✨｜sdxl

[craggy agate]

is sd next faster than a1111?

[digital shoal]

interested to try your upscale workflow when it's out sytan :P

[tribal moat]

unless you have long process of generating images like first text-to-image then img2img then upscale, you gotta do that all manually in Auto1111 but in ComfyUI, you connect all these nodes together, they would all be done.

[high escarp]

nope just use hi res fix

[cursive warren]

Looks like SDXL is familiar with this title:

[sharp robin]

Too much connection to the dots.

[earnest anvil]

what do you mean by solid?

[earnest anvil]

ahh makes sense now thank you

[cobalt kernel]

Feels more responsive in general ..
I like the built in " queue " and got latent previews on mine ..

[high escarp]

he means solidly confusing

[cobalt kernel]

Eg. if you cancel a job - its cancelled !

[placid coral]

Hey, anybody knows what might be causing Sdxl to take extremely long time to generate a single 1024x1024 image?

[cobalt kernel]

I like that part too , it makes me learn a lot more ..
About how the networks are set up ..

[sharp robin]

Gpu?

[placid coral]

I've got a 3060ti

[shy kelp]

anyone got an optimized workflow.json for SDXL 1.0?

[reef sky]

https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[heady vale]

open task manager and check the dedicated memory graph for your GPU. probably low mem

[high escarp]

ppl keep saying that what you learn

[tribal moat]

I got experience due to me using alot of blueprints in my game development process 😉 . Blueprints is Visual Node Programming, Looks similar to ComfyUI.

[shy kelp]

I dunno, I couldn't get C3Po to weild a lightsaber in sytanSD's workflow lol

[placid coral]

Oh yeah it says Utilization keeps bouncing anywhere between 4% and 98%, its strange

[sharp robin]

Feels like i need to learn programming for this.

[reef sky]

Sounds more like a prompting problem than a workflow problem

[cobalt kernel]

Good place to start with " comfy ui " ..
— Is to try out all the example setups here :
[ https://github.com/comfyanonymous/ComfyUI_examples ]

[placid coral]

Im using automatic1111 though

[reef sky]

Unless you tryin to inpaint the lightsaber or something

[placid coral]

Do extensions have an impact on how long it takes to generate 1 image?

Cause its taking like 10 minutes for me to generate 1 image

like a basic one

[heady vale]

change one of the graphs to CUDA and you will see it, like this

[shy kelp]

Possibly, though C3po weilding a lightsaber doesn't really leave much room for error in a prompt. I was getting robots and lightsabers being weild, but never a solid c3po

[boreal bough]

you ran out of vram, and are now using standard ram - which is slow

[tribal moat]

seems like you got alot of knowledge about comfyUI, do you know how to integrate deepbooru?

like interrogate deepbooru feature in auto1111 ui

[cobalt kernel]

I totally missed the " comfy examples " when I started ..
Since its on a completely different repo ..

[placid coral]

Im dumb, I dont see a CUDA option anywhere

[placid coral]

You think so? Should I run one of those lowvram commands then?

[heady vale]

the little karat symbol next to the graph name is a dropdown menu

[sharp robin]

Yeah. Same w blip/booru. Usually booru is for anime tho right?

[tribal moat]

yeah right. but It was included in readme section of main repo.

[tribal moat]

yes!!

[shy kelp]

How odd it's doing it today lol

[lean granite]

anyone know how to switch between workflows?

for comfy

[shy kelp]

theres a load button, it uses workflow.json

[tribal moat]

but wait, how to use BLIP in comfy? If I can know that, I can do same for booru.

[placid coral]

I dont see one 😦
Unless Im stupid blind, which could be

[lean granite]

o cool!

[sharp robin]

Right bottom click load open past workfllws. Make sure to save current one first

[shy kelp]

https://github.com/SytanSD/Sytan-SDXL-ComfyUI/blob/main/Sytan SDXL Workflow v0.5.json

[cobalt kernel]

Better still you can just drag your output images ..
Straight onto the ui and it'll load up the workflow used to generate it ..

— That includes all the prompts ..

[placid coral]

damn its taking like 5 minutes to even generate a 512x512 image now

something's definitely wrong with my zetup

[tribal moat]

no, use ComfyUI

1024res image took me about 8 minutes in auto1111

but just 30 seconds in comfyui

[placid coral]

why such difference?

[tribal moat]

I got same gpu, 3060ti

[tribal moat]

I don't know man....Magic just works.

[placid coral]

Hmm, I've got like a million loras and models and extensions installed on A1111, aw man

[placid coral]

Well, thanks for letting me know

[sharp robin]

Umm. Not quite sure. I’ve only used comfy for sdxl. And it doesn’t have clip interrogator for that yet and havnt see. A node yet.

[heady vale]

right click and change to multiple, then you can change

[vocal rapids]

dreamshaper XL +1.0 refiner without upscaler

[agile quarry]

what upscalers are compatible with comfy? all of these? just clone this repo into my upscale_models folder? https://huggingface.co/uwg/upscaler/tree/main

[shy kelp]

I seem to maybe have a comfyui bug after clearing and reloading my workspace too many times lol

[dusk mica]

i think a1111 is simply not optimized enough

[placid coral]

mine's already set to multiple engines....

[heady vale]

if you click the name "3D" do you get a dropdown?

[placid coral]

is this true? Should I not even bother with A1111 sxl

[dusk mica]

sdxl took 29 seconds for me on comfy, on a1111 is goes to 1min30 or more

[queen relic]

Loving Dreambooth LoRA + SDXL 1.0

Really solid results and much more stable likeness, even over SDXL 0.9. 🙌

[placid coral]

ah, yess, I see it thanks

[tribal moat]

You gotta wait. Auto1111 does things slowly nowadays.

[vocal rapids]

Prompt executed in 33.80 seconds in comfyui using colab ,lol

[high escarp]

I'm generating images at 30 seconds in automatic 1111

[inner ruin]

what parameters did you use?

[heady vale]

so the CUda graph will be the main manager for SD. if its maxed out 100% then its just waiting for vram

[placid coral]

my regular model image generations take like 5-15 seconds, I cant wait 10 minutes for 1 pic

[heady vale]

welcome to XL where technology is cutting edge

[placid coral]

yeah, cuda is being maxed almost instantly

[shy kelp]

SDXL takes about 5 seconds for me

[placid coral]

So I should probably run it medvram command you saying?

[tribal moat]

then switch to ComfyUI for SDXL....It's just 1.5gb download, then extract, copy sdxl model and you can quickly get started. Default workflow already loads for you, just set whatever prompt you like and get started

[dusk mica]

what default workflow

[placid coral]

Fair enough, I'll give it a try

thanks

[high escarp]

automatic 1111

Photo manipulation, a sailboat, sailing over a dense sea of morning fog, aged design with sunrise-reflecting gold sails, ethereal style popular on ArtStation, an 8K, highly detailed masterpiece in the style of Galen Rowell.
Negative prompt: Anime, cartoon, graphic, text, painting, crayon, graphite, abstract glitch, Noisy, blurry, soft, mutated, ugly, disfigured, hands
Steps: 50, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 2889460742, Size: 1024x1024, Model hash: 7440042bbd, Model: sd_xl_refiner_1.0, Denoising strength: 0.5, Hires upscale: 1.05, Hires upscaler: 4x_NMKD-Siax_200k, Version: 1.5.0

[soft zealot]

so ................

On the left is one originally generated in SDXL0.9 with a prompt specifically written there.

On the Right is from SDXL1.0 (usingthe SDXL0.9 VAE although the baked in V1 VAE is 99.9% identical)

SDXL1 just feels washed out in comparison

[queen relic]

150 training images, 500 reg images, about 20k steps, Learning Rate: 0.0004

[livid cradle]

anyone having issues with the masking function of SDXL?

[spice kettle]

Is a NVIDIA GeForce RTX 3070 Laptop (8GB VRam) enough to run the model?

[wet merlin]

Can anyone tell me how the training data for sdxl was captioned? Captioning isn't mentioned in the sdxl paper on arxiv.

[livid cradle]

init image is this.

[inner ruin]

seems excessive.

[shy kelp]

8gb is enough to store the models but it would depend on how fast your PC can move stuff from RAM into VRAM

[livid cradle]

the result always modify the text on my product...

anyone know how to fix this?

my mask is good.

so in theory it should not modify anything behind the black pixels.

[high escarp]

are you using inpain global-hormones

[heady vale]

ensure you have "only masked" option selected

[livid cradle]

im using REST API image-to-image/masking

im not using any public UI, i build the code for a product.

