio]

negative prompt is unnecessary now yes?

[rigid laurel]

largely

I only use it when something keeps appearing that I don't want

like this one time it kept putting mushrooms on my witch

or nipples on my forest sprite

outside of those rare occasions, my negative prompt remains empty

[soft zealot]

first time someone has complained about SDXL being able to draw nipples lol

[rigid laurel]

Oh I didn't say it was able 😂
just that it was trying

I'm trying to make some faecore body horror shit and it keeps slapping pepperonis on the poor woman

[woeful patio]

how do you all get such good faces mine look like they've had a horrible chainsaw accident

[sly jay]

the AI perfectly placing nipples on my SFW generations

[sly jay]

luck, impainting, codeformers, luck

[rigid laurel]

experimenting with samplers, experimenting with schedulers, using the refiner, using hirez fix

[lilac wren]

To me it works like 1.5, you have to upscale or simply refine. With A1111 for example you make a 1st gen, then an img2img with refiner

[trim totem]

how do I get the refiner in comfy ui?

[nimble heart]

hmmm im not too sure about that

[shy kelp]

Nipples?? I'm very disappointed SDXL 1.0 still can't seem to get hands right. But more importantly, at least to me, it still does feet poorly!

[rigid laurel]

use Searge's SDXL nodes. Look up Searge on civitai.

[rigid laurel]

a self-report the size of the sun

[sly jay]

hands are generally better than base 1.5, but still not great

[rigid laurel]

the government really will inject its propaganda anywhere, huh

[rigid laurel]

controlnet support is coming, so are anatomy loras of all kinds

[woeful patio]

do we have an estimated date for controlnet?

[still dove]

not working

[rigid laurel]

if you mean the json file, you need to install Searge's modded SDXL nodes. search his name on civitai.

[rigid laurel]

Joe posted a screenshot of it working, so we know they have a functioning prototype. But if anyone mentioned a date, I didn't hear it

[shy kelp]

Yes I admit hands are marginally better....but not great. I had assumed when the devs were promising hands would be better that it would also translate to feet. Seems that it does feet the same.....poorly.

[rigid laurel]

pretty sure SDXL wasn't trained in the wikifeet corpus, so there wasn't a large sample size

[shy kelp]

@rigid laurel I'm sure they are aware of this and in the future somehow it will be solved. But not with this model.

[sly jay]

for some reason this aint working for me but its too damn late so I am going to bed and will try again tommorow lol

[rigid laurel]

:/

if there's an error message, I'll try to see what the deal is

otherwise, gl and gn

[sly jay]

me when no sdxl feet 😞

[woeful patio]

i'm impressed by how it gets almost all the numerals right

[sly jay]

oh last quick comfybox question, is it possible to get gen data of previously gened images?

Restarted to try install the thing you linked and am trying to get back to where I was with my gen

[rigid laurel]

@sly jay @still dove @woeful patio My apologies, Searge was only one of the two extensions I used. I forgot about this one because it wasn't on civitai. One of the workflows on there comes with this file that you need to put in the custom_nodes directory. Put this in there, restart comfy, and that'll give you ther SDXLKSamplerAdvancedIterate node you're missing

I should probably replace it with searge's stuff eventually, but OH LOOK AT THE TIME

[sly jay]

yeah its fucking late lol

at least for me in pst

[woeful patio]

is a1111 gonna come out with a method to make the refiner work in 1 step?

[sly jay]

one can dream

[rigid laurel]

4am here. imma fuck off. hope this solves things for you tomorrow

[sly jay]

jesus, good night lol

[sly jay]

but yeah likely someone will make an extension

[delicate grotto]

is there an sdxl with a fixed vae?(that i don't have to load seperately because i don't have enough ram)

[spring fulcrum]

Do you also have an image you can post with your workflow embedded?

[rigid laurel]

https://civitai.com/models/117188/sdxlfixedvaefp16remove-watermark

[woeful patio]

isn't there a vae inbedded in the base model?

[delicate grotto]

its horrible

[rigid laurel]

there is a vae, yes

they released a fix

that civitai link has a merge someone made

[delicate grotto]

great thanks

[rigid laurel]

don't know how to do that king 👉 👉

[spring fulcrum]

you just drag an image into discord that you created while using your setup. it should have your setup embedded in the image.

[rigid laurel]

ballin. that's news. alright. will post.

[delicate grotto]

i guess ill use the same vae from the fixed sdxl vae on the refiner too

man it sucks that they didn't notice it early on, it hurt my eyes in the first few pictures i saw
i was sure it was pixelated somehow, and i did see some literal blocks

[spring fulcrum]

The image has to be something that you saved locally on your pc. otherwise it wont work.

[rigid laurel]

uh

[delicate grotto]

not really

[rigid laurel]

?

[delicate grotto]

you can drag it as is without zooming in on the picture on discord

[rigid laurel]

that's what I did, but apparently it didn't do the thing

[delicate grotto]

try this

drag this pic

as is

no zoom in, no open original, just from here

[spring fulcrum]

that worked

[delicate grotto]

don't click on it, just drag

good

[spring fulcrum]

I was saying to share your setup it has to come from an image saved

going into discord

[rigid laurel]

yeah, both the mushroom and the dragon were saved

[delicate grotto]

you can also copy it from the workflow

[spring fulcrum]

how do i do that?

[rigid laurel]

mind, I'm using comfybox and not comfyui

it might not embed

[delicate grotto]

right click on the image in comfy, and open in a new tab

then just right click copy img and paste it here

[spring fulcrum]

ohh ok... that might be it then

[delicate grotto]

man, civitai works fast

just yesterday sdxl was empty with 4 submissions

[pallid sail]

If I'm not getting black results from SDXL when using float16 is it worth using the fixed vae anyway?

[calm shell]

Done with ArtroomAI, 896x1152, total generation time was 3.5 minutes each. no loras, no textual Inversions (haven't figured out if ArtroomAI even handles those.

[calm shell]

my standard test prompt: ink, white background, scenery, mountains, trees, shrubbery, water, sunrise

[uncut steeple]

You can also try my workflow, its a little more complex tho

[woeful patio]

does anyone know if Ultimate SD Upscale works for sdxl?

[calm shell]

but the prompt I used for those two picture was : realistic, white background, scenery, mountains, trees, shrubbery, water, sunrise, dramatic clouds

[heady vale]

it does. I used it on the above image, but posted a resize

[uncut steeple]

Im using it in comfy UI

[woeful patio]

does it work with 4X ultrasharp too?

[heady vale]

whatever upscaler you used before

they work

[woeful patio]

cool thanks

[uncut steeple]

nice little prompt

[boreal bough]

its only passing along the denoiser value right now.
could be a mismatch between your ui & comfy version.

after work, will try not using the swarmui downloaded version of comfy, and instead connecting it to my normal install

[wicked frigate]

fwiw i tried shoving two ksamplers in, and it correctly yields two ksamplers in the param list. If only denoise is visible that sounds like you're looking at the first and not the second?

[uncut steeple]

Whats that menu

[rigid laurel]

oh, is that the new swarmui thing?

hoping to try that omorrow

[uncut steeple]

Another new UI

[rigid laurel]

eh, this one's endoresed by stability, right?

so maybe we can finally have a canon app to use

which may mean better support

lotta maybies, but I like the sound of it

[uncut steeple]

We will see where it goes

Gonna be hard to catch up with A1111

[rigid laurel]

...not really. that thing is made out of spaghetti

[uncut steeple]

Yes but it is like half a year of development ahead

[rigid laurel]

it's fantastic given its circumstances, but other apps can learn from its successes and not lose time to its failures

[uncut steeple]

True

[rigid laurel]

also auto is a great example of how sometimes starting over is faster than trying to fix things

I learned that the hard way a few times

so Swarm has every chance to be better as long as it actually does things right

which is kinda the gotcha

[soft zealot]

Its only full of spaghetti if you dont arrange it to hide the spaghetti.

This is my daily driver view

admittedly the "helicopter" view is a bit messier

[rigid laurel]

oh I'm talking about a111

comfy has it's own problems, but the backend code seems much cleaner

which makes said problems easier to solve

[uncut steeple]

My workflow right now

[woeful patio]

anyone know what this means?

[rigid laurel]

uh, pretty much exactly what it says on that last line

try each of the two options it suggests, they'll likely fix it

the only thing it doesn't mention is that sometimes that happens by pure chance

[woeful patio]

won't that fuck anything up, I'm not sure what any of that means lol

[rigid laurel]

so you can do nothing and just try again; it might have been a fluke

[rigid laurel]

the image is represented by a bunch of numbers. changing the cross attention layer from float16 to float32 means the numbers is uses will be more precise, and it's less likely to encounter the math errors it mentioned there. On the flipside, each number takes up more space.

SD generally uses "half-float precision" by default to save vram

but that can cause fuck-ups

it is much more likely to cause fuck-ups on certain cards, or with certain drivers

so some people have to disable the half-sized numbers altogether and use float32 all the time to keep it from exploding on them like that

anyway, I'd do what it says and change the setting in the menu it describes

[spring fulcrum]

@soft zealot Hey Winston. Can you post the json for your full setup... that looked pretty awesome.

