✨｜sdxl

[smoky patrol]

But yes, SD is great for bringing things to life

....it's alllllliiiiivvvveeee

[stone fossil]

Boombastic. 🙂

[soft zealot]

well thats an "intersting" T shirt

[zinc cargo]

2fast2kebabius

[zinc cargo]

carmen san diego

[ivory blaze]

it's like the reverse of this one

[ivory blaze]

and no it is not a nipple, it is a logo of some sort

[soft zealot]

This is what I like aboutteh CHatGOT enhancement.
Subtly modifies my original prompr (orange) on each generation

although de some reason SD decided this one would be more feminine

[ivory blaze]

I have always wondered.

do negative prompts affect positive inferences?

eg. no barbie, influence to be "feminine" but not barbie?

[soft zealot]

as a good friend of mine onece said "it's not the mind it comes out of, it's the mind it goes into"

[soft zealot]

pass, all i know is irs actually IMHO one of the better ones 🙂

[ivory blaze]

I only meant the thumb in chat looks like it, but zoom view shows clearly a logo

[soft zealot]

ah just looked at the prompt, the one in the screnshot must have been for the next generation.

This one was just" A picture of a young, fashionable skateboarder on a Californian boardwalk. Intricately detailed arm tattoos, happy and smiling."

with no gender mentioned .

(and this girls & boys is why I output a text file lol)

[ivory blaze]

fashionable probably leans more toward female training than male as well

[soft zealot]

treu, either way I do appreciate the small extra element of randomness the ChatGPT enhancement provides 🙂

or variation if you prefer

[ivory blaze]

I need to figure a controllable method of img2img loopback that allows panning and motive elements to a loopback animation

[stone fossil]

Linguistic Only :: Stripped.
Supporting Only :: Stripped.
Negative Only :: Stripped.
Base Prompt G (Linquistic) :: Stripped.
Base Prompt L (Supporting) :: Stripped.
BasePrompt G (Negative) :: Stripped.

[ivory blaze]

didn't mean to send that as a reply

[soft zealot]

I was wondering lol

[ivory blaze]

I have noticed if you offset the aspect ration just a bit, it'll move, and zoom, or opposite direction / zoom out,

which I can use to induce panning of sort, or I could pad the wall and kick it to the direction, with some scipting, maybe a control radio button set, up/down/none left/right/none in/out/none and use a quick image mask blank and shift method,

[steel mist]

Whomever threw a google earth image into control net you are a genius. This is super fun to play with

[ivory blaze]

ahahah

I wonder.

[ivory blaze]

google api, let's play

I'm going to img2img the entire earth

[zinc cargo]

what what what

this is amazing

canny?

[steel mist]

depth

just on a screenshot

[zinc cargo]

OMG

i love it

just finishing this one thing quick and on it

[steel mist]

I actually might try an upscale on these and print them

[soft zealot]

also Its nice the way it expands on some of the short Random Prompts (from Lexica using a WAS node)

random indeed

[indigo carbon]

batch size 1-4, res 64-2048 officially working with latest update of the node =]

[thin oasis]

are there any inpainting models at this point?

[stone fossil]

Rofl back bsod thanks electonic gods.

Where was we.

[clever verge]

Beautiful style!

[clever verge]

Is there more than prompting to generating this?

[stone fossil]

Love.

😄

https://www.youtube.com/watch?v=UBsgdl1W0vg&list=PLaWR0UtAgqduBbazYOcekhD0ghnob4D_C&index=4

[thin oasis]

are there any inpainting models for sdxl at this point? im not finding myself

[clever verge]

Not yet, only controlnet afaik. I hope that's what Thaevil1 is using. 😄

[stone fossil]

I can.

But all these images I post here are just txt2img.

Unless told otherwise. 🙂

[clever verge]

Good job! Especially those that are spherical.

[stone fossil]

Ive used CN XL and I love it Ive shown examples here before.

But for pure detailed SD love you need a pure txt2img.

[indigo carbon]

my favorite is easily copax_colorfulV2, second favorite is CCXL, then all the others are in the same level imo

[stone fossil]

I train also lora's here and help them get me what I want.

Thanks to @dapper dragon

They are not fully needed heh but if you have wild idea's that do not fit they can help a lot.

