✨｜sdxl

[vital ermine]

portrait art studio portrait = realistic

studio portrait and just portrait

[hardy cipher]

frog

really settled into that look there

[ionic gulch]

without ip-adapter, just pure text prompts with movie styler v2:

[hardy cipher]

lol

pretty accurate tbh

[hardy cipher]

ooh, that reminds me, I don't know if it was just because I was running it online or what, but something in your nodes goofed and was causing those sorts of errors that prevent nodes from loading. I should try redownloadingmaybe I tried to get it when you were in the middle of updating

[ionic gulch]

maybe try to remove and readd them through the manager. a lot of nodes changed with the 10.0 workflow.

[hardy cipher]

not gonna lie. once I get it working I plan on butchering your workflow a bit and mutating things

[ionic gulch]

you have to beat the checksum protection first :)

[hardy cipher]

pshh

[ionic gulch]

just joking

[wide crane]

@ionic gulch what is the node that shows the image as its being generated

[hardy cipher]

I have some ideas rolling around in my head though, and want to start looking at model data more. just to see what I can mess up

[ionic gulch]

there are multiple ways to do that. as a node or as extension to comfyui gui. as i don't use any of them, i don't know in more detail which solutions are good.

[hardy cipher]

should work out well

[heady vale]

monster incoming

[glad grove]

smallest dress in england

[hardy cipher]

warmup image

[wide crane]

@ionic gulch its extremely simple but here is the base thing i got

[vital ermine]

any able to get a lora, ipadapter, and refiner to work in comfy?

I can get two of those but not all three

[hardy cipher]

hmm, let me add some loras to see if they work

[vital ermine]

Love to know where they have to go in the chain if it does work cause I tried everywhere

couldn't get past a recursion error

[heady vale]

I added it between model loader and ipadapter loader. no problems. just need a low strength of 0.4 on the IPA

[vital ermine]

did that and no dice

[ionic gulch]

yes, but only if i use refiner without those addons. afaik the refiner model can't be modified by lora for example, as the loras are made for the base model nor for the refiner. this results in the refiner removing some of the lora effect of course. so maybe only use a 90/10 split

[vital ermine]

I can, and do, use base+refiner+lora but not if I add the ipadapter but remove the lora and it all works

again

[heady vale]

I was using comfyroll lora loader, for reference

[vital ermine]

I am just not feeling the refiner since I can change it to anything and get the same results

I tried CR and norm loader

[hardy cipher]

what I'll do when adding something new is just start it at close to 0

[vital ermine]

it should go right after the checkpoint loader and from the lora onward to the rest but recursion error

I rebuilt it from the ground up twice then said to hell with it

[ionic gulch]

depends on the style of the image. some styles get better with refiner, some get worse. i always try a few images with both options turned on in my workflow and after i know if refiner helps i turn off the base only or the refiner option to speed up image generation.

[hardy cipher]

I have all the things

[ionic gulch]

doesn't really help that a lot of extra features are not really tuned for the refiner or even not compatible. that makes the refiner even less useful if you do anything more than a simple 80/20 split txt2img workflow without extras.

[hardy cipher]

this is true

I have no idea how it's supposed to work with these other things. total third wheel

[vital ermine]

yeah, above about 20 the picture degrades fast

[hardy cipher]

makes it smooth skinned

[hardy cipher]

mighty fine hand ya got there

[ionic gulch]

clockwork orange

[wide crane]

ComfyUI
Prompt: 1girl, head tilt, green eyes, smile, looking at viewer, close up, face, blush
Negative Prompt: N/A
Model: pastelMixStylizedAnimeprunedFP16
VAE: vaeFtMse840000Ema_v100
Sampler: Euler A
CFG: 7.0
Steps: 25
Seed: 302725635828231
Width x Height: 512 x 512
Output:

Automatic 1111
Prompt: 1girl, head tilt, green eyes, smile, looking at viewer, close up, face, blush
Negative Prompt: N/A
Model: pastelMixStylizedAnimeprunedFP16
VAE: vaeFtMse840000Ema_v100
Sampler: Euler A
CFG: 7.0
Steps: 25
Seed: 302725635828231
Width x Height: 512 x 512
Output:

ComfyUI
Prompt: 1girl, head tilt, green eyes, smile, looking at viewer, close up, face, blush
Negative Prompt: N/A
Model: pastelMixStylizedAnimeprunedFP16
VAE: vaeFtMse840000Ema_v100
Sampler: Euler A
CFG: 7.0
Steps: 25
Seed: 302725635828231
Width x Height: 1024 x 1024
Output:

Automatic 1111
Prompt: 1girl, head tilt, green eyes, smile, looking at viewer, close up, face, blush
Negative Prompt: N/A
Model: pastelMixStylizedAnimeprunedFP16
VAE: vaeFtMse840000Ema_v100
Sampler: Euler A
CFG: 7.0
Steps: 25
Seed: 302725635828231
Width x Height: 1024 x 1024
Output:

Automatic 1111 time taken on 512x512 image: ~12 Seconds
Automatic 1111 time taken on 1024x1024 image: ~22 Seconds

[wide crane]

ComfyUI time taken on 512x512 image: 3.36 Seconds
ComfyUI time taken on 1024x1024 image: 13.82 Seconds

Conclusion: ComfyUI is considerably faster with equal to higher quality image generation compared to Automatic 1111, at the price of being much less user friendly off the bat and requiring a significant learning curve to use effectively compared to Automatic 1111
Side note (ComfyUI is extremely faster and significantly more capable of running SXDL models and producing high quality images)

[wet nacelle]

@hardy cipher How would I solve this?

[wet nacelle]

Fooocus?

[wet nacelle]

https://tenor.com/view/thats-right-correct-yes-right-yup-gif-16435470

[wide crane]

bruh

this is with an anime model

[shy kelp]

i cannot use control net, any solution

[hardy cipher]

hmm

one minute

[wet nacelle]

Maybe this? https://huggingface.co/h94/IP-Adapter/tree/main/sdxl_models

Trying to find the models.

[hardy cipher]

well firstly you do need those, yes. not sure which one you'll need to use. mine are named differently. but as for the other error, let me look into it further. might have forgot to put something in the requirements file

[wet nacelle]

I don't even know where to put these files yet.