[high escarp]

Whole picture or only Mask

[heady vale]

probably an option in code to change that option

[dusk mica]

which workflow

[high escarp]

automatic 1111

Digital art, a sailboat, cresting a tidal wave of thick storm clouds, antique design with deep indigo sails, a mystical style trending on ArtStation, an 8K, highly detailed masterpiece reminiscent of the works of Albert Bierstadt.
Negative prompt: Anime, cartoon, graphic, text, painting, crayon, graphite, abstract glitch, Noisy, blurry, soft, mutated, ugly, disfigured, hands
Steps: 50, Sampler: DPM++ 2S a Karras, CFG scale: 7, Seed: 2889460742, Size: 1024x1024, Model hash: 7440042bbd, Model: sd_xl_refiner_1.0, Denoising strength: 0.5, Hires upscale: 1.05, Hires upscaler: 4x_NMKD-Siax_200k, Version: 1.5.0

[dusk mica]

thanks

[high escarp]

Slightly different prompt for each

[royal fern]

I noticed that the suggested workflows from comfy don't make use of the specialized SDXL nodes
https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

[high escarp]

comfy is worthless

[half ivy]

yup

[royal fern]

wow...

[fleet harness]

https://tenor.com/view/baikinman-anpanman-takashi-yanase-soreike-anpanman-anime-gif-5519678

[fallow kestrel]

srsly? I contemplated switching to comfy

[dusk mica]

for me comfy is much faster than a1111

[wispy grotto]

whats the best ammount of steps for sdxl and is it reccomanded to use refiner to get best result or i can just use the base version?

[royal fern]

@visual glade is the person

[fallow kestrel]

^ that's what i heard too

[high escarp]

not for all

[earnest anvil]

30+

[craggy agate]

did anyone compare comfy and v;ad?

[fallow kestrel]

sdxl supposedly took like 8 minutes in 1111 and 30s in comfy

[high escarp]

same on my 3090

[craggy agate]

vlad*

[half ivy]

comfy is life

[wispy grotto]

thanks

[fleet harness]

comfy is like 10x faster for me

[earnest anvil]

refiner is hit or miss. I find it works best on images that are slightly blurry

[high escarp]

not true

[craggy agate]

not true also for me i have a 3090

[wispy grotto]

good to know that, have to test

[craggy agate]

comfy is slower?

not sure why

[vocal rapids]

comfy is lifeboat for me

[fallow kestrel]

i said supposedly - do you have any ressources on how both perform?

[fleet harness]

https://tenor.com/view/metal-gear-rising-metal-gear-rising-revengeance-raiden-senator-armstrong-source-gif-25038929

[high escarp]

where is yours

[fallow kestrel]

bruh who hurt you, I'm the dude asking you 😂

[twilit zinc]

how to install on 1111 I am getting error i put base model in model folder but its get error

[turbid garnet]

Is SDXL 1.0 allowed to be used commercially?

[craggy agate]

its fully open source

[queen relic]

It's easy to get likeness trained in 3000-6000 steps with 10-20 training images. But i'm yet to see a model trained which is as versatile as my setup.

It's a very accurate likeness, very versatile in composition and can be mutated into any style really easily.

[craggy agate]

which means yes

[inner ruin]

can you generate non-portraits of yourself?

like yourself doing random stuff

[placid coral]

Okay wow comfy ui dies run sdxl like 8 times faster than automatic1111

[turbid garnet]

@craggy agate I know, it's just that it wasn't very clear from the license on GitHub. Plus checkout ClipDrop AI's page -
Image

[heady vale]

models go into \stable-diffusion-webui\models\Stable-diffusion\

[high escarp]

This is the only proof I have

[craggy agate]

its been generally released yesterday

so as of now it is fully open source

whic will remain as such as far as im aware

[turbid garnet]

Others can you all please weigh in

This is important

[high escarp]

yep

[craggy agate]

just fyi stable diffusion is a big part of my actual proffesional job if it wasnt allowed for commercial use i would have known

[heady vale]

you can read the license on the HF page https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0

[queen relic]

Not tried specifically "doing things"... but heres some more examples...

(also these were 0.9)

[turbid garnet]

Yeah I read the one on GitHub, wasn't very clear

since clipdrop ai said otherwise

[twilit zinc]

hi guys getting this error while input sdxl base model

size mismatch for model.diffusion_model.output_blocks.8.0.skip_connection.bias: copying a param with shape torch.Size([320]) from checkpoint, the shape in current model is torch.Size([640]).

[heady vale]

open the license and read Section III

[shy kelp]

Where are these upscalers meant to go in SwarmUI

the comfy workflow isn't decteting them and I can't get the widget to locate them

[twilit zinc]

ya i put it in the same but geeting folowing error : size mismatch for model.diffusion_model.output_blocks.8.0.skip_connection.bias: copying a param with shape torch.Size([320]) from checkpoint, the shape in current model is torch.Size([640]).

is ther anythin i have to do?

[dense chasm]

does anyone test the celebrity generation with sdxl1.0 like Jackie Chan?

[heady vale]

Im not too sure what that error is. Best to ask in #🤝｜tech-support

[twilit zinc]

thnaks

[median mulch]

it says size mismatch, you prob generating images less than 1024 x 1024 maybe

[ornate elk]

how do I avoid doubling? by making square images?

[heady vale]

you can gen XL images smaller that 1024. they just wont look good

[small tree]

comfyui uses like 3gb less vram than automatic1111 for some reason

[median mulch]

i think a1111 has partial support for now, it's not optimized

[ornate elk]

man it just seems like there's no reason to use a1111 anymore

[shy kelp]

whys it complaining

[small tree]

nah a1111 is great

[shy kelp]

https://tenor.com/view/jackie-chan-meme-what-confused-huh-gif-5480485

[small tree]

most people just wanna put in a prompt and get an image

not learn the node science

[quaint mantle]

guys, is sdxl slower than normal sd?

[ornate elk]

ive been using it for a year but the whole sdxl support issues are getting tiresome. also why is there still no way to erase parts of a mask lol

[heady vale]

InvokeAI was updated recently also with more features and XL support in beta

[small tree]

yeah inpainting very buggy

[shy kelp]

I would love too but its not exactly easy to find your way around this community, there's like a billion different links to a million other different links

[median mulch]

people expect everything to work the first day... smh, give it some time

[shy kelp]

When a place suffers from TOO much documentation, you can't find what your looking for

[small tree]

yeah i can relate

brain exploding tryint to comprehend comfyui rn

[heady vale]

discord search helps a lot

[ornate elk]

bruh sdxl 0.9 has been out for how long?

[clever verge]

That's good new nodes. Funny it's quite similar to what I've implemented in custom nodes and also workflow.

[nova aspen]

They need to go into the models/upscale... Folder

[sullen adder]

maybe someone can create a real ui for comfy ui

[dusk mica]

comfybox?

[shy kelp]

okay so the 1.0 version was finally released. What's the general consensus? is it good?

[median mulch]

i mean my point still stands, give it some time until we figure out a good workflow, settings, etc

[normal geyser]

Is the SDXL style preset implemented through the following methods? Or text inversion?

Style: Photographic
Positive: cinematic photo {prompt} . 35mm photograph, film, bokeh, professional, 4k, highly detailed
Negative: drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly

[fleet harness]

I am liking it so far.

[vale eagle]

just prompt

[small tree]

its way better than 1.5 thats for sure

[fleet harness]

It's really good

[shy kelp]

So it can make legible text or no?

[small tree]

does anyone have a good comfyui workflow for sdxl

im using the example one but im not sure its any good

[shy kelp]

It's only been out a day. There should be everything you need for it soon though.

[fleet harness]

just take my photo I just posted and drop it into the comfy UI

[small tree]

alr thanks

[fleet harness]

I recently found this one

its my fav atm

has notes and everything

[small tree]

it has some nodes i dont have

[fleet harness]

you gotta snag those. there is a note in to workflow with the link

[ornate elk]

it's been over a month

[shy kelp]

I'm sticking with a1111. Is comfy that much better?

[delicate grotto]

Vae fix availqble anywhere yet?

[small tree]

alr

[ornate elk]

do you know how to use refiner properly?

[median mulch]

yes and people are still looking for a good workflow and settings

[shy kelp]

Haven't installed the model yet. Is the refiner really necessary?

