✨｜sdxl

[honest flint]

i have been working on my model for 8 months now and didn't even get to finish the last fork i was on because the architecture changed out from underneath me

[icy brook]

isn't that lovely? I am kind of in the same boat 😅

[boreal bough]

rip ❤️

[soft zealot]

#✨｜sdxl message

[high skiff]

Introducing the new Sytan workflow validation image

2048x2048 pixels of pretty kitty

[soft zealot]

#✨｜sdxl message

[high skiff]

Winston Woof is a mad man, I'll say that haha

glad I gave him early access

[soft zealot]

#✨｜sdxl message

[soft zealot]

hehehehehehehe

[boreal bough]

confused man is confused by how well the upscale works

[honest flint]

btw any one doing finetunes for sdxl of decent size should probably target 1280x1280 and not 1024x1024

[soft zealot]

#✨｜sdxl message

#✨｜sdxl message

[high skiff]

mad man did all my testing for me while I was refining things haha

[soft zealot]

#✨｜sdxl message

#✨｜sdxl message

[soft zealot]

glad to be of assitance 🙂

#✨｜sdxl message

#✨｜sdxl message

#✨｜sdxl message

[kindred bramble]

Roman statues of beautiful Irish woman with a hint of egyptian culture

[soft zealot]

#✨｜sdxl message

Last one

#✨｜sdxl message

[icy brook]

@soft zealot got any suggestions here? 🙂

[hollow hare]

i see, because the KohyaSS guide tab actually recommend using 4e-7 because it's the standard LR but so far i only used 4e-4 and perhaps if somebody trained using 4e-7 can share the findings too

[spring fulcrum]

@high skiff where do i get the upscaler from?

[high skiff]

oh, let me link

oh crap

their site is maxed out, what

here

[soft zealot]

trial & error lol

all i can say is look at the prompts Ive used , Im not speciying anything special although I am using Preconditioning and the offset LORA plus various "styles" , analog phot is aquite a good one for getting a good result IMHO.

Anyways going quiet as have to go out for a hospital appointment shortly

[high skiff]

This is the recommended upscaler

if you don't have it, 4x ultrasharp is almost as good

[soft zealot]

Personally I like

[icy brook]

ok. thanks for helping 🙂 noticed that it looks better with older people bc of wrinkles

[high skiff]

different models for different things

The pixel upscaler is not as important now as it was before

[honest flint]

lmao watda

[high skiff]

damn, I wish I could share with people all the hell and 1000's of tests I went through lmao

[soft zealot]

exaclty , as Ive said before , at the end of the day do/use whatever makes you happy 🙂

[high skiff]

I tested like 20 different pixel upscalers, and at least 5 major upscale revisions, with tons of different settings and stuff

over 3k image samples and tests

[eternal fog]

I won't have time to look over this until after work. But I see in the JSON a contrast fix. Nice, I was bitching about it messing with the contrast on img2img, but I'd not thought of doing this.

[high skiff]

Yup, I added it cause I saw it as well

the default setting should be good for most images

[eternal fog]

I'm interested to see what you've done with it.

[high skiff]

I mean, I can share now

[honest flint]

did they ever say why they changed the encoder on the vae between SD 1/2 and SDxl

[late marsh]

@high skiff Did you have your secret upscaling method shared somewhere?

[high skiff]

take final image, upscale 4x downsample back down using a decimal multiplcation, encode that to a latent, put that into an advanced K sampler with add noise turned off, starting at 50% or later in the diffusion pipeline (this is the magic sauce, its what keeps it faithful, but adds noise to greatly improve deformities)

Then that gets exported, and overlayed on itself to fix the contrast iossues

[high skiff]

its officially out

https://www.reddit.com/r/StableDiffusion/comments/15ed0o8/sytans_sdxl_offical_comyfui_10_workflow_with/

[honest flint]

its kinda annoying since i can't pre-init alot of the layers from my last fork

[eternal fog]

Pretty sure it's a requirement because of the model change.

[late marsh]

Thanks! ❤️

[honest flint]

nah, you can use the sdxl vae on sd 1/2 if you retrain

[high skiff]

of course, I am so pooped lmao

[spring fulcrum]

Is it not recommended to upscale to 4096x4096?

[high skiff]

you can, tho you need at least 20GB VRAM to do that in a decent amount of time

I have done it

[boreal bough]

# SDXL original learning rate <- was meant as in this is what SAI used to train the whole SDXL model, with a batch size of 2048. While it's a neat fact, it doesn't really apply to any training nor finetuning, unless you plan on training significantly above 1 million images, and have around 64 A100s to pull off that training.

for LoRA, typically you'd train around 50~100 images, and that can be done at 1e-3, since the sdxl model is a lot more robust than SD1.5 used to be. You can achieve better quality with lower settings, but by lower I refer to 5e-4, which is as low as you'd ever need to go, for like 5k image datasets. Rather than focusing on training speed though, I can promise that every improvement you can make in your captioning, will show significantly more improvement in the final LoRA, than settings will

[high skiff]

took 7 minutes on a 3080 lol

[eternal fog]

This sounds pretty much like what I was doing, but I couldn't get it quite right. I'll compare it with what I have tonight.

[soft zealot]

makes more sense if you pull things out in your workflow 🙂

[high skiff]

4096x4096 image

4096x4096 is very much viable on a 3090

[eternal fog]

I know you couldn't use it because it had to be stock nodes, but give it a go with the tiled ksampler

[high skiff]

with incredible fidelity and detail

[spring fulcrum]

I'll put my 4090 to work then... I have to say so far this is looking amazing... Thank you for your hard work.. Great job

[high skiff]

oh also, one of the biggest finds behind the advanced sampler offset

the base

[eternal fog]

Lmao it's annoying me now. I really want to see what you did differently. But work :(.

[high skiff]

using the base as the upscale model is insane

forget the refiner

its a lost cause

[eternal fog]

I found the refiner to be really good for eyes and noses. But it wrecks everything else

[high skiff]

base 100%

its way more consistent, more faithful, scrubs less detail, fixes deformities more, and also has higher pixel level detail cause its not overfitted

[honest flint]

out of morbid curiosity are you using the 0.9 vae or the 1.0

[high skiff]

😳

0.9 of course lol

[pallid sail]

No horizontal bands so assuming 0.9?

[spring fulcrum]

What values do I need to change to upscale to 4096x4096

[high skiff]

I just didn't ship it with that, cause I am not sure still why SAI is being all silent about it

[honest flint]

yeh thought so, didn't see any chromatic aberrations

[high skiff]

downscale to 1

[honest flint]

i personally prefer 1.0 except for the ca's... 0.9 is too sharp for me

[soft zealot]

@high skiff One little "under the hood" tweak I made was to chnage the final VAE Decode to a Tiled VAE decode as I noticed it was always (in the console) generting an Out Of Memory Error and falling back to Tiled VA (which is nicer from @visual glade than just crashing) . Saves about 20 seconds (on my 1080ti YMMV with newer cards) waiting for it to fail over

[high skiff]

0.375 is 1.5x upscale, 0.5 is 2x upscale, 0.75 is 3x upscale, and 1.0 is 4x upscale

[spring fulcrum]

Does this need to change at all?

Or should i keep that at 2048?

[high skiff]

you can do whatever you want with it, its not very crucial

That is the only vlaue I have not found a real purpose or use for in SDXL

its never better or worse, always just different

[pallid sail]

Replaced my first stage base img2img with a prompt only pipeline and went from the left image to the right image. Odd.

[hard fractal]

SYTAN

PLEASE

BE

KIND

❤️

[high skiff]

whut

[spring fulcrum]

He has been awesome

and kind

[hard fractal]

he's been awesome!

[spring fulcrum]

yessir

[hard fractal]

a little kinder please ❤️

[eternal fog]

He just means for the upscale pass I believe

[high skiff]

^

the refiner is not viable for the upscaler pass

[hard fractal]

https://tenor.com/view/the-office-michael-scott-leaving-turning-around-no-gif-21361019

[high skiff]

hell, I am very close to removing the whole refiner from the workflow after some recent finds

[eternal fog]

And it does really struggle with img2img, but don't think it's really designed for that.