[crisp owl]

Believe it's this one for the ipadapter node
https://huggingface.co/h94/IP-Adapter/blob/main/models/image_encoder/model.safetensors

On mobile so don't have my links handy

[wet nacelle]

I think I have the node?

[crisp owl]

Yah for that node

Install that file into the custom_nodes folder for the ipadapter

[wet nacelle]

I have the models but don't yet understand where to put them.

[crisp owl]

Should be a models subfolder or something

[wet nacelle]

I have no clue which.

[crisp owl]

In the custom_nodes

[hardy cipher]

have you successfully loaded an image into one of those yet?

I haven't had that problem, and haven't heard of anyone else having issues, not to say they haven't

[wet nacelle]

As in this?

[hardy cipher]

well then

there you go

[hoary saddle]

where can i find a workflow for ipadapter?

[wet nacelle]

Yeah but the error is still present.

[crisp owl]

#✨｜sdxl message

[crisp owl]

If my memory serves right, the file path is comfyui/custom_nodes/ipadapter (or whatever that pack is called)

[wet nacelle]

@hardy cipher alright new error.

[wet nacelle]

I was being dumb actually.

I got that part worked out just now.

[crisp owl]

Nice

[wet nacelle]

Now I am running into an error I wouldn't know how to fix.

[crisp owl]

No clue on the Pop node issue, I don't use those

[wet nacelle]

@hardy cipher If you manage to find the issue and find a way to fix it please let me know. I'm gonna make some drowning art in photoshop.

[hoary saddle]

pretty cool stuff

[shy kelp]

if jerma, you're bound to run into errors

[wet nacelle]

bitch please

[shy kelp]

[hardy cipher]

wait, so are some of them loading and some aren't?

[wet nacelle]

brother!

The issue seems to be that the image being read by the canny is not "ANTIALIAS"

[hardy cipher]

give me a minute. I'll figure out what's going on

[wet nacelle]

wicked

[dense chasm]

sdxl T2I adapter then input in pika image2video is cool

[heady vale]

new fashion earrings

[uncut fiber]

not sure you solved it, when unrolled it upscaling, when not it do not upscaling (Hiresfix)

[hasty smelt]

Hi guys! My 16 GB RAM system prevented me from continuing to study SDXL and Comfy UI. I am determined to learn everything about SDXL and, therefore, I just bought 64 GB of RAM, 2x32. I hope this is enough! hahaha

[crisp owl]

64 GB ram is nice to have

[shy kelp]

unexpected

[fierce compass]

I'm trying to do an XY plot over the CFG with ComfyUI, SDXL+Refiner and Efficiency Sampler Advanced. I only want to control the CFG of the first stage sampler, not for the refiner. I can't get the Refiner sampler to properly integrate with the CFG script. Any pointers?

the problem is that the refiner only gets the first image sampled in the previous script

[shy kelp]

https://tenor.com/view/patrick-bateman-why-not-you-stupid-bastard-gif-23978200

[wide crane]

yo can someone help me

im tryna upscale my image to 2x

but i have no damn idea

how to do taht

[crisp owl]

send to ksampler, and output as usual

[wide crane]

this is the model i have

rn

i mean

workflow

[crisp owl]

take the last vae decode image output, put to the upscale image node, follow those in the image I screenshot, send from the encode to another ksampler with a minimal denoising, .15-.35ish

[crisp owl]

you can just do a rough upscale, but that won't give you best results

[nimble heart]

anyone know if the comfyui caching is smart enough to know which specific values are changed for children nodes?
so take something like

A -> B ⤵
⤷ ---> C

if the output of A changes in such a way that it only affects C, will B still be re-rendered?

[wide crane]

okay 1 sec

[crisp owl]

can also use a 4x model and use an upscale image by node with a value of 0.5 to bring that down to 2x, but that requires more processing time.
I've just started using a 2x model to begin with

I usually go with about 10 steps for the upscale

[wide crane]

what is upscale image by?

i dont have that

[crisp owl]

It's hard to search to find it, search "by"

[wide crane]

ah

[crisp owl]

second on the list

imagescaleby

[wide crane]

i see

[uncut fiber]

i think it is i think usual method, bicubic bilinear....

This you can make in batch in any program.

[wide crane]

does the scale by thing matter

[crisp owl]

see this comment#✨｜sdxl message

[static wigeon]

How can I add weight to prompts? Is it like (Red Hair : 1.0) or some other way?

[static wigeon]

Ahh thank youu. What is the maximum weightage?

[wide crane]

this is what i got

[crisp owl]

I don't usually go beyond 1.5 personally

[uncut fiber]

@crisp owl tried chaiNNer, and it has very super advanced resize algos. It is basicaly father of Comfui.

[static wigeon]

I see, will try it now! Thank you! So by default, each prompt will have a weightage of 1 right?

[crisp owl]

Looks good, just gotta decode that final image output

[uncut fiber]

For signatures @vital ermine went up to 2,7 to not see them

[wide crane]

ye ye

[crisp owl]

I'll have to check into that, I've been working on getting my main workflows in place first, was gonna look into upscaling options later. Thanks for the heads up

[uncut fiber]

np 🙂

[wide crane]

@crisp owl thank you man

i never woulda figured this out

how do i rename these buggers

[crisp owl]

right click the little blue dot

[wide crane]

thanks man

alright cool

@crisp owl what ksampler settings do you use?

this is my goto

if it would send

damn

[uncut fiber]

if it has not blue dot, i believe you can do this rmb on it anywhere and change Title

[crisp owl]

Generally the same, but I use the advanced ksampler nodes, so more playing with the steps and such.
But same sampler_name and scheduler generally. I occasionally switch it up for fun. Steps though, I'm usually down in the 30-40's

[wide crane]

if i go to 30-40 steps my output looks fucked

lemme show

[crisp owl]

cfg varies based on what I'm prompting

[wide crane]

or not

what the hell

ive been tricked

curse you comfyui

[uncut fiber]

didnt you mistake it with CFG?

