✨｜sdxl

[quasi remnant]

but not worth an upg from 3090 imo

[placid edge]

48 would never have been possible, now it would be because vram is cheap but nvidia is making too much bank

[zealous horizon]

4090 does make me happy 🙂

[placid edge]

also 48gb would cannibalize their professional card

[quasi remnant]

yeh, it's still a beast of a card for consumer AI and the speed is like 2/3rds faster than 3090 for sd

i do kinda regret not waiting a month but my pc died and i needed to replace it right before 4090 came out xD

[shy kelp]

I got a 4080, its pretty damn fast. uses scraps of power too

its like alien technology the amount of work it does for just 100 watts

[placid edge]

im still going on the 3090, i have owned 1080ti 2080ti and now 3090, first time skipping a generation, dont feel like im missing out too much

[zealous horizon]

I use it for so many things and compared to an H100 there isn't too much difference for regular tasks (bark, Sadtalker, 4bit LLMs ~13B)

[fair crow]

Anyone tried running sdxl on tpus?

[abstract comet]

3070 with 8gb.. still works with SDXL 🙂 I think I wait for AMD to come up with some competition

[polar epoch]

Yep. Manual nonsense it is with navigator lol

[zealous horizon]

As long as it's CUDA most things will be fine if you don't intend to train or do insane upscales.

[solar magnet]

where are people chatting about about fine tuning sdxl? curious how people are fine tuning and which hyperparams are working for them

[boreal bough]

by not using 1/16. that exponentially increases it

5e-4 is fine - but either with 1/8 or 2/16.
just wanted to test it for science!

also this is not me saying that 5e-4 is the new correct setting - since it's still way overkill

just testing around

[thorny frost]

Hi! Is there a way to avoid InvokeAI of unloading models after generating? it makes generation so much slower!

[wet nacelle]

Hey guys I made an SDXL 1.0 upscaler workflow from a 9.0 and standard upscaler workflow.

[bold osprey]

Is there a way to stop processing a run with Comfy?

[boreal bough]

view queue - then cancel the job

[wet nacelle]

yes.

[bold osprey]

okay thanks going to try it out

had a process running for 400seconds now

[wet nacelle]

large res?

[bold osprey]

no usual setup but another process occupied some VRAM and it seemed to have crashed

[wet nacelle]

zamn

those zamn external applications!

[bold osprey]

yeah I'm not one for closing my apps

also seems to be because I minimized the window while it was running

[wet nacelle]

I do it knowing that I use hardware intensive programs everyday.

weirirsdaioh

weird

[bold osprey]

everyday and always

[wet nacelle]

yes

[soft bone]

oh got it. what about 8/16?🤔 i'm still not familiar with how the dim and alpha act in relation to each other, I just know a 1/2 ratio used to be standard

[wet nacelle]

I only run AI now

I do nothing else

[bold osprey]

yes

spent months every day and all day on sd and other softwares

was a hell of a ride

[boreal bough]

alpha same or lower than dim. it basically protects the lora from being overcooked/breaking/many bad things.
for the sdxl ideal, you'll want dim:alpha of 8:1 ratios

[wet nacelle]

I've been apart of the AI art community since last year. I was in MJ since v2.

I now run several generative AIs a day.

[bold osprey]

never used MJ tbh but was there for most stuff from day 1 anticipating it beforehand

[soft bone]

Is that based off your findings or is that ratio related to the architecture or something

[wet nacelle]

interesting.

[bold osprey]

rock solid render

[hoary saddle]

come across this while setting things up?

[deft stream]

that looks cool!

[wet nacelle]

I'm running it through my model upscale setup

[boreal bough]

originally recommended to sai by kohya himself.
I then tested it, and at least on 0.9 I had a 100% success rate for all lora types I wanted to train, and got my time down to 14min minimum for a good lora of any type

[bold osprey]

just switched from colab to own gpu and damn it was necessary

the boot times alone

[wet nacelle]

yuupers

[boreal bough]

while there are niche cases where that ratio doesn't work - they probably wont be relevant to you, unless you want to make a bad lora, to generate a few good images, with which you'll then properly train a good lora

[wet nacelle]

I keep thinking that this is the tech support channel whenever I get into a conversation.

THis channel helps more than the correct channel.

[bold osprey]

what do you mean there is proper channel for that xd

[wet nacelle]

https://discord.com/channels/1002292111942635562/1002602742667280404

[bold osprey]

it's all lies

the tech support rests here

[boreal bough]

its like when people point out that general chat or offtopic exists. it just ends the conversation - but no convo ever moves chat XD

[wet nacelle]

It really does

I've seen that

[shy kelp]

dumb question.....how do i manually update a1111?

[edgy sierra]

she's sexy

[exotic path]

open repo

git pull

[shy kelp]

ok

[terse sedge]

I’m looking for the same. I tried using LastBen’s Dreambooth Colab to no avail

[wet nacelle]

https://tenor.com/view/brian-regan-derp-what-weird-crazy-gif-11376849

[tribal jackal]

Pretty solid for a simple test of the base model. Didn't even use the refiner.

[zealous horizon]

nope

[exotic path]

Someone know if the training parameter for SDXL 1.0 are available ? Stable Diffusion staff tell on the stage conf that they'll be publicy available but i can't find them

[hushed hornet]

All sentient beings are welcome to my swamp.

[bold osprey]

he looks stoned

[wet nacelle]

opaque (ohpayka)

how does that make sense

[hushed hornet]

He is.

[deft stream]

Smouth and Mash 😎

[bold osprey]

smouth and mash

[hushed hornet]

Smouth & Mash

[shy kelp]

great British cuisine

[lilac wren]

Or you can create a txt file in stable-diffusion-webui directory, set this in :
`@echo off

git pulland call himupdate.bat`

Just double click on update.bat

YOu can do a shortcut on your desktop

[shy kelp]

this is what my webui-user.bat file looks like: @echo off
git pull

set PYTHON=
set GIT=
set VENV_DIR=
set COMMANDLINE_ARGS= --autolaunch
call webui.bat

[wet nacelle]

https://tenor.com/view/jerma-jerma-sus-jerma985-twitch-unedited-andy-gif-23569746

[frozen inlet]

is this for comfy ui?

[wet nacelle]

fuck no

I'm just fucking around

[wet nacelle]

Is that real life?

Did he go to the buig gulp?

[bold osprey]

no it is not

[frozen inlet]

ok so its a virus good good

[wet nacelle]

Bu da big gulp

[lilac wren]

https://tenor.com/view/crazy-eyes-steve-buscemi-big-eyes-gif-5312326

[hushed hornet]

Yooo I made one of Keanu drinking ayahuasca! Let me find it

[bold osprey]

HAHAH SHOW ME

[wet nacelle]

no u

[deft stream]

Keanku Reefs

[tall pulsar]

https://tenor.com/view/virus-computer-matrix-humanity-gif-22263680

[bold osprey]

you should always get the cpkt instead of the safetensors :)

[wet nacelle]

\

\

[]]

[deep dagger]

XL indeed

[cyan crown]

still problem with hands......

[wet nacelle]

and???

it's a base model

it's going to be bad

[fading haven]

how much time it takes for you guys to make a single image?