[indigo carbon]

SDXL is easily smart enough to create any style, you just need to know how to prompt

[stone fossil]

Not rly.

I had to make ice / pcb / blueprint.

To make them shine.

It depends if you have something out of order.

[indigo carbon]

idk man, I might just know how to prompt very well. I never got to a point where I couldn't get what I want with good prompting

[stone fossil]

Make this.

[indigo carbon]

give me the basics of that prompt, and I'll do it

[stone fossil]

I am counting down.

How many counts would u like. 😉

[stone fossil]

Thats the whole idea.

SX is bad at a few things.

This is one of them.

Also ascii art.

Try some ascii art.

[indigo carbon]

looks fine..

[stone fossil]

Not bad.

We can not deny tho the lora one rocks

lol.

It rly tries to explain how a tree works.

Which is funny.

[pure crystal]

@stone fossil https://civitai.com/models/5271/style-info-an-embedding-for-infographic-style-art

[native knot]

Thing is Tree Tree

[stone fossil]

xD

[native knot]

top-right...first thing I saw. 🤣

[stone fossil]

Neat. 🙂

Like said with a lora it help to push it towards.

[indigo carbon]

idk. this image took 14 seconds to generate, and I didn't use any embeddings

[pure crystal]

https://civitai.com/images/339250

[indigo carbon]

TREE WORKS

[native knot]

Bluppint!

[stone fossil]

🙂

[wet nacelle]

@zinc cargoI'm working on drawing the Lora base images now and just wanted to share these. They look hella funny.

It just gets progressively darker and deeper.

[stone fossil]

How about a mona lisa GPU die?

[native knot]

Nice!

That's a great idea.

[pure crystal]

https://civitai.com/images/2073070?period=AllTime

[stone fossil]

Mona lisa gpu die makes the AI art. 🙂

[stone fossil]

Noice bob. 🙂

[pure crystal]

not mine, from that link above.

[pure crystal]

I love diagram stuff, I never tried one

https://civitai.com/images/1948016?period=AllTime

[clever verge]

RTX 5000 series. 😄

[wet nacelle]

https://tenor.com/view/mona-lisa-painting-gif-9441058

[stone fossil]

Yes then AI will make AI all we haveto do is sit on the toilet.

[native knot]

That's like 50% of what I do now, so...

[indigo carbon]

I love me some water cube

[stone fossil]

https://www.youtube.com/watch?v=hmTOM2fuFwk&list=PLaWR0UtAgqduBbazYOcekhD0ghnob4D_C&index=11

[pure crystal]

@ionic gulchany easy way to install the missing nodes from your workflow? I have comfy manager

[zinc cargo]

generated this in MJ last septmber... still one of my favorites

[zinc cargo]

click on install missing nodes

[ionic gulch]

@bob they should all be available through comfy manager if you use "install missing nodes"

[indigo carbon]

even comparing MJ to SDXL is like comparing a bogy to the titanic

[zinc cargo]

regardless of my opinion of what you said, it was in context of the diagrams @pure crystal posted

[pure crystal]

thanks, I thought install missing nodes was supposed to work. My install must be screwy, thank you

[indigo carbon]

I never saw MJ get even remotely close to this, idk man

[pure crystal]

I love this haha, napkin doodle cookie monster diagram

[ionic gulch]

you find my custom nodes if you search for JPS in comfyui manager

if you miss other nodes i could look them up - most are from common sets, so you should have most of them if you used comfy and other workflows for a while.

[wet nacelle]

Can someone right now supply me with a simple SDXL Lora workflow? The one on the Comfy examples git page runs extremely slow for me for some reason.

[zinc cargo]

here MACK, this has 2 positive lora and 1 negative

[wet nacelle]

would it be simple to bypass one of the loras?

[zinc cargo]

and @soft zealot 's prompt

[wet nacelle]

or both

[zinc cargo]

yea, just click on the node, then ctrl+b it will change color and that's it

[wet nacelle]

I'll look into it. Thank you.