[crisp owl]

https://tenor.com/view/a-night-at-the-roxbury-chazz-palminteri-i-know-your-tricks-benny-zadir-gif-21898451

[wide crane]

nah

[static wigeon]

Is there a LORA or an extension that can fix the abnormalities in The anatomy of some people in my generations?

[wide crane]

loras usually cant fix that

most of the time i have to run it through inpaint

[static wigeon]

Ohh thenn? Negative Prompts

[wide crane]

in theory but they also change the image

sometimes really heavy

so i dont use them

in fact i dont use any negative prompts

[static wigeon]

Is it not possible without inpaint? I want to automate the process so post-generation stuff apart from ADetailer doesn't work for me

[crisp owl]

I use like three words in my negatives, low quality, dull, boring. No reason why, I just do lol

[wide crane]

i dont think so

i have to inpaint

and even for me

adetailer is unreliable

[static wigeon]

I see

[wide crane]

a lot of the times ill straight up load the messed up image into photoshop

make rough edits

[static wigeon]

It was generating good images for me but when I started using Weightage, they kinda become cursed

[wide crane]

and run it to img2img

[static wigeon]

😭

Like this

I see

[wide crane]

yeahhh

trying to fine tune is really cancer

time consuming

but if you really want the output to look good then you have to

😭

[crisp owl]

Are you generating 512x512 for SDXL?

[static wigeon]

These are my previous generations without weightage

But it really ignores the main idea

[static wigeon]

512x512 with a custom model called Comic something, I dont remember the name

[crisp owl]

SDXL doesn't like that, you'll get weird stuff happening.

1024x1024 or variants of that

[static wigeon]

Oh so 1024 is recommended

[static wigeon]

Okayy lemme try

[crisp owl]

1MP is what those all calculate to, which is what SDXL was trained on

[static wigeon]

512x512 takes like 7-8 minutes

[crisp owl]

A1111?

[static wigeon]

So i guess 1024x1024 might take double of that time 😭

Yeah

[crisp owl]

yeah....

I don't use A1111 anymore

[uncut fiber]

4x

[static wigeon]

I am not comfortable with comfyUI

[crisp owl]

I have a 2060S card, I generate a 1024x1024 image in like 25 seconds

[static wigeon]

Oh damn

I have 1650

With 4 GB VRAM

[uncut fiber]

i am playing with storyboards 🙂 , it is funny specific movie.

[wide crane]

@crisp owl do you know of any good anime upscalers?

[static wigeon]

but I will prolly host the WebUI in a VM with 3090 or something

[crisp owl]

Not really, I stick with realism mostly

[wide crane]

damn

[uncut fiber]

https://openmodeldb.info/
@wide crane here you can find yourself

[wide crane]

finally getting comfortable with comfyui

[crisp owl]

last good one I made before I started breaking stuff in my flow trying to get it working properly

[wide crane]

this looks awesome

[crisp owl]

realistic take on Shoto

[wide crane]

i guess its time to fuck with upscalers until i get something good

😢

[uncut fiber]

Rambo II. 🙂

[static wigeon]

Does Comfy UI require more Resources or is it lesser resource hungry when compared to WEBUI. If so, I might try to learn it and convince my boss to shift to that instead

[crisp owl]

less

more bang for your buck

[static wigeon]

Oh then that's cool

[wide crane]

bro i was using webui for a while

but comfy

is so much faster

[uncut fiber]

@wide crane page i sent you there in some cases you can see differences. And you can use keyword Anime to have less to choose from

[wide crane]

like insanely faster

[static wigeon]

Do you use Storyboard as a prompt for that format?

[wide crane]

like it absolutely smokes webui

[wide crane]

thanks mna

[uncut fiber]

@static wigeon yes

[static wigeon]

I see, will maybe shift to that. I hope that has an API

[static wigeon]

If you don't mind, can I get the prompts you use for that format?

I am trying for a similar output but mine comes like Comic Panel network

[static wigeon]

Like this

[uncut fiber]

Rambo II. movie (Storyboard:1.4) layout. Quick pen stroke only sketch. With outlines with black micro pen. Seen from top. Speech bubble.

[static wigeon]

Okie thank youu, will try it now

[uncut fiber]

bubbles spam all image with speech bubles

[wide crane]

@static wigeon the only problem i can see with comfy ui is that there are less extensions supporting it however im sure as it gets more popular people will make more extensions

[static wigeon]

I see, the only extension I am using is ADetailer. I hope this is supported there

[wide crane]

it is

[uncut fiber]

for 3xxx and 4xxx there is priceless thing called AITemplate

[static wigeon]

I don't require speech bubbles snice it's impossible for me to get another AI or software can fill in the speech bubbles for me

[uncut fiber]

yes

[static wigeon]

masterpiece, best quality, lora:style05:1, multiple views, comic, line art, comic network

This is my default prompt for the comic style

The Lora is chibi comic style (https://civitai.com/models/123374/chibi-comic-style-q)

[wide crane]

@uncut fiber thanks for the upscaler website but i think imma just use ultraupscale 4x

it looks to just have good results

[uncut fiber]

yes, can be handy other day 🙂
You can find already 2x upscalers there as well, or not upscalers but enhancing photography (jpg) and other things like denoise 🙂

[static wigeon]

Right, the 1024x1024 version is generating a very suspicious output

I interrupted it 😭

Not NSFW but still looks very eh

This is not the fully generated image

[wide crane]

phasing in and out of reality

[static wigeon]

Ikr

[glad grove]

reduce weight of lora to something like 0.6

[static wigeon]

WHY ISNT IT INTERRUPTING

Yeahh I will try that

[wide crane]

it never does

[static wigeon]

It usually does for me, but now it is weird

[wide crane]

its never worked for me

[uncut fiber]

to upscale, what do you think about this?

[wide crane]

thats good

[crisp owl]

phew that's an 8x model

[uncut fiber]

yes on icon it is quickly very, on large image it take some time using tta workflow

[soft zealot]

ComfyUI Impact Pack has "Switch Any" & "Inverted Switch Any"

The former is one to many outputs, the latter is many inputs to one output

ANd provding you configure thing correctly it is possible tocontrol multiple switches with onne Input

[static wigeon]

Shiny Tortoise

I am now hosting Webui in Google Colab, let's see how fast it is

[static wigeon]

Damn it takes 5-6 seconds for 1024 x 1024

[shy kelp]

ill give you $5 for the turtle

[floral island]

lets see if i can make it more expensive then 😮

[uncut fiber]

They walks so quietly 😄
Man watch your back 🙂

[shy kelp]

$1 now

[wide crane]

man

i cannot get

loras to work

[uncut fiber]

are they for SDXL?