[cyan crown]

20 sec with 50 steps

[fading haven]

it takes 15 to 20 minutes for me, but my GPU is not the best

D:

[bold osprey]

so I developed a blindness for that kind of thing

cant see it anymore

[hushed hornet]

Love the workflow @wet nacelle thanks homie

[bold osprey]

nsfw works fine for me tbh

[frozen inlet]

sdxl is very impresive indeed, prompt was frog at the supermarket photography... what is wrong here? im running it inside auto1111

[wet nacelle]

no probs

[bold osprey]

what res is it

[hushed hornet]

Looks right to me

[frozen inlet]

512

[bold osprey]

what the fuck

[wet nacelle]

This is so funny

[bold osprey]

100k seems overkill

[wet nacelle]

the fuck

[frozen inlet]

it is ok except its not a photography and frog looks stupid

[bold osprey]

take your power bill into perspective

and the time it takes to train

[wet nacelle]

A story

[frozen inlet]

what resolution should it be?

[wet nacelle]

yoyoyoyo

[bold osprey]

at least 768

[wet nacelle]

yo guys

porn porn porn

https://tenor.com/view/bitcoin-cryptocurrency-blockchain-crypto-moon-gif-19439861

https://tenor.com/view/doge-doge-coin-to-the-moon-crypto-cryptocurrency-gif-14563637

https://tenor.com/view/hoezaay-sup-yo-yo-rap-rapper-gif-25521647

https://tenor.com/view/bro-review-imax-laxman-bhibatsam-bro-crying-gif-25874631

[frozen inlet]

okay 768 thats the way

[uncut bison]

how am I supposed to find this missing node

[shy kelp]

i get something similar but different nodes lol

[bold osprey]

the higher the better

I like adding the prompt "color grading" to it

[uncut bison]

googling for "comfyUI text box node" really doesn't lead me anywhere where I would find it

[shy kelp]

its just a text box too ... like come on how can a text box node be missing

[bold osprey]

makes it look so insane

[wet nacelle]

BROS!! 50 50 I SWEAR!! BRI BRO

[bold osprey]

bro what is your dataset in the first place

[wet nacelle]

brother what!!

[bold osprey]

how would you get so many high wuality pics of your step sis

[wet nacelle]

the fuck!

[bold osprey]

100k is too much

[uncut bison]

when I click on "queue prompt", I get this, is this because of the missing text box node, or a different issue?

[bold osprey]

nah bro you are delusional xd

for loras you dont need that much

[wet nacelle]

this was funny now it's fucking gross man

[bold osprey]

bro

[orchid oar]

1024x1024 on rtx 3070

[orchid oar]

thx bro

[primal hatch]

bro

[frozen inlet]

is this base sdxl?

[orchid oar]

yes

with refiner

no

bro

i just say frog in supermarket

bro

[frozen inlet]

aaaaahhh the refiner thing I need

[wet nacelle]

shut the fuck up you spaz

[bold osprey]

shush

[orchid oar]

maybe just try 1024x1024 first

[bold osprey]

use roop and shut up

[wet nacelle]

naw not that far man

[orchid oar]

nice

Looking good

[bold osprey]

damn man

these frogs are getting out of hand

[meager pebble]

How do you keep the seed from changing in ComfyUI? It keeps rolling a new one every queue, and I can't link a primitive in to override it

[bold osprey]

soon they are claiming the cheese aisle

[lilac wren]

Baguette !

[frozen inlet]

you all stealing my clever prompts it took me months to craft pay me

[orchid oar]

U can buy a better gpu

[orchid oar]

That will make it faster

[meager pebble]

loll I'm an idiot, thank you

[orchid oar]

Maybe think about buying a second one

[cyan crown]

yeah billions and billions

[potent summit]

Getting there with text, kind of...

[hushed hornet]

better than anything I've gotten! Nice

[frozen inlet]

I like this one

[wet nacelle]

Why does it look awful? like ddim awful. ddim is good just not for 1.0

[candid spindle]

Is this whole 1.0 VAE thing resolved now? Was it a watermark or what was it?

[uncut bison]

so can anyone tell me where I can download this "text box" node that is missing?

[sullen glade]

no negatves

[bold osprey]

kohya

[wet nacelle]

naw

can't be \

dude no

[orchid oar]

Yes, it looks very real

[sullen glade]

try (childish, imperfect, incorrect, poorly drawn:1.4) as negative

[orchid oar]

Damn, text working really well

[frozen inlet]

why me frogs not good i dont understand

[wet nacelle]

Is that 1.0? You changed the style of one off the outputs.

How did you do that?

[bold osprey]

lmao

[frozen inlet]

i need the refiner thing

bro

[orchid oar]

Download it

[turbid gulch]

Nothing in those is underage lol. That's all 1.0 polished up in 1.5

[wet nacelle]

who did that?

where did it go>

[orchid oar]

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

[frozen inlet]

ty

[soft bone]

get comfy node manager or whatever, it has a button to get all missing nodes

[frozen inlet]

got em

[uncut bison]

ah, thanks!

[orchid oar]

I dont think so

[uncut bison]

and where can I set the path where comfyUI saves the image to?

[wet nacelle]

Do you know what it even does? I don't mean to sound rude.

It jsut runs the output again and refine the details of the image.
Your output won't change with the base if that's what you're thinking.

[solar magnet]

going to start a seperate discord and start dumping my findings in there if you are interested

[vital wolf]

Can someone provide a workflow with lora+ upscaler?

[terse sedge]

👍

[wet nacelle]

Joe Biden neters the server and only talks about training nudes and his step sister.

[frozen inlet]

okay thanks for the info well as long as its a little bit more refined its good, have anyone test sdxl with roop faceswap?

[wet nacelle]

shut

[terse sedge]

Post the link 🙏

[hushed hornet]

When I was in middle school, I'd try to join chat rooms where they talked about rap music, and tried to be condescending to them. I know the phase 🤷‍♂️

[soft bone]

roop is very low res, not sure it can handle XL sizes

[wet nacelle]

I don't think this persons an american. I don't think they live in MAerica either. I'm almost certain that it's jsut a guy in Niger.

[solar magnet]

i'll dm you looks like I can't post the link

[solar magnet]

anyone else interested dm me 😄

[wet nacelle]

Hey devin I sent you a DM

[orchid oar]

Almost ("I wanted to be a Horse")

[quasi remnant]

learning how to utlize the text_l encoder is still taking a while xD

[kind pewter]

then, he burped alot of hot gas

[solar magnet]

idk if you are welcome joe lol

[wet nacelle]

I don't think he understands where he is.

[hushed hornet]

does LemonParty still have that training option?

[terse sedge]

Sent you a friend request

[frozen inlet]

downloaded refiner but this is not a model is a small file... sorry im a newbie where do i drop this refiner thing in auto1111?

[solar magnet]

lol

[wet nacelle]

I actually don't use Auto anymore. I'm using comfy and understand that file set in general.

[wet nacelle]

Try to search online for a 0.9 tutorial. It likely has the same workflow type.

[quasi remnant]

would not recommend using base+refiner on a1111... can give you a simple workflow for comfyUI if you need one.

[hushed hornet]

I believe you generate an image with txt2img with base, send it to img2img, and then use refiner as the model there in img2img

But agree with the above - stick with ComfyUI for now. I learned it just for this and it was worth it.

[wet nacelle]

Sounds like hell.

[hushed hornet]

Oh and if you do use A1111, make the denoise 0.2-0.4 or something.

[wet nacelle]

same would go for comfy

[hoary saddle]

@zealous horizon get to the last step in that video where he clicks Print training command, and get the TI error, will research it more, thanks for the link! very cool stuff

[hushed hornet]

But yeah, it sounds like hell which is why I'm using the comfy for now.

[wet nacelle]

trye

'tyure

true

[quasi remnant]

the refiner is a model BTW, it still goes in checkpoint/model folder, but you need to be able to pass a still noised image from the base sdxl 1.0 to the refiner to finish last 20% of steps.