[ornate elk]

yes but there is NO workflow. a month is long enough

[delicate grotto]

6gb vram is enough to render at 1440p in sdxl plus refiner
In au1111 i can barely do 768 in the models, not even sure if sdxl would run

[fleet harness]

there are many workflows

[spice sonnet]

If you can't get used to Comfy you can try InvokeAI, it supports SDXL.

[ornate elk]

0.9 came out a month ago, it works the same way

[fleet harness]

I collect workflows like people do prompts

https://tenor.com/view/yy-smoking-counting-money-anime-gif-17852058

[small tree]

thanks ill look into it

[shy kelp]

@delicate grotto Wait, are you telling me using comfy saves on vram???

[small tree]

how are you doing 1440p with 6gn

[delicate grotto]

It uses A LOT LESS

[small tree]

i have a 3080 10gb and it uses almost 8gb of vram

[fleet harness]

Comfy saves A TON of vram

[small tree]

doing 512

[fleet harness]

I sped up by like 10x

[delicate grotto]

Tiled decoder

[shy kelp]

@delicate grotto WEll damn.......I guess I should check out comfy then

[small tree]

oh

is it like controlnet tile?

[delicate grotto]

No idea

[amber fulcrum]

how do you find SDXL 0.9 vs SDXL 1.0? Are there any breaking changes? I have seen so far one post with comparison and 0.9 won in that case in my opinion

[celest jungle]

Does invoke use the same as Auto1111?

[shy kelp]

specs are a 5800x3d, 64GB ddr4, 3090......and sometimes with a1111 it still feels like it's dragging with image generation

[celest jungle]

of vram

[ornate elk]

1.0 is designed to be more tweakable. 0.9 will look better out of the box

[shy kelp]

@amber fulcrum I didn't even bother with 0.9. It was just for testing anyways. I ignored it.

I got seages workflow working but its not upscaling

some how

[median mulch]

the only thing good about 0.9 is the VAE as far as i understood so far

[shy kelp]

Is there a book I can read that explains how all the nodes work and why each thing does what it does so I can build my own workflows

[celest jungle]

Has anyone tried SDXL on invoke?

[shy kelp]

Been sick with the flu delirious. But gonna get back in teh swing of things soon. I was expecting a few more days until 1.0 release. But bam it hits now.

[delicate grotto]

Look at other people workflows basically
Takes a day of playing around

[shy kelp]

@celest jungle no not yet. not me anyways. What I want to know is working with nodes in comfy any better?

its like a noodle soup lol

[median mulch]

i actually never tried Invoke before, il try tomorrow :3

[shy kelp]

it looks like noodles ya

[spice kettle]

Any ideas as to what I'm doing wrong?

[shy kelp]

lol damn look at that old school flowchart 😄

[median mulch]

people generating waifus already

[shy kelp]

Can someone confirm that Comfy uses less vram?

[clever verge]

Yeah, I think that's what the bots use and that's what I use too. SAI shared them, was in a sticky.

[half ivy]

ofc

[spice kettle]

I was completely unable to use A1111, but Comfy generates just fine

[shy kelp]

@median mulch People don't waste any time! 😄

[nocturne dove]

You need to be at 1024x1024 for starters.

[celest jungle]

I just don't know how to use comfyui and heard that Auto1111 isn't optimized for SDXL yet and thought maybe invoke might be better, I don't use invoke either but it's better than learning comfyui for me

[spice kettle]

Though my quality has been really bad

[spice kettle]

Didn't know that, thanks!

[shy kelp]

@spice kettle under what circumstances were you not able to use a1111? how big an image did you try with a1111 that you couldn't do and what are your specs?

[median mulch]

just give it like couple weeks idk, then everyone will be back to a1111

[shy kelp]

If you don't mind me asking

[spice kettle]

512, 10 steps

[spice kettle]

Laptop Nvidia RTX 3070Ti 8Gb Ram

[shy kelp]

@spice kettle I had no problem generating images especially at 512 with a desktop 1050 Ti 4GB vram. Not sure about cut down laptop stuff.

[nocturne dove]

You should probably scroll back through this channel and grab a couple of images, If they came from comfyui you can drag and drop them into your comfyui window and it will recreate the workflow. Save your current workflow before trying this though...

[livid cradle]

where can we download this workflow?

[shy kelp]

Plus I have 64GB ddr4 that may be a factor

then last week in preparation for 1.0 I upgraded to a used 3090

[spice kettle]

When I loaded SDXL on A1111 it instantly took up all of my laptopts VRAM only with the model

Well 7.8Gb/8Gb

Then when trying to generate it gave me a not enough memory error

[shy kelp]

wait I'm sorry yes we're talking about 1.0. I was referring to sd1.5

[spice kettle]

I have a suspision it wasn't even using my laptops ram

[vale eagle]

to who want a decent workflow. https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[median mulch]

i wonder if most people use the Sytan one

[shy kelp]

@spice kettle Have you tried comfy with sdxl 1.0?

with the laptop

[spice kettle]

Yes, that works fine

[clever verge]

The nodes he's using haven't been released yet I think. It's a tease to what's probably coming.

[spice kettle]

Takes a long ass time to load, but it renders

Though it's not rendering quite as well as I was hoping

[shy kelp]

Then I got to investigate comfy then.

[median mulch]

it's time to upgrade that potato PC for some people

[spice kettle]

My PC is a year old

[shy kelp]

It's early days. on civitai there's going to be a ton of custom models soon enough

[median mulch]

Dreamshaper XL baby

[small tree]

all rtx 30 series users are crying rn

[median mulch]

but yea gonna be awesome to see all the models and features coming :3

[shy kelp]

yeah, it took a week's pay...but was able to upgrade to a 3090....used....in preparation for 1.0. Works great. founders edition kinda loud though

[celest jungle]

wait why is that

[shy kelp]

It's the driver

[small tree]

only 3090 has more than 12gb vram

[shy kelp]

Nvidia is aware of the problem and is working on a stable diffusion specific driver

[spice kettle]

I bought a laptop with 8GB Vram and 16Gb ram and figured it'd be great for gaming for the next 10 years considering I play almost entirely indie. Then AI models started coming out and I cried.

[small tree]

this is why you dont immediatly upgrade nvidia drivers

[shy kelp]

@spice kettle yeah with that either you need a super laptop or a reasonably priced, but specced out desktop

[normal geyser]

Does this mean that the official is making a style preset Textual Inversion?

[celest jungle]

I mean you can always just downgrade right? is it the latest driver?

[shy kelp]

3090 with 24GB of vram can't be beat right now still. That 4090 is too stupidly priced

[small tree]

yeah

only used 3090s make sense to buy rn

[shy kelp]

do not immediately upgrade drivers. If anything, use an older driver

[small tree]

if you wanna dabble with ai

[livid cradle]

is there any comfy ui workflow that uses image to image masking function?

[spice kettle]

1.6k for a graphics card is still insane.

[small tree]

i got a used 3080 10gb for 400 bucks thanks to mining crash lol

[spice kettle]

nice

[shy kelp]

@spice kettle and with that moronically sized cooler because they thought the card was going to be running with 600w.

The entire 4000 series nvidia sucks

[fickle jacinth]

1.0 is powering the bots on this server right?

[small tree]

nah

4070 is decent

everything else completely sucks

[shy kelp]

@small tree yeah not bad....for the money.

I overpaid for mine cause i wanted to win the bid....$740

[celest jungle]

hopefully the 50 series is better...

[small tree]

it better be

its 1.5 years away

[shy kelp]

Hopefully. But Jensen has no motivation not to be greedy and stupid

[small tree]

amd just needs to be more competitive

[shy kelp]

yeah not till 2025. The beotch has the gaul to actually skip 2024

[boreal bough]

@wicked frigate cannot get any custom comfyui workflows working :c

[shy kelp]

oh lord. when i see pics like that i know it's going to be a pain to learn comfy

[median mulch]

i think that is error from swarm

[cunning geyser]

Anyone got a config for training sdxl loras? x-x

[shy kelp]

and i just got a good handle at using a1111

[boreal bough]

yeah, that's swarm intergration into comfy

comfy is innocent

[median mulch]

[shy kelp]

I wanted to train stuff but I'll leave it to you guys and look forward to your work on civitai 😄

[fleet harness]

idk why everyone has issues, it booted right up for me

never looked back

.>

[median mulch]

the big bois will train all the nice models just give it couple days

[boreal bough]

it works just fine - its only when you use the comfy workflow editor tab that things crash

[short marsh]

is it possible to connect to ComfyUI remotely? from another pc or mobile device on the same network?