[ionic gulch]

mine has 3 loras with on/off switch, but it's not that simple :)

but most options have on/off switches so you can use it for a simple txt2img with 3 loras easily

[indigo carbon]

if you would try to generate this with MJ, it will look lower quality and the sign would say something like: "NWNIWINWNWHEHEHRERHWDU"

[indigo carbon]

like, not only SDXL has its own league of quality, it even does text 80% of the time (at least with the AIT workflow I made)

[indigo carbon]

even full sentences =]

[eager tide]

WTF THIS MODEL IS INCREDIBLE

[zinc cargo]

lemur tail secret tecniuque

[glad grove]

real

[terse pewter]

Hi, I need a concept art for a video game, and I am finding it extremely hard to generate this image.
All I need is a garage with few snowmobiles, but we also need a gas station alongside the garage where a person is lying on the ground(dead), it should be a night sky, and the environment should be snowy...

I have tried to generate this image on SDXL, Leonardo AI, blue willow, everywhere, I cant generate where all these conditions are met
Can someone help me with the prompt?

[steel mist]

ya'll depth canny is just too much fun

[terse pewter]

how did you generate these? are these from SDXL bot?

[steel mist]

composition wise, how far off is this

may have told it to make him sleepy instead of dead but

[glad grove]

bros takin a nap

[terse pewter]

very close to what I am looking for, it's for a video game so the environment can be colorful, moody

what prompt you used for this?

[steel mist]

I'm assuming you're not using comfy, but if you are the workflow is in there

sec

pos: a photo of a person sleeping outside a gas station, next to a garage with snowmobiles

[terse pewter]

yup, I am aware of comfy, will watch its tutorial today. I heard its amazing

[steel mist]

neg: art, illustration

[terse pewter]

thank you so much, i will try to use this in the bot

[steel mist]

To refine that though really, I'd use that as the base image and then try to get the snow/night stuff right

[terse pewter]

We can also use images as base? Can you kindly tell me how I can do that? Do I need to install SDXL on my pc?

[glad grove]

yes if you use comfy u have to install on pc

[steel mist]

Yes, if you're running comfy locally you can do things like use an image as a control for a generation

It'll take a bit to get setup for sure

[terse pewter]

Awesome! I guess I need to download it because I am heavily involved in this stuff. I want to create game environments using SDXL

[steel mist]

I will say that's a pretty complicated scene

[steady grove]

i'd plan it, conceptualize it, controlnet it, inpaint it, add a bunch of post, and call that a concept draft.

for a professional product, you'd want to have a documented process. imo

[steel mist]

lol somebody needs to make a snowmobile lora, I dunno what this is but a snowmobile is it not

[glad grove]

tasty

[indigo carbon]

fighting fire

[stone fossil]

Zwaar master

[glad grove]

new animal found on steven seagal table

[wet nacelle]

Actually false. It did get better. It isn't consistent though.

[glad grove]

hmmmm hippo

[steel mist]

gotta be honest not sure I'd eat at that restaurant

[indigo carbon]

eh, what about stuff like this? not to mention that the quality of SDXL is miles ahead of MJ..

[indigo carbon]

yeah, jesus. comparing SDXL to MJ is futile at this point. MJ is similar to a 1.5 model, SDXL is a league of its own

[west breach]

sometimes it's too eager to write text in the prompt. especially if you have 'sign' in the prompt

[native knot]

ANK COEFFE ON RIIVERIDI

[eager tide]

prompt?

[indigo carbon]

ComfyUI workflow in metadata

[eager tide]

not on discord

[indigo carbon]

yes it is. anyways, prompt was "cube made of water"

[eager tide]

ho ok thx

[eager tide]

i got this

[native knot]

You have a different set of dimensions, #1.

[indigo carbon]

did you import the ComfyUI workflow?

[eager tide]

no

i use automatic1111

[native knot]

lol

[eager tide]

with my own prompt

[native knot]

a1111 isn't going to hold a candle to comfyui

[native knot]

SDXL is far better on comfyui right now

[indigo carbon]

why? ComfyUI is better on almost every single aspect

[glad grove]

only use a1111 if u are using 1.5 checkpoints if u are on sdxl stick to comfy

[eager tide]

i got pretty good result

with sdxl

[indigo carbon]

yeah, especially when you use AIT

[eager tide]

bro this is fire u do better on comfy?