[wicked frigate]

see but I'm cheating - it tightly integrates Comfy as a backend, so all development work on ComfyUI that's already happened is provided to SwarmUI for free out of the box :D

SwarmUI is a frontend and a midlayer, doesn't reimplement the backend

[rigid laurel]

https://tenor.com/view/the-secret-ingredient-is-crime-jbone666-mouthful-building-walking-the-street-gif-16400025

[uncut steeple]

Yeah I saw that, looks really promising tbh, looking forward to see more of it

[wicked frigate]

if by support you mean user-support (answering questions / fixing issues / etc), ftr in addition to being an employee of stability, I've also spent years of my life running FOSS projects and have had many people tell me my projects have the best user support handling of any project they've seen, so... yeah I think support'll be good :D

[woeful patio]

thanks for explaining I restarted sd and it seems to work now, thanks

[rigid laurel]

you'll find it uses more memory now, but it should be more stable 👍

pun not intended

[rigid laurel]

I more meant things like not having to trawl through 4chan to find tutorials on how to use it on a 7900xtx like I did 😩
that and community extensions

[wicked frigate]

olol

[wicked frigate]

yeah that's part of it too. I gotta spend some time writing docs

I have placeholders atm (because alpha stage)

but like

the UI is full of ^ little clickies that tell you what the thing does

[rigid laurel]

yeah, that shit's good

[rigid laurel]

that'll be helpful for answering the questions people keep dumping in here 😂

[little gazelle]

@candid pond Did you manage to use the FP16 VAE on Kohya? I'm getting the error RuntimeError: Error(s) in loading state_dict for SdxlUNet2DConditionModel: Missing key(s) in state_dict:

[candid pond]

I was ablo to start a training with it yesterday.

[candid pond]

Do you think we will be able to fine-tune checkpoints with 24 GBS of Vram?

If so I'm seriosly thinking about buying a used 3090 24

[vocal stream]

I believe people said they were finetunning 0.9 with 24gb

[candid pond]

I see

Then with furter optimization we should be able to do it with 24

Time to buy a card that has been used in mining 😄

[vocal stream]

I guess most of those were sold a while ago given that mining hasnt been profitable for a year but yeh that's the best case

[candid pond]

Oh we still have people in Türkiye that try mining

3090 goes for about 15k tl

[vocal stream]

then maybe you can get a decent price, here (germany) used 3090s aren't listed much and the last one is 850e 2nd hand which isnt great

[candid pond]

Damn that is expensive

It goes for around 600usd here

[vocal stream]

yeah that's worth

[candid pond]

Time to upgrade then

[vocal stream]

I'd get one if it was that much here but at this point Ill probably just pull the trigger on a 4090 next week

[lilac wren]

A1111 works not so bad with SDXL

[candid pond]

This SDXL is causing me upgrade my entire system one by one

I just got a brand new 32 Gbs Ram yesterday

[candid pond]

Damn how I envy you

[lilac wren]

Don't tell me about it, I'm looking for a used 3080

[candid pond]

I work on Fiverr which is like a goldmine in Türkiye but even 4090 costs me about a month of work

[modest mulch]

isnt stable diffusion use v ram, why you need pc ram?

[vocal stream]

I already had to spend like 400 on a bigger PC case and a bigger PSU to even be able to use a 4090

[candid pond]

XL supposedly requires 32GB's of system ram for refiner+base

[delicate grotto]

the models runs on the vram, but it is being loaded into the ram

[candid pond]

Yeah constantly shuffling models was too much of a hassle

[vocal stream]

at this point I'm treating my PC how people treat having a car and will probably just start buying the latest one (skipping TIs) and selling the old one immediately when a new one comes out to recup and just budget like 500e or whatever it works out per year on average for gpu upgrades

[candid pond]

Aahaha

[candid pond]

I think that is definetly the best way

[thorny frost]

I got my 3090ti new at 1200€ about 8 or 9 months ago. Didn't know anything about AI generation back then and i thought so much vram was overkill for games. Best unconscious buy ever. Lol

[candid pond]

Ahahaha, yeah

1200 EU is like 3x minimum wage in TR

[wraith apex]

Tried automatic1111 with sdxl and says most times not enough memory on 8gb card, really? 1024x1024 :/

[candid pond]

Try comfy, it seems way better optimized

[thorny frost]

Still it was a great price. 4090 had just releases and 3090 went down for a few days.

[vocal stream]

tbf with how bad the lira is doing it was probably only 2x minimum wage when he got it

[wraith apex]

Thanks, will do

[candid pond]

Auto1111 really needs a better workflow with SDXL

[lilac wren]

I have the terrible choice between repairing my car or buying a 3080 TI/3090.
As we say in France: The plague or cholera

[candid pond]

Ahaha, for sure

Gotta love recession though

About 5 orders from Fiverr a week pays like 3x minimum wage now

1USD=27TL

[candid pond]

What is wrong with your car?

[candid pond]

Is this with Comfy or Auto?

I couldnt get any decent images in Auto

[thorny frost]

Nobody uses imvokeai anymore?

[indigo carbon]

auto

[candid pond]

I think people are using it with the latest update

[lilac wren]

The wishbones, stabilizer bar supports, discs and brakes, power steering pump, power steering belt, 2 front tires

[candid pond]

Than it's a me problem since all the images I had generated in Auto have immense noise in them

[candid pond]

Damn lmao

Seems the car is more urgent

[thorny frost]

It support xl now. You can get nice results easily.

[lilac wren]

yeah

[candid pond]

Will check that today, Invoke supports simpler inpainting then comfy correct?

It had like a canvas feature if I remember correctly

Been ages since I used it

[thorny frost]

Yep it has, don't use it much myself though

[candid pond]

I do personalized images on Fiverr so inpainting is a must for me

But I mean I can also wait for the optimization of Auto11111 for few weeks

1.5 still works fine most of the times

[lilac wren]

:o, with A1111, tiling is not more necessary for me, with my poor 2070 SUper, i can generate 1920x1080 > upscale to 1,2x, with no Tiling 😮
No crash CUDA or out of memory error !

(SDXL)

[thorny frost]

There are some free photoshop plugins which work great for modifying images using SD

[thorny frost]

And for gimp too iirc

Stability.ai or something.

[lilac wren]

Nothing amazing, but no CUDA crash

[candid pond]

Still looks great

[static prawn]

In Auto i need 20minutes for 1 picture 😄

in Comfy 1-1:30 Minutes

im still super overwhelmed with all those extra settings in comdy

comfy

[lilac wren]

Yeah, usually to go up to such resolutions I have to use TiledVAE, here I did 1920x1024 70 Steps > Upscale (HiresFix) x1.2 with 30 steps. With such a small card, incredible. I don't understand how it's possible, with 1.5 it's impossible.

[rigid prawn]

I've downloaded the WAS node suite for comfy but can't find any style for the node "prompt style selector". How do I get them?

[static prawn]

and i dont like the idea of all those workflows in general , imo would be nice if anyone would be capable doing consistent results

[wraith apex]

If buying new gpu for SD, probably better gp nvidia, instead of amd, even amd might get more vram?

[soft zealot]

Ok that smildly annoying.

Its trying to upgarde my existing version of Git with the same version of itself

[static prawn]

anyone has a nice and clean workflow for comfy?

[soft zealot]

depmds on your definition of "tidy"

#✨｜sdxl message

[static prawn]

i just dont wanna scroll around to set everything up

need something that fits on 1 screen xD

[soft zealot]

Mad now its hung , grrrrrrr @wicked frigate

[candid pond]

There is no such thing in comfy 😄

[soft zealot]

did you look at my screenshots in the link ?

[candid pond]

Sytans workflow is pretty easy to use

However it lacks upscaling and img-to-img

[static prawn]

yes 😄 ty , but the second one is much more isnt it?

[vital wolf]

Are you using stableswarmui?

[candid pond]

You can set both of them quite easily yourself though

[soft zealot]

Thats the "under the hood" stuff that doesnt need touching on a daily basis.

[static prawn]

oh ok

[soft zealot]

no, I was just5 going to have a play

[static prawn]

just have to find links for all those upscalers

that are mainly used with sdxl

[candid pond]

Dont forget to download the noise offset vae aswell

[static prawn]

is ur workflow capable of a highresfix too?

[soft zealot]

ah lets just git clone it lol

[candid pond]

Most workflows I've seen online uses the vae and wont work without it unless you remove or use it too.

[soft zealot]

there is a HRF workflow in there , if you use my workflow please ensure you read the credits & notes boxes

[candid pond]

Can you send the json file? I'd love to try it aswell.

[static prawn]

ok allright 😄 thanks a lot

[peak dove]

Try Sytan's

[soft zealot]

Just download one of the (generated) images Ive posted today and dragndrop into comfy. Workflow is included by default

[static prawn]

oh ok ty

[candid pond]

Oh ok

Thanks

[static prawn]

i just hope they get a better version of auto1111 soon, absolutely dont like the idea of those workflows at all 😄

[soft zealot]

Once Youve gone COmfy theres no going back ;o)

[static prawn]

nah, i hate that everyone has other workflows which messes up any consistency in the results

[thorny frost]

Comfy is pretty easy to use in fact.

[soft zealot]

sigh

[static prawn]

a clean ui that is capable of anything sounds better to me

[woeful patio]

what do you think is a good amount of sampling steps, currently have it set at 45

[wicked frigate]

as a user with technical experience, you can skip the installer and just git clone + launch-windows

[thorny frost]

The difficult part is to use it right.

[wicked frigate]

the installer is mostly just ultra easy mode for noobs

[soft zealot]

already doing that lol

[soft zealot]

you mean "12 o'clock flashers" 🙂

[edgy sierra]

how to remotely run comfyui?