[shy kelp]

it's 3:30am here for me. I just wanted to see what the word was on 1.0 before i go to bed. I'll check in again tomorrows. peace

[boreal bough]

yes

[celest jungle]

I'm more worried about it not working with Auto1111

[fleet harness]

SDXL 1.0 in comfy

[median mulch]

a1111 support is being worked on, it's still partial

[dim basalt]

a1111 sdxl is still partial?

[celest jungle]

I just hope that my 12gb will be enough... cause I have no idea about anything to do with comfyui

[craggy agate]

vlad has full support

[wicked frigate]

oh that one is a really straightforward oopsie i think - git pull, restart, and try now

[craggy agate]

just need to add sd_refiner on the quick options

[shy kelp]

Where's Olivio Sarikas' video on 1.0 is what I want to know. okay gotta force myself to sleep.

[fleet harness]

I run 8GB and it runs wayyyy better than A1111

[tropic turret]

How could I run ComfyUI with command arguments? My generations are pretty slow on my RX 6800 XT. :c

[shy kelp]

So A1111 doesn't fully support 1.0 yet?

[celest jungle]

I mean if support for auto1111 gets better does that mean vram usage will drop drastically?

[fleet harness]

no, its a little wonky i hear.

[shy kelp]

So Comfy is that much better for me to ignore a1111 and relearn a new system?

[fleet harness]

like starting an old NES

[median mulch]

blowing that NES cartridge :3

[shy kelp]

Or should I just wait until a1111 figures stuff out?

[fleet harness]

meh, It took me like a day to learn the basics just from messing around and collecting peoples workflows

[sterile ermine]

I'm using Diffusers and I have the red dots as well 🙂

[fleet harness]

you learn a lot from observation

[sterile ermine]

idk if there is a way to turn these off

[median mulch]

comfyui is not NASA, i mean come on man, you can learn it

[ornate elk]

I haven't seen them since I switched to 1.0

[celest jungle]

You probably have experience in something similar to it though right?

[shy kelp]

I'm sure I can learn it just fine. But I don't wanna learn a new system! sigh

[boreal bough]

@wicked frigate error still there

[sharp patrol]

My prompts renders a slight rgb shift in the output and are not sharp. any leads or working UI which solves this? What am I missing?

[sterile ermine]

I pulled 1.0 from Hugging Face

[small tree]

you should wait

[lilac wren]

Didnt have a single good image since 1.0

[sterile ermine]

are you using Diffusers as well?

[small tree]

base sdxl isnt really worth over using over current 1.5 models anyway

[fleet harness]

I have used nodes in other apps but it's not hard. you can take my png and drop it into the UI and immediately get my entire workflow

[ornate elk]

no idea. im using a1111 and I let it do most things for me

[small tree]

a1111 will probably be all figured out by the time good sdxl models release

[shy kelp]

@small tree I'm wondering if the "wonkiness" is what was causing that guy with the laptop to have image generation issues.

[celest jungle]

I'm guessing I will have to look to youtube to get a tutorial?

[shy kelp]

@celest jungle Olivio Sarikas is who I go to on YouTube.....if you can overlook the quirky nerdiness. They all are really. lol

[wicked frigate]

... it's having a stroke about the file separator symbol :(

[boreal bough]

@wicked frigate also, I swear its not a rocket science workflow XD

[wicked frigate]

idk why but you're sending the model through in a weird way

and... when it does

... s- stuff happens??

these are identical?

[ornate elk]

the same prompt game me this lol

[shy kelp]

No not rocket science at all. sigh

[wicked frigate]

but these aren't??

[median mulch]

can you generate a chocolate bunny :3

[wicked frigate]

also dustin pointed out in vc, you should have Save Image not Preview Image

[ornate elk]

I am sure that'd be heaps easier than what I am trying to do

[median mulch]

haha

[sharp patrol]

can anyone help how do I fix this blurriness? Even after I apply upscalers, it's there

[sterile ermine]

Even Hugging Face's own example seems to have them

[boreal bough]

good find. didn't see that. but it doesn'

doesnt change the model checpoint not loading

[shy kelp]

Someone generate an everything bagel, toasted, with cream cheese and lox.......I'm hungry. lol

Seriously, it's a good test

[celest jungle]

It looks like it's trying to make a painting

[fleet harness]

it can be complicated if you want

[median mulch]

ordering food with AI 🙂

[sour obsidian]

dont do that to me, im hungry too

🥯

[sterile ermine]

no, it's there on all images

[tropic turret]

1.0 refiner kinda iffy 😦

[boreal bough]

I let your app download everything. no custom comfy nor custom folders of anything

[sterile ermine]

definitely feels like a watermark but I don't know why it's very easy to see by eye and distracting

[shy kelp]

yeah I'm going to wait and see if they fix a1111 and if it will generate images with less vram. If they can't, I guess I'm going to comfy.

[wicked frigate]

can you send me the workflow file so i can try to replicate it?

[median mulch]

just give it couple days

[celest jungle]

I'm guessing you tried putting "painting" in the negative prompt?

[shy kelp]

snap to it people! I want an everything bagel with cream cheese!

[livid cradle]

can anyone share with me an image to image with masking capabilities for comfy ui??

[celest jungle]

same I guess

[boreal bough]

it happens on all/any workflows

[wicked frigate]

it... it doesn't for me .-.

[livid cradle]

I wanna test out masking and getting the subject out in different background.

[celest jungle]

unless I can use invoke

[boreal bough]

this was just the simplest one I had - to test it

[shy kelp]

That's not bad. This is more like what I wanted the cybertruck to be. Elon isn't as innovative as he likes to think he is.

[sterile ermine]

as I said it has nothing to do with painting. Here is with the prompt "cat"

[tropic turret]

Theres a span of green pixels on most edges and straight lines in this image.

[boreal bough]

maybe cause I installed it in a different subdirectory, then moved it to root?

[sour obsidian]

"it works on my machine"

[sterile ermine]

not just this though, all images I generated so far has it. I'm using the diffusers version

[boreal bough]

the whole swarm folder

[wicked frigate]

wait

is it

[celest jungle]

Beats me then 🤷‍♂️

[wicked frigate]

a trailing slash problem again?

[tropic turret]

Should I use v0.9 refiner instead?

[wicked frigate]

there's an unexpected trailing slash and it doesn't compensate properly?

eg in your settings -> model dir?

[sharp patrol]

Help with this please. Its giving me major OCD

[shy kelp]

gn im out was fun 😄

[celest jungle]

gn

[ornate elk]

I tell it to make cybernetic bunny rabbits and it makes terminators with bunny heads

[clever verge]

I think it's your prompt if I check your image.

[void loom]

have some artists been scrubbed between 0.9 and 1.0 ?

[sharp patrol]

why is it so different?

[clever verge]

I used only this: beautiful korean woman, kpop idol, white hair, ponytail, olive skin, tight all white sleeveless body suit, red lips, small breasts plus a style prompt.

[wicked frigate]

i'm so confused

where is the secondary slash coming from?!

[sharp patrol]

it looks nice.. this is SDXL? and do you also get the red green blur?

[boreal bough]

and why does it work when I use your built in workflow XD

[wicked frigate]

... so it does work on some workflows??

[deft quiver]

Any comment on the watermarked VAE?

[clever verge]

Yes, this is SDXL. Not sure about the blur.

[ornate elk]

now make it a bunny rabbit

[median mulch]

armored core leaked pics 😮

[ornate elk]

finally giving me an actual cuyborg undead bunny

[clever verge]

If you mean the VAE-lines I see in your image then yes, but that's fixable with the old SDXL 0.9 VAE.

[ornate elk]

plus floating ears because I am not square

[boreal bough]

nop. only when I disable workflow editor

[indigo vine]

using 0.9 vae, green artifacts disappear

[sharp patrol]

yeah those. do I use 0.9 VAE?

[cunning geyser]

Anyone got a lora training config for what they use

[boreal bough]

also fun fact. everything does work when I generate inside the workflow editor XD

[clever verge]

Not if you don't load a separate VAE. Then it uses the 1.0 VAE from the SDXL-model.

[boreal bough]

@wicked frigate but try it yourself XD

[sharp patrol]

can you share your workflow?

[short marsh]

Do i need to open the port in firewall? im having trouble currently

[tropic turret]

Is this so?

[wicked frigate]

heh, yeah, okay, that one errors for me too lol

[sharp patrol]

I have just moved to comfy from A1111. working with it feels like if i switch a wrong node, it might trigger a nuke somewhere

[clever verge]

I got custom nodes in it so probably not fully usable but you might get an idea.