[native knot]

if you're happy with the results, then good on you, but you're missing out on a lot of additional flexibility that comfy allows

[indigo carbon]

yeah, also in 4th the time =\

[eager tide]

ye maybe i need to try

[indigo carbon]

heck, maybe less, I made a workflow that uses something called AIT that makes SDXL over twice as fast @eager tide

[eager tide]

links?

[native knot]

It's embedded in the images like he said.

[eager tide]

okok

[native knot]

Just download the full size image and drag it into comfy.

[indigo carbon]

https://civitai.com/models/133818/tdg8uu-aitworkflow

[native knot]

Then use ComfyUI Manager to download any missing stuff.

[eager tide]

if i got same quality in less time i will go comfy

[eager tide]

thx

[indigo carbon]

if your GPU is at least 3000+ series, it will at the very least double speed

4000 series gets an even more massive boost from AIT

[eager tide]

gtx 1080 💀

i will try on my friend pc with a 4090

[indigo carbon]

oh.. sorry, the AIT nodes won't work on that. to still use that workflow you must remove the load AITemplate nodes

[eager tide]

Is there hires fix on comfy?

[indigo carbon]

hiresfix isn't a node on it's own, it's a technique that passes the latents between 2 K_samplers, the workflow I linked uses that.

[eager tide]

ok

[native knot]

Who needs highres fix when you can render directly to 1920x1080 just like this?

😉

[eager tide]

nice

hires fix add detail

[indigo carbon]

if you use the refiner, you did use hiresfix. hiresfix is the term of passing a latent between 2 stages of K_sampling

smartest way to do this to pass the latent from K_sampler1 to K_sampler2 without encoding or decoding in between, then upscale using the upscale node. again, this is what the workflow I made does

[eager tide]

ok

[shy kelp]

how to have a good result sd1.5 in comfy ui

[quasi crag]

Probably not the right place to ask, but is there a way to get the AI to do anthro furries?

[indigo carbon]

just took a look, you used AIT in an inefficient way here; why would you need 3 passes? also that seems like a modified version of an older version of my workflow.

[shy kelp]

you have a different taste sir

[quasi crag]

Thanks, super helpful

[glad grove]

theres a lora for sdxl with that style

also a checkpoint

[quasi crag]

Okay, what is a lora?

[glad grove]

sorry you would have to learn how to use all of that with a video

[native knot]

I like to experiment with the outcomes and see what occurs. I also found that doing a third pass with the original checkpoint gives some interesting results...sometimes cleans up minor oddities from the refiner. But I play around with workflows all the time.

That was a render from almost a week ago anyways.

Just made this using your current one.

[glad grove]

that one looks exactly like him,is it a lora or inpaiting?

[indigo carbon]

https://tenor.com/view/noice-nice-click-gif-8843762

[glad grove]

😬

[indigo carbon]

I don't know if I'm stupid for asking this; who tf is that guy and why are there so many images of him in here

[native knot]

Steven Seagal.

[eager tide]

bro midjourney is farrrr behind

[vital ermine]

LOL

Steven not happy

[glad grove]

he means business

[glad grove]

he ate the blueberry

[hardy cipher]

please rate this canny

[vital ermine]

hahahaha

[native knot]

11/10 - IGN

[glad grove]

remake of walker texas ranger with seagal

[vital ermine]

🙂

[steel mist]

8/10 would recommend

[glad grove]

kronos before he ate all of his offsprings

[vital ermine]

Tasty

[glad grove]

real

[steel mist]

"a visual representation of the pile of connections one needs to use comfy, cables, power"

[glad grove]

this doesnt look like him

[vital ermine]

https://civitai.com/models/135354?modelVersionId=149181

[native knot]

NSFW, bro.

[glad grove]

gotta protect our users against evil images like abs 🙏

[native knot]

Damn right.

[vital ermine]

LOL

[glad grove]

speaking of nature

[eager tide]

wtf

[steel mist]

Dude what model/Lora is that? Looks amazing

[eager tide]

https://civitai.com/models/128607/nightvision-xl-photorealistic-or-portrait-or-photography-or-hyperreal-or-architecture-or-interior-design-no-refiner-needed?modelVersionId=146882

https://civitai.com/user/xppenTC

this is my page for prompts

[wet nacelle]

Here are some more goofs for the Lora.