[uncut steeple]

Add --listen to the batchfile

[soft zealot]

@wicked frigate does dotnet require a PC reboot
(probably is the answer t that lol)

Oh well later

[wicked frigate]

it shouldn't but, yknow, microsoft technology

might need to close console window + reopen or something

[south igloo]

Why did they call it comfy UI when it's anything but comfy lol

[wicked frigate]

it's been weirdly inconsistent about installing for a lot of people

i'm not sure why

[wicked frigate]

it's the developer's username

[south igloo]

Ahh that makes sense now

🤣

[wicked frigate]

same reason for the "automatic webui" that's a very manual process to use usually lol

[soft zealot]

aye , its going to be a sessionthing. Closing the console & starting a fresh one works

[wicked frigate]

yay

[thorny frost]

Webui is the surname

[static prawn]

i probably just wait with comfy i dont get all the extra stuff

[queen scaffold]

is there any workflow with roop and controlnet together?

[wicked frigate]

It can download the refiner now, and use it out of the box - https://github.com/Stability-AI/StableSwarmUI/discussions/11#discussioncomment-6574816

[uncut steeple]

@wicked frigate Btw do you work in hand with comfy?
Just curious

[soft zealot]

@wicked frigate Feedback/Question.

I've already got all the models I need ( so presuming I can just skip this step/going to skip) .

Maybe needs rewording ???

@wicked frigate just cancelled it as it wants to downlaod a COmfy Backend. Why cant I just link to my existing install??

[wicked frigate]

oh - yes you can skip

[wicked frigate]

just select 'None' on that page

[abstract comet]

SDXL only works for me in Comfy, so started playing around with 2 days ago. Super confusing at first, but got into it pretty fast. Kinda neat that you can setup your whole workflow, save it all, and so on

[wicked frigate]

you can link the existing one yes

[wicked frigate]

yes i do, he's on our team!

[static prawn]

i dont get all the extra resolutions and that stuff, its super overwhelming for me

[uncut steeple]

Thats really cool

[static prawn]

unlucky, stable swarm installer always crashes here. its weird

[soft zealot]

errm there wasnt a "None" Option, just went into downloading it

[wicked frigate]

there... there was, go back a page or two

page 4? i think

and then after install is done you just

[uncut steeple]

👁️

[wicked frigate]

add it and link it via address or filepath to your preference

[abstract comet]

yeah not sure about a lot of stuff in there either, so far have only base and refiner hooked up, also did some img2img stuff and upscaling, pretty basic

[wicked frigate]

and link the models dir in server config

[static prawn]

@wicked frigate i just get a folder stableswarm, but it always closes when i try to open

[wicked frigate]

open a command prompt window and run the .bat through it (so it doesn't close and you can see any error message)

[zealous horizon]

swarm should create a shortcut on your desktop

[soft zealot]

Ok Feedback:

This i smisleading, needs to be split into two clear blocks IMHO , One Titled something like "New Install Options" and the Other as "Existing" . Like may people I selected ComfyUi as the question asked which backend and its the top option so naturally you stop reading

[static prawn]

it does

but i cant open it

just closes

[zealous horizon]

any notable Errors during install?

[wicked frigate]

actually: if you git pull in the install folder and try again, it should keep the window open for you

[static prawn]

it works now

[soft zealot]

@wicked frigate Feedback ThThere is no settings tab listed. Ok I've worked out its under Server but maybe the message should be consistent with the Tab Names

[static prawn]

might be a more usable experience than comfy, i just cant get along with comfy.

[cinder aspen]

@static prawn tell me pls how it is ^^

StableSwarm

[static prawn]

dunno i hate that there is no structures, its cancer in my eyes xD

i know its capable for so much stuff, but all this stuff doesnt look nice at all

[static prawn]

i installed a wrong SDK

just download the right sdk

now its working

[cinder aspen]

Qualitywise or the UI?

[wicked frigate]

how bout dat:

[soft zealot]

Much Better

[static prawn]

the UI

its just so unstructured

[cinder aspen]

Oh

Well yeah, alpha phase

[zealous horizon]

IMHO SwarmUI is heading in the right direction but ATM I use ComfyUI standalone as a consumer. I think where swarm shines is when you want to distribute and handle huge workloads that are scaled over a swarm of GPUs.

[static prawn]

and for swarmui

Install DotNET 7 using the instructions at https://dotnet.microsoft.com/en-us/download/dotnet/7.0 (you need dotnet-sdk-7.0, as that includes all relevant sub-packages)
Open a shell terminal and cd to a directory you want to insta

[wicked frigate]

oh oops, will fix

[static prawn]

this helped me

[zealous horizon]

the nice thing is that you can easily use any Comfy workflow in swarm once it's more stable

[soft zealot]

@wicked frigate GrrrrrrBack ;o)

FFS just tried loading y custom workflow (which runs in comfy no issues) after setting the Server as API to existing and its a Red Wall on (mainly) WASSuite Nodes

[zealous horizon]

Test, test, test.. it's up to us to give @wicked frigate good feedback and contribute.

[delicate grotto]

where is the k-sampler located in the comfy folder?

[cinder aspen]

I only need one easy setup where i can generate nice and good quality images, no fancy settings since i have Firefly/PS already for the customization part

[wicked frigate]

uhh are you sure you linked the correct comfy install? It shouldn't have changed your node list at all

[wicked frigate]

you mean SwarmUI? If so, can you explain more what you mean

[static prawn]

nono comfy in general

[wicked frigate]

ah okay

[zealous horizon]

check the server tab to see which Comfy instance you're using. If it's in Swar it will be inside dlbackends

[static prawn]

ur UI is the solution i was looking for

[soft zealot]

yup, ahh just noticed its also running a self starting one as well, lets stop that

[static prawn]

getting all that stuff in a more structured and clean way

i feel absolutely unable to work with comfy , especially on 1 monitor

[wicked frigate]

oh, lol, you didn't stop the install of comfy fast enough I guess? Can delete that and nuke the folder

[static prawn]

its just dragging stuff around, editing 5 resolutions, setting way to much samplers, etc.

i still hope auto1111 will bring a performance update, i customized anything so good for sd 1.5

dont wanna miss this

[wicked frigate]

auto's definitely working on it

got some good progress too last i looked

[soft zealot]

Lets just nuke the whole thing and start again after my conference call

[static prawn]

hopefully for lower end hardware aswell 😄

[wicked frigate]

that works too lol

[static prawn]

i always need 20 minutes or more for 1 image in auto at the moment

[normal rampart]

hey guys can i not use embedings with SDXL?

[edgy sierra]

and then?

[tribal moat]

Hey guys, is there any good prompting guide for SDXL??

[static prawn]

is it possible adding custom nodes and workflows in Swarm?

[zealous horizon]

main problem of auto1111 are the >2k issues on github. Thinggs just take a lot loger in Auto1111. Some of it is due to Auto being the first bbig app, suffering from issues that happen when ytou're the first one to implement things.

While I still love and use Auto on a daily basis I now moved to 90% Comfy for SD (100% for SDXL)

[normal rampart]

i guess but i like auto1111

[candid pond]

32 inch monitor is your friend

[static prawn]

i just dont understand how people are able to work with comfy 😄

i feel like its super unproductive

[candid pond]

I have an absolute monstrosity of a monitor and most comfy workflows still doesnt fit 😄

[static prawn]

I cant go for more than 24 Inch

atm

[candid pond]

It's a really great investment if you can afford it

Saves you a lot of time if you're working from pc

[static prawn]

i need the monitor for gaming too, and playing cs:go on a 32 is not possible

[lilac wren]

To me, since 1.0, results with comfy are shit. Greatly better with A1111

[static prawn]

im still happy i have a tn panel, highly prefer this over ips

[candid pond]

Comfy generates better results for me actually

[normal rampart]

yes but i cant figure out how to use the node for like roop and stuff comy is very confusing

[soft zealot]

@wicked frigate some more feedback for you.

There are a number of areas where you cant read the info boxes

@wicked frigate Yup clean install is better 🙂

[wicked frigate]

I'm tempted to remove those entirely cause they rarely fit lol

click the ? to actually show the info properly

[zealous horizon]

I felt the same. But you really get used to it super fast. When you use Comfy,you understand way better how SD works under the hood, thus improving your workflows and the final product.

[wicked frigate]

comfy is great for the technically-minded power user that wants to understand the internals

[static prawn]

i just dont actually understand how i get all the stuff for comfy i need

[molten breach]

Some people love node based workflows....I really struggle with them

[static prawn]

nodes / upscalers etc. probably with a good detailed guide i would understand

[soft zealot]

@wicked frigate Question:

(and yes I know Im pinging you a lot but............lol)

As Ive loaded an existing workflow why doesnt the Front end pick up on basic things like my Size settings etc

Actually the front end is only showing a Positive & Negative Prompt and not my Supporting Terms or Randomiser

(maybe my expecation is wrong)

[zealous horizon]

don't get hung up on custom nodes and extensions. You only need the basics to get great results.

[static prawn]

but the whole structure is terrible imo

i just want upscalers/highres fixes

i generate a lot on 1024x1024 with sdxl and doenst really like it, without some fixes 😄

[soft zealot]

Ok lets see what this does then 🙂

[wicked frigate]

you haven't applied the workflow, gotta clicky

that... should say 'Generate Tab' not text2image fuck

[soft zealot]

doh

[wicked frigate]

and yeah don't worry about the pingspam, basically every ping you've given has been something i need to fix lol

i really appreciate it

[static prawn]

but im on low end hardware anyway, probably wont be able to use sdxl in a good way i guess

[zealous horizon]

@regal vessel, try this one (it's vanilla, no extensions) I only added upscalers and highresfix (sd1.5 workflow)

[soft zealot]

I'll probably go quiet for a bit, technically I'm supposed to be working lol

[thorny frost]

Can't do it myself now but has somebody tried to use the refiner as base model reducing resolution to 512 or 768 to avoid double images and such to see what happens? You could then refine without having to change models.