[boreal bough]

...
lemme just remake it. maybe its using an outdated node

[clever verge]

Keep it simple with comfy, there's no need to go overboard with it.

[median mulch]

keep it comfy

[humble wigeon]

it doesn't start in automatic1111.(wrote a parameter --no-half-vae ) the model is not loading.
can someone give a configuration file for sdxl 1 ComfyUI ?

[sharp patrol]

nothing in there

[frosty pumice]

anybody can help me with searching room where I could generate something, I can't find this

[fleet harness]

oh, didn't you know? They are piloted by bunnies.

[molten crypt]

Just how do we limit the GPU usage of ComfyUI, in the same style as "--medvram"

[ornate elk]

that is a very large rabbit

[sharp patrol]

https://huggingface.co/stabilityai/sdxl-vae/tree/main
Do these help in any vae(way)?

[fleet harness]

or the least imposing mech ever

[median mulch]

FromSoft dev leaking pics again

[vocal rapids]

tried to upscale to 1536*1536 in comfyui,

[ornate elk]

with an action figure by his foot?

[boreal bough]

@wicked frigate remade it. nop. even tried different sampler. no idea why every comfysetup crashes for me :c

[fleet harness]

yes lmao

🐇

[median mulch]

🐰

[sour obsidian]

this is more unsettling than I anticipated 🐰

[sterile ermine]

@ornate elk did you figure out the red dot problem?

[fleet harness]

they should battle

[sterile ermine]

opened this: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/discussions/15

[sour obsidian]

https://tenor.com/view/bunny-heir-fight-rabbit-gif-5950408

[median mulch]

Finish Him!

[sterile ermine]

All images on Hugging Face's docs also have it

[fleet harness]

https://tenor.com/view/machi-bunny-punch-face-fight-gif-10128531

[ornate elk]

yeah it went away in 1.0, or maybe it was my prompts, I am doing different stuff today

[high skiff]

red dot was a way of watermarking the images

1.0 now has a new way that is less intrusive, but still very annoying

[median mulch]

watermarking

[sterile ermine]

🤔

[ornate elk]

same for me really...struggle to get it to make the character bunny shaped

[sterile ermine]

but it uses invisible watermark

those don't seem invisible at all to me 😄

[high skiff]

they are not even remotely invisible lol

[deft quiver]

not a very good name

[sterile ermine]

I agree very much 😄

[high skiff]

its in the 1.0 VAE

you can see it here

[ornate elk]

I told it no more humans or humanoids

[high skiff]

1.0 VAE with watermark, vs 0.9 VAE without watermark

its a sort of stair stepping compression and chromatic aberration

[dusk mica]

i think the baked in VAE has an issue

[high skiff]

its a lot more visible on some images

[sterile ermine]

any way to get around it with diffusers?

[high skiff]

from what has been suggested to me, it might be a watermark on the images

[deft quiver]

switch to the 0.9 VAE

[high skiff]

I just use the 0.9 VAE lol

[sterile ermine]

this seems like a very bad way of "invisibly" watermarking images

by destroying them

[boreal bough]

it really likes to affect areas with 'extra attention' to detail, like eyes & lips

[median mulch]

0.9 vae baby

[sullen glade]

Try "a mechanical bunny"

[sterile ermine]

it has research only license

[ornate elk]

bunny ears are the new fingers

[high skiff]

yup, and also high contrast areas

like white against black

also, @boreal bough

[median mulch]

i generate my images for research purposes of course

[ornate elk]

it's actually more complex than that, I am trying to get one that's got flesh over cyborg, terminator style

[high skiff]

did you see my new upscale workflow?

[dusk mica]

12k, hdr, ultra realistic

/s

[ornate elk]

pretty sure that'd give me something that looks like a wind up toy

[boreal bough]

oh cool its online? 😄
I've been fighting with swarmui to load my comfy setups XD

[high skiff]

not yet, still testing

but I have been flexing it here :p

[dusk mica]

gonna try invoke next

[high skiff]

and it works damn good

[boreal bough]

yeah. loving the high detail raccoons XD

[clever verge]

Have you shared the workflow?

[high skiff]

not yet, still working on it

expected release for SDXL mixed diffusion, fractional offset, and upscale workflow with documentation is hopefully within 3-4 days

[celest jungle]

Let me know how the vram usage is

[void loom]

Just trying SDXL 1.0 a bit now with an old prompt from 0.9 using the DDPM sampler, this sampler was always 'different' some strange contrast with it etc, has this been fixed now as loading this prompt in has considerably changed the picture (as expected) but it doesn't have that unique flavour that the ddpm sampler had, does anyone know what im talking about? was the sampler 'fixed' ?

[high skiff]

It does even better on some prompts

[boreal bough]

remember to make a patreon by then!
or at least a ko-fi if patreon is too much for you

[clever verge]

Care to share a tease in PM? 🙂

[high skiff]

have a look at that

[celest jungle]

damnnn

[high skiff]

or this

[lilac wren]

This errors are driving me crazy, i have all fucking custom nodes required

[fleet harness]

lol

[high skiff]

4x the total pixel count upscales

[ornate elk]

has anyone figured out how to use refiner in a1111 with sdxl 1.0?

[boreal bough]

maybe they were renamed? but just reload them, and replace the old ones

[median mulch]

refiner is basically img2img

[vocal rapids]

imgtoimg and choose refiner model

[fleet harness]

get the comfy ui manager extension. itll auto download them for you

[high skiff]

thats the less ideal way to use it

[lilac wren]

Its not my workflow, i don't know what strange nodes were here

[median mulch]

i know, but i keep telling people to wait for updates

[woeful patio]

from what i heard, generate an image and use img2img on .2 strength with the refiner

[ornate elk]

oof, yeah I tried exactly that, didn't go well

[quartz sequoia]

i know topaz gigapixel when i see it sir

[woeful patio]

ah yeah I havent tried it yet

[high skiff]

its not gigapixel lol

[boreal bough]

you can open the json file with a text editor, for the input values
and if you have the custom nodes, you can just place them there yourself

[ornate elk]

i mean it wasn't terrible. just not that great. although I am doing weird stuff ig

[quartz sequoia]

has anyone attempted to train the refiner?

[median mulch]

doing weird stuff with AI?

[west breach]

yehaw!

[fleet harness]

weird stuff guys

[median mulch]

stop the madness

:3

[boreal bough]

#✨｜sdxl message

[fleet harness]

this is normal

[ornate elk]

moar pls

[median mulch]

that dino one is actually nice

[ornate elk]

I have to say....sdxl is fuckin fantastic

[fleet harness]

haha ty

yeh it is

[quartz sequoia]

yeah, but but for now... i would like to train it. to fix those weird faces from distant shots

[high skiff]

dayyummm

[boreal bough]

correct negatives solves that. I've seen multiple people having it solved already.

[ornate elk]

try it on one of mine

[woeful patio]

why do my sdxl come out looking like shit?

[boreal bough]

sampler + cfg + step choice

[quartz sequoia]

for trained face LoRA?

[high skiff]

it works on current generated images, not img2img style

At least, I have not messed with it

[ornate elk]

send metadata

[median mulch]

lol

[quartz sequoia]

pssst whats ur comfy workflow

[fleet harness]

this is the future

[quartz sequoia]

asking for a friend

[ornate elk]

oh shit do that but make the feet tank treads

[lilac wren]

@high skiff
Did you finish the workflow you wanted to share? I don't have good results with what I was using on 0.9

[boreal bough]

ah. yeah no - then you only have the option of either overfitting your face data - then use only base model
or use refiner only as a gloryfied sharpener, with super low denoise

[high skiff]

I am actively updating it

Some things changed from 0.9 to 1.0

Need to change some values, documentation, and continue to test my upscale workflow as well as my fractional offset

[quartz sequoia]

well thats why i want to train it.. if it knows the face it might do the fixing for me automatically, without the need of overfitting

[pastel spire]

i'm kinda bummed. I got it working, and its definately a better base model, but I was expecting a lot more. seems to have some pretty big issues. oh well. back to 1.5 until sdxl matures and has some solid models.

[boreal bough]

finetuning only then. so either a long time on 3090/4090
or renting an A100 or A100 stack on runpod

[quartz sequoia]

only for now...

[fleet harness]

https://tenor.com/view/pawn-stars-yes-yass-awesome-gif-21935666

[boreal bough]

?