[orchid oar]

1. Frog sad alone
   2.Frog searching other Frogs
   3.Frog found Friend Frog
   4.Frog happy

[quasi remnant]

which is really easy to setup in comfyui

for auto1111 idk how you do it

[civic sigil]

Interesting the base model has 23k downloads and the refiner has 13k

[orchid oar]

I mean u dont really need the refiner to make good images

[turbid gulch]

Size

[frozen inlet]

time to switch to comfy ui, just wonder if we can use dynamic prompts there

[quasi remnant]

polls say different

[civic sigil]

And Stable Beluga 2 already has 100k 😂

[quasi remnant]

like prompt weighting? you can do that.

[civic sigil]

LLMs are a lot more popular than diffusion models I guess

[quasi remnant]

https://github.com/Stability-AI/generative-models/blob/main/assets/001_with_eval.png?raw=true

[orchid oar]

Oh, i though it was only a little bit better

[quasi remnant]

the refiner greatly improves quality

even sdxl 0.9 with refiner>sdxl 1.0 no refiner

[turbid gulch]

Many said screw it an are using playgroundai.com instead

[lunar patrol]

everything I'm trying to do in sdxl looks absolutely terrible, any ideas why? Also what's the deal with the refiner? This whole thing feels like a massive step back in terms of ease of use

[orchid oar]

oh okay, i mixed the sdxl models up. sorry

[quasi remnant]

ye that 2nd one is 0.9 refiner

[turbid gulch]

Forward lighting helps with photorealism

[orchid oar]

Are u using Comfy?

[quasi remnant]

Now you CAN still get good base images but

[civic sigil]

You are doing 1024 resolution right

[quasi remnant]

the refiner often gives better outputs on average

like a big average

[lunar patrol]

I'm using auto1111

[orchid oar]

Yes, it does look really good with the refiner

[civic sigil]

I would skip the refiner if you're just getting started

[lunar patrol]

and yes, 1024

[orchid oar]

That was the problem for me

[lunar patrol]

I'm just getting started with 1.0, not with SD in general

[frozen inlet]

not prompt weighting I mean saying {red|green|blue} frog and if you generate forever some come out blue some come out green some red

[quasi remnant]

i think that works fine

[turbid gulch]

If you're just getting started just use www.playgroundai.com. You can clean them up and full edit in their Canvas there. Very powerful kit.

[quasi remnant]

i recall using it

i don't do stuff like that very often tho

[frozen inlet]

okay lets test this

[civic sigil]

I dont use the refiner and I've been doing SD since before auto lol

[wet nacelle]

I run mine through the refine then upsacle it with 4x-UltraMix_Balanced

[frozen inlet]

grisk?

[lunar patrol]

I've been running SD locally for ages, I have no desire to use any remote services

[quasi remnant]

i'm still in the middle of prompting something so can't test it rn

[civic sigil]

Idk what that is but I just used the command line back then

[wet nacelle]

Scary real

[frozen inlet]

thats real og couldnt get it working like that

[turbid gulch]

Well, I can understand that. But PlaygroundAI doesn't do that "cedits" and "tokens" crap. It's 1k images a day free. Paid version is like $15 bucks and adds more toys. Worth it. It's the only one I could find that does that.

[runic lark]

I don't remember Pixar monsters being this scary looking?

[hushed hornet]

Rewatch Up, the monster of grief is far scarier in the first 10 minutes of that film

[lunar patrol]

I spent over £2k on a 4090 and I'm damn well going to use it 😆

[turbid gulch]

And no, I dont work for or represent playgroundai.com. Just highly recommend it. For both pro's and newbs.

[turbid gulch]

I get it.

[austere ferry]

Why is this happening? I'm using the refiner checkpoint on automatic 1111 It seems like i'm getting worse results than models based on SD 1.5

[frozen inlet]

its actually not bad, no need of gmail account, its easy works on a potato laptop

[hushed hornet]

It's hit or miss for me. A lot of the same issues randomly.

[hoary saddle]

maybe it means i dont have enough vram (3080 10G)

[austere ferry]

I haven't been able to get a single prompt that isn't doing something weird :S

[civic sigil]

As in only refiner?

[wet nacelle]

1. WIth
2. Without
3. Upscale

[orchid oar]

u mean with or without refiner?

[turbid gulch]

Try more one word prompts and less 30 word prompts. Less is more with 1.0

[wet nacelle]

I DFUCJKED IT UP

SHIT

[orchid oar]

Noooo

[wet nacelle]

sorry hold on

[turbid gulch]

And neg prompts are often ruining my images

[wet nacelle]

I sent the wrong images

[orchid oar]

Np

[turbid gulch]

I honestly think some neg prompting is being added to the prompt lol

[lunar patrol]

I still don't get why I have to do this half vae thing either

[hushed hornet]

Ugh but what if I NEED the upside banana shaped anthropomorphic doorknob wearing a Tony Robbins outfit to have twelve trees and the sun baby from Teletubbies behind him?!?!? Idk how to do that in one word

[wet nacelle]

with = refine
without = no refine

[orchid oar]

Thats hard

[civic sigil]

I dont get what you're doing, it sounds like you're generating the whole image with the refiner model

[austere ferry]

Am I not supposed to?

[civic sigil]

No

[austere ferry]

ah

so what then? i generate using the base model and then img2img to refine it with the same prompt?

[orchid oar]

Im thinking about it, need a sec

[turbid gulch]

That's what I've been doing

[civic sigil]

SAI should have worked on A1111 themselves instead of antagonizing it, now people are doing text2img with the refiner model only 😭

[turbid gulch]

Adding lighting prompts has helped too

[austere ferry]

AH! ok. I kind of half understood that from what i saw y'all saying, but this kind of confirmed it. Thanks

[austere ferry]

can you give an example?

lit from side?

[civic sigil]

Idk if A1111 has the "official" workflow, you can do it that way but there are other ways that are generally better. If I was using A1111 I would just use the base model only tbh

[turbid gulch]

"natural light", "torch light" candle light" spot light" flash light" "flash" etc. Just be creative. Light sources clean up those 1.0 dark shadows.

[eternal fog]

No you just can't train TI's yet. Do a LoRA that works.

[austere ferry]

where can I find the official workflow?

[orchid oar]

no refine, refine, upscale (thats gonna be soooo wrong, im not good at these)

[wet nacelle]

correct

[orchid oar]

It gonna be all wrong

[turbid gulch]

You can literally tell the AI to give me an orange spotlight in the top left corner. Aqua top light. Bottom right corner glow. Etc

[orchid oar]

oh really?

[wet nacelle]

I should've told you that

[civic sigil]

Idk if there's an A1111 version or not

[austere ferry]

oh damn, let me try that

[austere ferry]

that's fine, i'm curious as to the official workflow independent of a1111

[orchid oar]

Told me what? (my english is bad im sorry)

[austere ferry]

getting better results on base now. Thanks, will try passing it thru refiner

[potent summit]

It's like one out of 50 tries...

[wet nacelle]

I forgot to let you know the roder at which the images were presented.

[turbid gulch]

Spotlights will often alter the image base colors. This is red spotlight on text saying "YouTube"

[indigo carbon]

@visual glade can you share your .trt file for SDXL? I remember you said it worked for you

[high skiff]

whats the struggle?

[civic sigil]

Ask @high skiff I dont have it personally

I just use the base model

[high skiff]

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

here you go

[orchid oar]

Im sorry lmao. I dont understand. I thought i should guess them? The order was left to right top to bottom

[high skiff]

its not the official 1.0 release, but that will be out soon

[wet nacelle]

I was trying to actually show you what it looks like when comapred.