[abstract comet]

nice can you share prompt? I've been trying to get clean colorized pencil drawings, no luck though only good results are black and white

[soft zealot]

@wicked frigate having said I'll go quiet in best Columbo fashion "just one more thing"

(and in fairness thi smay be me just skimming at he moment, need to sit down properly)
Its still showing the image size as 512x512 rather than the 1024x104 in my workflow.

I can se that linked to the aspect ratio button so it it ignoring whats specifed in the workflow???

It also only shows 1 Model Loader not 2 Model loaders (for Base & Refiner)

[static prawn]

where do u get the upscalers for comfy?

[zealous horizon]

this is a custom model. sry. It is merged to do children book illustrations but i's SD1.5

[soft zealot]

same places you would for A1111

[static prawn]

oh those are the same upscalers used?

[zealous horizon]

drop them in models/upscalers

[static prawn]

i thought they are different 😄

[soft zealot]

in fact if you already have them in an A1111 install just point COmfy at them in the COmfy Config file

[zealous horizon]

I just dopped them all in models/upscale_models

[wicked frigate]

refiner works as a base model but much worse at prompt-following and structure, wouldn't recommend it. Better to use Base alone than refiner alone if you don't want both

[wicked frigate]

you need to select a model in the Models tab at the bottom.

i'll fix the res bug

[static prawn]

i dont get scheduler, that option isnt in auto 1111

[soft zealot]

ah , see I said it was me skimming 🙂

Shut it for now and will pick up properly later

[shy kelp]

I don't know what everybody says about SDXL but regarding fingers and humans that are not close up it shucks. Can't understand why #midjourney is always years ahead

[wicked frigate]

oh, wait, no, it doesn't copy over base param values at all atm

i needa make it do that

[thorny frost]

Nobody can

[pearl shoal]

SDXL works fine for first 2 generation but crashes for 3rd generation. Can someone help? I have 12 GB Vram, 16 GB ram. Sd1.5 worked fine. I am using comfy ui. I tried using pytorch-cross-attention and dont-upcast-attention. They too crash after exactly 2 generation. Need help.

[static prawn]

at least the speed of comfy is fine, i need 2 minutes for 1 gen

[pearl shoal]

I am using sdxl base

[eternal fog]

How does it crash? Any errors what are the symptoms of the crash.
Do you have any overclocks or undervolts set on your system at all?

[pearl shoal]

Setting up MemoryEfficientCrossAttention. Query dim is 640, context_dim is 2048 and using 10 heads.
model_type EPS
adm 2816
making attention of type 'vanilla-xformers' with 512 in_channels
building MemoryEfficientAttnBlock with 512 in_channels...
Working with z of shape (1, 4, 32, 32) = 4096 dimensions.
making attention of type 'vanilla-xformers' with 512 in_channels
building MemoryEfficientAttnBlock with 512 in_channels...
left over keys: dict_keys(['cond_stage_model.clip_l.transformer.text_model.embeddings.position_ids'])
torch.Size([1, 1280]) 1024 1024 0 0 1024 1024
torch.Size([1, 1280]) 1024 1024 0 0 1024 1024
100%|███████████████████████████████████████████| 20/20 [00:20<00:00,  1.00s/it]
Prompt executed in 155.30 seconds
got prompt
torch.Size([1, 1280]) 1024 1024 0 0 1024 1024
torch.Size([1, 1280]) 1024 1024 0 0 1024 1024
100%|███████████████████████████████████████████| 20/20 [00:13<00:00,  1.50it/s]
Prompt executed in 21.74 seconds
got prompt
Killed

[pearl shoal]

It says got prompt and killed ^

[eternal fog]

What about overclocks or undervolts? Especially GPU.

[pearl shoal]

I dont do overclocking, what is undervolt? I dont think I did that.

[eternal fog]

You know if you did

I'm not sure then. Only time I've had comfyui crash is when I was using an undervolt, which is apparently stable in games, but not with SD

[pearl shoal]

SD1.5 works fine

[wicked frigate]

that was surprisingly easy :D it now transfer your base param values as expected (excluding prompt)

[pearl shoal]

it is SDXl which casue problem

[soft zealot]

cool, will be back with it this evening after a kebab

meanwhile

[static prawn]

i dunno why but my sdxl results are always super pixelated/ grainy

[indigo vine]

Hi, my image is 3072x1280. If you zoom on the plants, you will see it is really low rez, undefined. I upscaled by mixing 4xultrasharp and 4xNMKD from a 1536x640 image...Do you achieve a nice level of details at this rez? If yes, how?

[thorny frost]

I think that guy has worse problems than being pixelated.

[elfin cobalt]

Check dmesg

[tall lion]

chk out hugginface. There are some good ones. I have used this x4 upscaler with good results on SDXL 1.0 - https://huggingface.co/stabilityai/stable-diffusion-x4-upscaler

are you using base+refiner?

[thorny frost]

I have noticed some pixelation in human skin too, mainly in mid distance portrait, as if the model doesn't know how to do skin pores or something

[heady vale]

Im often seeing leathery skin texture when upscaling

[thorny frost]

Yep we have some skin issues.

[pearl shoal]

[ 8469.985905] audit: type=1400 audit(1690544506.803:5647): apparmor="DENIED" operation="ptrace" class="ptrace" profile="snap.discord.discord" pid=11098 comm="Utils" requested_mask="read" denied_mask="read" peer="unconfined" [ 8469.985908] audit: type=1400 audit(1690544506.803:5648): apparmor="DENIED" operation="ptrace" class="ptrace" profile="snap.discord.discord" pid=11098 comm="Utils" requested_mask="read" denied_mask="read" peer="unconfined" [ 8469.985910] audit: type=1400 audit(1690544506.803:5649): apparmor="DENIED" operation="ptrace" class="ptrace" profile="snap.discord.discord" pid=11098 comm="Utils" requested_mask="read" denied_mask="read" peer="unconfined" [ 8469.985959] audit: type=1400 audit(1690544506.803:5650): apparmor="DENIED" operation="ptrace" class="ptrace" profile="snap.discord.discord" pid=11098 comm="Utils" requested_mask="read" denied_mask="read" peer="unconfined"

it says this a hundred more times

[elfin cobalt]

None of those are related. You’ll want to filter out the discord stuff.

[wicked frigate]

if you have a very strong GPU you can do a highres step in the refiner at the full scale and improve detail that way

[uncut steeple]

You start it, the link probably wont work so you have to take the A1111 URL for example and switch the port to 8188

[tall lion]

I agree...in some of my gens so far, the fingers are kinda mixed up

[indigo vine]

I have a 3060 so I suppose it will not work.But I'm going to try a 1.5 ControlNet Tile img2img to see what it does...

[fringe barn]

SDXL is making my GPU cry blood 😭 🩸
I'm dreading switching to Comfy
Are there any public directories of workflows that I could use?

[wicked frigate]

if you're dreading it because the UI is complicated, https://github.com/Stability-AI/StableSwarmUI uses comfy's better backend, but has a clean n easy UI

[vital wolf]

Please do a video guide for it.

[fringe barn]

Appreciate the tip. I've looked at it. No clue how to get it installed. (I'm an old non-technical artist), so I'm waiting for some chum to make a video tutorial.

That said, I'm afraid it won't have all the plugins I need.

My current a1 workflow includes Txt2Img w/ Hi-res fix, Tiled-Diffusion, Tiled-VAE, triple ControlNet's, Latent Couple, and an X/Y/Z plot script

[woeful patio]

when can we start expecting models for sdxl?

[wicked frigate]

txt2img: yeah obviously
Hires fix: yep, added today
tiled-diffusion/VAE: not in the easy interface but i believe those are both available as Comfy nodes so you can do it that way
controlnet: SDXL controlnets aren't ready yet
latent couple: idk
Grids: StableSwarmUI has the best grid tool in the world lol. A direct port of the Infinity Grid script for auto webui, with some added upgrades

[wicked frigate]

The official models are public, community models are already starting to get posted on civitai

[fringe barn]

Thank you for that

[woeful patio]

amazing thanks 🙂

is there a way for hires fix to work with a1111 on 1024px. My poor 8gb gpu just keeps giving me memory exceeded errors.

[wicked frigate]

SDXL starts at 1024, hires fix would take you even higher (like 2048 or so)

SDXL at 1024 works within an 8GiB GPU fine on ComfyUI or StableSwarmUI

(or, the base does at least, refiner is more needy)

[woeful patio]

yeah it works great on a1111 too, only when i try to add the hires fix on top off that do i get the memory error. I was just wondering if there was a way to upscale the images within the 8gb vram, but thats probably asking for a bit much

[vocal stream]

isnt it way worse without refiner though?

[uncut steeple]

SwarmUI doesnt load the ComfyManager for example, does it?

[delicate grotto]

it also works on 1440p at 6gb vram

[solemn scroll]

Yes it does : #✨｜sdxl message

[uncut steeple]

Or any extensions

[wicked frigate]

It literally just runs a comfy install, you can load anything onto it that you want

you can even use an external copy of comfy

it all just works ™️ do whatever you want

[uncut steeple]

But it doesnt load my comfyUI fully then it seems

[wicked frigate]

?

that sounds like ya did something wrong then?

[uncut steeple]

you see how it doesnt show the manager button?