[wide crane]

nah im not doing sdxl rn

[uncut fiber]

o.k.

[wide crane]

@uncut fiber help

when i do lora

am i doing this right

i take the model

put it in the lora

take the clip and also lora

and then put the model into the ksampler and the lora into the text prompt

?

[uncut fiber]

i used lora only once in A1111 it was easy.
how about clip output to cliptextencode?
I must be wrong

ou you probably have it.

[wide crane]

like this

start at checkpoint

link thge clip and model to the lora

link the clip from lora to the positive prompt and model

should be it?

[uncut fiber]

have you not must properly mentioned it in clip text?

[wide crane]

?

what you mean sir

[uncut fiber]

something like Lorafor_spacesuit:1,3 ?

[wide crane]

uhhh

i doint think so

the lora module has a strength

[uncut fiber]

in a1111 it was must
Yes it has strength, but it may in some cases dont know where it belongs. For example you want woman wearing guardian suit, not anybody else, so you put it where you describe woman?

[wide crane]

yea probably

[uncut fiber]

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

in a1111 it works differently

[shy kelp]

i find doing something like this helps (not comfy):
something, something, woman wearing lora:something:0.9 guardian suit, something, something
ofc, () and [] as needed

[wide crane]

god

comfyui destroys my skull

[uncut fiber]

@shy kelp thank you. I am not using it, and cant recall clearly. Once i probably will have to use it. Not using it loosing big part of AI generating

[delicate grotto]

anyone used the "wuerstchen" model yet?

[crisp owl]

toss this image into your comfyui for connected lora

then switch back to your models and such

[nimble heart]

4k attempt #589

[analog roost]

Nice! Not perfect but very close to the stock image in atmosphere/meaning!

[wide crane]

its gotten

soi much more complicated

[shy kelp]

[wide crane]

this is why i like automatic1111 so much

but its soo slow

😭

[nimble heart]

its only as complicated as you make it

[crisp owl]

ALl part of the fun 🙂

[shy kelp]

try this in your webui-user.bat file, it makes it not take a million years to produce a single image lol

[wide crane]

i did that

nothing changed

[shy kelp]

hmm

idk how fast comfy is, is it that fast?

[wide crane]

over 2x faster

not even jokng

[nimble heart]

optimize your nodes so its actually visible on one screen

[shy kelp]

hmmm

[nimble heart]

30% faster for me personally ever since the upcasting changes when XL released

[wide crane]

damn

automatic 1111 inpainting is better by a landslide

[wide crane]

15 images on comfy ui takes me that amount of time

when i say its better by a landslide

there is no competition

[nimble heart]

XL @ 1024² a1111 was about 2.55 it/s
comfy about 3.17 it/s

[shy kelp]

am i gonna have to get comfyui to see and play this visual novel to understand how it works?

[nimble heart]

3090 boi over here

assuming that 5.66 is 1024

[wide crane]

holy shit comfyui is

brain aids

[shy kelp]

hmm

[wide crane]

why is this happening

random ass distortion

coming from

god knows where

now the face is fucke dup

gaaaaaaaaaaah

[nimble heart]

skill issue

[wide crane]

im about toi fucking

[shy kelp]

[wide crane]

skill issue off a bridge

im in hell

of my own making

[shy kelp]

do the positions of this save onto next load?

[wide crane]

satan has his clutches around my soul

pulling me

to the depths

[nimble heart]

if you save the workflow as a .json or import it from an image yes

[shy kelp]

ty

[nimble heart]

fyi images in if uploaded directly to this discord or elsewhere will also have embedded workflows

so if you wanna see what other people are doing you can just open the image in your browser and click+drag

[wide crane]

look at my silly workflor

😛

[shy kelp]

if i wanted to add more than one lora, do i just make another lora node before plugging it all into the sampler? im assuming so

[nimble heart]

yea

[shy kelp]

i wonder if my pc will blow up if i shoved a 1.5 lora into SDXL stuff

[wide crane]

it will

anyways

dont use loras

im getting

terminal cancer

forever

[nimble heart]

they're just nodes

they cant hurt you

[wide crane]

they can hurt my

brain

if webui was any faster i would switch back

😦

[shy kelp]

do you have to call the lora in the prompt or is it just there?

[wide crane]

dont have to call it

[shy kelp]

[nimble heart]

the node is the lora call

[shy kelp]

interesting

[nimble heart]

auto's lora-in-prompt is just its way of auto-adding the node

if you wanna think of it like that

same with BREAK/AND and the prompt switching

in comfy those are just nodes, not prompt bits

[shy kelp]

i have no idea what BREAK AND does

[nimble heart]

Comfy does have the dynamic prompts thing included by default

so you can prompt photo of a {dog|cat|horse} and get a random choice

also has weights with CTRL+up/down to adjust word weights

[wide crane]

maaaan

i want automatic 1111 inpainting in comfy

[nimble heart]

what about it do you miss

[wide crane]

a1111 inpainting is so good

[wide crane]

being able to base the inpainting off the original image

[shy kelp]

png info is what i want, does comfy have that? (pls)

[nimble heart]

?

[shy kelp]

i often refer to my old gens with png info

[wide crane]

when doing inpainting on a1111 you can use the original image as reference

with comfy

it just nukes it

and restarts

[nimble heart]

if its an image made in comfy you can drag it onto the canvas to see the workflow. If it's an a111 image you need to look at it in auto or a program that lets you see png metadata

[shy kelp]

damn it

[wide crane]

maan

[nimble heart]

not sure what you mean by this. like clipvision?

[wide crane]

the faces are so fucked

[wide crane]

what is clipvisio

[nimble heart]

idk what do you mean by "image as reference"

[wide crane]

as in

it takes the structure

of the original image

and just changes some features

but also keeping the structure intact

[nimble heart]

painting a mask? controlnet?