[woeful patio]

nvm, its probably because I can't write prompts for shit lol

[fleet harness]

this one has no lizard but still good

[high skiff]

reasonable assumption haha

[median mulch]

prompt engineering school :3

[woeful patio]

the refiner goes in the models/stable diffusion folder too right?

[west breach]

0.9 vs 1.0

[lusty prairie]

prompt please!

[wicked frigate]

@boreal bough git pull and try now.
The key issue was your model wasn't, uh, the original model node, so it added as its own node, and then the input for model nodes got confused by windows filepaths having \.
I fixed it to not get confused about \ for now, for using the proper model node, uh - use 'load default' in the workflow view and then don't delete the checkpoint loader there. I need to make it actually process that when you only have one model, it's the model

[boreal bough]

rip<3 that was a lot of changed files XD already did it when I saw you typing. all works now ❤️

[fleet harness]

its in the meta ❤️

[lusty prairie]

is that for comfy?

[fleet harness]

[lusty prairie]

xD

thx

[median mulch]

[uncut steeple]

Me trying to figure out a new best workflow for me in comfy

[fleet harness]

[woeful patio]

I find comfy to be very confusing compared to a1111

[lusty prairie]

hmmm there are missing nodes @fleet harness

[fleet harness]

do you have comfy manger extension?

[uncut steeple]

Usually depends on if youre used to working with nodes

[lusty prairie]

nope

[fleet harness]

get it!

XD

[lusty prairie]

oki

[woeful patio]

I'm not so that might explain it

[tropic turret]

"A red cube and a blue sphere"

[median mulch]

nice

[fleet harness]

once you have that you can have it auto install missing nodes, sometimes you just have to google them if it doesnt catch them

[ornate elk]

bro I just asked for pics of pink bunnies and it gave me 50 images where everything but the bunnies were pink

[quartz sequoia]

i swear they said SDXL listens to you more

[median mulch]

(((pink bunnies)))

[tropic turret]

This is what I am sorta nitpicking about

[ornate elk]

it's got so many ))))

[quartz sequoia]

its just trolling you

[tropic turret]

(pink bunnies: 1.5) /s

[median mulch]

what if you type pink colored bunnies instead of pink bunnies

[ornate elk]

the problem is I've got it mashed in with a bunch of other adjectives

I can make it make a pink bunny, it's just kinda funny

[median mulch]

ye

[ornate elk]

they all had this colour scheme

[lean granite]

somebody know which folder to put the sdxl refiner in? is it just checkpoints?

[median mulch]

ye

[lean granite]

thx

[ornate elk]

models/stable diffusion

[lean granite]

using comfyui i dont think there is a sd folder

[fleet harness]

SDXL can make your dreams reality

[peak dove]

I have noticed that 1.0 often goes on a dissipated spree rather than follow the prompt 🥳

[ornate elk]

o sorry, meant for a1111

[lean granite]

all good!

[median mulch]

sdxl controlnet waiting room

[hollow dirge]

So when using sdxl 1.0 I do use it in the model slot? I mean when I was using Dream Shaper, I do this now instead?

dont understand it yet

[median mulch]

Dreamshaper XL tho

[west breach]

I would love if the history in ComfyUI was sorted descending. Since I usually want to redo the last gen and not the first one.

[woeful patio]

Does anyone know what sdxl has been trained on?

[lean granite]

not sure how we would know unless it was anounced somewhere

and if it was anounced somewhere i think the media would be all over it xD

[median mulch]

"on all sorts of things" :3

[west breach]

Probably one of laion's data sets

[median mulch]

i mean isnt it mentioned in the research paper?

[soft zealot]

did you restart Cofy after installingthe Custom Nodes?

[median mulch]

☕

Cofy

[desert fox]

I suddenly want to know what "XL" in SDXL means. Does anyone know?

[median mulch]

extra large

[shy kelp]

@high skiff are you still using your github workflow?

[vale eagle]

means the size

[clever verge]

It would be easy to save all prompt history as a json and order in whatever order you like. I should implement that in my styler.

[high skiff]

yup, current goal is to release a very rehauled version in the coming days

[woeful patio]

godzilla at the nuremberg trial,
reality can be whatever i want

[west breach]

not prompt history, when I'm tweaking with samplers and CFG and what not

[median mulch]

oh shit LOL

[west breach]

I already have a style node

[uncut steeple]

Thats where I am right now
We really need better color profile management for generated images

[desert fox]

thx.

[clever verge]

That should be possible too, to save samplers and cfg and such.

[lean granite]

there's been some people talking about invokeAI 3.0 as well, did anybody try that in comparison to a1111 and comfy?

[median mulch]

i think so far, comfy is leading the SDXL race

[clever verge]

I've done a custom primitive so it's faster to change settings. I guess that could log too.

[hollow dirge]

is sdxl worth it over other models from Civitai?

[lean granite]

up to debate I think. With current sdxl u dont have controlnet so it also depends on ur workflow

[median mulch]

the original sd 1.5 model was kinda boring as well... give it time until people fine tune

[woeful patio]

that one is awesome

[hollow dirge]

@lean granite @median mulch so you think they will make new models based on SDXL?

[median mulch]

of course

Dreamshaper XL already out

[lean granite]

^

its an alpha tho i think

[median mulch]

ye

[void loom]

how to get consistent realistic shots with sdxl? im throwing in some 1.5 prompts i had for realistic images and it's giving me animations/paintings/cartoons :\

[hollow dirge]

It's time to get comfy with comfyUI....

[peak dove]

Question: If I use embeddings based on 512x512, they will work inSDXL? Or must I remake them at 1024x1024?

[median mulch]

i think civitai has a xl tag, so maybe refresh that page and see what's new haha

[peak dove]

Thank you

[woeful patio]

was there a rule on what width and height the canvas has to be?

[clever verge]

Yeah, expected I'd say. SDXL requires more natural prompting.

[median mulch]

chatgpt can help with prompts i heard

[west breach]

try to keep it as close to 1024*1024, and go in multiples of 64

[proven moat]

Does comfyui stable work with sdxl?

[peak dove]

So I just load base and refiner into comfy, and hey presto?!

[uncut steeple]

Old embeddings wont work with SDXL afaik

[peak dove]

Thank you, good to know

[uncut steeple]

For a simple workflow yes, but there is more to it if you want to invest some time

[median mulch]

some very good workflows on discord, experiment with couple

[trim totem]

do you recommend comfyui or a1111?

[median mulch]

right now, comfyui is leading

[peak dove]

ComfyUI so easy to setup. A1111 has more features

[median mulch]

for sdxl

[trim totem]

ok thanks

[vocal stream]

a1111 still has stuff like deforum which I dont think comfy has so depends on what you do

[peak dove]

I cannot get refiner to work in A1111

[median mulch]

use it as img2img

with very low denoise str

[peak dove]

Txt2img with base, then img2img with refiner?

[median mulch]

yes

[peak dove]

OK

[median mulch]

that is what the refiner is anyway

[peak dove]

Will try, thank you

[vocal stream]

does the refiner not take the prompt at all? beyond what it sees in the base iamge

[median mulch]

i think you dont need the prompt, but it can help i guess if you keep it the same

but a1111 is not optimised yet for sdxl, wait couple days

[void loom]

i don't understand though, i've got the typical negative prompts telling it to not do anything cartoony or artificial, this sort of thing "(3d, render, cg, painting, drawing, cartoon, anime, comic:1.4)" and (detailed photo) in the main prompt, the thing is i've done a batch of images and some come out cartoony and some more realistic, some painted.. like it's really varying and lacks consistency :\ you know of any write ups on prompting for SDXL?

[clever verge]

Got rid of a decent amount of blur / bokeh with no prompting for horizon.

[median mulch]

time for Kill Bill vol 3 baby

[leaden fable]

Hello did they updated the 1.0 vae with artifacts?

[shy kelp]

can I use the same ComfyUI workflow of 0.9 on the new model 1.0 ?

[median mulch]

use 0.9 vae for now

they have to fix

[peak dove]

Imma winnah!!! 🤩

[leaden fable]

Someone on reddit told me they updated the 1.0 an hour ago

[median mulch]

hmm maybe... i didnt check haha

[formal jasper]

System requirements for sdxl 1.0 pls I have i7 12700 and rtx 3060ti will it work for me??