[pearl shoal]

Out of memory: Killed process 13718 (python3) total-vm:51390488kB, anon-rss:14263076kB, file-rss:78232kB, shmem-rss:4108kB, UID:1000 pgtables:35332kB oom_score_adj:0 Why does first 2 generation work and it is only out of memeory for 3rd?

[vital wolf]

Yeah, I also tried, mine also doesn't show the manager button.

[uncut steeple]

Also the custom script settings dont appear as well

[edgy otter]

Thank you ❤️

[wicked frigate]

Is the backend Selfstart or API by URL?

[uncut steeple]

by URL

[edgy otter]

~1.5 iterations/second, 1024x1024

[wicked frigate]

... actually also, are there errors in your browser devtools?

[inland sandal]

BROH WTF IS THIS SHIT? since when SD has become nodes like blender?

[wicked frigate]

the manager might be broken on subpathed connections? That was only added to comfy recently

[inland sandal]

Please clarify what is that shit o:?

[uncut steeple]

No errors nope

[wicked frigate]

that's ComfyUI, a node-based UI for SD

[inland sandal]

WTF????

waht for?

such complexity

[wicked frigate]

for people who like complexity

[inland sandal]

i need an AI to manage that

[uncut steeple]

🫠

[inland sandal]

explain briefly what can youa chive with that?

[edgy otter]

for complete control over how an image is generated

[wicked frigate]

stuff that normally you could only control by writing your own python code

[inland sandal]

image or prompt?

its like generating minimal changes like scale and iterations with suttle changes in propt at the same time ?

any video explanatory?

sounds and seems like quatum computing

[woeful patio]

in ultimatesd upscale do i leave the tile size at 512 or do i up it to 1024 since i'm using sdxl?

[uncut steeple]

Im using it with 768 for SDXL, 1024 didnt seem too good for me

[smoky bough]

After you install comfyUI you can drag and drop the pics in this link for a workflow you can start from: https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

[pearl shoal]

Can sdxl work with cuda 11.8?

[inland sandal]

thanksy

is a 4090 good for 512x512 iamge generation?

how fast it is with sdx;?

[uncut steeple]

SDXL is made for 1024x1024

about 8s in a basic setup with a 4090 for 1 image

At least for me

milage may vary

[static prawn]

someone here using a 3060:?

[uncut steeple]

Still trying to get better skin texture

[pure hazel]

Does this work with 1111's?

[soft zealot]

why shouldn't they ?

[uncut steeple]

Cause low VRAM probably, but should be fine

[static prawn]

im just interested in performance

wanna upgrade to a 3060 probably in the future

[rigid prawn]

there is a way to add the date on the name of comfy output files? If I put [time(%Y-%m-%d)] in filename_prefix it always overwrite the 0001

[uncut steeple]

[time(%Y-%m-%d)]

Oh I should read it fully

[soft zealot]

I did some trawling of the data at https://vladmandic.github.io/sd-extension-system-info/pages/benchmark.html a while back to build a relative performance comparison chart (just added the 4060 as well for refeence)

Also at https://docs.google.com/spreadsheets/d/1lCFek25kHy0ENxjvr8sLkF-uBX_QVT0X/edit?usp=sharing&ouid=109983571622180941171&rtpof=true&sd=true

[inland sandal]

another question if I have 25 its per scond ? that means i can make a 25 step image in 1 second?

what means it/s?

[uncut steeple]

For me that worked, but im using a custom image save node

[echo pine]

I trained a lora for sdxl and put it into the comfy lora folder but can't seem to get it to show up in the UI, are there any other steps required after training or something? my lora is 870mb but the test lora from stability is +-40mb so I'm clearly doing somethingg wrong

[molten gull]

how did you train your lora ?

[uncut steeple]

your LORA is 870 GB?

[molten gull]

and whats the file extension ?

and yes, you pretty clearly did something wrong

[rigid prawn]

me too, the was suite one. can you send me an IMG of the node settings?

[echo pine]

mb* 🤣

[molten gull]

whats the complete filename of your lora ?

[echo pine]

.safetensors

[molten gull]

with which program did you train it?

[echo pine]

y2k_ootd.safetensors and I used bmaltais/kohya_ss SDXL -LoRa finetuning phase 1(and 2)_v1.1

[molten gull]

you're using comfyUI, i guess ? what node do you use to load the lora ?

[echo pine]

just the load lora node

[glad fulcrum]

When will you optimize Lora for highvram? Without highvram my generations take so long because the model loads every time I make one image and it takes forever. Will it take a lot of work / time? Thank you

[molten gull]

do a F5 reload of the gui, and check if it's in there then

[grave crane]

can i run Xl with 6 GB on NVIDIA GeForce GTX 1660 SUPER. also
how do i even install stablestudio when no shell knows git as program

[echo pine]

nope, if I click on the arrow it just says [object object] and goes back to the sdxl offset example lora

[molten gull]

you did a reload?

[echo pine]

yeah

[uncut steeple]

Enough eyes for now

[molten gull]

try train one with this here: https://github.com/derrian-distro/LoRA_Easy_Training_Scripts

and can you send a screenshot of your comfyUI pls? with the part where you load stuff

[molten gull]

that looks like it did load your lora ?

by the way, i m pretty sure that lora+refiner = not working

[echo pine]

no that's the sdxl example lora, not the one I trained

[molten gull]

ah, ok

[tawdry trellis]

I'm using automatic 1111. Is is possible to let automatic run the refiner automatically? Or do I need to switch from base to refiner everytime?

[fierce hollow]

size depends on the dimensions, sai's is only like 4, you probably trained a locon with 128 dim and 128 conv dim or even something bigger

[molten gull]

in what directory did you put the lora file ? can you do a screenshot of this, too ?

[echo pine]

I trained with 128 dims yeah

[fierce hollow]

yeah, hence the size, it shouldn't be an issue when loading though

[fringe barn]

In Swarm's Server Config settings, do you guys know the difference between ModelRoot and SDModelFolder?

[echo pine]

yea its weird, I'll try the other trainer, maybe that will work

[molten gull]

well, it looks weirdly huge for a lora

mine are all about 100mb

[glad fulcrum]

Thats a big Lora 🤣🤣🤣🤣

[fierce hollow]

sdxl loras are bigger than sd in general, it's normal with 128 dims

[slow marsh]

Hi, do anyone know where to download SDXL beta (2.2.2 version)? I found good results from that model

[glad fulcrum]

Is there a way to utilize styles in comfyUI like auto1111 does?

[sharp robin]

? That don’t exist?

[echo pine]

it does, it was a version of a couple months ago, just never released (i think)

[fierce hollow]

never heard about it

[slow marsh]

It does I believe, check this out, the XL version on this website is 2.2.2 versionhttps://platform.stability.ai/sandbox/image-to-image

[primal vault]

Experimenting with prompting the same for both CLIP models in base vs. separate prompts for CLIP G and CLIP L.
1st image: prompted both CLIP in base with a CLIP L optimized prompt (SD 1.5 style)
2nd image: prompted both CLIP in base with a CLIP G optimized prompt (new prompt style, better for SDXL)
3rd image: prompted base with separate prompts for CLIP G and CLIP L, each of them optimized for their respective CLIP model
I generated both prompts with CLIP interrogator, on a screenshot from the movie Oppenheimer

[slow marsh]

Exactly, I can't find that online sadly, because I really find good results with that model

Does anyone know where can we download the SDXL Beta (2.2.2v)? I would be really appreciate!

[fierce hollow]

if nobody leaked it in some way then you're out of luck

I see references to it existing but not in the form of a checkpoint you can just download and run

[slow marsh]

damn, does midjourney v4 open source then?

[fierce hollow]

midjourney and open source don't mix well

[slow marsh]

means if I host midjourney v4 to my google cloud machine, it won't perfect well?

[shy kelp]

@tall lion I have to take the bad gens to photoshop beta and fix the fingers. It does a great job

[wicked frigate]

ModelRoot is models/, the root
SDModelFolder is Stable-diffusion, the sub-folder (in the near future other sub-folders like Lora will exist)
You can click the purple ? on any of the options to see an explanation of them.

[wicked frigate]

the old beta model is not public, but the refiner model is closer to the beta than the base is, so you can use the refiner on its own if you want something close.

[slow marsh]

Thanks Alex! Can I know where got refiner XL 1.0 to be test online? Because I'm not sure do the refiner known as SDXL v1.0 on https://dreamstudio.ai/generate or it is a base, do you know this?

[wicked frigate]

er, if you want to try it online you can use swarmui's colab https://colab.research.google.com/github/Stability-AI/StableSwarmUI/blob/master/colab/colab-notebook.ipynb

there might be sites that offer it but it's not really a normal way to run

[indigo vine]

did some similar tests today. From what I found, CLIP_L= OpenAI = telegraphic prompt style like: english man, 19th, peaky blinders etc. And Prompt_G=OpenClip= sentence prompting like "an english man from the 19th century in front of a house in the style of peaky blinders etc. And you repeat approximately the same thing between the two. Correct? Did you find some noticeable differences in your tests? for me, I had better result copy pasting the prompt between the two...And it's also confusing to have those two prompts plus style.

[slow marsh]

Got it, let me try, can I message you personally if I find any updates?

[peak dove]

Is there any way to start A1111 so that it automatically sets to 1024x1024? It always starts at 512x512!

[tawdry trellis]

I get this error when using the refiner, what could be the problem? RuntimeError: The size of tensor a (768) must match the size of tensor b (640) at non-singleton dimension 1

[primal vault]

Just using the same prompt can already lead to great results. I just found that I have more control over details with prompting G and L separately, once I find a seed with a good compositon as base to iterate on

I'll often start with the same prompt and then refine each of them separately for adjusting details. works pretty well most of the time

btw, comfy workflows are embedded in the pictures I posted, but they need my custom node extension for comfy to work

[woeful patio]

is it me or do hands come out a lot more consistent too in sdxl?