[high skiff]

need more testing tho

[hard fractal]

we've already removed it.

[peak dove]

I'm loving the artistic value of the horizontal bands 🙂

[pallid sail]

Where's a good place to find upscale models? I've been using RealESRGAN4x but I'm keep to try alternatives

[honest flint]

i think both the refiner and base model need some work but i am super happy they were released in the state they were... arch is probably pretty stable, just some more steps and rhlf inputs

[hard fractal]

that's base only

[woeful patio]

Ah I'm currently training at a 100, I guess that won't come out right then?

[high skiff]

you guys have good quality upscales yet?

[high skiff]

Cause I cracked the code on my end

[boreal bough]

@hard fractal but while you're here. could I ask for this message to be pinned?
it will make redirecting people a lot easier than constantly posting the link
#🔧｜finetune message

[pallid sail]

Imagine AI art archaeology decades from now. Those bands will be like sediment layers in rock

[ember cargo]

amazing work @high skiff !!

[high skiff]

thank youuu

[charred beacon]

Pou coming to Fornite??!!!!

[high skiff]

wouldn't be possible without the support of this community

[ember cargo]

i tried to generate at 4k on my 3090 and it takes ages though, 2k is fast

[spring fulcrum]

My 4090 went OOM with 4x upscale 😦

[high skiff]

oh weird...

hmm

[peak dove]

Some more stunning beauty from the Triple-Process R-B-R

[high skiff]

I did 4096x on my 3080

[pallid sail]

@high skiff So how do you end up using the base model as an upscaler anyway? I thought it would only work up to 1024x1024

[hard fractal]

will have my team look it over!

[brazen patrol]

When you say decades from now, you mean like next week right?😁

[hard fractal]

what was the crack?

[ember cargo]

here is topping at 23,6gb when VAE decoding at 4k

[high skiff]

so the way I use it is to limit its detail focous

Its refining the small details, not ruining the big details

[boreal bough]

I mean if you save every 5 epochs, it doesn't matter - that's why we use repeat = 1, so we can save the model often enough

[hard fractal]

oh, you're actually changing the size of the image to 4096x4096?

[pallid sail]

Dog years versus internet years versus AI years in ascending order of vroooom

[icy brook]

that's cfg 7.5?

[high skiff]

more of my mixed diffusion, but this time img2img with upscaling

[icy brook]

or 7

[boreal bough]

thank you ❤️

[hard fractal]

yeah

[icy brook]

would be cool to get the prompts for it

[high skiff]

yes, I am getting clean and dramatically better looking 2048x and 4096x gens compared to base SDXL

[hard fractal]

tom cruise looking serious at the camera Style: Cinematic

[hard fractal]

yeah, ditto!

we wanna add upscaling to the bot

[icy brook]

why is style seperate? is that some comfyui thing?

[hard fractal]

images are ridiculously better

[high skiff]

After seeing how much it fixes all of SDXL's glaring issues, I can see why

[spring fulcrum]

Am I supposed to switch the refiner model to just base?

[hard fractal]

but would likely be a subscription or something

[hard fractal]

SYTAN

KINDER

PLEASE

[woeful patio]

I have the repeat set to 1, No idea what that means but I assume it'll be fine then 🙂

[pallid sail]

Oh I see, Is that the area value in the downscale node then? Do you manually move it around to specify areas to refine?

[hard fractal]

That's warning number 2.

[high skiff]

ok, so I just can't say literally anything thats not just 100% blindly praising SDXL?

[hard fractal]

"glaring issues"

no

[high skiff]

it has glaring issues, its a base model

its not perfect man

[dense chasm]

what's your prompt?

[high skiff]

anyways

[hard fractal]

#✨｜sdxl message

[hard fractal]

didn't say it was perfect.

[honest flint]

its the best base model we have ever seen

[dense chasm]

your workflow of sdxl 1.0 is awesome

[high skiff]

for sure 100%

[high skiff]

glad you like it

[hard fractal]

i agree!

[high skiff]

its been a shit ton of work

[icy brook]

the eyes are better in the first image?

[ember cargo]

4k looks amazing

[high skiff]

@hard fractal

[dense chasm]

i use ultrasharp 4X model instead

[high skiff]

left is base SDXL, right is after my high res fix

[hard fractal]

give 8K a try 🙂

[high skiff]

that one works nearly as good as the one I settled on, likely not enough to care much

[hard fractal]

those were fun tests!

[high skiff]

no need to flex that you have better hardware, we all know lol

[hard fractal]

you don't need better hardware

8K still runs on 16GB

[spring fulcrum]

@high skiff So when I do x4 upscale it goes OOM then it says its trying a tiled approach and then it takes a while but still works... Should I use --lowvram or something?

[hard fractal]

with comfy

[woeful patio]

me just chilling with 8gb 😦

[high skiff]

try swapping out the last VAE decode for a tiled VAE decode

thats weird that 24GB VRAM would be having that issue

oh wait

are you running --gpu-only?

[spring fulcrum]

yessir

[icy brook]

must be something else you're doing here in comfy. can't get that sort of result on 7.5 elsewhere

[boreal bough]

it switches to tiled vae, since otherwise it would die at that resolution - but it doesn't do this for all users. that's why for some people they just go way OOM
manually putting in the tiled vae solves that instantly though

[dense chasm]

left 1024 res refine model, right 4X ultrasharp with mix diff and downscale 0.5

[spring fulcrum]

oh wait I can do that? --gpu-only

Ill try that

[hard fractal]

it is

[honest flint]

when i was toying around with the idea of trying to merge weights from sd 2.1 partially into sdxl i noticed most of the parm counts per layer were 2x'd... thats a pretty large size increase... i imagine we are still really really far away from overtraining either model based.. that or we need to start doing layer-wise learning rates

[hard fractal]

we're looking to see if it holds, and if it does, we'll release

there's a glaring issue with your workflow, Sytan.

let's see how long it takes for you to find it 👀

[high skiff]

and that would be?

[icy brook]

can't use comfy very much locally. 8 gb vram and 16 gb regular ram seems too little

[hard fractal]

yeah, at least 24GB recommended

[icy brook]

for which one type of ram?

[pallid sail]

My poor 3080

[high skiff]

system

[peak dove]

I can see that Sytan uses a robust turn-of-phrase - but its grazing the deepest sensibilities of a lot of people who've sweated blood for SAI ... 🙂

[hard fractal]

oh, I'm not the one to ask.

[high skiff]

24GB recommended for system

[hard fractal]

that's a comfy question.

[pallid sail]

And here I am trying to run SDXL and Unreal at the same time lol

[hard fractal]

lol

[high skiff]

I have talked with comfy, 24 is the recommended minimum, but 16 works

[eternal fog]

16gb system ram works fine with comfy, but it gets a bit laggy sometimes.

[icy brook]

hehe. ok. a few people that will be resorted to online solutions then 😄

[high skiff]

32 is the actual recommended

[hard fractal]

by whom?

[high skiff]

comfy

[hard fractal]

ah

nice

[peak dove]

(Am I the only one dedicated to using Scott Detweiler's R-B-R Triple-Process for ComfyUI?) 🙂

[high skiff]

he and I talked about this during the research

[hard fractal]

i know.

[high skiff]

back when I first got my weights

[eternal fog]

Increase your page file size if you are having issues with 16gb RAM

[hard fractal]

you had a whole reddit post about how beautiful it was.

[high skiff]

man, was that already a month ago? jesus

[honest flint]

have they announced what controlnets they are going to release for sdxl?

[icy brook]

what's that?

[peak dove]

Virtual Memory

[spring fulcrum]

@high skiff OOM

[hard fractal]

it'll do it.

[high skiff]

are you using --GPU-only in your args?

[ember cargo]

same

[pallid sail]

I've been playing around with something similar in Diffusers. Trying to get some good A-B comparisons but I'm sure I'm screwing something up since I'm getting wildly different outputs

[high skiff]

same for you, are you using --gpu-only?

[spring fulcrum]

yes... Should I take that out?