[clever verge]

try this template, describe what you want to see in the {prompt}. positive: cinematic photo {prompt} . 35mm photograph, film, bokeh, professional, 4k, highly detailed negative: drawing, painting, crayon, sketch, graphite, impressionist, noisy, blurry, soft, deformed, ugly

[boreal bough]

it will run

[median mulch]

if you use comfyui, you can do with like 8gb vram

not sure about 6gb vram

[formal jasper]

I have 8gb

[median mulch]

well you are set

[void loom]

will try thanks

[delicate grotto]

6gbvram is enough for sdxl 1440P

[clever verge]

Try to keep it simple and try to experiment with little or no negative. SDXL won't react the same way to prompting as 1.5 did. SDXL is more LLM-like.

[vital wolf]

does comfyui have clipskip option as well? @visual glade

[hollow dirge]

does anyone know a good comfy UI tutorial on YT?

[void loom]

seems like something is broken

[median mulch]

no dark theme

[soft zealot]

OG & Upscaled from your Prompt + my tweaks to it (workflow in Images NB this is setup for myuse so not necessarily easy to follow unless you start dragging it out, Image of workflow as I use it daily included)

Extra details not necessarily commented in the workflow

Credit & Notes:

modified using original from https://github.com/SytanSD/Sytan-SDXL-ComfyUI

HRF modified from https://civitai.com/models/107144/sdxl-09-semi-official-workflow

Seed with Text from https://gist.github.com/alkemann/7361b8eb966f29c8238fd323409efb68

Multiple Nodes from WAS Suite, Efficiency nodes & Deerfu ::
https://github.com/WASasquatch/was-node-suite-comfyui

https://github.com/LucianoCirino/efficiency-nodes-comfyui

https://github.com/Derfuu/Derfuu_ComfyUI_ModdedNodes

styles sheet with SDXL Styles + others https://drive.google.com/file/d/1IZq_0CGTbfxlAdIMsjz3VwRzHjl3JL-n/view?usp=sharing (config in WAS Node Suite fpr location)

you need to add the Path to the styles2.csv (or whatever you decide to call it) in "\ComfyUI_windows_portable\ComfyUI\custom_nodes\was-node-suite-comfyui\was_suite_config.json"

So in my case its

"webui_styles": "Z:\AI\ComfyUI_windows_portable\ComfyUI\styles2.csv"

SO you can save it wherever you like

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

[delicate grotto]

use this workflow, drag this to your comfyui window, it has a lot of stuff connected, you can look how it all connects

[median mulch]

haha nice 🙂

[delicate grotto]

this one has a prompt added on top of a prompt(style in au1111)

[fast vector]

set last clip layer I believe

[hollow dirge]

will Discord not compress the image?

or is it still embeded into it

[median mulch]

you dont compress metadata

[hollow dirge]

oh cool

learned a new thing :D

[lean granite]

not sure if you know about it and would rather not use it but there is a custom extension you could get that adds undo and redo

https://github.com/bmad4ever/ComfyUI-Bmad-DirtyUndoRedo

only thing about it is what he wrote on the git

[vital wolf]

where is it?

[west breach]

that's cool 🙂 I just meant this history part, if it was sorted descending that would be so much easier to redo the last workflow. Especially since I'm usually using randomize on everything

[lean granite]

ahh, misunderstood then! mb !

[fast vector]

you can search nodes. just click twice your mouse and type clip.

[vale eagle]

if comfyui could have a formatter node or function would be very good.

[vital wolf]

found it. sadly, i don't know how to connect it xD. thanks though

[heady vale]

A1111 needs the backend image gen

from Comfy

[spring fulcrum]

Does anyone have any suggestions on how to get these images to come out a little less blurry? Using SDXL 1.0 ComfyUI with Refiner, and Ultimate Upscaler

[woeful patio]

when can we start to expect controlnet for xl?

[weary idol]

Can one of the devs here confirm or deny if the 1.0 vae is utilizing watermarking for c2pa purposes?

[fierce hollow]

just take the clip from checkpoint and connect it to input and then output to encoder, that said sdxl doesn't need clip skip, it was already trained with -2 afaik

[fiery cargo]

I have a gtx 1660, I can run SD 1.5, will I be able to run SDXL?

[uncut steeple]

Getting somewhere

[spring fulcrum]

I would think so maybe you will need to add --lowvram to the commandline arguement

[sudden lodge]

you possibly can, but good luck

as a 1660S user. it is pure agony

[fiery cargo]

Thanks, I will try

[median mulch]

gtx in 2023

[fierce hollow]

you can modify comfy to force the text encoder on cpu, if unet fits on gpu it should work

[rough kayak]

masochism

[fiery cargo]

Also, what is the difference between Auto and Comfy performance wise?

[fast vector]

if not mistaken, you connect small dot to checkpoint loader and big one to your prompts aka clip text encode.

[sudden lodge]

comfy > auto imo

[ornate elk]

im having a hard time trying to get it to draw people with spiders crawling on their faces

[median mulch]

comfy is leading the performance right now for sdxl

[fiery cargo]

Ok then, I will try that. Thank you.

[clever verge]

You have got to effective negatives (bokeh and shallow depth). Maybe something with stuff regarding the camera system in your positive.

[vocal rapids]

dreamshaper checkpoint + xl refiner work pretty good

[median mulch]

dreamshaper xl or?

[vocal rapids]

dreamshaper xl

[median mulch]

nice

[vocal rapids]

DreamShaper XL1.0 Alpha2

[median mulch]

that looks very nice

[spring fulcrum]

Ahhh I see it now I think its the f-stop... It was set to f/1.8 and for greater depth of field it should be more like f/8... I'll give that a try

[woeful patio]

very stupid question but 1.5 lora's dont work on sdxl right?

[clever verge]

Not had much luck with f-stops in positive, what have you found out?

[sullen glade]

@ornate elk#1100170312106127410 message

[spring fulcrum]

Well it might work but at this point it seems to be a negative, I'll play around with it and let you know

[fierce hollow]

nope, the structure is too different, probably impossible to convert

[woeful patio]

cool, thanks

[fierce hollow]

(happy to be proved wrong though)

[woeful patio]

is there a way to train loras forxl yet?

[fierce hollow]

with kohya's scripts, just like sd

you have to switch the branch to sdxl

[sullen glade]

lykon (creator of dreamshaper) already did. it's on his civitai

[nocturne dove]

Yeah you can train loras with kohya script like normal. Many people have done this already. Works well, just like in 1.5.

[lean granite]

anybody know why im getting results the size of one pixel?

thats literally the img xD

[woeful patio]

great, i'll give it a try. And probably fail horribly

[nocturne dove]

Have you trained lora for 1.5 before?

[spring fulcrum]

Have you had any luck or found any difference when adjusting these numbers?

[nocturne dove]

If so, you basically just need to set the base model to your sdxl-base model and check the box for an sdxl model. Everything else is the same. You probably should set the res to "1024,1024" though because that is sdxl's base res.

[fierce hollow]

you're using some weird-ass sampler without any vae specified, probably why

[west breach]

I'm surprised it's working without th vae connected

[forest wedge]

In KohyaSS, why am I getting a train_data_dir must be the parent of folders with images error when I'm specifying the parent folder? ie. .../LoRA Training/rosuuri/images

[delicate grotto]

you can still use it, just dont zoom in, drag it as is from discord

[lean granite]

ohh so you need a vae for sure huh, guess i was used to the a1111 auto vae thing

lemme try plugging it in

[median mulch]

sdxl has baked in vae

[clever verge]

I stopped using the SDXL clips and went back to basic a while ago. I felt there were way to many parameters that could affect the image. And when I saw a simple example from comfy I went back. There was also an workflow pasted by jon that didn't have it so I guess it's not totally required to get good images.

[fierce hollow]

you still have to connect the vae to decoder

[median mulch]

ye

[delicate grotto]

sdxl vae is borked, use the 0.9 vae instead

[median mulch]

i know that, but thanks.... again

[spring fulcrum]

Can you post an image I can use with your current setup?

[nocturne dove]

In kohya the folder setup is an image folder that you pick. In there, there is supposed to be 1 or more subfolders with names like "10_my_favorite_person" or whatever where the "10" would indicate that you wanted 10 repeats of each image for one epoch. You should watch the video here if you have not done this before https://github.com/bmaltais/kohya_ss/#tutorials

[west breach]

1.0 vae is fine if you look at the picture at 100%

[west breach]

also make sure the res is 1024x1024

[delicate grotto]

it still hurt my eyes

i noticed it when it was minimized, zoomed in is just terrible

[digital plaza]

at 100% you still see the staircase/aliasing effect

[delicate grotto]

i thought i saw it pixelated for a reason, but then i noticed the shockwave function artifact

[digital plaza]

but 0.9 VAE works fine with 1.0 model

[delicate grotto]

yeah

[lean granite]

eyyy got it!

thank you both!