[orchid oar]

ohhhh

I thought i should guess

[soft zealot]

darn am I going to have to rework mine now as well lol

[orchid oar]

lol

I thought it was game

[potent summit]

Getting actual text in an image

[high skiff]

likely haha

I have some slight new changes, of which should hopefully include a crude but effective upscale workflow

[tacit basalt]

Anyone have a idea why all my generations look like this? (only with SDXL)

[high skiff]

bad VAE

[bold osprey]

do you have the separate vae

[high skiff]

can I see your nodes?

[turbid gulch]

Orange spotlight changes these shoes via 1.5

[orchid oar]

These Images look very good

[quasi remnant]

Vae mismatch?

[high skiff]

the model has the VAE baked in

[cunning geyser]

AYO THATS SOME FRESH FUCKING J'S

[tacit basalt]

Gonna check the Vae, one sec

[wet nacelle]

@orchid oarI remeber you asking about the benifit of the refiner. I wanted to show you it.

[bold osprey]

then it might be the base sdxl without refiner

[quasi remnant]

But don't use the built in one, use the one here: https://www.huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors

[turbid gulch]

Or use two spotlights. Here's red and aqua. Notice the scene tho. Clean.

[orchid oar]

Ah, yea now i see what you meant. Thanks for the presentation

[visual glade]

I can't do that

[orchid oar]

Its actually a bigger difference than i thought

[wet nacelle]

no problem.

yeah

[cunning geyser]

o o f

bro dropped the freshest pair of timb jordans i ever seen

then downgraded

[indigo carbon]

damn. can you at least explain what did you use to convert it? is there like a ComfyUI script for TensorRT conversion?

[cunning geyser]

bro was gonna create the jortimbs

And then dropped it

[wet nacelle]

https://tenor.com/view/jordans-shoes-new-shoes-oc-gif-12930939

[fleet veldt]

as some have said.. VAE but.. mine was setup in the command line args on the .bat.. so check there too.. that overrode any settings made in the gui for me

[visual glade]

yes I have scripts but it's not something that is in a state to be released to the public

[quasi remnant]

to keep it simple, just override with the vae that stability ai posted (from link i posted)

it's better than the one built into 1.0 too

[wet nacelle]

does comfy use comfy ui?

[indigo carbon]

[soft moth]

Who trained the SDXL lora?

[tacit basalt]

It was the Vae, thanks everyone that help me 🙂

[quasi remnant]

np

[austere ferry]

Prompt:
Man in white Suit walking down alley towards viewer

I can't seem to get SDXL 1.0 to make the character in the scene face me. Is there a specific prompt for that?

[soft zealot]

is it as crude as mt original crude upscale lol

[wet nacelle]

wicked

[quasi remnant]

"looking at viewer" maybe?

[high skiff]

pretty close lol

[wet nacelle]

lol

[high skiff]

oh wait, I wanna throw up my old 0.9 research workflow lol

[olive hearth]

How are you people getting such good faces? She's melting 😄

ComfyUI 0.9 workflow, base 1.0, refiner 1.0, vae 0.9

[lunar patrol]

I hate this. Think I'm just going to hold off until there are some good models available for 1.0

[high skiff]

[austere ferry]

prompt:Man in white Suit walking down alley looking at viewer surrounded by neon lights in chinese and japanese,digital photography, 1man

😦

[quasi remnant]

that's not denoised.

you can try my workflow instead if you want

[shy kelp]

if you're getting a complete image by your base model and feeding it to the refiner, ur not doing it right

[high skiff]

are you using my workflow?

[wet nacelle]

zamn

[quasi remnant]

no they said the 0.9 workflow

the old one that was going around

[olive hearth]

Thank you ❤️

[high skiff]

Wasn't sure if they menat my 0.9 workflow

[wet nacelle]

Can you show what you mean?

[civic sigil]

I would try "looking at camera" or "looking at us" and then put "(face, lapel, knees:0.8)" at the end

[olive hearth]

Yes! The 0.9 one (downloaded last night)

[shy kelp]

you need to end at x step and start at x step for the refiner to takeover in the generation process

[azure oxide]

probly wouldnt work for you anyways since youd have the have the same GPU as comfy. he also mentioned that he did it on an A100, which I doubt many of us here have access to

[quasi remnant]

there was this one specific workflow that had SDXL 0.9 that a lot of people were using, it's a pretty bad workflow and did nothing regarding the new optimization tricks we learned along the way & what not

[high skiff]

I see

[shy kelp]

ive got a simple workflow here but im not sure if it works out the box.

[quasi remnant]

116kb jeezus

[high skiff]

ok, if thats the case, 1.0 needs some changed settings, of which I am working on

[quasi remnant]

how is that simple xD

[austere ferry]

Man in white Suit walking in alley looking at camera, surrounded by neon lights in chinese and japanese,digital photography, 1man, (face, lapel, knees:0.8) this worked 😄

[wet nacelle]

oh interesting.

can you send me an example of how I can accomplish that in comfy?

[orchid oar]

Drag and drop this image in comfy for a good workflow (not made by me)

[indigo carbon]

I just don't understand why it's so impossible

[shy kelp]

just did. you can just copy my ksampler nodes and paste them in your own workflow

[olive hearth]

Yes, I've seen your messages and I'm excited, thank you for your incredible work 😄 I'm just comparing images posted here and mine.

[quasi remnant]

np; if you use mine and are under 24GB vram be sure to reduce batch_size. it should fully denoise any prompt you do.

[wet nacelle]

ah sorry

https://tenor.com/view/ahhh-ahhhh-gif-25957248

[quasi remnant]

because the reason your image output looked bad was the remaining noise

[wet nacelle]

ah sorry

[shy kelp]

oh maybe you cant 😅

[civic sigil]

Probably the prettiest version of the default prompt I've seen so far

[shy kelp]

anyways its a good workflow

[olive hearth]

This community is amazing

[hushed hornet]

Choose your class

[azure oxide]

comfy mentioned that it does work, you might have just messed in a step somewhere. shouldnt be impossible at all

the .trt files not working for other graphics cards is just how it works

[shy kelp]

Where do you have your vae stored?

[orchid oar]

https://comfyanonymous.github.io/ComfyUI_examples/sdxl/

[ember cargo]

is there any workflow base + lora + refiner? I'm not sure how to use the refiner with lora

[shy kelp]

in the vae folder?

can someone explain why comfy is apparently so loved?

[bold osprey]

I am still using the A1111 default paths and have pointed the paths there

[solemn scroll]

This model is crazy, it can do things that would normally need a Lora or finetune without one. I mean, it even knows about Belle Delphine

[shy kelp]

I don't know where that is

[civic sigil]

Oh didnt know that

[shy kelp]

ComfyUI_windows_portable\ComfyUI\models\vae

[fast vector]

you don't you use lora with base model

[civic sigil]

The other one looks nice to

[ember cargo]

i use lora with base model, but i'd like to know how to use the refiner too, with lora

[bold osprey]

seems like we need to do some research

[orchid oar]

Yea, i found out about that website today. Its pretty helpful

[bold osprey]

for science

[shy kelp]

cause people are huge nerds and auto1111 dosent have proper sdxl support yet

Hey another dumb question, who makes a1111?

[azure oxide]

tbh im curious to see if we end up with some good finetuned refiner models

[high skiff]

Automatic 1111

[fast vector]

you can't

[high skiff]

thats the devs name

[shy kelp]

ohhh"

ive been mislead by the another models folderlol

[wet nacelle]

It loads with 80% of it in red.