[high skiff]

yeah, thats whats doing it

[hard fractal]

@high skiff where's your newest at?

[high skiff]

https://www.reddit.com/r/StableDiffusion/comments/15ed0o8/sytans_sdxl_offical_comyfui_10_workflow_with/

[eternal fog]

https://learn.microsoft.com/en-us/troubleshoot/windows-client/performance/introduction-to-the-page-file

Basically if you run out of RAM it puts it in a file on your hard disk. It's extremely slow, but it stops OOM or crashing in edge cases

[ember cargo]

I'm not using any args iirc all default

[dense chasm]

diffusers sdxl lora can't run on 16G RAM GPU,it require more highly

[high skiff]

that is very strange

[honest flint]

(honestly if i was stability ai, i would charge $5 or $100 or something for the controlnets, i would buy it just to feel like I contributed somethint to the cause and get something I would use)

[eternal fog]

This is normal it will switch to tiled.

[boreal bough]

even a 4090 will complain if you try to play a game + run kohya training + use comfy to generate images all at the same time 🤣

I would know after today

[icy brook]

Thanks!

[ember cargo]

then it uses tiled VAE decode and takes a long time, VRAM almost filled

[high skiff]

Give me a few days and i will be able to fix it on my end

my 3090 gets here on wednesday I believe

[eternal fog]

How much VRAM?

[high skiff]

yeah, wednesday

[eternal fog]

I've been able to do 2048*2048 on a 3080

[high skiff]

once that is here, I can run the 24GB VRAM tests

[ember cargo]

generating 4k it almost tops, on 2k i didn't check

[boreal bough]

pretty confident that comfy doesn't account for overhead. but just manually changing the decode vae for tiled decode should solve that

[hard fractal]

DM me tomorrow

I'll get you some compute

[boreal bough]

(since in the end it defaults to tiled vae anyway)

[high skiff]

I have done 2552^2 on my 3080 without OOM, used 9.8GB VRAM

[hard fractal]

for now, i sleeb.

[eternal fog]

Yeah 4k won't work though

[high skiff]

oh? that could be amazing

[eternal fog]

Unless you use a tiled ksampler

[honest flint]

you are giving away compute? ears perk up

[hard fractal]

I can try.

[ember cargo]

if I generate 2k, as it's the workflow default, I use almost 10gb of VRAM

[spring fulcrum]

That would be nice... I've been working on my cloud architecture cert lately

[eternal fog]

That's about right

[hard fractal]

I'll ask the SAI computemaster.

I mean, we're hiring.

I'm looking for 2-3 community finetuners.

[high skiff]

interesting

[boreal bough]

[high skiff]

lol

[hard fractal]

hi Caith!

[honest flint]

my day job is reliability engineering, I just do this fine tuning stuff cause its fun

[high skiff]

gotta love caith

[spring fulcrum]

That would be an awesome job 🙂

[high skiff]

I'd offer up Caith, hes goated lol

[hard fractal]

DM me some links!

[eternal fog]

Same. Different day job. But I just do it to see what happens XD

[dense chasm]

which VAE model u recommend? i choose https://huggingface.co/stabilityai/sdxl-vae/resolve/main/sdxl_vae.safetensors currently

[high skiff]

I have no recommendations for VAE's at the moment

[eternal fog]

I'd use the 0.9 one. They re-uploaded it I believe.

https://huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors

I think that's the one you linked

Which I believe was reverted back to the 0.9 one

[dense chasm]

Lora train on cloud is a pain in the neck, some options on github run locally

[eternal fog]

Kohya_ss scripts are the best for training

I can do LoRAs on a 3080 10gb with them.

[dense chasm]

i prefer to diffusers

[spring fulcrum]

@hard fractal I send you a DM

[soft zealot]

I use it

[hard fractal]

https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_base_1.0_0.9vae.safetensors

here

use that

[peak dove]

How does it compare to @Sytan setup?

[dense chasm]

how u run kohya like on colab?

[pallid sail]

Dug up one of my first test scenes and ran it through SDXL. Such a massive improvement over when I was using SD1.4

[peak dove]

I have d/loaded Sytan's but yet to implement

[eternal fog]

Just run the scripts locally

[peak dove]

Your layout is tidier than mine 😉

[eternal fog]

@dense chasm mcmonkey posted some settings to use
https://gist.github.com/mcmonkey4eva/0f0bd074c17802213817a9a5a50098df

Try it out with them

[soft zealot]

I presume that was directed at me.

My entire workflow is based on @high skiff 's work ad then tweaked

#✨｜sdxl message

[hard fractal]

geez you built a whole front end in there.

[dense chasm]

i know a bunch of apps run locally,but i have to run on cloud,cuz i have not yet bought any GPU for my computer

[eternal fog]

I think @boreal bough has posted some settings too.

[heady vale]

@high skiff do you have any commands for vram on startup? Mine blows past 10gb on the 2k upscale

[hard fractal]

We tested it on the bot.

[peak dove]

So you have Sytan/R-B-R?

[soft zealot]

you could say that lol

[eternal fog]

I think there's some collabs for it. But not something I've looked into.

[hard fractal]

RBR was slightly under-preferred over BR.

[peak dove]

How does R-B-R compare to @Sytan at all?

[high skiff]

@peak dove Winston has been working with me in DM's for a while now, he uses the core stuff from my workflow

[soft zealot]

I have Sytans flow with a Pre COndition STep added befor the normal 2 Step Ksample process yes

[high skiff]

I have absolutely 0 clue, it runs just fine for me

I'll have to sleep on it

I am so burned out after this 72 hour crunch I have been doing to get all this out

[hard fractal]

Also, fixed conditioning (including 4096x4096) was underpreferred.

[spring fulcrum]

--lowvram

[heady vale]

no probs. it can wait and I will do some reading

[hard fractal]

that's your glaring area of improvement there, sytan

[eternal fog]

RBR being Refiner, Base, Refiner?

That acronym is Red Bull Racing to me lmao.

[high skiff]

when did that change 💀

[hard fractal]

mm?

[peak dove]

Try R-B-R Triple-Process ? 🙂

[soft zealot]

this is the "helicopter" view of it 🙂

[eternal fog]

I've been setting the conditioning to match the aspect ratio. That seems to work well.

[hard fractal]

how are your noodles so straight?!

[high skiff]

you told me when I dropped my first workflow that SAI concluded that 4096x4096 is the best Latent size

And then I tested it myself and found no evidence supporting that

[urban breach]

I looked at Winston's workflow and promptly went out to buy alcohol

[high skiff]

damn

my phone died

[hard fractal]

????

why would that be best?

[high skiff]

I have no idea

[soft zealot]

My Name Is Max and I am Inevitable

[eternal fog]

I don't think latent size is the correct word there.

[eternal fog]

No 😦

[high skiff]

@hard fractal

[hard fractal]

resolution

4096x4096

what you just did with the owl

[high skiff]

yes, did I misunderstand something?

oh

[hard fractal]

yeah

[high skiff]

is this not latent size instead of res?

[soft zealot]

oh and @peak dove I also use the offset example LORA

[high skiff]

ah man, my body hurts

[eternal fog]

That's got nothing to do with the latent does it?

[hard fractal]

latent size is 128x128x4

[soft zealot]

Checo should be embarrased being so far behind him inthe same car

[hard fractal]

well

depending on aspect

that's the square latent size

[eternal fog]

I know. Marko is going to make him vanish.

[high skiff]

alright, I will have to catch up with this info tomorrow, I am too trashed from all this work for tonight

[soft zealot]

Swim with the fishes lol

[high skiff]

(had to redo my launch post to reddit 3 times, cause reddit wasn't saving my drafts properly )

[soft zealot]

grab a few hours shut sye after pouting a stiff one

[eternal fog]

Orange

[high skiff]

Im 19 💀

ORANGE YES

[woeful patio]

what are the realest looking images y'all have seen from xl?