[wide crane]

maskl

[nimble heart]

you can paint masks in comfy

[wide crane]

yea

but they nuke the image

and completely replace

[nimble heart]

?

[wide crane]

whatever you masked

[nimble heart]

just like

turn the denoise down

[wide crane]

how

what

[nimble heart]

idk on whatever sampler node you use

[shy kelp]

hmm so -1 isn't random seed, 0 then?

[nimble heart]

"control_after_generate: randomize"

every time you click 'add to queue' it randomizes the seed

[uncut fiber]

3070 i have speed as 3090 probably using AITemplate. One of many things keeping me in Comfui.
A1111 make big step in performance in last update imo.

[shy kelp]

will it still if batch count is more than 1?

[nimble heart]

batch > 1 just presses queue prompt for you

oops didnt mean to reply

[uncut fiber]

Also you can switch to increase to have x+1, x+2, x+3 and so on

[wide crane]

@nimble heart I COULD KISS YOU ON YOUR LIPS

YOU BEAUTIFUL PERSON

YOU SEXY MAN

[nimble heart]

did you figure out the denoise

[wide crane]

yes

[nimble heart]

if you're using the (advanced) nodes for refiner BS just change the start sample to be > 0

[wide crane]

ok now time to find a way to make the upscaler not give me cancer

[nimble heart]

?

[wide crane]

the upscaling never looks good 😦

[shy kelp]

how do you cancel a generation? lol

[uncut fiber]

cancel in bottom where batch are. View queue and cancel

[uncut fiber]

yes

[nimble heart]

upscaling how

[uncut fiber]

try SwinIR, I like it.

[shy kelp]

running but those options aren't there

[wide crane]

click view queue

[nimble heart]

lotta people use UltraSharp 4x too

[uncut fiber]

view queue

[nimble heart]

it's a tune based on realESR

[wide crane]

i will try it

[shy kelp]

interesting

[uncut fiber]

i keep sometimes in batch count 1 and auto queue checked. When i want last image, stops, i simply unchecking autoqueue

you cant get same image in two different UIs

[shy kelp]

i figured as much, was reading how weighting is different in a1 and comfy since a1 avg's it out with others or something

[nimble heart]

if you're comparing seeds, auto uses GPU seeds by default while comfy can only use cpu seeds

if you set auto's RNG source to "cpu" and make sure the casting/cross attention/sampling is all the same the images are almost 1:1 with comfy

[shy kelp]

wasn't too concerned if it gave me an exact copy but happy that it's giving me generally what i was looking for

[uncut fiber]

@shy kelp what is your gpu? If recent you can have some boost in generating

[shy kelp]

2080

in comfy, doing about 2.x and in a1, was doing around 1.x so definitely faster with pretty much the same stuff going on

[uncut fiber]

hmmm, pity it should be only for 3xxx and 4xxx 😦

[nimble heart]

if its ait you might be able to manually compile for 20 series

[uncut fiber]

yes i have it running with all problems i have with it to get it work properly today again.

[nimble heart]

is it still going to be officially put into ComfyUI or is it forever an addon now

[uncut fiber]

funny is it is generating still different images, on same seed, but difficult to learn when Comfui, dont let you generate same image with same settings in row.

Still addon

[nimble heart]

wdym "dont let you generate same images"? is that AIT or just comfy? My comfyui images are perfectly reproducible

[uncut fiber]

but i have to have 531 nvidia drivers 😦
AIT

not identical, just little bit different

[nimble heart]

ah nvidia xformers right

something about xformers rounding means they can be slightly different

I cant use xformers so everything's basically pixel perfect

[uncut fiber]

yes just each generation you have different image, supposing rounding should be same, not forever different.
Will try to make some images to show you. And yes i am not IT expert.
I thought xformers generate different image but still same.

[shy kelp]

wait, images generated in comfy, how would i grab png info on that? shove it into a1?

[nimble heart]

drag it onto the canvas and it rebuilds the whole workflow and settings used in-place

[shy kelp]

that's pretty cool

[nimble heart]

was it pretty cool last time too lol

[icy brook]

anyone got a suggestion for the best inpatining model to inpaint realistic eyes?

[heady vale]

about the best Ive been able to generate #✨｜sdxl message
using nightvision

[shy kelp]

idk why it does this, it'll lag really hard and go super slow

[uncut fiber]

yes 11 second per iteration. Dont you have running both UIs? To me it throw error

[shy kelp]

wish that 11 seconds was super fast

damn it lol

[nimble heart]

vram overflow? the new nvidia drivers push stuff to system ram to avoid mem errors but in SD it just causes it to run @ 1/50th the speed cause it doesnt unload the parts it doesnt need

or something like that

in task manager you can check if your gpu's non-physical memory is also being hit

[shy kelp]

think this may be it, it'll run super super slow but eventually finish, could never figure out how to fix it

[nimble heart]

ComfyUI has options to make models unload aggressively. forgot the exact names. could try them

[shy kelp]

this upscaling with esrgan is so fast compared to a1

[nimble heart]

#✨｜sdxl message

that one

[shy kelp]

fucked eyes but pretty nice

[nimble heart]

fucked eyes is SDXL in a nutshell

so many images look good except they got chameleon eyes

[shy kelp]

im guessing this goes in the batch file, after the --windows?

[nimble heart]

yea

test it with a different ways make sure it works. I don't have an Nvidia card so IDK if it'll fix it properly

[shy kelp]

welp, time to fuck around and find out

[nimble heart]

last resort could also --low-vram

oh lmk if it fixes it so I can tell future people

[shy kelp]

[nimble heart]

well the flag worked see if it's fucked up still

2080 should get way more than 11s/it

more like 1.5 it/s

[shy kelp]

i cant use controlnet, any solutions?

[nimble heart]

or more idk

make sure it's an XL compatible controlnet else idk AIT is buggy so try removing it

[shy kelp]

i remember it triggering when i tried to add clip skip to it and it changed to 11 then hmmm

[shy kelp]

THANK YOUUU

i just remove the ait and its worked

[uncut fiber]

@nimble heart here examples of same AIT and seeds, and all same simply.
I think if Comfui allows generate same image with same seeds with AIT in Wf, it could make interesting animation things.
Just examples and difference in this case is realy very low, maybe play a bit with cfg

[nimble heart]

looks like xformers

I had similar on my gtx 1070

[uncut fiber]

xformer vs xformer?