[shy kelp]

Guess that makes sense.

[orchid oar]

Is Comfy faster than Automatic1111?

[shy kelp]

lol

[solemn scroll]

I will generate those pics instead of buying them, and they will go bankrupts because of it. That's an absolute win.

For the AI research field, by the way.

[orchid oar]

Because for it seems like that

[high skiff]

if you have enough VRAM, way faster

[hushed hornet]

Good question! Comfy and Automatic should race each other in a 5k

[shy kelp]

i think you can get the manager to download the missing nodes?

[bold osprey]

you'll have more value that way

[wet nacelle]

Where is that?

[sly jay]

really happy with this image and SDXL in general, but wondering how I may fix the perspective on the sword to make it look more natural(using a1111 with impainting)

[wet nacelle]

fuck

[ember cargo]

I can lol

[shy kelp]

you gotta install it.

[orchid oar]

8gb and i can see the diffrence

[hushed hornet]

Real answer: I'm actually unsure so far. Other people claim Comfy is faster

[wet nacelle]

From where?

[turbid gulch]

Use www.playgrounai.com until it's released. 1k images a day for free. SDXL 1.0 + editor.

[shy kelp]

from where you download it 🙂

[orchid oar]

Yea i think so too, but im not quite sure. Maybe i will race them in some time

[bold osprey]

it definitely is waaaaaaaay faster

[wet nacelle]

from the git?

[shy kelp]

https://letmegooglethat.com/?q=comfyui+manager

[quasi remnant]

Comfy is incredibly faster, espec since it has python 3.11 cross-attention support

which is even faster than Xformers

[bold osprey]

my setup struggled with a1111 at low res and now I can run fullhd in the same time

[fast vector]

Oh, so it loads actually

[quasi remnant]

• it does cuda-malloc stuff

which makes it like 30% faster again

[hushed hornet]

The reason I'm unsure is because I haven't touched A1111 since SDXL0.9 came out

[soft moth]

I never got the cuda malloc stuff

What does that do?

[quasi remnant]

nvidia's recent drivers introduced a kind of... bug?

cuda-malloc stuff fixes that and puts it back to old performance

[dusty turtle]

man this multiple text encoder stuff is really begging to be messed around with, as far as I can tell XL hates <768 res for more complex prompts

[soft moth]

Ahh that's spicy

[quasi remnant]

so like 30% faster less VRAM usage, etc

[cunning geyser]

bro

[quasi remnant]

basically, comfy is just way better than a1111 in every way except maybe ease of use for inpainting/img2img xD

[cunning geyser]

whoever uses comfyai like its an entire hacking program holy moly

[cunning geyser]

[brazen patrol]

Does usisng Python 3.11 make a diff? I have 3.10

[bold osprey]

love it

[cunning geyser]

WHY DO YOU DO THIS TO YOURSELF?

[quasi remnant]

Yes.

HUGE diff

[cunning geyser]

WH YDO YOU NEED THESE MANY BOXES

[quasi remnant]

3.11 + you have to get the cu121 nightly pytorch build

[bold osprey]

I need more

[quasi remnant]

and you will get the new cross-attention

[cunning geyser]

A1111 WAS ALREADY COMPLICATED ENOUGH FOR ME

[dusty turtle]

idk its pretty simple if you know what the stuff does

[cunning geyser]

kaibse show me your setup

[dusty turtle]

just annoying to get started since everyone is starting with XL

[bold osprey]

with lora or without

[cunning geyser]

w

[soft moth]

You can try it on Artroom

[cunning geyser]

wwith

[cunning geyser]

with lora

[soft moth]

Same backend but really easy UI

[cunning geyser]

[brazen patrol]

Can you link this for me? The last. Python I know

[high skiff]

I love him

he is so happy

[orchid oar]

Me too

[quasi remnant]

i can dm you the command to put in terminal

[shy kelp]

I'm about to try a1111 1.5.1 anybody else tried it?

[high skiff]

dare I say the happiest

[brazen patrol]

that would be awesome

[orchid oar]

Yes🤩

[cinder aspen]

Someone tried the SwarmUI?

[brazen patrol]

I am really enjoying this new SD

[bold osprey]

the latest without though

[orchid oar]

Whatever floats your goat

[quasi remnant]

check message requests, i gotchu just need the answer to that 1 q

[soft zealot]

Im using a HRF/Upscale that I sourced from HRF modified from https://civitai.com/models/107144/sdxl-09-semi-official-workflow (and credited in my workflow for reference along with a nod to @high skiff for the bits I took from his original work

Cyan is Image Settings, RAG is the prompts, Purple is the OG Image, Blue is the HRF/Upscale

[bold osprey]

just noticed I dont have any sfw loras welp

[shy kelp]

[lilac wren]

Saw this on Civitai, apparently put a lot of steps paid on SDXL

[fast vector]

were do you save styles for comfyui?

[ember cargo]

my mini base + lora WF

[inner imp]

hi , a question where i can find sddxl for 1111?

[shy kelp]

i dont have any legal ones

[steady grove]

Is there controlnet out yet?

[shy kelp]

(its a joke relax)

[soft zealot]

I use WAS Nodes Suites Style Loader so :

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

[steady grove]

Bad joke

[shy kelp]

ur bad

[soft zealot]

same place you find SDXL for everything else 🙂

https://huggingface.co/stabilityai

[shy kelp]

i only have loras certified by allah himself

[inner imp]

thanks

[steady grove]

Recognizing that illegal NSFW stuff is a bad thing, does not mean I'm a devout religious person. Hyperboles.

Be less edgy. Talk to your mom more.

[shy kelp]

i already talk to your mom plenty

[orchid oar]

Tried to do 2048x2048

[eternal fog]

Did an interdimensional being sit on your Lamborghini?

[orchid oar]

Looks like it

[slender hinge]

Any ideas why my image is coming out like this with SDXL on automatic 1111

[tender birch]

is there a way to include Styles like in A1111 for ComfyUI? with a dropdown for quick selection?

[orchid oar]

My Lambo before that :(

[slender hinge]

seems to be working fine, until last final step

[eternal fog]

wrong vae

[kind pewter]

interested to know too

[lilac wren]

There is no exif data in your image, so impossible to know what you did

[eternal fog]

People are doing this a lot with the wrong vae, I wonder if they are using A1111 and just leaving the old vaes enabled

[slender hinge]

yes, that is what im doing whooops

[lilac raven]

This was my first experience

[dusty turtle]

people still gen with default vae lol

[sweet abyss]

wow good job with sdxl, im getting good results ❤️

[eternal fog]

I mean they've been using SD 1.5 models so leave that vae loaded and then try SDXL

[quasi remnant]

not using the right vae.

[rustic surge]

[quasi remnant]

https://www.huggingface.co/stabilityai/sdxl-vae/blob/main/sdxl_vae.safetensors get the vae here

[minor glacier]

Why is very one having issues with the default VAE ? Is it an app issue.
This is what I'm getting with the default VAE and no refiner from a 'simple' python script and Diffusers

[kind pewter]

damn, this looks fantastic

[orchid oar]

now im hungry

[eternal fog]

They aren't they are not changing the vae they have set already, which in most cases will be an SD 1.5 Vae, so it doesn't work.

[rustic surge]

I found SDXL to be very good at drawing food

[high skiff]

ok, 1.0 workflow with upscaler is approaching

[minor glacier]

Ah, right that explains it. thanks.