[eternal fog]

Live in a country where you can drink at 18 😉

[high skiff]

hmmm, thats a hard one

[soft zealot]

The Future is bright, The Future is Orange (as the old ad tag line went)

[high skiff]

cause some images have an insane level of realistic detail, but they don't look like a "real" photo, if that makes sense

wow, this is the longest I have been orng

[spring fulcrum]

@hard fractal Did you say you had some links for a Position at SAI?

[woeful patio]

I think this one looks pretty good

[high skiff]

have my tiger

My pride and joy

[soft zealot]

depends on you definiton of "real". I, rather fond of this one

[high skiff]

ok, I have to go, I am starting to feel sick again. This happened after I stayed up too late yesterday

[woeful patio]

that is a good looking tiger right there, the only thing giving it away on that one is that he is neck deep in leaves

[urban breach]

I've been sending that to my friends as one of the best examples of how amazing these pics are getting

Well done

[hard fractal]

just my DMs

[hard fractal]

get some sleep!

[spring fulcrum]

Get some rest... Great job on the release... Its awesome

[trim totem]

Hey all, Searge's nodes are not working here.

I'm running comfyui in colab and the default nodes are working, other custom nodes like from Sytan are working as well. Unfortunately, Searges nodes are neither loaded in comfyui (only the default) nor can I open the nodes from the file when I load it in comfy directly. I followed the workflow https://github.com/SeargeDP/SeargeSDXL/blob/main/README.md and my colab script is this https://github.com/comfyanonymous/ComfyUI/blob/master/notebooks/comfyui_colab.ipynb

Had anyone else here this issue? Thanks!

[high skiff]

https://tenor.com/view/daeth-funi-periflight-dies-from-gif-22212023

I am off to sleep

[boreal bough]

the ones we made as a joke while trying to create a whole storyboard about & with chaz
context and situational timing adds a lot to images, I feel.

what XD i swear I replied to the other message

[spring fulcrum]

You mean only people you messaged?

[soft zealot]

also like this one

[high skiff]

discord gaslighting

[boreal bough]

#✨｜sdxl message

[woeful patio]

another contender?

[urban breach]

What's the deal with this new VAE, sd_xl_base_1.0_0.9vae.safetensors

Is that some improvement over the 1.0 one?

[boreal bough]

fixes issues. use the new 0.9 one

[brazen patrol]

You know there are ways to handle your problem. The first is no blue light and therefore no device such as phone or computer. I don't know if you read but I cannot live without my e-reader. I have an 8 inch one which I love to death and is the last thing I'm using with my lights off set at 2%. Some people recommend melatonin supplements but I find they leave me a little doped so instead I found a great alternative with L-tryptophan. Which I highly recommend. Just friendly advice my friend

[soft zealot]

ok a cute kitten before I leave (standard res only as I never did an upscale on this one)

[urban breach]

another dumb question, why do some workflows use the VAE, and some not at all? for instance this latest one from Sytan and friends does not have it?

[brazen patrol]

I admit I haven't really figured out the upside of these upscalers. It's not that the results are bad or anything. It's just that the time invested, and realize that I'm doing this on a souped-up laptop, is just absurd compared to the result. So instead what I opt for because I do want a bigger and higher resolution image is Topaz Gigapixel which I upscale to four times and then downsize to whatever size I'm really looking for. It's been very satisfactory overall and is insanely fast.

[soft zealot]

is it the same as the OG 0.9 VAE (which Ive been using with SDXL1) or subtly different does anyone know ?
@hard fractal maybe?

[rustic garnet]

@hard fractal hi, short question: what is the noise offset used for SDXL training. Is it still 0.0375 as for the SDXL 0.9 version?

[dense chasm]

do u know some datasets i can use with kohya lora testing?

[soft zealot]

I use a seperate VAE

I like the ability to switch

[soft zealot]

Just realised I have something to tweak there

[peak dove]

As many a glitch as VAE 1.0 is giving, it adds artistic value 😄

[hard fractal]

oh I'm not sure about diffusers weights, that's been patrick

i uploaded the 1.0 base weights + 0.9 vae

[brazen patrol]

And we very much appreciate it!

[peak dove]

Overall, SDXL 0.9 is better artistically speaking; 1.0 looks a tad pastel/washed out - but that's just me 🙂

[boreal bough]

exact same

[rustic garnet]

I found that 0.9 looks better but 1.0 has more details and realism. But maybe its just the prompting behaviour is slightly different

[boreal bough]

emphasis on vae though - not model

[peak dove]

I'd say 1.0 was a photographer's dream, and 0.9 is fine art

[brazen patrol]

There was a comment in the Reddit by one of the Developers saying that if you use 0.9 you should raise the CFG by 1.00 to get more equivalent results.

[rustic garnet]

@hard fractal but really: what are the correct settings for training SDXL. Its really confusing that in your paper you write you used a noise offset of 0.05 while the kohya script is using 0.0375 as default. Also, the technical paper was never updated since 1.0, but I assume that you changed something between 0.9 and 1.0. Is there any report of what is different in 1.0. And again: what is the noise offset setting?

[spring fulcrum]

These are looking pretty awesome 🙂

[west breach]

My new Prompt With Style V3 node now supports adding multiple styles and multiple loras using the <style:style_name> <lora:lora_name> syntax. Also supports __wildcards__ with the syntax for adding multiple lines from the same wildcard file https://github.com/bash-j/mikey_nodes

also has ratio selection, option for custom size (including fit to 1024 res) and the clip conditioners are built in

[peak dove]

What node package gives me tidier (and square) noodles/node connexions?

[peak dove]

Awesome

[urban breach]

where's the police car with the military Humvee turret from yesterday, from the guy who kept posting that sea monster lol

[soft zealot]

wanst that @late marsh ?

[late marsh]

Ummm I posted a couple Cthulhus but the other dude got an inspiration and kept doing them a lot 😄

[late marsh]

@urban breach I guess it was @timid sonnet

[boreal bough]

Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn

SDXL understood the assignment 🤣

[late marsh]

Haha, nice 😄

[timid sonnet]

@urban breach

its uploading

[late marsh]

Oh god that red one is terrifying 😄

[boreal bough]

the red one is actually lore accurate. well done

[late marsh]

Mine is just faded black inked to my shin 😛

[smoky patrol]

Cthulu heard the call

[late marsh]

Oh man I love SDXL ❤️

[smoky patrol]

Burb! Love it!

I've been working on a massive prompt list for everyone.

[boreal bough]

for sytans new workflow, I can recommend this style tag centered, dreamworld, sunset, bokeh, f1.8

adds nice colors and patterns to every image

[boreal bough]

weeeeeee ❤️

[late marsh]

I just run One Button Prompt randomness, it's fun 😄

[polar epoch]

Same lol. This was one of the last ultrawide wallpapers i generated before i hit the sack lol

Need a few eye doctors, but once those kinks are done with

[smoky patrol]

I love your pfp, btw

[urban breach]

I love this style, very striking

[timid sonnet]

Here's your prompt generated using my workflow

[smoky patrol]

Those eyes! P e r f e c t i o n

[polar epoch]

Gothic, but high cfg

[smoky patrol]

Very intense look

Yeah, high contrast lighting...might look interesting with some intense shadows or fog

[eternal fog]

I've only got my own datasets, don't know about any public ones.

[peak dove]

<------- just trying-out Sytan's Implementation ...

[indigo vine]

👍 good tip

[peak dove]

U can take DulcoLax for that! 😄

[turbid garnet]

Error occured while processing model:Input type (c10::Half) and bias type (float) should be the same

Getting this error while trying to run SDXL

[boreal bough]

jonah about to be swalled by the shoggoth kraken, Ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn

[peak dove]

First Sytan example - clear and sharp!

[late marsh]

Oooh, nice composition ❤️

[turbid garnet]

Error occured while processing model:Input type (c10::Half) and bias type (float) should be the same
Getting this error while trying to run SDXL

[peak dove]

But no real advance/disadvantage to the Triple-Process R-B-R imho

Need to add a Preview Image to Sytan's w/flow

[peak dove]

Moby Dick!!!

[smoky patrol]

Oh....you know. I totally forgot all about....