[nimble heart]

try the sub quad attention see if it still does it

[dense chasm]

elementary school celebrate the traditional Chinese mid-autumn festival,with SD it's a piece of cake for asthetic painting lol..

[nimble heart]

xformers isn't fully deterministic

same with the men efficient SDP attention

the other cross attentions should be so see if it still has differences when rendering the same image twice with something like sub quad

[wide crane]

@nimble heart how do i do

clip skip

comfyui

[uncut fiber]

it is nothing i should be worrying about 🙂 just would be nice to allow make such animation.

[wide crane]

i saw it somewher

e

[nimble heart]

no idea never used clip skip in my life. just search "clip" in the node search menu when you double click

[uncut fiber]

clipsetlastlayer?

[dense chasm]

with controlnet save gpu memory ,has to add pipe.enable_xformers_memory_efficient_attention()😅

[dense chasm]

i recommend use T2I adapter instead of controlnet,less params and inference faster

[shy kelp]

i was just looking at this stuff earlier lol

[nimble heart]

yea it's the last layer

[shy kelp]

it goes negative i believe, so -1 is 1, -2 is 2? (i think)

[nimble heart]

that's the right one I'm pretty sure

also

[shy kelp]

comparing to a1* that is

[nimble heart]

combine is equivalent to auto AND, and concat is equivalent to auto BREAK

I think

there's also a conditioning timescale node somewhere that lets you do things like the prompt editing in auto

[heady vale]

I wouldnt recommend using clip skip with SDXL, you will get worse gens

[shy kelp]

hmm with batch count >1, you'd have to look in the folder for results and it won't show as grid huh?

[nimble heart]

I might write my own node eventually that automatically makes a graph of combine/concat/etc nodes based on your prompt like auto does

[nimble heart]

yea. or you can load history items.

batch size >1 makes a grid though

[static wigeon]

I was using the base 512x512 Resolution and my output was amazing. The moment I went to 1449 x 2048 since that is my desired res, it became horrific 😭 Like cursed outputs, any ideas on how I can fix it.

[nimble heart]

low res pass first then upscale and do another high res pass

[shy kelp]

neat

[static wigeon]

The before and after

[nimble heart]

yep

pixel upscale the first image to the desired resolution then denoise it at like 0.3 strength or so

or if you're brave you can just latent upscale and denoise at 0.6 but it's really fiddly

[static wigeon]

So generating at my desired res won't work?

[shy kelp]

hmm https://github.com/receyuki/stable-diffusion-prompt-reader

[shy kelp]

is it node or model?

[dense chasm]

xformers,it's facebookresearch,anyway,if pytorch version above 2.0 not recommend

[shy kelp]

oh this works fantastic

[nimble heart]

that looks like an sd1.5 model so highest you can really go before it gets Eldritch is 1024x768

[static wigeon]

I am using Storyboard as a prompt but it doesnt work 😔

[dense chasm]

pre-trained model,T2I is much more lightweight than controlnet,run faster i just tested in google colab

[static wigeon]

Ohh I see

[nimble heart]

any image editor that can view all metadata works too

[static wigeon]

It is a model I downladed from civitai so I assume it uses 1.5

[nimble heart]

models are trained on a specific size so the further away you get the less coherent the images are

most civil models are 1.5 yes

[shy kelp]

most 1.5 is 512x512

[nimble heart]

you can make high res images just not in one pass.

[static wigeon]

That is quite unfortunate

[shy kelp]

https://tenor.com/view/flapping-catherine-ohara-moira-moira-rose-schitts-creek-gif-20774524

[nimble heart]

#✨｜sdxl message
I made an okay-looking image in 3840x2160 on SDXL by first rendering 1368x768, upscaling, then denoising the upscaled image.

I have a few 1920x1080 images I did using that method in 1.5 too

[static wigeon]

So upscale uses img2img if I am right. Do I need a model for that?

[nimble heart]

no

just img2img

same model

set denoise to like 0.3 or so for pixel upscale, 0.5+ for latent upscale

[uncut fiber]

i am in comfui, so no xformers. But AIT, very not polished

[nimble heart]

yea I'm gonna hold off on trying to compile AIT for my GPU until it's less buggy lol

[uncut fiber]

try negative color? @static wigeon

[dense chasm]

it's insane if you know how to use with python library,i use aria2c for downloading and uploading all kinds of huggingface sdxl models in colab and designate google drive where the comfyUI located,it only took a few seconds finish the job,then i can run comfy nodes faster and smoothly with refiner model,so cool

[static wigeon]

Okie will try

Wdym by -ve Colour

[uncut fiber]

negative color/s

[dense chasm]

around 10G safetensor file

[static wigeon]

Ohh

[shy kelp]

alright, the hell is AIT

[uncut fiber]

if her hands, he has both right?

[nimble heart]

sketchy performance hacks

[static wigeon]

I chose resize and fill but it did not fill

Just resized it like this

[shy kelp]

i see

[static wigeon]

Denoising is 0.3

[uncut fiber]

this is resize and fill.

[static wigeon]

Ohhh

[nimble heart]

yea you changed the aspect ratio so it's making stuff up

keep the aspect ratio the same

[rustic garnet]

hm, in the past there was quite a noticeable performance difference between T2I adapter an Controlnet. How good is the SDXL T2I? Anyone tried? I don't care about inference time, I want high quality images.

[static wigeon]

What is the maximum res for a 1.5 model?

[nimble heart]

depends on the model

just try a few and see where it starts turning into soup

[static wigeon]

Okiee will try

[nimble heart]

most are trained at 512 or 768

[floral island]

going to see if can make some more interesting model mix for my idiot prompting 🙂

[dense chasm]

https://huggingface.co/spaces/TencentARC/T2I-Adapter-SDXL

i just generate T2I sdxl image to video with pika

look the eyes of game of throne ice dragon

[shy kelp]

i like albedo and nuclear, though nuclear is a bit hazy for my taste

[floral island]

i now forced the protovision vae on it, it looks it just has a poor vae

[shy kelp]

i always hate when i prompt stuff like this because not sure if i should do mountain or mountains or if it even matters

[static wigeon]

Ah I see

[floral island]

but the image from nuclear bit distorted too, so i can't use it, the weights in that model probably won't fit well with my style

[shy kelp]

what the, HEY, THAT VAE IS FOR PROTO!