[glad fulcrum]

Hands are horrible still most of the time

At least in my experience

[indigo vine]

I am using it already! thanx a lot for the git repo, it's a good start to experimenting!

[meager canopy]

It mentions Peaky Blinders a lot, for Oppenheimer 😄
...and what's that secondary prompt about?!

a close up of a person wearing a suit and tie, peaky blinders, teal color graded, focus on two androids, interstellar movie

[primal vault]

I hope the reborn workflow also makes it a bit easier to quickly experiment with different prompt styles and ways to combine prompts

[slow marsh]

Anyone know where to test SD refiner model online?

[primal vault]

as I said, I got those prompts from clip interrogator and copy/pasted them for this test

[wicked frigate]

you're welcome to @ me in this discord with any questions

[slow marsh]

Thank you!

[peak dove]

I am using Sytan's workflow, I just have to move safetensors to match his file placement

[primal vault]

in this workflow the G and L clip models are prompted by main and secondary prompt, the style prompt is used for both clip models

[wicked frigate]

it can be helpful to emphasize the subject in the G input, and the style in the L input - L tends to focus on stylistic information and G on content/structure

[meager canopy]

Is there a written explanation of how the different prompts (L, G, R) are supposed to be used?

[wicked frigate]

... what is R lol

and: generally when in doubt don't worry about it and just use one prompt

[dapper current]

Who added a third encoder

[wicked frigate]

beyond that is an experimental topic

[primal vault]

that's pretty much my experience as well, in my older workflows the L was considered the style prompt, but with the new approach I get more consistent results across multiple seeds and don't have to cherry pick as much from larger sets of generated images

[meager canopy]

@primal vault ?

[wicked frigate]

oh, refiner

[primal vault]

someone asked me on Github about G, L, and R, here is my reply: https://github.com/SeargeDP/SeargeSDXL/issues/5

[deft coral]

does anyone know where 1.0 will be trainable

[meager canopy]

Thanks 🙂

[upbeat summit]

@wicked frigate Not a question. Just wanted to say thank you for your work and I'm looking forward to trying out StableSwarm! 🙂

[sleek quest]

I am, works fine with Comfy,

[meager canopy]

What happened to the other model candidates that were being considered for 1.0? Will they ever be released or used?

[deft coral]

i might make this in SDXL

[sharp robin]

They are likely not. I mean the difference cant be that drastic?

Would just add confusion

[meager canopy]

@wicked frigate do you know?

[ember cargo]

is it possible on comfy to load all loras and load them in the prompt with the weight and all (like in auto1111)? 🤔

[timber quiver]

yes, check these examples https://comfyanonymous.github.io/ComfyUI_examples/lora/

[ember cargo]

i already do that and adding more and more nodes for each lora seems like a pain

it's not all that bad if the lora is trained with captions though

[wicked frigate]

I don't, that wouldn't be my decision. You could beg Joe or Emad about it if you really want em.
It's possible some of the theories used to make the other candidates will go into future models (eg an SDXL 1.1 or 2.0 or an SD 3.0 or whatever it is we publish next)

(Some of the theories were really cool)

[runic lark]

I keep getting Downloading model.safetensors when i restart A1111 (3rd time since SDXL installation). Do you know if there are updates daily in the beginning, or is it something with my arguments that redownloads the models?
--medvram --no-half-vae --opt-sdp-attention

[cunning geyser]

Do sdxl loras work on 1.5

Contemplating testing my anime lora on 1.5 instead since sdxl has no anime rn

[ember cargo]

This is my very compact and tidy workflow, it has base + refiner + 2 lora + SXL Aspect Ratio and SXL Prompt Styler nodes

[upbeat summit]

thank you for sharing it!

[ember cargo]

you're welcome!

[soft zealot]

@wicked frigate random question:

Do you know why I keep seeing this line whenever the SDXL models are loaded (in COmfy) please?

Doesnt seem to impact anything just call me hastag curious

missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'}

[indigo vine]

@primal vault

• 1.5 prompt: photo, marble statue, Artemis, museum entrance, Fujifilm X-T3, marble veins, intricate details, sharp reflections, great perspective
• SDXL prompt: photo of a marble statue of Artemis standing in the entrance of a prestigious museum shot on Fujifilm X-T3 film. The veins of the marble are visible and the statue if full of small nice details. The reflections are sharp and the perspective interesting.

using Reborn:

• image1: promptG: SDXL, promptL: SDXL
• image2: promptG: SDXL, promptL:1.5
• image3: promptG: 1.5, promptL: 1.5

[wicked frigate]

that's a question for @visual glade rather than me, but I think it's normal/fine

[cunning geyser]

anyone got a config for training loras?

SDXL loras that is

[soft zealot]

NPs I was just curious if it was something you guys had noted during testing as it seems to be model related 🙂

[wicked frigate]

i've seen errors about that key being discussed repeatedly... usually unexpected extra key rather than missing key

i never actually learned myself what's up with it, i assume it's something silly based on different tools having different opinions of whether it should be there

[soft zealot]

fairy snuff 🙂

[soft zealot]

oh I forgot the second line lol

missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'}
left over keys: dict_keys(['denoiser.log_sigmas', 'denoiser.sigmas'])

[primal vault]

nice comparison, in those images the 1st prompt combo seems to create the result I like most

[sage basin]

I started seeing that error back on 0.9 when I converted the CLIP model to fp16 (there's an option to force position ids, but I never bothered to try it seeing how everything worked fine)

[cunning geyser]

Anyone make loras on SDXL yet?

[polar epoch]

I would love to find a esrgan_2x so i can double my res and not having to upgrade to 128GB ram

[polar epoch]

https://github.com/derrian-distro/LoRA_Easy_Training_Scripts You can try this one, don't remember if i found it up in chat just now as my short term memory needs a replacement lol

[civic forge]

I have a question about comfyUI and LORAS, the standar workflow is to load a checkpoint, then a lora, then pass to clip text encoders and then to the ksampler. But since SDXL has two checkpoints, Do I have to load the lora for the refiner model too?

[cunning geyser]

Which version of torch do i install

[upbeat summit]

@soft zealot thanks again for sharing your workflow and all the info! I guess (hope) I'm not a total noob but that workflow has so many features - at first I thought I was looking at blueprints for a new LHC installation 😄

I'm still analyzing it and have not yet figured out all the routes but it's great to learn more about ComfyUI and workflows in general.

[indigo carbon]

duck ball

[uncut steeple]

Amazing

[trim orbit]

i often feel that way with many node editors. i recognize they're powerful but wow do i just gloss over when i see a new one.

they're not intuitive ui's

[civic forge]

They are intuitive if you know what you are doing, but most of us don't know.

[trim orbit]

yeah i suppose if you need to rely on experienced knowledge instead of intuition, they're intuitive

[wicked frigate]

lmao

[still dove]

Can I upload your image to load this workflow?

[trim orbit]

i'm grinding. i'm gonna get there. i'm no newb either. i just need some time to acclimate

[ember cargo]

yep, it should contain the WF

[still dove]

Hope so

[meager canopy]

It doesn't

[trim orbit]

There was a node i saw recently that lets me use lora's in the text prompts. i want this node.

wiring up a lora everytime i want to use one just makes no sense to me.

[upbeat summit]

I know what you mean. I always wanted a node based tool for AI images because you can build your own workflows and are not locked in into a process. but there are always pros and cons depending on what you want to do.

[trim orbit]

here it is. going to set this one up and get it in my workflows

[meager canopy]

You can add a switchable node for lora
https://github.com/RockOfFire/ComfyUI_Comfyroll_CustomNodes

[civic forge]

I mean they are not intuitive If you have zero knowledge, but they become pretty usable once you start to know a tiny bit.

[uncut steeple]

[civic forge]

It's like dark souls, It seems imposibly hard but once you learn the basics is not that hard

[cedar locust]

incredible, what's the prompt lol?

[ember cargo]

This one should contain it. And the other did too

[trim orbit]

i saw that pack, but it just complicates things it seems. adds like 18 new nodes that all have to be used in tandem.

[uncut steeple]

You can drop it into comfy

[trim orbit]

i don't think node based editors are like dark souls. pretty sure this doesn't need elaborating

[meager canopy]

No they don't. You just add the ones you want to use, ie. "CR Load Lora"

[indigo carbon]

it won't be the same, I'm not on Comfy

the prompt should be there tho

[uncut steeple]

Yep it is

[trim orbit]

when i first found CR nodes they were all dependent on each other. CR had to be used with CR. i guess they've changed that. yay intuition.

i'll look into it

[meager canopy]

I'm about to try it, so I'll let you know

[uncut steeple]

[trim orbit]

i first looked into it for the switches it implements. would be nice to have a workflow like that

[indigo carbon]

oh no, you did something horribly wrong lol

[uncut steeple]

Or very right

[indigo carbon]

whatever you want to call it -_-

[civic forge]

Well, there are patches for both dark souls and node based editors 🤣

[meager canopy]

Works just fine 🙂

[trim orbit]

well i dont know the point of this argument, insisting that ds and node editors are the same. i feel like the metaphor is falling apart but you're grasping. i'm annoyed and i'm out.

[indigo carbon]

however, I'll switch to Comfy in a heartbeat if it will support TensorRT before A1111 does

[uncut steeple]

mood

[trim orbit]

thanks!

[civic forge]

Don't get angry, I'm just kidding.