[rustic surge]

china food

[patent badger]

you use koyha? if so, (not trying to be annoying) u mind sending a screenshot of the settings? (I'm not super technical so I'm not sure what you mean by 20 repeats - my guess is the file name is "20_something") - did you do captioning / regularisation?

[orchid oar]

japanese ramen

[wet nacelle]

@shy kelpI got the nodes installed and now I'm facing this when I try to prompt.

[hybrid loom]

you're a fucking g for this. Thank you.

[shy kelp]

can you screenshot me the model area?

[wet nacelle]

SHould I put the base on top?

[astral jay]

Hi guys, sorry to interrupt, I've been dealing with this error while trying to train an sdxl lora on runpod, any ideas?

[shy kelp]

wait, ur not using sdxl?

[wet nacelle]

No it just autoed

[astral jay]

I've had success training loras for sdxl in my personal pc, just started with runpod

[wet nacelle]

I'm wondering if base should be on top and refiner on bottom

[shy kelp]

oh you need the base ontop and the refiner on bottom 😂

[orchid oar]

Someone hungry?

[wet nacelle]

Alright it's running now.

[bold osprey]

spaghetti porschonese

[hushed hornet]

Someone hungry?

[bold osprey]

ayo

[patent badger]

I got same error / something similar and fixed it by playing with the resolutions settings a bit. I had to give up on using different res / buckets but idk if that's the most suitable fix for you (of for that matter, if it will fix anything at all)

[bold osprey]

looks like my left nut

[candid pond]

What is the best Comfy workflow we have acces to right now?

[high skiff]

doctor

now

[candid pond]

Can someone share the config for it?

[wet nacelle]

@shy kelp Was it meant to be this way from the start?

[candid pond]

Oh, I've heard yours is the best one. 😄

[hushed hornet]

Isnt everyone expecting information about UFOs/aliens right now?

[bold osprey]

don't worry it is not haunting and destroying whole villages

[candid pond]

Is it private or do we have acces to it?

[wet nacelle]

the "run half steps" then "run other half steps with refiner" @shy kelp

[azure oxide]

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

[high skiff]

thats me :D

[candid pond]

Noice, thanks!

[patent badger]

you are famous now

[high skiff]

1.0 workflow release may be out sooner than expected :p

[wet nacelle]

@shy kelp Wait now I'm confused. I ran it and it only rendered the base.

[shy kelp]

yeah thats how its supposed to work according to stable foundation. thats why theres no support in auto1111 currently

[wet nacelle]

What do I do from here?

[shy kelp]

it wont show you the result from the base cause the base it just a noised image

[wet nacelle]

yeah but I don't have a refine output. I don't know what to do from here.

[azure oxide]

unless you use a preview image node of course

[shy kelp]

so what you are seeing here is the complete image. next it goes into some upscalers but you need to load some upscaler models.

[azure oxide]

use sytan's workflow as a general layout to understand, his is fairly basic and shows you the steps

[wet nacelle]

@shy kelp Should it had upscaled?

[shy kelp]

this is what i use. you can just skip all these if you dont wanna use em.

[wet nacelle]

I still frankly don't understand. If I had more than one upscaler would it have outputted the upscaled image?

[azure oxide]

only if you sneak a preview image node in the middle

[shy kelp]

i only use one upscaler. the others are face restore and skin detail models

[hushed hornet]

It's happening. Photo from South Jopsmit, Hasmin Islands

(I do not condone misinformation with AI imagery plz don't take this seriously)

[orchid oar]

OMG A UFO

Is that real?

[lyric hill]

I watched a video about SDXL, and he said you guys recommend comfyui or something, but he preferred A1111, is one of them faster or is it just a preference?

[polar epoch]

Judging by the left side, it's a UFIO Unidentified flying inflated object

[hushed hornet]

Yooo if you haven't seen it, watch the movie NOPE...

[polar epoch]

Aye, know what it's about, but i'm not a horror guy :P

[astral jay]

Hey guys, can anyone help me with this error? I'm training a lora on sdxl on runpod, I've already trained multiple ones in my pc but this is the first on runpod:

[hushed hornet]

Neither am I!!! It's more like Jurassic Park than horror...but he sure knows how to make a fabric alien thing look scary

[lyric hill]

are you using a spot instance?

[astral jay]

I thought it was due to the accelerate config, but I just added the config and still got this

[astral jay]

wdym?

[rotund cliff]

the face restore node looks daunting, any tutorial on it?

[orchid oar]

Can i post zombies in here? Or are they NSFW

[high skiff]

if they aren't too gorey

[orchid oar]

ok

[high skiff]

if they have gore, spoil them with a warning

[orchid oar]

how do i do that?

[astral jay]

Heres my pod

[shy kelp]

so what i have is, i save 2 versions of an image. a non upscale and an upscaled. the non upscale is for having png metadata to recreate the image, and the upscaled is just for better quality. but both images have gone through the face restore and 1x skin detail model before being saved.

[orchid oar]

||hi||

Ahhh

[polar epoch]

@orchid oarDm me image, i can judge

[lyric hill]

on runpod, if you rent a spot instance, it means if someone pays more, then it will turn off the machine (much more complicated, but that is what it is simply)

[orchid oar]

ah thx

[quasi notch]

hey guys do you know where i can find the exact buckets used for sdxl 1.0?

[lyric hill]

well, that is on demand, idk then, sorry

[wet nacelle]

Why is there a second image preview next to the first one? Would this not also be filled?

[lyric hill]

anyone know about this?

[wet nacelle]

I'd say pref

[orchid oar]

Blood warning or something like that idk

[lyric hill]

ok, i heard somewhere that the backend for comfyui is faster so idk

[wet nacelle]

It is

[hushed hornet]

Check, please!

[orchid oar]

Check what?

[polar epoch]

Oh, that's not bad at all lol. Just a bit of starved stanley ipkins

[shy kelp]

there is one raw straight out of the refiner, one for face restore+skin detailer (saves png) and one last for the upscaled version (saves jpg)

[hoary saddle]

best way to put TEXT into prompt? thoughts?

[orchid oar]

Yes, i wanted to post other ones, but i think they are too much. Like a drawing but much blood

[wet nacelle]

So the left side is for the image that is not yet upscaled correct? If this is true then how would I upscale it so that I can have the upscaled image?

[polar epoch]

Stanley ipkins for those who doesn't know the reference :P

[shy kelp]

you need to load the same upscaling models as i have for that

[polar epoch]

Indeed.

[shy kelp]

right now its not even saving the image

[wet nacelle]

Ah okay I see.

[shy kelp]

or whatever you want to use

[shy kelp]

if you dont wanna upscale you just plug the vae decoder straight into save image

[orchid geode]

Most of those are not even using the refiner, great great base model! im hooked

[polar epoch]

Man, can't wait to continue my SD adventure with nodes once windows is reinstalled. Don't wanna bother generating more as my windows install is busted lol

[hoary saddle]

yay!

[shy kelp]

if you do it properly you can make based chads like this guy

[polar epoch]

Based on what though?

[zealous horizon]

https://gist.github.com/pixelass/446c3cafde6bd81d4f2de01f8ce9b8ee There you go. (look for su8j3ct and paths)

[shy kelp]

absolutely nothing!

[vital wolf]

How soon? Like today soon?

[polar epoch]

Then don't use based lol. That's already a word for something It's like when kids says "cap" when "lie" exist :P

[high skiff]

not likely today, as I do have a lot more documentation to write, but within the next couple days if all goes well

[polar epoch]

Oh damn. Now with a spoon of realism and i'd love it! :D

[vital wolf]

Anyone using stableswarmui?