[sinful dirge]

[peak dove]

That phrase should go in the Supporting Terms box? 🙂

[late marsh]

Haha, the randomness ❤️

[boreal bough]

yep

[peak dove]

Thank you

[pallid sail]

Melon cat

[urban breach]

Seems like the new Sytan Workflow is adding some nice details and fixing faces

[peak dove]

^..^<

Can you do a melon collie?

[vital wolf]

Share the workflow as well for noobs like me?

[smoky patrol]

I love melon cat

[boreal bough]

trying to make a xena fighting a named devil for you... but all I get is them weirdly dating instead 🤣

[west breach]

Just fixing an annoying bug with the regex pattern. I'll let you know once that's done 🙂

[smoky patrol]

Well, you see, go look at my prompt lol

The full is far too long, but you can see what I posted haha--it's close enough

[smoky patrol]

Love the warmth behind

[peak dove]

I am generating images using Sytan's w/flow - how do I get it to Upscale at all?

[urban breach]

do you have the one he released like an hour ago?

[peak dove]

Yes

[urban breach]

it's already upscaling to 2048x isn't it?

the second output image

[peak dove]

No, then mebbe I'd better re d/load 🙂

[urban breach]

should look sorta like this

[peak dove]

Yes I have that look ...

[urban breach]

Also you have to download the upscale model he linked, if you don't have that

[ionic dragon]

@eternal fog why do these both image nodes points have the same input?
whats the point of having 2 image nodes with the same input?

[timid sonnet]

@ionic dragon one is preview one is save

You disable Save and enable preview

if you dont want the files saving

[urban breach]

This one

[polar epoch]

Huh, do share the workflow! And what addons would be needed?

[ionic dragon]

whats the point?
anyways its getting saved

why would someone want to preview it

[woeful patio]

any idea why khoya is making multiple files?

[urban breach]

Sure, I did change the VAE Decode to Tiled because he recommende that, something about memory. Here's the PNG. This was before upscaling, just for workflow.

[boreal bough]

yep - cause the script says 'save every 5 epochs'. that way you can compare them to each other

[eternal fog]

My workflow is a mess I use for testing. I move the nodes on or off depending on if I'm testing or saving.

[boreal bough]

I usually end up using 20/25/30 - but its good to try with all of them, and see & learn from the results

[ionic dragon]

oh ok

[polar epoch]

Aye. Tiled means it will use less memory to instead of generate one large and instead make puzzle pieces of smaller gens instead

[woeful patio]

so do they all have different results when using and you just pick the best one?

[urban breach]

Also here is where he uploaded it - https://github.com/SytanSD/Sytan-SDXL-ComfyUI

[polar epoch]

Thanks!

[pallid sail]

There's your melancholy

[boreal bough]

exactly. the bigger the number - the more it has been trained. But at some point, more training starts becoming a bad thing. this way you can find just the right sweet spot

[woeful patio]

ah that makes sense thanks

[peak dove]

Werf! 🙂

[ionic dragon]

@eternal fog why does it have an extra sampler?
what does it do and it starts at 18 and ends at 22?

[eternal fog]

Use sytans workflow. I've not seen his new one yet. But I've had a quick look at the JSON and he's doing similar to what I have now, but better.

[peak dove]

OK, I got the freshest release at Sytan that I could, will re-open the json file ...

[eternal fog]

Not at my pc so I can't look at it. It's probably old as well.

[ionic dragon]

ok

[peak dove]

Sytan's implementation gives a very crisp and detailed finish - perhaps a tad too sharp for my more artistic tastes - but that's just me 🙂

Prompt = a vivid watercolor depiction of diverse Rococopunk Afrofuturist women with beautiful and bold head wraps walking toward a huge moon in the background, holding hands, walking away from the camera, collaborating in a creative and productive environment, women empowerment poster, photography, inspired by the styles of victo ngai and vladimir kush

[peak dove]

sd_xl_base_1.0_0.9vae.safetensors - is this used as VAE or Checkpoint?

[unreal plover]

Nano robots

[eternal fog]

Vae

Well both

It's the SDXL 1.0 checkpoint with the 0.9 vae

[peak dove]

I thought vae and ckpt were separate?

[eternal fog]

Because people didn't like the 1.0 vae

[eternal fog]

No you can embed them with the model as well.

[peak dove]

The 1.0 VAE opens up so many distinctive artistic possibilities over here at Torcello Towers ... 🙂

[eternal fog]

It's why the comfy model loader has a VAE output

The 1.0 vae adds weird lines and colours for me.

But both are there to choose from.

[peak dove]

<--- gone to Tescos - L8r

[vital wolf]

Can someone share a ss of sytans workflow, not at my computer

[honest flint]

Its not sdxl but it seems good enough for a 2.1 model to go through the effort of releasing probably?

[woeful patio]

sorry for the questions, do you have any idea what the 20 to 40 number changes?

[proud stirrup]

SDXL Now available on Mage.Space. Not a shill, just a fan.

[urban breach]

it's like a video game, how do i disable depth of field, bokeh in all these pics 😄 just put that in the negative prompt and hope for the right seed.

[honest flint]

could probably do a no bokeh lora pretty easily

[silver dock]

too sharp because upscaler model i think, for base without upscaler pretty nice, i try your prompt, it's beautiful

[urban breach]

i think somebody was working on one yesterday

[unreal plover]

Is here a way to reduce the size of a lora ?

[urban fjord]

Yes, use a lower dimension for the network and conv.

[rustic garnet]

or skip conv at all 😉

and use bf16

[ionic gulch]

would it be better to do x2 + x2 or directly x4 with your upscaling method?

[late marsh]

Oh!

[spring fulcrum]

@soft zealot are you still online?

[soft zealot]

Out away from desk for 2-4 hours

[azure oxide]

so the answer is yes 🙂

[spring fulcrum]

Would it be ok if I chat with you privately when you get back?

[soft zealot]

Drop me a question and I’ll look later

[eternal fog]

2x + 2x would just make it take longer.

However it's possible it might prevent possible artifacts.

[west breach]

@vital wolf you will need to git pull to update the code for my nodes to get V3. Here is an example workflow that uses the new node.

[trim totem]

How do I get photorealistic hands?

[ionic dragon]

click a picture of your hands

[trim totem]

in sdxl all the hands are fucked up compared to Realistic Vision 5.0 for example?

left sdcl searge nodes left Realistic Vision 5.0

[boreal bough]

wont solve it, ironically enough. it's that many words have bokeh & blur baked in. the trick is to avoid those words. especially cause when you accidentally stack them, the blur starts overtaking all other aspects (4k, 8k, 16k <- are top offenders)

[urban breach]

hmmm very good point

[upbeat summit]

but but 4k, 8k, trending on artstation are my most important tokens

[urban breach]

i was indeed using 4k lol

what about "cinematic photo"

[boreal bough]

thank god 'artstation' wasnt affected. I actually need that one for concept art styles, until I get a better bigger prompt

[boreal bough]

oh yeah. cinematic lenses have amongst the biggest DOF, so insane blur

[upbeat summit]

it works much better again in XL compared to SD 2.1

[urban breach]

what keyword would you recommend for the style to get a realistic photo

[ionic dragon]

@high skiff whats your opinion on DreamShaper XL1.0?

[urban breach]

without the blur i mean

[azure oxide]

all hail sytan

id just keep shallow depth of field, blur, bokeh in negatives

works fine for me

tbh idek if shallow dof works

[late marsh]

❤️

[urban breach]

but aren't you supposed to kind of specify a style at the beginning of the prompt to avoid cartoons, anime, drawings, etc. something that is "photorealistic" but without the blur

[azure oxide]

ohhh righ this reminds me to test out how well sdxl can draw supercell thunderstorms

[cedar lark]

I get crisp and sharp images, maybe you should up your prompt game 😛

[late marsh]

It does great job with this

[soft zealot]

Use “hands in pockets” in all your prompts………….

[pallid lintel]

In sdxl I found that it’s better to not use “photorealistic” and the like since that imply a style

[urban breach]

wow that is ridiculous

[azure oxide]

idk thats missing the charactaristic supercell look

nice lightning flashes though

[urban fjord]

I find hands to usually be good, just make sure to not draw too close attention to them.