[indigo carbon]

is this with the refiner?

[uncut steeple]

Yeah but not with the upscaler

[indigo carbon]

damn. if you used the same configuration as I did the refiner screwed it up lol, the detail is way off with your gen.

[wicked frigate]

i was going to say that comfy's mentioned it before but #✨｜sdxl message it was to you lol

[indigo carbon]

he didn't make it public, publish it or it never happened

[civic forge]

I've been experimenting with comfyUI and I don't know If the refiner model accepts LORAs, I've tried but loading a lora in the refiner model doesn't seem to affect the result.

[trim orbit]

you didn't read it did you

[uncut steeple]

My setup is tuned for realistic stuff atm, I just threw this in

[meager canopy]

Where did all the square lines appear from? I updated custom nodes and comfyUI, so not sure which may have caused it

[uncut steeple]

Its in the settings

Link Render Mode

[indigo carbon]

I did. I don't agree that it shouldn't be on local inference. it can even triple speed when configured correctly without losing precision. I just don't understand why no one thinks this is a big deal.

[meager canopy]

Thanks!

[trim orbit]

its unreliable and hard to set up. benefits are better at larger scale

[wicked frigate]

the refiner can use loras in theory but they need to be separately trained from base loras

[indigo carbon]

still probably way worth it. but this is just my opinion. I'm sure the TRT conversion can be automated though,

[trim orbit]

you're sure . hm. sounds easy. great thing about FOSS....

[soft zealot]

ahh has @visual glade integrated this (or similar) then ???

Link Tidier. Save in \ComfyUI\web\extensions (0=straight 1= slight curves 3 = Standard) https://drive.google.com/file/d/11pAM5HW3S72DQv4Qw6cXQXTwCrLFlN2o/view?usp=sharing

[uncut steeple]

Nope its from an extension

[indigo carbon]

I used it all the time when 1.5 was still being used, it was incredible- I got to 30it/s with TensorRT. the conversion wasn't that complicated.

[civic forge]

Thanks! so I was doing the right thing but the LORA I've used for base model is not enough, I need the same LORA but trained for the refiner model too.

[trim orbit]

unreliable doesn't mean it never works

[woeful patio]

what do you guys think is the best way to upscale an actual low resolution image in sdxl?

[meager canopy]

Isn't the offset example lora supposed to make things look a little darker?

[soft zealot]

Thats what I mean I use one I DLd sperately but @meager canopy seemed to be saying they had just appeared

[meager canopy]

It's part of an extension, but it seemed to enable itself after an update

[indigo carbon]

what's unreliable about it? you would just set min resolution and max resolution, and it would convert itself. I don't see why this can't be an automated process with some kind of extension

[trim orbit]

if it was reliable and could be done on most configurations, it would be out by now.

[wicked frigate]

the best way for that rn is a bit limited from what it will be in time.
Cause the best way to do that in SDv1 is: use a controlnet. SDXL controlnets aren't released yet :(

[trim orbit]

perchance there are other factors in play other than one single vanilla system configuration

[indigo carbon]

it was.. It only worked on 1.5 though. NVIDIA said they are working on a specialized version of it built specifically for SD

[woeful patio]

everytime i tried doing it in sd 1.5 I got that uncanny look, I hope it will be better with XL

[meager canopy]

These images were same prompt/seed, the only difference was one used the SDXL offset Lora and the other didn't. The one with the lighter face and jacket used the Lora 🤷🏻‍♂️

[hearty ginkgo]

Dosnt the background have some kind of weird blur effect as opposed to the one with the Lura looking more normal effect for blur?

[lilac wren]

Have a link about this?

[meager canopy]

I think that's just the upscaling

[wicked frigate]

what was the prompt?

[indigo carbon]

from the extension's repo.

[dapper current]

Could be the refiner as well

If it was used

[hearty ginkgo]

Was both of them upscaled?

[meager canopy]

Yes

[tribal gale]

an offset doesn't inherently mean a darker scene

[dapper current]

A disadvantage to using the refiner is it can change the output of a trained LoRA on the base.

Also offset can produce more contrast and not necessarily just dark

[wicked frigate]

the lighter one does have more contrast visibly

[tribal gale]

yep

[wicked frigate]

it has spots just as dark as the non-lora one does, but also much brighter sections

[hearty ginkgo]

This is without a Lora for base and refiner and also 4x upscale

[visual glade]

yeah that's normal, that's just debug info for me

[glad fulcrum]

@visual glade When will you optimize Lora for highvram? Without highvram my generations take so long because the model loads every time I make one image and it takes forever. Will it take a lot of work / time? Thank you

[uncut steeple]

She looks kinda long

[visual glade]

I will rework a bit how memory is handled soon to take less ram so once I do that

[meager canopy]

This is with the Lora

[civic forge]

I bet prompt is: Protagonist of all modern videogame/netflix production/young adult media. I have no proof but I have no doubt either 🤣

[glad fulcrum]

She grew a ear

[meager canopy]

I think it was always there 😄

[normal geyser]

Has anyone tested apple Mixed-Bit SDXL on MacOS 14?
https://huggingface.co/apple/coreml-stable-diffusion-mixed-bit-palettization

[azure oxide]

wait i missed this message, this is so hype

[meager canopy]

Both of those images used female actress names in the prompt! 🤷🏻‍♂️

[uncut steeple]

Anyway to give comfyUI output images a unique name like A1111 does

[dapper current]

That's probably the first I've heard that it's preferring men - what was the remainder of the prompt? There must be a bias

[meager canopy]

a close-up of a beautiful __actress__ wearing a suit and tie, peaky blinders (2018), attractive, wearing teal beanie, beige fedora, in roger deakins style, fire behind him, jaguar, red ascot and a field cap, robe, perfect face, england, beautiful face, long dark shadows, late 30s aged, incredible wide screenshot, diamond, fujifilm, highly detailed, vintage photography

[uncut steeple]

peaky blinders probably doing its part

[sharp robin]

Yep. Peaky blinders all men mostly

[compact flax]

Woah, how do I use this on CUDA?

[azure oxide]

someone linked this comment earlier, what's the problem with tensorRT for local inference due to this? are you just referring to the fact that LORAs dont work with it?

[meager canopy]

I changed the prompt around, mainly removing in roger deakins style...looks better! 🙂

[ruby marsh]

first SDXL video 😈
https://youtube.com/shorts/U3-LmYx5IvY

[primal vault]

cool, I didn't look at the example lora so far, but now I'll have to put that on my todo list

[static prawn]

A tensor with all NaNs was produced in VAE.
Web UI will now convert VAE into 32-bit float and retry.
To disable this behavior, disable the 'Automaticlly revert VAE to 32-bit floats' setting.
To always start with 32-bit VAE, use --no-half-vae commandline flag.

what does that mean, i only get this for sdxl

[soft zealot]

@wicked frigate Ok I broked it lol

Im guessing there s a config file hiding somewhere I can edit to get past this rather than a fresh install???

[hearty ginkgo]

No I was just bored and tried to do a girl from the future with furistic cloths that looked like futuristic body armor something like that idk

[leaden marlin]

any idea when controlnet will be available for SDXL

[uncut steeple]

maybe

[hoary saddle]

Z:...X:... Got a bad path somewhere, two drive letters in one

[soft zealot]

@wicked frigate thought I had fixed it by editing Settings.fds but now got this

[wicked frigate]

heh, well i should fix it to die less hard in that case - but, yes, StableSwarmUI/Data/settings.fds

[wicked frigate]

fuckin what

OH you put the 8188 comfy port as your swarmui port

and it got very confused

[soft zealot]

was that fo rmy benefit lol

[uncut steeple]

No cursing on my christian minecraft server

[soft zealot]

somewhere I have fucked thi sup, lets go wipe and start from scrathc thi stime paying attention lol

[wicked frigate]

you're giving me more actionable feedback in 1 day than anyone else has in months btw, cannot say enough how much i appreciate it

[soft zealot]

I have a reputation in work for breaking things as well lol

want something tested, give it to Mike

[wicked frigate]

are you telling me this whole time your name wasn't actually Winston ;-;

[soft zealot]

yup lol

[wicked frigate]

a deception... gaslighting...

lol

[soft zealot]

this is The Real Winston

this is me 🙂

[azure oxide]

nice glass

[bold osprey]

ruperts drop

[bold osprey]

you look great man

[trail obsidian]

whats your guys favourite sampler for base image and high res fix?

for SDXL of course

using the dreamshaper alpha model right now

[uncut steeple]

Im currently with dpmpp 2s ancestral for base and upscaler and euler a for refiner

[soft zealot]

@wicked frigate one more for you I just noticed
It sayas Demo Test Page on my Tab

[soft zealot]

@wicked frigate ok so I know whey I eneded up with that stupid path error.

I stuidly though that I could use this sectiion to link to my Shared Models folder on My X Drive.

I relaise now that was erroneous and that I should leave thi salone and do the same as I do in A1111 & COmdy and use SymLinks 🙂

[trail obsidian]

alright yeah same for me dpmpp 2s ancestral has been amazing at like 30 steps and i was using DPM 2 ancestral without PP for high res as PP changes the picture and morphs things weirdly. Im trying Euler a for the high res fix now

[static prawn]

looks kinda ok, still always getting oversharped images / or images that look a lil pixelated/grainy

[ionic dragon]

For real?