[hoary saddle]

image2image on A1111, can't wait for a good setup with vae/refiner/i2i on COMFY

[polar epoch]

Here we go!

Dragon looks a tiny bit off though

[mellow tendon]

They do exist.

[candid pond]

Damn, comfy works way better than Auto1111 with SDXL

[wheat moat]

why do I get blurry images like these?

[patent badger]

and you used the "20_something" naming format, correct?

[candid pond]

I hope Auto1111 gets updated with the better workflow

[hoary saddle]

find any links, lemme know, i can't build one that works lol

[pallid path]

are you upscaling by any chance?

[candid pond]

Or Comfy gets a proper inpainting support

[wheat moat]

nope

[hushed hornet]

Y'all ever have one of those dreams where

[pallid path]

then it could be the VAE

[lyric hill]

how do I run sdxl on colab? I used the notebook by camenduru (the A1111 one) but it crashed

[dusty turtle]

hmm I must be missing something when trying to use sdxl (base 1.0) as a refiner/img2img for 1.5 outputs since it just degrades the image lol

[shy kelp]

refiner dosent work for img2img. your results are blurry

[wheat moat]

just disable them?

[pallid path]

no, you need a previous one, hold on

are you using AUTO1111 or Comfy?

[wheat moat]

a1111

[pallid path]

oof, idk if it works then, but I'll give you a link

https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/tree/main

[hoary saddle]

that's very possible 🙂 just using base on a1111 atm

[pallid path]

get the .safetensor one

[wheat moat]

thanks

[pallid path]

and apply the VAE and see if it works

[wheat moat]

ok

[shy kelp]

ur better off just using base model for img2img, or a sd1.5 model

refiner is most likely never going to be useful for img2img

Hi guys, how much second it takes for you to generate a image? For me it is 18 seconds on rtx 360

[candid pond]

About 15ish with 3060 12gb and 32gb system ram

[shy kelp]

1.6 it/s

[candid pond]

1.9 it/s here

[shy kelp]

Did you tried to optimize

[candid pond]

Oh but I'm using euler ancient right now

[shy kelp]

depends how many steps 😂

[nimble heart]

shameless plug, my Principled node does i2i with either refiner, base, or both if you just hook an image into it
https://github.com/Beinsezii/bsz-cui-extras

[shy kelp]

And make it under 10 secs?

[shy kelp]

https://www.youtube.com/watch?v=JluH7Wc-EyA&ab_channel=ImpactFrames

[dusty turtle]

2.5 s/it

[shy kelp]

I am watching this

[candid pond]

Nope, Im fine with 15 secs on 720x1280

[mellow tendon]

https://files.catbox.moe/vdjtgr.json this is one I have built, the img2img upscale is a bit inefficient but it works.

[candid pond]

Ooh, but I'm using xformers yes

[shy kelp]

I want 12/s

[hoary saddle]

thank you! will give it a try

[hushed hornet]

I love da soulcard

[hoary saddle]

ty!

[bold osprey]

yes first time I tried it out

[candid pond]

How do I download this? 😄

[hoary saddle]

i opened it in browser, then right click save as

[candid pond]

Oi

Makes sense

I was trying to use stringify in python lmao

[solemn scroll]

Hi, just installed comfy to use it with SDXL, the extra models file to SD webui. But will it detect the default SDXL vae? Also, how can I use my SD styles in comfy, if at all?

[orchid geode]

and then its just a script? how do I use it 😄

[hoary saddle]

drag it onto your comfyui screen

[safe spear]

What notebook should I use for kohya-ss? I'm using right now the one linked on the repo but I have no idea how to add SDXL there, and I don't see it on the list of model quick pick

[orchid geode]

Oh my bad I thought its an automatic1111 script for some reason

[shy kelp]

It takes 4 second with 4080 for a image

[mellow tendon]

Its a ComfyUI workflow file, have you go ComfyUI Installed?, you can drag it or click load

[orchid geode]

Yes il try it! thanks

[soft zealot]

name removed to protect the innocent.

Pleas eplease cn we have this information as a sticky !!!!

TIS

:o)

[hoary saddle]

got the upscale node working, can't find the save as webp node

[dusty turtle]

the hybrid workflow is kinda interesting

[solemn scroll]

I can just drag and drop my styles.csv? this is wonderful

[dusty turtle]

xl->1.5 model

[hoary saddle]

not sure, i'm a noob with this stuff

[eternal fog]

@high skiff I think I might have got this upscaling workflow almost working well. Left is before, right is after.

Needs a little more work on getting the eyes right

[soft zealot]

No

You can dragndrop images containing workflow, styles csv you need to look at the insturctions for the custom node you are using

[mellow tendon]

Sorry you will probably have to replace the save_as_webp node with a standard image\Save Image node

[shy kelp]

I will test xl speed with a100

[candid pond]

SyntaxError: Unexpected non-whitespace character after JSON at position 4

[candid pond]

I'm getting this error using your workflow.

[mellow tendon]

Why does the scratch get less detailed?

[eternal fog]

Because getting a img2img highres fix to work is a pain in the ass.

The refiner adds detail to the eyes and removes it from everywhere else.

So I'm trying to use the refiner to detail the eyes and then the base to put detail back in, without fucking up the eyes again.

[orchid oar]

Crowded Shibuya Crossing (not good img)

[vocal cosmos]

what are the steps for installing sdxl 1.0?

[mellow tendon]

I like it.

[orchid oar]

Its a funny picture

[sharp patrol]

The eyes and fingers issue 😦

[solemn scroll]

What kind of dark magic is Comfy performing ? I went from 5 minutes per image to 20s

I'm not joking

[orchid oar]

Yes me too

Its awesome

[soft zealot]

download the models from Huggingface and put them in the relevant models folder of your UI of choice

https://huggingface.co/stabilityai

[dusty turtle]

hmm never thought I would have a use for ~.40 denoise but the effect is kinda nice on XL outputs to make them more anime like

[mellow tendon]

Sorry I don't know, I have a lot of custom nodes installed, but didn't think I was using many there, maybe you are missing ControlNet? Maybe remove stuff from the end until it starts to work!

[wet nacelle]

I have installed all of the files that you have and still it is not upscaling the image.
I also can not find the image within my comfy folder even after changing the path that was already set up.

[shy kelp]

did you load em into the loaders?

[wet nacelle]

@shy kelp

@shy kelpI thought that I had.

[shy kelp]

so it should give you some kind of error if its not working

[orchid oar]

how? Mine doesnt get better that this

[astral jay]

Is anyone getting errors with accelerate on runpod? I wanna train an SDXL lora 😦

[dusty turtle]

cursed workflows

XL into 1.5 model

[obtuse slate]

in comfyui how do you organize the flow nodes to be a specific shape? Do you need an extension like in blender?

[obtuse slate]

Like this

[alpine pine]

the nodes or the noodles?

[smoky bough]

nvm organize

[wet nacelle]

Do you know how I might solve this? I have them all installed.

[obtuse slate]

The noodles

[alpine pine]

I, too, am curious. Flat noodles would make everything so much cleaner

[obtuse slate]

Blender has this too. There are some extensions that give you more ways to organize.

[plain flame]

When i try to generate images in clipdrop in mobile browser the file size is coming out around 1.5 mb where in the pc browser its coming out in 200kb. Why is this happening with the change of device thats being used to generate?