[static wigeon]

Upscaling is working and it is actually pretty good, thank you! but there is some loss of details. Any recommendations on how I can avoid that?

[floral island]

protovision has pretty nice image quality, but it isn't really creative 😦 need me some more of that back

[static wigeon]

Original and Upscaled

[shy kelp]

previews kind of screw my speed

[floral island]

which preview do you use?

you should swwitch to the taesd one, the other one is really slow

[shy kelp]

the taesd one yeah

it came with the install as i was getting ready to install it lol

[floral island]

dunno if comfy has a setting for "poll once in N steps"?

[shy kelp]

oh yeah, there was something with upscaling i was wondering

because this is just upscaling it to 4x or w/e but doesn't have the other options

[floral island]

oh my... when all the resulting images are... actually pretty damn nice. the nsfw model is out tho, too little coherency. i wonder why my merge with it does produce nice images though

[shy kelp]

like if i wanted to do this

[wide crane]

what an elegant

workflow

i have trancended

this mortal plane

[shy kelp]

so clean

[shy kelp]

that nuclear

[floral island]

yeah, was the vae that's borked. the model has some incredible creativity

but albedo and zavychroma are pretty damn sweet too 😮

[shy kelp]

yeah those two are nice too
is that vae being broken a happy unexpected positive bork though? looks good

[floral island]

nah, if the vae selection in A1111 is on automatic, it uses the vae that's inside the model

[wide crane]

baked vae

[shy kelp]

bleh

[floral island]

and with the grayish /faded output it had compared to the rest in the first comparison shot, i figured the vae that's inside that model is probably the cause. so just manually use protovision vae. and that fixed the dull color/contrast

[wide crane]

is it just me or are orange eyes like

[static wigeon]

😟

[floral island]

hmmm... let me try something stupid after this result

[obsidian rock]

What Vit-L-14 model is used in sdxl?
('ViT-L-14', 'openai'),
('ViT-L-14', 'laion400m_e31'),
('ViT-L-14', 'laion400m_e32'),
('ViT-L-14', 'laion2b_s32b_b82k'),
('ViT-L-14', 'datacomp_xl_s13b_b90k'),
('ViT-L-14', 'commonpool_xl_clip_s13b_b90k'),
('ViT-L-14', 'commonpool_xl_laion_s13b_b90k'),
('ViT-L-14', 'commonpool_xl_s13b_b90k'),

[shy kelp]

can you make these comparison type grids in comfy? i just found out about them for webui after using it for like a year lol

[floral island]

not that i'm aware of

[uncut fiber]

what is that x and y in comfui, isnt it it?

[shy kelp]

i literally did it the hard way, open chkpt, change scale, pretty much all manually lol

[shy kelp]

with clip skip

without

wait, wrong seed

[heady vale]

using SDXL? generally its worse

[shy kelp]

hmm, it's the same
i think i read that sdxl or something doesn't use clip skip, maybe i read it wrong

[dense chasm]

it's like a video game scene view The Legend of Zelda🤣

[wide crane]

sick as hell

[shy kelp]

this one has negative prompts in it from these positive prompts, this one had no negative prompts

[wide crane]

you're ai generating sick ass landscapes, im ai generating cute women

[shy kelp]

losing some QoL with using comfy instead of A1 but so far, don't mind if the little things in A1 are not immediately available

[wide crane]

we're not the same

https://tenor.com/view/transitions-kinemaster-black-guy-suit-tiktok-gif-25279479

[heady vale]

plus thats a 1.5 gen

[shy kelp]

i like landscapes though

[wide crane]

mine isnt

lol

[wide crane]

if you are talking about mine its made with sdxl

[heady vale]

looks the same as 1.5

[wide crane]

yea except its way higher resolution with better colors and face

like

[shy kelp]

i like how this jinx generation turned out

[torn lark]

your daily prompt quota. Please use /subscribe to a plan that allows you to get more prompts. ¯_(ツ)_/¯

[rustic garnet]

clip-skip does only makes sense for SD 1.5 in my opinion, because all later SD versions do clip skip automatically anyways

[wide crane]

do you already have a lora?

[shy kelp]

the sdxl jinx lora is out there ye

[wide crane]

br

[dense chasm]

where is fuji mountain Imao

[glad grove]

still looks the same to me

[wide crane]

not sure

what to tell u mang

[floral island]

I'm generating an even more huge-ass comparison sheet 😄

[shy kelp]

this is my favorite jinx generation tbh

[glad grove]

u can generate that on 1.5 to push sdxl furter u would have to generate something thats extremely hard to do on 1.5

[shy kelp]

if you really wanna get wild, generate on 1.5, img2img to SDXL

[wide crane]

ah yes well you see currently i am insanely tired because it is almost 7 am so i will be going to sleep

https://tenor.com/view/transitions-kinemaster-black-guy-suit-tiktok-gif-25279479

[glad grove]

with 0 denoising strenght

[wide crane]

and ive been learning this mind numbing comfyui all night

[shy kelp]

gotta do a couple passes

god its so fucking blurry lmao

[wide crane]

@glad grove rate the workflow

[shy kelp]

im used to the node stuff because of unreal

[wide crane]

ive literally never done anything node related in my life

so no

fun

[glad grove]

it looks sick and slick

[shy kelp]

its pretty cool in unreal if you understand blueprinting, can make some pretty cool games

pretty fast prototyping

i want an assistant type plugin though for unreal, that'd be real nice

like what roblox is coming out with

[wide crane]

whar

roblox node?