[static prawn]

using 1.0 vae

[soft zealot]

last Time I looked in a mirror yes (althugh my hairs got longer since then)

[ionic dragon]

Use 0.9 vae for now, its better
1.0 would be fixed

[static prawn]

oh i think there was allready a fix for 1.0

i downloaded a second version of it

[ionic dragon]

You're the oldest mam I visited on discord

[unreal grove]

nsfw filter pretty much kills this

[ionic dragon]

Ohz I don't know about,
On vacation from 2 days

[static prawn]

i still dont know how to handle the prompts the images never look really promising

[ionic dragon]

But NSFW would be soon bypassed by models

[soft zealot]

60 next March 🙂

(now y'all know why I can be a grumpy old git at times lol)

[unreal grove]

yeah but running it locally isn't an option for all. this used to be the best

[wicked frigate]

HAHAHA ohmygod i never changed that

[soft zealot]

and noone has noticed before?

sigh

[wicked frigate]

who looks at tabs? tabs aren't real, never heard of em

.>

<.<

[azure oxide]

it was the first thing i noticed when it was revealed but i assumed it was intentional with it being alpha and everything

[wicked frigate]

that was literally my day 1 placeholder and i just forgot about it

when the first version was just a concept demo

[soft zealot]

thats better

[civic forge]

How did you do to get good text on the t-shirt... Oh! It's a real photo!

[wicked frigate]

you should be able to in theory

I think you'd have to redirect ModelRoot rather than SDModelFolder though?

[soft zealot]

is Ok I done my usua trick of symlinking and then changing if reqd in the UI config

[wicked frigate]

i mean... sdxl base can get close...

[wicked frigate]

symlinks are magic

[idle pelican]

Will SD add the functionality of their bots to be able to add them to our own servers ?

[soft zealot]

thats annoying although its now pulling inthe Resolution the options all seemto presume a 1.5 model it feels like

[idle pelican]

much like midjourney, blue willow etc

[wicked frigate]

public discord bot from us might happen - but even if not, other people can and have made SD bots

it'd be really nice to have an official bot you can easily add tho i agree

[soft zealot]

aye I just keep Batch fi8les in my UI folders to resotre them if a UI upgrade overwrites them

[idle pelican]

yeah ive seen some of them but a official one always feels the best and looks the best

especially on a server with 36k+ members

[wicked frigate]

lol. for text-gen-webui i've just learned the practice of, rename it, git pull, rename it back .-.
if swarmui ever kills a symlink please tell me asap, i don't want it doing that

[wicked frigate]

it chooses res based on model type

so Type: Stable Diffusion XL 1.0-Base should be relative to 1024 if selected

[soft zealot]

ah I havent select ed a model yet

[wicked frigate]

oh, lol, well there's ya problem

[civic forge]

I'm surprised, that's a pretty good text for stable diffusion standards. It still looks the one who was typing had a seizure when he was halfway done, but It looks far better than 1.5

[wicked frigate]

it assumes 512 as a default on the grounds that if it doesn't know what you're running it's probably sdv1.5, since SDXL models are clearly identified

[soft zealot]

This is annoying, cant I just have a simple drop down list sorted alphabetically, this all takes upp too much space and has silly pictures ;o)

[wicked frigate]

(the reason they're clearly identified is https://github.com/Stability-AI/ModelSpec which i'm so happy i got officialized)

[static prawn]

i like the result, but i cant fucking get rid of that pixelation

[wicked frigate]

click that, change to list

[versed locust]

Where do i put the refiner? In with the models?

[wicked frigate]

yep

[soft zealot]

nope dont like this view either

[wicked frigate]

oh that's a lil jank ain't it

[boreal bough]

XD

[wicked frigate]

yknow what hold up i got you

[boreal bough]

is the ui part template based html/css, or all in C?

[sharp robin]

Yeah. Basically it’s just like a model built for adding details.

[static prawn]

when i crop u see the pixelation extremly

[wicked frigate]

HTML/CSS/JS

[boreal bough]

❤️
might send you push request or code snippets for styling

[soft zealot]

meanwhile still cant see an option to add the refiner as a second model

Its not even mentioned in the lsit on the left and it is in th eGraph View (think theres a load of others missing as well but jsut about to have a Kebab to settle my beer)

[spark sinew]

can you share the workflow?, thk

[vital cairn]

having a lot of fun just playing with SDXL.

toothless wearing a chefs hat and cooking a cake with gordon ramsay:

[marble dew]

https://tenor.com/view/gordon-ramsay-idiot-sandwich-angry-mad-what-are-you-gif-4169547

[soft zealot]

and in fairness some of it my faulyt for not changing Node Titles , look at all those primitives lol

[marble dew]

one day we will recreate gordon sandwiching folk

[boreal bough]

So I'm not sure what to think about the base 1.0 model :/
it works great with refiner, but that invalidates a lot of lora possibilities.

on the other hand, if I finetune my own set of facedata (to use base only), I'll have to add that dataset to each lora I make, or keep training from that checkpoint :/

not really ideal.

[ionic dragon]

Btw how did you discover discord?

[marble dew]

use ultimate upscaler

[ionic dragon]

Oh, yeah, if you can afford you can rent GPUs

[soft zealot]

I typed http:discord.com ina web browser

stupid question lol

[ionic dragon]

No, I mean who introduced you to discord?

Nvm

[soft zealot]

you pesky kids think you invented the internet lol

[ionic dragon]

No

[soft zealot]

discord is effectively a 21st Century Usenet or IRC, its nothing special that hasnt been done before

[autumn forum]

Hey @boreal bough ! Guess what! My Lora actually worked! So I’m gunna move on to a higher amount of images and train that as it’s not completely what I want yet

[soft zealot]

actually more IRC than Usenet

[sharp robin]

this looks like VAE1.0 try VAE 0.9

[ember cargo]

drop into comfy (i deleted this image form the original comment by mistake)

[static prawn]

nothing to do with that

[sharp robin]

do we need to feed the lora into the refiner aswell?

for SDXL?

[boreal bough]

@hard fractal
(Context: base only workflow, due to refiner not accepting unknown/new concepts)
currently the only Idea I'm actively pursuing for complex loras, is to save the training data in kohya, which has been trained on a large enough face/hand dataset - then provide that kohya checkpoint as community checkpoint to continue training from there. It works, but is far from ideal.
Do you have any ideas internally that I could pursue other than this?

I mean at this rate I'm just making a glorified hyperdreambooth with more parameters.
Relevant because loras that modify any body parts or face weights currently do not play well with one another :/

[azure oxide]

due to refiner not accepting unknown/new concepts
wait is this true?

[ember cargo]

I had some luck by only doing a few refiner steps before base + lora processing

[sharp robin]

ur doing refiner first?

[wicked frigate]

tharr, git pull - the place where it displays the current model is now clickable and it gives you the classic dropdown listing

[boreal bough]

a bit too short formulated.
Refiner can be finetuned just fine - this is only in context of loras

[ember cargo]

load the image i shared earlier and you'll see

[wicked frigate]

oh yes please! if you're good with CSS gimme PRs

[sharp robin]

it seems with a lora that the refiner has no clue what is going on

[sharp robin]

guh u stole my name

[boreal bough]

webdev/webdesigner for 10+ years. I will ❤️

[sharp robin]

🫂

[wicked frigate]

it seems to have read your refiner as the main and your base as the secondary, i see base model there as a custom param

[soft zealot]

cool will check out (no pun inteneded) shortly , meanwhile its been sat like this for nearly 10 minutes now, that aint right (and why no mesages in Console? Is that Normal?)

[ember cargo]

I do base + loras, then the last few steps I run the refiner, the more refiner steps, the less it'll look like the lora

it depends on the lora too tho

[soft zealot]

eh!>?

[boreal bough]

if you change the input text, you can turn it into a sharpener - which then works for most situations. Only when faces or face accessories get involved, do you have to 100% give up on refiner

[wicked frigate]

i still know a few people who use IRC to bridge into discord channels

[soft zealot]

ah It finaly died

[ember cargo]

yeah that's right, although maybe 5 or less refiner steps is good too

[wicked frigate]

it goes in the order the nodes were added. So to force it to recognize it, uh, delete the refiner node and replace it and it should be in order again

[hard fractal]

oh look.

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

[wicked frigate]

woo

[hard fractal]

oh look look:

https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/tree/main

[uncut steeple]

New Vae?

[hard fractal]

um... nope?

not new.

😅

[wicked frigate]

check comfy console (or if using self start, launch with --loglevel debug)

[boreal bough]

thanks for the update ❤️ will solve a lot of newbie questions

[wicked frigate]

oh wtf

Why are the worst issues always the ones that happen in somebody else's code instead of mine

i can't fix when dotnet doesn't know how to install itself or when imagesharp breaks its own pointeras asjfgasjags

[soft zealot]

Im off to eat my kebab, feel free to chiuck my workflow in your die and see if its just me lol

[sly jay]

What VAE for SDXL?

[boreal bough]

because the worst allergies, are the ones caused by Cross-Reactivity. They only show up when two factors are present at the same time XD making them super hard to diagnose

[wicked frigate]

yeah bout that

[sharp robin]

that's how i have it, but refiner just eats away the lora

damn

[boreal bough]

0.9 - a new update was just posted with it pre embedded

[sly jay]

Okay!

[soft zealot]

see the Credits & Notes box top LH side................

[wicked frigate]

oh - probably swap Image Save to the official Save Image

the Save Image node is the one comfy node that is really strongly expected

[soft zealot]

theres a few custom nodes in it lol

[ember cargo]

then do very few refiner steps, or make the refiner steps start at the total of steps (then you'll do no refiner)

[sturdy tulip]

New models have been uploaded to HuggingFace the comment "add weights with 0.9 vae" anyone know what those models are about? I don't see any announcements.