[late marsh]

Oh.. let's see

[azure oxide]

hands are always easy, just edit them in photoshop afterwards 🙂

perfect hands everytime

[urban breach]

personally i am so overwhelmed by the style of pictures my brain just filters out bad details. it's not until i send it to somebody and they are like, "what's with the extra 2 fingers" or "legs don't bend that way" that i'm like... oh yeah lol

[woeful patio]

why do the faces of my lora come out looking like they've been in a chainsaw accident?

[spring fulcrum]

Great Tater

[late marsh]

This one is pouring 😅

[barren bluff]

My first gen with SDXL

[ionic gulch]

can you explain in more detail? i already thought that 4096x4096 couldn't be best for all aspect ratios. that's why i used some math nodes to do width * 4 and height * 4 instead - from limited tests that reduces problems with stretched content. but maybe there is a better calculation for those values?

[peak dove]

My latest gen using Sytan's w/flow

[ionic gulch]

oh, i had the same idea. how do you calculate the values? i was just using some math nodes and use height x 4 and width x 4 to get close to the recommended 4096x4096, but consider the selected aspect ratio.

[ionic dragon]

what does this end at step:1000 mean?

980 steps?

but it isnt doing that

[late marsh]

What's with the spaghetti thing I keep seeing from time to time? 😂

[ionic gulch]

if end_at_step is higher than your steps, it just ends at your last step (30 in that case). just avoids that it get's stopped too early if you have some calculation or rounding errors in your workflow.

[shy kelp]

No shit Sherlock, if you over-prompt like that it's perfectly normal that the end result will follow your direction. That wasn't my point, try "a close-up photo of a beautiful woman" in SD 1.5 and then compare it to SDXL, let me know

[peak dove]

I have the upscaler in my setup, but my images are still a tad too sharp but not upscaled. I will prefer Sytan's w/flow for a more phoographic result, and use R-B-R (or even SDXL0.9) for art 🙂

[ionic dragon]

oh ok

[static prawn]

also downloaded the sytan workflow but not sure how i can edit the steps in general

[peak dove]

You have a sublime touch to your art!

[peak dove]

Oppenspaghetti?

[floral island]

thx ^^ just re-tracing my steps and see what made it click

[ionic gulch]

Only use Tiled VAE if your VRAM is too small. It will/may introduce small errors (color differences/lines between those tiles) and shouldn't be used if you have enough VRAM.

[vital wolf]

Titled vae?

[ionic gulch]

it's a different node that some people use instead of the regular vae node.

[ionic dragon]

which one do you prefer?

[peak dove]

The sharper on right side

[ionic gulch]

left has less broken finger

[full spruce]

Secınd one look better but just slightly

[young mantle]

Not really

The fingers

Left has less

Broken ones

[azure oxide]

right has more details, fingers can be fixed with editing

[young mantle]

Exactly

[hoary saddle]

where can we download the 2 nodes in sytan's workflow?

did git pull on comfy already

[full spruce]

Both images has broken finger to be honest

[young mantle]

Left one has less

[ionic dragon]

which one from these?

[azure oxide]

yeah so if youre already going to edit either one, you may as well choose the one that has more details

[barren bluff]

Still learning how to tame this powerful beast

[young mantle]

I mean, is harder to see the broken one in the left img than in the right one

[static prawn]

why my upscales always use 11 steps?

[young mantle]

Left one

The background is more detailed

But honestly, both are cool

[full spruce]

I am kinda new with stable diffusion, what should be steps for base and refiner? Refiner seems to break faces especially eyes

[young mantle]

Too cool

Did u use an upscaler?

[ionic dragon]

@azure oxide @peak dove your opinion on these

[peak dove]

1st one

[ionic gulch]

80% base + 20% refiner

[barren bluff]

Nope, just native 1024 x 1024

[nimble marten]

Anyone try pruning the SDXL base model? Is it even possible?

[young mantle]

I see, pretty cool tbh. But the eyes in some imgs are... Awkward XD, if u use a good upscaler it would be fixed

[late marsh]

Tanks!

[floral island]

anyone got it working in vlad? i just haven't been able to get decent output from sdxl in vlad

[shy kelp]

its already pruned to 6.5gb. it's got a larger size for fp16 because of it using two clip models

[floral island]

ThANKS!

[nimble marten]

Ah, I figured.

[peak dove]

Vlad explodes my 8GB VRAM RTX2070 - pushes it into error after error

[nimble marten]

hehe let's prune it to fp1 and see what happens 🤣

[young mantle]

Whoa that's nice

[spring fulcrum]

Did you do that with ComfyUI SDXL???? Those actually look like real tanks.... Awesome job with those.

[late marsh]

This is made in SD.Next. Thanks!

[full spruce]

Do you have recomendation for upscaler?

[spring fulcrum]

Excellent job with those.

[late marsh]

Thanks.. good 50% random prompts 😄

[peak dove]

I got Vlad to work Base only (using Diffusers) - so so results, softish finish

[indigo carbon]

new minecraft update hits different

[floral island]

yeah, this is the issue i have too, either base only, or it just disintegrates

[brazen patrol]

Minister Gordon Ramsey went public in stating: "Officially a state of war does exist with Italy"

[ionic dragon]

so the 1st post 1st image & 2nd post 2nd image had Sytan's default's samplers and schedulers.
1st post 2nd image & 2nd post 1st image had dpmpp_2m and karras for all ksamp.
but we cant always say that its better, have to try for other styles too
but ddim produces better fingers

[late marsh]

That's what I'm using

[barren bluff]

Pink Slime! I need that to make my mining laser lens...

[indigo carbon]

FTB?!?!?!

[peak dove]

Seeing that Comfy actually works at SAI - I'd say that ComfyUI is the present-time leader in the field of SDXL ... !

[barren bluff]

Ah yes, good old days

[indigo carbon]

man, I feel old now

[peak dove]

Both the ones I chose I liked because crisp and sharp

[ionic dragon]

yeah, the texture was better

[peak dove]

This kind of romantic art needs Base only ...

... it lives to be soft focus

[barren bluff]

So how do you use the refiner? just chuck the image I got from t2i to i2i?

[floral island]

actually, i do base+refine, upscale and do a base/base+refine, and then get the best upscaled image

[indigo carbon]

If you're on A1111 there's an extension for it

[floral island]

sometimes base gives better output, somtimes base+refine gives better output

[static prawn]

idk why but my image is stuck

in vae decode

[peak dove]

OK - I've yet to implement Upscale - seeing as I use GigaPixel 🙂

[ionic dragon]

i think for all pastels we should only use base

[peak dove]

SDXL1 has that kind of pastel finish compared to the more saturated 0.9 look

[barren bluff]

Where do I find that?

[young mantle]

The best I have seen is the @high skiff one. Literally fixes any output. But idk if it's already out, he said yesterday that he is gonna make it out today but idk if it's already.

[indigo carbon]

https://github.com/wcde/sd-webui-refiner

[urban breach]

Oh wow thanks, I will change that back cause I wasn't having issues before hand

[peak dove]

Sytan released his latest w/flow already

[young mantle]

Yayy

[static prawn]

and i dont know where i can see the outputs of comfy

my pictures dont get saved

[young mantle]

Do u have the link?

[indigo carbon]

I don't like using SDXL on Comfy, it wants 2 positive prompts

[peak dove]

From final VAE Decode set a Save node

[heady vale]

they go in ComfyUI\output

[nimble marten]

Comfy is just annoying to use in many ways.

[young mantle]

Woah pretty cool

[peak dove]

I only ever use 1 positive prompt, and almost always leave the negative prompt empty

[indigo carbon]

however if ComfyUI will have the AIT optimization, I'll switch to it in a heartbeat and make a workflow that I like

[ionic dragon]

@peak dove do you use Sytan's workflow?

[peak dove]

Sytan's w/flow uses 2 positive prompts to good effect

[peak dove]

Yes, but it gets over-sharpened imho

[static prawn]

would like to adjust the steps in sytans workflow, but dunno how

[urban breach]

Oh it does this anyway, I guess it makes no difference for me? Warning: Ran out of memory when regular VAE decoding, retrying with tiled VAE decoding.