[shy kelp]

https://www.youtube.com/watch?v=GHZZffgRvkw

[glad grove]

now rate this workflow

[wide crane]

[wide crane]

what even is the use of this

[shy kelp]

just waiting for the inevitable unreal engine one because i am super super lazy (or even godot or any other engine (not unity, fuck unity))

[glad grove]

after playin with nodes i was finally able to generate a pregnant steven seagal 👍

[wide crane]

man

i cant be the only mfker

who thinks like

intricate eye designs

are just so cool

[wide crane]

https://makima-thighs-are-the.best/D4RkMXwq

this is literally my wallpaper

[shy kelp]

i like these eyes

[wide crane]

literally douma eyes

[shy kelp]

oh those are nice colors

keeps generating volcanos

[shy kelp]

i need to work on my hyper color characters

[wide crane]

good morning bro

im going to sleep

[shy kelp]

good morning lol

[wide crane]

this one has

heart eyes

i didnt even prompt that

[shy kelp]

heart eyes best eyes

yeah

[wide crane]

now this

is sick as fuck

[floral island]

could be straight of out my random prompts 😮

awesome stuff!

i... i'm model mixing again 😦 i thought i was better than this 😮

[shy kelp]

i saw something about previewing model mixes with comfyui today, definitely want to look into that some more

[analog roost]

Damn, this one is nearly perfect, pity for the lighting

On her legs I mean

[floral island]

casually chatting with eldritch overlord?

[uncut steeple]

I see similarities

[uncut steeple]

Bro is harvesting my random prompts

[west breach]

you were prompting for whoopi as ursula?

[uncut steeple]

[uncut steeple]

nightmare fuel model

[heady vale]

hehe

[floral island]

trying to see if my model mix is any better than the base models, that being that SDXL trained models everyone makes are awesome by default

[west breach]

looks good

[floral island]

that's just their base models being good

any decent sdxl finetune can do that

[uncut steeple]

[floral island]

testing out my model mix on protovision

allnuclearprotonreversed is the one i'm actually really happy with

[uncut steeple]

bottom one is my fav

[floral island]

same here

i noticed protovision has incredible visual fidelity, but it just isn't... creative

am trying to fix with with some of the more creative merges

[uncut steeple]

you gonna release that?

[floral island]

dunno of socalguitarist is ok with that, should check the license

[uncut steeple]

Otherwise let me know what you merged

[floral island]

well, now that i have small success, i'm probably going to end up testing all kinds of merging weights now XD

keep visual fidelity, but boost the creativeness and details to the max

@uncut steeple creativeml openrail license, thats the one where retrainting/mixing and all is ok?

[floral island]

is selling images of that do not sell phone boook allowed?

[wet nacelle]

DON'T DO IT!!!

[native knot]

Do nnot
Sell!

SELL

[stone fossil]

Sell or die.

[floral island]

the important question is, do i get a beer because it's wweekend?

i'm pretty sure all drunkards in channel will shout yes

[stone fossil]

There are many stupid questions but this is one of them, ofc you do.

[wet nacelle]

https://tenor.com/view/yippie-yay-catbug-bravest-warriors-gif-11992784

[floral island]

no don't book? 😛

[stone fossil]

Plz do not book.

[floral island]

do phone book!
noit sell!

[stone fossil]

You can place that on so many things that would make it akward. 😛

[floral island]

"book of grapes" -> you know what will end up on the cover, given enough generations

[uncut steeple]

I'm really bad with that stuff and not the right to ask

[floral island]

it'll probably ok, i think?

[uncut steeple]

They'll probably let you know if it isn't

[noble shoal]

The Phonebook company seems to be pretty serious about their intellectual property. PS: Do! Niele nout do sellay pone piciona.

[wet nacelle]

catbug

[shy kelp]

1.5 on comfy seems better than a1 for me

[shy kelp]

what the

[uncut fiber]

Some words knows AI pretty well 😄 Would be fun make a list of them. Short length words it knows realy well

[wet nacelle]

https://tenor.com/view/life-is-roblox-life-roblox-is-dj-khaled-gif-8271039727770837203

[stone fossil]

Made a wool lora cause why not, the previous output was ok but this is better.

[wet nacelle]

Really nice and isolated

[uncut fiber]

@stone fossil instead of tie the shoelaces, you tie whole shoes 🙂 <- not sure it make sense 😄

[stone fossil]

SD is a sick sneaker gen, try it.

[uncut fiber]

what to try?

[stone fossil]

It will make anything, nike, rebok, adidas.

[uncut fiber]

i know

[strange mist]

not an expert on license stuff either. but another dude, merged SCG's models and created Realities Edge XL.
but you could simply ask him in here @delicate kelp (sorry for the ping) 🙂

[floral island]

Well you just pinged him, so imma just write it below, you ok with people merging/mistreating ur model, socalguitarist? (amazing model btw)

[shy kelp]

is it possible to have SDLX powered editing offline, like what clipdrop offers online?

[obsidian rock]

I have a question about the revision feature in comfyui. Am I correct in thinking it basically does the clip image encoding using the same clip model that is used to do the text encoding?

[zinc cargo]

do you mean to generate images?

[shy kelp]

erase elements, background, generate new backgrounds

[noble shoal]

You could use LaMa Cleaner to erase elements. It's also open source afaik

[uncut fiber]

Moonlike cabin 🙂 or cottage?

[native knot]

Nice

[uncut fiber]

they were kicked as black passengers from titanic.
How in English, deaf or blind passengers?

[vital ermine]

I had the one at 3. Some I have not been able to get rid of at all and now I am getting urls.

[uncut fiber]

😦

[noble shoal]

In some cases "watermark" in the negative helped me with that problem

[noble shoal]

Excuse me. You have a dead pixel over the eyelid. You might check your camera.

[noble shoal]

How about a sheep? 🤔

[amber musk]

hey, can anyone help me getting started with pikalabs with sd?

i'm using auto1111webui

[amber musk]

@stone fossil Hi

is it possible to create a video (movie /reel kind of using sd auto1111?

suppose i have a story "turtle jumps from parachute" or "naruto escaping from jail" (in a cinematic style?)

[stone fossil]

Wait lol.

Well you got quite some subjects right of the bat there.

[delicate kelp]

I'm fine with it. You find a way to improve my models, I'm all for it... though I may borrow back on the next update if it's good enough 😉

[stone fossil]

So we have to order them.