[digital plaza]

1.0 VAE vs 0.9 VAE

[delicate grotto]

you almost got it, SDXL has a G and L prompt, and also a negative one

[fierce hollow]

it hurts my heart more than eyes since it supposedly functions as a watermark, bad move like usual

[lean granite]

the sampler is still p much the same it just lets me output the input nodes directly from it to reduce overall connections

[lean granite]

G and L prompt?

[spring fulcrum]

Are these separate downloads?

[delicate grotto]

drag this to your ui @lean granite

you will see it there

[forest wedge]

Thanks for this, it works!

[void loom]

ddim strangely goofy

[fallow kestrel]

Dear devs, if you're reading this: Thank you all so much for the release, I've rediscovered my unhealthy addiction to generating images. Fantastic work!

[digital plaza]

That's not the watermark.
The watermark is a separate thing in Diffusers, it's not baked into the model

[rigid prawn]

are Euler a in a1111 and euler_ancestral on comfyui the same sampler?

[clever verge]

Yeah, one is on my github and mod-downloader (the styler, the small missing one). The other is unreleased but only connects primitives and sets resolution.

[digital plaza]

yes

[fallow kestrel]

yes

[trim totem]

I have found this tutorial https://www.youtube.com/watch?v=09Qq3U7LDs4&ab_channel=AIMarvelverse

[lean granite]

u got 2 negs then the G is the regular prompt and the L is for extra details?

[fierce hollow]

invisible watermark is just some garbage you can remove with 2 lines, people are saying the vae itself functions as a watermark here

well, here, on reddit, everywhere

[weary idol]

Can you post some reddit links?

not that I don't believe you

[delicate grotto]

G is the main, L is the style - there is a notepad on the bottom right with all of the styles
neg is the negative

[digital plaza]

Yup, the VAE is broken, it's not intentional

[deft quiver]

are you a developer?

[digital plaza]

I'm a developer but not from Stability

[fierce hollow]

there was a reddit post from the maker of dreamshaper or something, lemme see if I can find

[lean granite]

ahhh cool cool! Thanks for sharing, ill try to reverse engineer it to figure things out

[deft quiver]

so this is not an official developer statement

to be clear

[digital plaza]

I know those artifacts from ESRGAN

also

why would they do that

the watermark function already exists

[fast vector]

They reverted changes on hugging face.
|https://huggingface.co/stabilityai/sdxl-vae/commit/c1b803cdacd792f788384449b8e1d539da4297ad

[digital plaza]

see

[weary idol]

@digital plaza https://twitter.com/StabilityAI/status/1647969418624790530

Stability AI (@StabilityAI)

Stability AI is proud to support the CAI in its initiative to authenticate AI generated content.

Our CTO Tom Mason @nin_artificial will be representing us at this event, please register to watch

https://t.co/AzjgLFPlZU

Twitter · Apr 17, 2023, 2:25 PM

[digital plaza]

I've seen the exact same thing with ESRGAN 3 years ago, they probably messed up a loss function or something like that idk how they train the VAE

[digital plaza]

As I said they already have a watermark system in Diffusers

everything else is unofficial

[weary idol]

@digital plaza they seem to be avoiding the question

[deft quiver]

but you are just speculating

[fierce hollow]

https://www.reddit.com/r/StableDiffusion/comments/15al0j3/dreamshaper_xl10_alpha_2_fixed_the_vae_issue_of/

[fallow kestrel]

should I use this with a VAE loader or pull the VAE from the checkpoint loader?

[full drift]

do we know what for

[digital plaza]

oh they, nvm

[weary idol]

there was were some rules regarding not disclosing or sharing details regarding 0.9 finetuning, I don't think this is about fingerprinting

[fierce hollow]

their new post lacks info about the vae though

(you can see the new one on their profile)

[spring fulcrum]

I got the one from your github... Is there a way to get the other one? It wont work without it and I don't know how to put other things in its place?

[leaden fable]

hello please i need a little help, i use the basic sdxl workflow on comfy, i only downloaded the 1.0 base and refiner, how do i update the VAE with the change with artifacts reverted ?

[clever verge]

It's not ready for release yet but just convert the nodes back to widgets on the ksampler should do it

[fast vector]

Sorry, I am little bit confused. In comfy I believe you use it with loader. Since SDXL was baked with the faulty vae

[digital plaza]

If comfy supports Diffusers, you can just swap the folder out

[fierce hollow]

last I checked a couple days ago only 1.5 diffusers work with comfy, 0.9 didn't at least, might have changed

[hollow dirge]

are there any good plugins for ComfyUI? I heared something about "Manager"

[fierce hollow]

oh seems like it was fixed

[delicate grotto]

didn't try any, personally i think the vanilla is good enough

[fast vector]

Manager is great. It's like extension tab for ComfyUI. It's great if you download worklow that uses custom nodes you don't have

[fierce hollow]

most of the nodes just add unnecessary clutter and load time tbh, only truly useful one are the preprocessors

[delicate grotto]

the only one i would advise is the one that can add or remove prompts mid denoising run, but i don't know its name

[split dock]

how many steps should i be using for the refiner

[fierce hollow]

maybe math nodes too if you can find them outside of some super-huge-node-suite

[fast vector]

Experiment. The example one uses 5 with total steps of 20. I suspect that 75-85% of job is done by base and finishing touches by refiner. This if for comfy

[delicate grotto]

between 0.3 to 0.5 denosing, 20 steps 60 steps, sometimes 8 are enough

[shy kelp]

this is why I won't trust anything said officially, they also silently updated the VAE https://www.reddit.com/r/StableDiffusion/comments/15al0j3/dreamshaper_xl10_alpha_2_fixed_the_vae_issue_of/jtlkzvl/

[delicate grotto]

Realistic stuff demand more denoising, anime and shit like that less

[weary idol]

So far there's a lot of speculation going on, would love some clarification from the devs

this rumor seems to have been started by 1 guy, so yeah just take it with a grain of salt

[shy kelp]

downloading 0.9 and comparing there is a clear downgrade in 1.0 using the same seed and settings

[fierce hollow]

one of the devs said they're not allowed to talk about the vae earlier, if you saw it you saw it, not gonna post the screenshot

[heady halo]

Sooo, how good is sdxl 1.0 at nsfw?

[woeful patio]

why do my faces come out looking like this? any way to fix that

[shy kelp]

but apparently the official word is the downgrade was required for better fine tuning

[lean granite]

that pic u sent had me thinking, is there anyway to not have a pic have all that data in it? screengrab it?

[shy kelp]

https://www.reddit.com/r/StableDiffusion/comments/15aolyj/sdxl_10_seems_worse_than_09/

[delicate grotto]

Personally idc if people grab stuff i make so i didnt explore it

[near barn]

Anyone know how to install seed with Text?

[lean granite]

fair enuff

[delicate grotto]

Au1111 has an optipn for that

[fallow kestrel]

that's exactly the answer I was looking for, thank you :)

[soft zealot]

download https://gist.github.com/alkemann/7361b8eb966f29c8238fd323409efb68 to your custom_nodes folder

[near barn]

Thank you

[chrome flicker]

hi, guys, how to use reference-only with sdxl 1.0 refiners model only?

Because I want to keey the same style with the given img

[west breach]

I use windows powertoys to remove metadata if I need to. It puts a handy shortcut in your right click menu when you click on the file. Then I use these settings:

[fallow kestrel]

I just love how accurate the Depth of Field is. Even with fine-tuned 1.5 models, it was tricky to get it right for the entire image - most of the time, single objects had the wrong "depth"

[uncut steeple]

[fierce hollow]

I made a node that just saves the image without any metadata but that's a neat idea too

[spring fulcrum]

Ok I figured out most of that... Where do I get this?

sd_xl_offset_example-lora_1.0.safetensors

[lean granite]

I just used gpt to create python script which removes metadata and seems to work as well

tried it with this

[uncut steeple]

or just use this node

[lean granite]

2nd one is cleared but I think I might just use the node xD

[clever verge]

I just tested it, it's disabled in my flow (0 strength). It was supplied by SAI to improve contrasts (true black and white) I think.

[west breach]

Most of the time I keep the metadata. Actually I have a custom node that adds the prompts into the metadata separate from the workflow

[fierce hollow]

bot going bonkers

[fallow kestrel]

You could probably just use FFMPEG to remove metadata easily

[uncut steeple]

I keep it too, since I got a library of images I might want to recreate any day