[glad grove]

our saviour

[eager tide]

niceeee

[terse pewter]

the problem with these ai image generators is that i think they can only focus on 1 point of action, for example, it cant focus on two different things in images at the same time which I think is a major drawback

[pure crystal]

@terse pewter generate art assets in SD --> compose in other tools. inpainting can help too

[terse pewter]

That's a good idea

[pure crystal]

if you use the same seed and change the subject of the prompt, you can get nicely cohesive art assets

[terse pewter]

How to use inpainting with SDXL?

[pure crystal]

https://www.youtube.com/watch?v=_Qi0Dgrz1TM this video has a workflow for that

[terse pewter]

I see, actually I only use SDXL bot that is in this discord server, so I am not much familiar with these terms such as seed, subject, controlnet.
Currently I am following a course that will teach me ComfyUi

[terse pewter]

Thanks a lot. I will watch this

[pure crystal]

awesome good luck. if you have access to google colab this works for comfy https://colab.research.google.com/github/comfyanonymous/ComfyUI/blob/master/notebooks/comfyui_colab.ipynb#scrollTo=jjjjjjjjjjjjj

taking a break while framing this deck. looks like a storm is rolling in

[glad grove]

bro was crucified on an amish cross

[native knot]

He was just checking that it was squared when some idiots come by and nailed him down.

[terse pewter]

I just opened it, connected my GPU, and pressed on run all, will this help run comfy ui without installing it on my pc?

[pure crystal]

@terse pewter yes https://www.youtube.com/watch?v=RBVVU_BKNKQ

[terse pewter]

Wow! Thanks again! 😃 Really ❤️ !

[pure crystal]

I would get familiar with a workflow and plan of action before starting that colab up so you dont burn hours while you learn

[terse pewter]

I am going to watch all these videos, once I perfectly understand all these terms. I will start my art

[pure crystal]

hahahahah my browser looks the same

[glad grove]

real fight scene

[terse pewter]

Haha i see 💖

[pure crystal]

was this in a movie or real life

@terse pewter another good COmfy channel. probably in your tabs lol https://www.youtube.com/@ferniclestix/videos

[glad grove]

its from the noah movie where he has to fight 2 of every single animal

[pure crystal]

my mother's favorite

[vital ermine]

Nice, I used the segal lora and it changed to

[glad grove]

lora knows hes the messiah

[native knot]

New Mortal Kombat character:

[vital ermine]

LOL

added segal

[glad grove]

seagal true form

[vital wraith]

I can't believe how well XL upscales..

[vital ermine]

He did that in one of his movies I think

[glad grove]

another one from the noah movie

[eager tide]

BRO WTF THIS IS INSANE

[pure crystal]

bond villain Seagal

[pure crystal]

Add cosmic meatballs to this prompt, it's a hack

[native knot]

yeah, you gotta get creative sometimes....alternate word combinations that would put things in a new perspective.

When you finally hit those max level pauldrons for your heavy armor set.

[glad grove]

real life animal/man fusion

[native knot]

Steven is a powerful cleric

[glad grove]

is there a god emperor of mankind lora for xl?

[shy kelp]

the hippo playing with the wrong balls

[native knot]

Arcane wizard, Keanu Reeves.

[native knot]

Amy Lee is one spicy rogue.

Damn...I think I like this version better:

Or this...

I can't decide!!

[glad grove]

i want both 😳

[native knot]

Me too.

[crisp owl]

uh

[glad grove]

happiest roblox player

[shy kelp]

HAHAHHAHAH

[indigo carbon]

that's not good

[native knot]

oh, damn...

[azure oxide]

.........................

[hardy cipher]

wow. such beauty

what a healthy looking lad

[glad grove]

hes the real samurai

[hardy cipher]

peak genetics

[heady vale]

Crikey!

[crisp owl]

He left us too young. Miss that goofball

[heady vale]

indeed

[hardy cipher]

not really vibing with that background leg

[native knot]

Just a little case of the elephantitis.