[obtuse slate]

I just want my workflow to look sexy 😔

[dusty turtle]

honestly even after only A day using comfy the spaghetti's ways make some cursed sense

must be the factorio brain cells activating when I see lines going everywhere

[sharp patrol]

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

[obtuse slate]

Damn it does it automatically

Kinda sad, blender let's you manipulate the shape of the noodles manually.

This is still cool though

[shy kelp]

does it say anything before this?

did you get codeformer from here?

https://github.com/sczhou/CodeFormer/releases/

[wet nacelle]

yup

[shy kelp]

did you put it in models/facerestore_models?

[wet nacelle]

That's all it says.

[plain flame]

When i try to generate images in clipdrop in mobile browser the file size is coming out around 1.5 mb where in the pc browser its coming out in 200kb. Why is this happening with the change of device thats being used to generate? Any idea guys?

[orchid oar]

Thx

[plain flame]

sdxl 0.9 seems better to me than the new one

[wet nacelle]

How would I replicate this if it isn't in the upscaler folder?

[plain flame]

im using the same prompts that i used for 0.9 in the latest 1.0 , the images looks bad and low in quality. speaking about clipdrop

[shy kelp]

just move codeformer into facerestore_models and reload it

[primal vault]

The latest update to my workflows and custom node extension for ComfyUI is now released on https://civitai.com/models/111463?modelVersionId=127446
The brand new Searge SDXL Reborn workflow requires the latest update of the custom node extension to run, but it has a cleaner interface and improved prompting. If that's too complex for your needs there are also the simple and advanced workflows. Everything tested with SDXL 1.0 of course.

[wet nacelle]

yeah. It isn't there. @shy kelp

[shy kelp]

it dosent work cause it thinks its loading an upscaler when its actually loading a facerestore model

[shy kelp]

click refresh

if you put it in the right folder

[wet nacelle]

It will not change.

@shy kelp

[shy kelp]

oh lol i think mine is located in \ComfyUI\models\upscale_models

[trail obsidian]

where will this be posted if you are done? 🙂

[shy kelp]

so move it there and it should appear

[high skiff]

It will be released on my github, and linked here

And if it still works out right with @visual glade , it will be on the wiki

[wet nacelle]

That's what I told you the first time.

I have moved it back

[obtuse slate]

Does SDXL base model works for inpainting? I've heard you don't need an inpainting model for it.

[shy kelp]

yeah well i thought you had already loaded it

[wet nacelle]

ok

[shy kelp]

move it back, click refresh and it should appear

[trail obsidian]

awesome ty! using a simple setup that works reall well but trying to get high res fix and some other things working...

[solemn scroll]

Do you have any idea why I get the error "Error occurred when executing KSampler: mat1 and mat2 shapes cannot be multiplied (154x2048 and 1280x768)" with refiner? Here's my workflow:

[zealous horizon]

how do you get the cables to be straight?

[shy kelp]

what do you mean? it isnt on the list?

[obtuse slate]

Someone posted the extension already

the one with 3 reactions

[wet nacelle]

I've done that and tried to run it. It gave back this

[wet nacelle]

@shy kelp

[dusty turtle]

I think it needs a seperate clip for the refiner

[shy kelp]

idk man cant help you further. it worked right out the gate for me. maybe you can try following the steps from the install https://civitai.com/models/24690

[inner ruin]

which scheduler should I use to make the images pop more?

[soft zealot]

another member of the group the nodes tightly together to hide the lines club. I commend you sir 🙂

[soft zealot]

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

[south igloo]

High res fix works in comfy UI????

[primal vault]

Yeah, I just started doing that after I realized how much time I spent zooming and scrolling in my older workflows

[soft zealot]

so much easier to daily drive

[primal vault]

It's also placed so that whenever I load or reload, it centers on the main UI elements

[viral narwhal]

i have a good feeling SDXL is a huge step forward!

[soft zealot]

I really need to that with mine, it defaults to the "under the hood" part of the flow

[gloomy barn]

anyone know in sdnext how to use a vae with SDXL 1? Noine is mentioned in the output. The images aren't crisp.

[south igloo]

Is there an idiots guide for comfy UI? I just want to use basic features like auto1111 facerestore an high-res fix and how to apply Lora's etc

[shy kelp]

Does it play nicer with AMD cards?

[dusty turtle]

if nice is a relative term yes

[shy kelp]

Right... so no 😦

Sad.

[dusty turtle]

I mean it works its just very, very slow on windows

768->1024 on comfyui with the refiner, probably needs a bit of tweaking

[pure hazel]

Are we able to use SDXL with controlnet or do they need to update the CTRLNET models? Also any idea on how to select speciic VAE's with comfy?

[static prawn]

where do i finde the custom node extension?

[zealous horizon]

loaders > load vae
put VAE in models/vae

[static prawn]

is the custom node extension the same as comfy ui manager?

[gloomy barn]

can the output dir be changed in comfy?>

[civic sigil]

We need to wait for someone to train controlnet models for it sadly

[vast narwhal]

Thanks for sharing, can you tell from where this node is?

[zealous horizon]

there are several custom node repositories. They can be cloned (or downloaded) in the custom_nodes folder of ComfyUI

[high skiff]

new 1.0 version should be shipping with a 2048x high res fix workflow built in

cropped in 1024x gen vs cropped in 2048x high res fix

[static prawn]

but im searching for that extension

[high skiff]

SS of the full image

[civic sigil]

Nice demo image

[dusty turtle]

hmm seems to be following the prompt but in a weird way

[soft zealot]

Looks like the Seed With Text Node.

See Credits & Notes Top LH corner for instructions on what I use

[static prawn]

never mind i understand it now , sorry 😄

[autumn forum]

are you using sdxl?

[nimble heart]

made some changes just now to make img2img a bit easier when scaling. Example of loading an image using the Principled node

[dusty turtle]

base+refiner ye

[autumn forum]

those dont look right lol. i wouldnt do anything lower than 1024. also the refiner sampler may be wrong.

[nimble heart]

base+refiner should use Euler

maybe DDIM

everything else makes hella artifacts

[dusty turtle]

its probably the deepdream prompt if you see the noise patterns everywhere lol

[autumn forum]

i use 2m and works great

[nimble heart]

karras?

[autumn forum]

yeah dpmpp_2m, karras schedular

[rigid prawn]

can you also send it to me please? I can't find it

[autumn forum]

wanna send me the prompt and see what i get just for giggles?

[vast narwhal]

It was there, ty

[quasi remnant]

check message requests

[nimble heart]

same settings, euler normal on left dpmpp_2m karras on right.
Refines into chunky soup

if you're just img2img-ing the refiner it might work on dpm karras ig

[quasi remnant]

recommend dpm2++ 2m normal rather than karras

[autumn forum]

i guess it depends on prompt then. i like the look on the right. and i dont do i2i

[quasi remnant]

karras is good for more anime-style images

but in general normal scheduler is better

[autumn forum]

2m karras got me this

[nimble heart]

actually open the images in your browser to see them full size. the right is just all artifacts and distortion

[quasi remnant]

it is nice, but karras smooths stuff out a bit too much

[static prawn]

i always get this starting comfy after adding custom nodes :/

[autumn forum]

beinsezii was saything the opposite of what you just said lmao

[quasi remnant]

well in general. it depends on the seed as always lol

[autumn forum]

2m karras

[clever verge]

Is that node as running on the image committed to github? Curious how you got randomize working.

[nimble heart]

yea

randomize always works

[azure oxide]

ya SDE karras is king for anime but when it comes to just the base models for both 1.5 and sdxl, ive found that euler a is way better

havent tested out the other ones

[nimble heart]

unless you mean on your own node. idk it just shows up if the max value is high enough I think