[nimble marten]

it's kind of annoying how SDXL doesn't seem to handle genning at 512*512 at all

as in it produces nonsense

usually

[peak dove]

As you go from one sampler to the next, 1st sampler say 0 to 3 steps; 2nd sampler 3 to 12 steps; 3rd sampler 12 to 20 steps ... rinse and repeat!

[static prawn]

i think this is actually the most significant, and most important benefit over sd 1.5 lol

[indigo carbon]

still, I hate the fact that it needs 2 positives. If ComfyUI will have AIT optimization I'll make a workflow similar to my A1111 one

[eternal fog]

You don't NEED 2 positives. But it can help.

[ionic dragon]

i want to understand the workflow can you explain it, the ksamps?

[brazen patrol]

It doesn't need two positives. That might just be the one by Sytan. Mine doesn't have two positives for example

[nimble marten]

@floral island i love these styles

[ionic dragon]

you have your answer, comfy is very flexible

[ionic dragon]

double exposure?

[floral island]

yes 😄 trying to figure out how to get it to work with other cool stuff

[peak dove]

base ksamp produces a small image from text - the second ksamp take img2img from base ksamp into refiner ksamp - or so I believe?

[late marsh]

Generation data stripped, had to fix it a tiny bit

[indigo carbon]

idk man, if comfy UI will have that AIT optimization I'll make an opinionized workflow

[peak dove]

AIT?

[floral island]

WE COME IN PEACE

[peak dove]

Turn our swords into ploughshares

[indigo carbon]

tensorrt's big mama

[nimble marten]

[late marsh]

Oh no👆

[floral island]

where do we even start to address the ... things in this image?

[young mantle]

The heck hahahha

[peak dove]

Thomas The Telescope?!?!?! LOL

[young mantle]

It's smoking lol

[urban breach]

What is that Lora if you don't mind me asking?

[static prawn]

is this new dprk propaganda?

[nimble marten]

Auto1111 seems to be struggling with SDXL img2img

[barren bluff]

No Lora

[indigo carbon]

capable of insane performance with no degradation

[urban breach]

it says offset 0.2 but mine is throwing errors

[nimble marten]

like it dropped to s/it

[static prawn]

use --medvram

[indigo carbon]

Comfy said they are currently working on implementing AIT to ComfyUI

[young mantle]

Send the prompt 🥵

[nimble marten]

"a battle tank with a main cannon with the face of Thomas the Tank Engine"

[nimble marten]

I tried the inverse "Thomas the Tank Engine as a battle tank with a main cannon" and it wasn't as cohesive

[young mantle]

That's disgustingly amazing

[spring fulcrum]

That is an awesome minecraft blob

[urban breach]

this is what the Lora comes up as, but the one they released with SDXL is saying Erorr occured when executing Loraloader.. HeaderTooLarge

[ember cargo]

there's a AIT VAE decode node but when I tried it, resulting image was all black :/

[young mantle]

Wow it's rly cool tbh

[indigo carbon]

it's not released yet, whatever you're using isn't the official one

[eternal fog]

You would still use 2 positives to get the most out of it. The model has 2 text encoders. It's your choice if you use them both or not.

On automatic it will just use the one. Or just use the same in both and mask it from the user.

[indigo carbon]

I want to punch it

[brazen patrol]

You can do whatever you like I'm just saying that there's no two positives that you keep complaining about

[uncut steeple]

[spring fulcrum]

Neon LoFi Girl in the city

[ember cargo]

probably using it wrong, so idk

[peak dove]

Sytan's "two positives" seem to be 1) a prompt and 2) a style modifier (rather than a full-blown prompt). It matters not whether you fill-out one or both!

[brazen patrol]

Well I've seen workflows with two actual positives. But it's a personal workflow thing and nothing to do with ComfyUI

[peak dove]

Somebody try this in Sytan's two positive boxes - BLACK - WHITE - the result should be GREY?! 😄

[young mantle]

Wow it's really accurate

Just tiny issues

[brazen patrol]

For example I found a great workflow for Loras, but it imposed some upscaler which I absolutely did not want. Outside from that it was a great workflow. So I deleted the nodes that I didn't want and left the rest. No big deal

[peak dove]

It has a soft watercolor finish ...

[hoary saddle]

What does this mean? Renders still work
missing {'cond_stage_model.clip_g.transformer.text_model.embeddings.position_ids'}

[spring fulcrum]

Thank you... Its a work in progress

[static prawn]

idk im still not a fan of the refiner, always destroys most of my images

[peak dove]

https://github.com/comfyanonymous/ComfyUI/issues/818

[floral island]

"you might wonder how i ended up here..."

[urban breach]

sorry guys does anyone know why it errors out with any value other than 0 for both these strengths?

[late marsh]

All my images are with refiner.. works nicely

[eternal fog]

What does it do if you use less prompts?

[urban breach]

same thing

[mellow tendon]

Yes I have had that issue, not sure how to fix , other than have it higher.

[eternal fog]

Can you show the stuff you have plugged into the LoRA loader?

[peak dove]

She's cool, but she'd slowly rust!!!!! 😦

[urban breach]

Alright I redownloaded it from here, I think the file was corrupt 🤷‍♂️
https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/sd_xl_offset_example-lora_1.0.safetensors

Now it's working 🙂

[eternal fog]

Ok cool

[peak dove]

Some far out stuff (TingaTinga Art)

Street Art

[hoary saddle]

thanks Torcello

[uncut steeple]

[peak dove]

Luke Skywalker

[proud stirrup]

Wutno.

[peak dove]

Gandalf?

[proud stirrup]

Beksinski.

[peak dove]

OK, the artist

[woeful patio]

does the refiner fuck up loras? because it looks fine until the refiner comes in

[solemn scroll]

Hi, any idea if () still emphasizes some keywords or not? Using comfyui rn

[vale eagle]

I feel the same but I kinda changed my mind

[hoary saddle]

do you have a github to grab those custom nodes from?

[proud stirrup]

I'm starting to warm up to SDXL.

[peak dove]

https://www.artstation.com/blogs/kaddoura/bKn1/stable-diffusion-text2img-prompt-modifiers

[wet nacelle]

You could just try it.

that's a dumd ass question

[compact obsidian]

How much faster can SDXL get with optimizations? 2x? 10x?

[young flume]

Hey guys

How can I use roop extension in SDXL?

[wet nacelle]

What would "optimizations" be referring to? I would also like to know now.

[solemn scroll]

I can try, but how do I measure for sure that "clown" is emphasized in my prompt "image of a ((clown)) eating pizzas"?

[solemn scroll]

Thx, so it doesn't change between old models and SDXL?

[peak dove]

(clown:1.3)

[compact obsidian]

Basically the Unet is the slowest part, so going through those layers and figuring out what is taking the most time, and pruning from there

[wet nacelle]

From what I understand that would be the way.

I never understood it fully but instead used the (word:1.0) method.

[peak dove]

try it - Stable Diffusion = A111, Comfy, NMKD etc etc

[wet nacelle]

Ah interesting.

[peak dove]

(a:1.1) (brown:1.2) (cow:1.3) (jumped:1.4) (over:1.5) (the:1.6) (moon:1.7)

[compact obsidian]

People have done optimized attention implementations for previous SD models, seems like the same concepts should work

[peak dove]

More Street Art

[static prawn]

using sytan workflow

without upscale

[static prawn]

with upscale

[peak dove]

Using R-B-R w/flow

[static prawn]

just erase a lot

:/

the first one without the upscale looks way better imo

[uncut steeple]

maybe give my setup a try, its embedded in this pic

[static prawn]

allright, ill try soon 😄

[peak dove]

I can see little difference upscaled/non-upscaled

[static prawn]

no its a big different, the upscale erases a lot of detail

look especially at the cats nose

[uncut steeple]

Its mostly about the proper settings

[peak dove]

Nasal upscales are tricky 🙂

[peak dove]

Cool!