[dense chasm]

the founders of BERT created it with Tensorflow initially, than huggingface team reproduced with Pytorch based on the Tensorflow's data and pre-trained model,it's called Transformer project,nowadays,Transformer has reproduced all kinds of different transformer models structure,the code is neat and readable,easy to use and modify,BERT is a strong competitor of GPT,open source is good for us

notebook is a good tool when you learn Transformers or diffusers,that's the reason why huggingface is an excellent choice of machine learning

BERT is excel at semantic knowledge learning,but not suitable for generative tasks,meanwhile opanAI's GPT insists in the languange model pretrained method,so the gpt model grows bigger and bigger until multi modals

Facebook sequentially released XLM,RoBERTa and LlaMA,learn from BERT encoder structure and surpass the performance of BERT eventually

the above is NLP history

[twilit nexus]

hey where i put the base model with vae? Does it go to model or vae folder in automatic111

[soft zealot]

If anyone is having issues witht he ComfyUI double click on canvas search bar dissapearing off the RH side of the screen it's likely this (removed this pack & problem went away)

https://github.com/giriss/comfy-image-saver/issues/5

[visual glade]

https://comfyanonymous.github.io/ComfyUI_Blog/comfyui/update/2023/08/27/Weekly-update.html

[indigo carbon]

@visual glade just git pulled, speed is significantly lower. is there a way to make it use VRAM like it used to?

[bronze arch]

is it just me or are the control-loras kinda bad

only tried sketch and canny

Prompt: a cat girl with a tail and a massive head of hair kissing a man, illustration, anime artstyle.
negative prompt: black and white

(The sketch, the output without any control, output with control sketch, output with control canny):

[indigo carbon]

I think its because when 4000 series cards that are under 4080 use memory bus technology from 2010; because NVIDIA thought it's a great idea. so when cards like 4070ti get full the speed decreases by about 500%

[visual glade]

if speed is slower it's most likely because you have custom nodes that take too much vram

[bronze arch]

Sketch control without any preprocessing (directly using the sketch as input for controllora)

[visual glade]

and nvidia that keeps messing with memory management in their windows drivers doesn't help at all

[indigo carbon]

no, that's not the reason. I used the exact same workflow I did before I pulled. it does use slightly more VRAM; but that's a very bad thing on cards like 4070ti. didn't change drivers or anything. it's because it goes over the threshold 4000 cards have due to NVIDIA being sooo smart.

the new memory management techniques shouldn't be enabled on all of the 4000 series under 4080 @visual glade

[visual glade]

if you want the previous behavior you can: --disable-smart-memory

[indigo carbon]

did that, didn't change anything. but that's probably an issue on my end

I'm sure the new update works good on 3000 series because NVIDIA was sensible while making those cards; but that's not the case on 4000 series except for cards above 4080.

anyways, since ComfyUI identifies GPU while starting up; I think it's a good idea to make it use the old management on all 4000 series cards under 4080 @visual glade

might not be a big deal when not using SDXL though, because SDXL is the only model that floods VRAM

[visual glade]

--disable-smart-memory is the old management

it will unload stuff completely when loading new models

but the code changed so some custom nodes might have some issues

[indigo carbon]

I don't know, my inference used a commit from 2 weeks ago and speed was massively higher than now, also when using that flag. it's also not the node.

[thorny frost]

That doesn't use to work. You need the prepossessing

[bronze arch]

yeah thats not exactly working either

[visual glade]

which custom nodes are you using in your workflow?

[bronze arch]

i was just using the workflow in the controllora hungginface page

downloaded the nodes using comfy manager

[visual glade]

not you

[thorny frost]

Your esketch has too many secondary lines, do another layer with only the main ones.

[bronze arch]

oh wait were you talking to the other guy lol

haha

[sand pier]

there is a comfy manager 0-0?

[indigo carbon]

after one of the recent updates, ComfyUI tends to use 100% of VRAM, which is usually great on 3000 cards, since normal mem bus. but when 4000 series uses 100% of VRAM, speed becomes a tenth of what it used to be.

[indigo carbon]

just the AIT node I helped develop, it's seamless; we made sure.

same even without the node

[visual glade]

can I see the memory graph in task manager?

[indigo carbon]

yes, I'll try to capture it. if I remember correctly, it used to do 11/12, and with newer commits it does 11.7/12, which is VERY bad on 4000 series. I'll screenshot.

[maiden forum]

重点词汇
11/5000
传统翻译模型
通用场景
Make this picture anime style blob:https://discord.com/0c333ec7-e139-450e-81de-26aaad63e584

[fathom hound]

👀

[indigo carbon]

11.5*

it used to do 11, and that makes all the difference on 4000 series

[visual glade]

what is it doing when it goes to 100%?