[heady vale]

read the UMD notes. you need to change the steps to tune it

[static prawn]

i did

and set it allready to 25

for less changes

[heady vale]

try 29 or 30

[urban breach]

This one is really good, one of the best I've tried so far. It creates some awesome detail, and these eye pics are crazy 😄 Nice work

[uncut steeple]

Thanks man, appreciate it

[peak dove]

Wonderful love the wide-angle and dof

[brazen sapphire]

Hey, a bit new to this. Noticing my generations lose all of my lora-ness when they pass through refiner. Any ideas of what I could try?

[cedar lark]

default short prompt images look way better in SDXL than 1.5, I don't know what you're trying to say

[vocal rapids]

Hi-res workflow of both Sytan and Searge work fine

[static prawn]

missing nodes for ur workflow

[uncut steeple]

Yeah I have a few custom nodes installed, with the comfy manager it should tell you what you need

or maybe this helps

[static prawn]

last time i tried to install ultimate sd upscale everything crashed

and i had to reinstall my whole comfy

xD

[uncut steeple]

Works fine for me

[brazen patrol]

Uses Frank Frazetta Lora. workflow in image

[late marsh]

Birb! (it might have a double face.. kinda. You decide!) 😄

[peak dove]

Art At Home

[hoary saddle]

anyone have a good image2image workflow to share?

[mellow tendon]

Are people getting better results putting the same prompts into both text_g and text_l encoders? or having different prompts,
I have had some really good results with just leaving the text_l encoder blank.
This seems to make more difference to the output image than what number of steps or sampler is used, but I haven't seen many people talking about it!???

[crisp owl]

Supporting terms to L if you're gonna split

[rustic garnet]

definitely something that should be investigated.
I found both happening. Some prompts get better, others get worse

[mellow tendon]

Yeah but what does that even really mean!??, I guess I need to look at some more good examples, it is like learning to prompt in different way/new language.

[peak dove]

I made a moonlight picture in Sytan's w/flow - the second prompt specified sunlight - so the result was a very sunny moon (if that makes any sense?)

[rustic garnet]

so far I prefer to just use same prompt for both

[rustic garnet]

using different prompts for both is very awkward and strange. It seems to work, but it's definitely a hack

[crisp owl]

G and L use different versions of Clip, which you can go down that rabbit hole as deep as you'd like

[late marsh]

Are these just prompting or some other wizardry? 😄

[peak dove]

I suspect what it does is it places emphasis on the words used early on in each prompt - SD takes heed of the earliset words in a prompt, de-emphasising the later words - so a 2nd prompt is a form of re-emphasising (that's my surmise at least...)

[indigo carbon]

the prompt literally had 4 words in it, so other wizardry

[rustic garnet]

the problem is that prompts are aligned, i.e. the k-th token in clip-l is concatenated with the k-th token in clip-g. Using different prompts totally destroys this alignment and might end up with strange results

[peak dove]

.. as in a "sunny moon"

[late marsh]

That's what I thought 👍🏼

[crisp owl]

Everything with SD is a "your mileage may vary" situation lol

[indigo carbon]

also, that's on A1111. the main reason I didn't move to Comfy yet is because it follows prompts in a different way

[rustic garnet]

if your prompts are "moon" and "sun" then this makes sense, as you end up with a token that is associated with both, moon and sun.
But if you use "moon" and "by greg rutkowski" then the word "moon" is associated with the word "by" which does not make sense. Of course the tokens are somewhat contextualised,but still it is strange

[trim orbit]

i spotted a wild vae today. supposedly is a lot more memory efficient. https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/discussions/7

[vale eagle]

it is identical in a1111 and comfy, like 98% of the image.

[indigo carbon]

weird, for me it's not like that. can you replicate this in Comfy?

[vale eagle]

The only way makes image different is the seed which calculated from GPU or CPU

If you make both using the same way. It would be identical

It is in the setting

[late marsh]

Testing Cyborg Style lora ❤️ | https://civitai.com/models/119405/cyborg-style-sdxl-or-goofy-ai

[indigo carbon]

I tried copying the workflow and params of the images I make using A1111 to Comfy, it's not the same.

[trim orbit]

I think my best results come when i give different prompts to the G and L, where i treat L like a support. Both clips are going to go out and invoke the latent space. I think if you prompt them seperately, it can give a higher quality focus across a wider attention. If that makes sense. Prompting is all about the attention you're drawing to various parts of the latent space. using 2 prompts is like, 2 wheel drive vs 1 wheel drive

[indigo carbon]

If it is the same I would love to see someone prove it

[vale eagle]

This setting

[indigo carbon]

ohh, is this possible to change also on ComfyUI?

[ionic dragon]

are any 3rd party control nets out?

[hoary saddle]

https://github.com/comfyanonymous/ComfyUI/discussions/118

[hoary saddle]

someone has to have a dope img2img workflow to share 🙂

[fading haven]

dope, not so much, but i got this one

ops

[hoary saddle]

thanks Grougarawr, will snag the upscaler and try it out

[fading haven]

np o/

good luck

[urban breach]

finally starting to get rid of the boka look, having a subject and detail in the background is awesome. just don't look at her fingers lol

i pretty much got rid of every photograph "high quality" related term before it stopped the depth of field thing with a photographic subject

[vast galleon]

Does anyone know how to add a toggle button to a custom comfyui node?

[vale eagle]

ctrl + m

[indigo carbon]

still not the same. A1111:

comfy

[vale eagle]

Could you give me the prompt and let me try?

[vast galleon]

Thank you. I'm looking for something more intuitive though

[indigo carbon]

all in the metadata

[crisp owl]

The way COmfy is currently written it does not allow for a toggle to be added into a node because it expects a connected line to have an output.
You can write a custom receiving node and custom output node to allow for that, but you'd have to do the same for every single node then.

Alternatively, yeah just mute the node, or drag the output to a reroute, following another reroute, and use that as the toggle

[thin nova]

It would be nice to have a switch junction. A reroute with two inputs, a toggle, and one output

[vast galleon]

Could it be added with code? Looking into litegraph and they have it on this project.https://github.com/jagenjo/litegraph.js?files=1

[soft zealot]

hmm this nose is suspicous

[twin bluff]

can anyone give me a good prompt to test if my sdxl works on auto1111 ? 🙂

[crisp owl]

In a way.

I fiddled around with trying to code something for a while, ended up abandoning it.

My process was to have an image output (like vae decode) go into an output switch, so one image input with two image outputs, controlled by a boolean. That way I could have a "switches" row of just booleans changing between 0 or 1, which would control which image output I wanted.
So I could have one for saving only, or preview and save.
I got them working by making a custom save image node which normally it would yell at you when there is a line connection but no image to be received and would redwall you, but I made it so that it would ignore an empty image.

This worked, except ONLY for that. If I wanted to do the image output to any other node, I'd also have to make that a custom node to ignore the empty input.

So basically the NoneType would have to be fundamentally rewritten by Comfy to allow empty inputs. That's above my python coding, no expert here

[indigo carbon]

did you figure out how to reproduce A1111 images in Comfy? I have been trying and I couldn't recreate the gens I make with A1111 with Comfy.

[vast galleon]

Thanks. Me neither but I'll mess around with it later tonight and see what I can do.

[vale eagle]

trying, have some interest result.

[crisp owl]

This was essentially it that I did
I can't completely show it because currently I'm testing some other custom node ideas, but yeah this is basically it. Could have also added an image input switch to the Preview image so you'd always get the preview.

[stray mantle]

I'm really liking @high skiff latest workflow 👍🏻

[peak dove]

beauties R-B-R Triple-Process

can I use my A1111 embeddings in ComfyUI?

[hoary saddle]

can comfy default, add a transparent png overlay onto a final output image?

like for adding a company logo in the bottom corner

[thin nova]

image/postprocessing/ImageBlend set blend mode to overlay and load a transparent image. might not always be in the exact corner if you change resolutions - haven't tried it

[hoary saddle]

will try it, image node (transparentpng) --> imageblend node --> save image node?

interesting result

[vale vapor]

i was just curious cause i lowkey love stats...whats the difference between discord bot users now that 1.0 is out and when it wasnt out yet.