[indigo carbon]

22s/it

when at ~80-90% it's 9it/s

[visual glade]

which resolution?

[indigo carbon]

speed is critical

1024^2

[visual glade]

and what happens to the graph after?

does it drop to zero?

[indigo carbon]

after gen?

I didn't wait long enough to find out

I'll interrupt it, hold on

[visual glade]

give it 1 step and let it finish

[indigo carbon]

still 11.5

[visual glade]

yeah that's a problem

oh is that your build with the high batch size?

[indigo carbon]

if it's 11, which it used to be on older commits, speed is 9it/s. and with 11.5 it's 30s/it..

nope, BS1 =\

[heady vale]

with AIT when I cancel it generally doesnt release vram

[fresh path]

windows hip waiting room

[indigo carbon]

the new versions are unusable on 4000 series with SDXL @visual glade , I'm reverting to a few weeks ago, will share details.

[vital ermine]

@glad grove V2 and so much better.

[visual glade]

the AIT nodes override a bunch of comfyui behavior so they need to be updated

[indigo carbon]

the nodes are fine, the screenshots I sent weren't even on an inference that has any custom nodes.

[visual glade]

rename the ait nodes folder to: .disabled

[indigo carbon]

there weren't AIT nodes in my inference with the newer commits.

[visual glade]

I know but try disabling the nodes completely

I see it patch stuff in ComfyUI

even when you don't use them

[indigo carbon]

yeah, I did.. the nodes aren't the issue, there were no custom nodes in the installation I tried the new commits.

[fathom hound]

Lol

Lmao

Somebody tell him

[vital ermine]

Why am I getting this now in comfyui?

[visual glade]

are you using 3M with normal?

[vital ermine]

damn, lol. ty

[sand pier]

I don't need to specify comfy paths in extra paths if i use it right to include other models right?

[visual glade]

no the comfy paths are always searched

[sand pier]

ty

[visual glade]

comfy.model_management.free_memory(1000000000000, device)

Add this here: https://github.com/FizzleDorf/AIT/blob/main/AITemplate/AITemplate.py#L214

[indigo carbon]

tell this to FizzleDorf. I only provide the modules/compile scripts

[visual glade]

I will but I think he's asleep right now so you can try it

[indigo carbon]

I'll change the code locally, give me a few minutes

[indigo carbon]

that line of code causes ComfyUI to not be able to load AIT node. but this has nothing to do with the issue. the issue is that the last few commits make the UI overload VRAM; which is a VERY bad thing on 4000 series.

even without any nodes that occurs

[visual glade]

yeah it will only work on latest comfyui code

and it will make it free all memory before the AIT stuff is loaded

[indigo carbon]

I was on latest commit

anyways, I'm going back to the old commit because this isn't ideal.

[thorny frost]

what is the matter with Steven seagal? is it deliberated or stable diffusion generates him when the words fat clown are in the prompt?

[uncut fiber]

painting or reality 🙂

[ionic gulch]

@visual glade regarding your discussion: @indigo carbon do you use the current nvidia driver or the one before they changed memory management a few months ago? i would highly recommend to go back to 531.79 for everything AI related. the newer versions push things to shared vram much too soon, so you won't be able to use your vram completely and applications that try to make best use of vram by keeping things in memory will suffer and speed goes down to 10% like you said in one of your posts.

[thorny frost]

sea is real, hull is a mix, sail is old impasted oil painting, sky is a mosaic.

[uncut fiber]

yes 😄

[hardy cipher]

JPS

how the heck do I set a widget so it will only be positive odd integers? it seems so simple, but then

[vital ermine]

@glad grove https://civitai.com/models/135544?modelVersionId=149424

[primal vault]

The first public test of v3.991 is now available, it's the next step towards releasing the SeargeSDXL v4.0 workflow for ComfyUI. The test version is here on CivitAI and can now be downloaded.

Please make sure to run python_embeded\python.exe -m pip install opencv-python at least once in the ComfyUI folder to install a required python dependency.

And there are more instructions in the readme file for additional information about the models that need to be downloaded and installed in ComfyUI.

[indigo carbon]

Nvidia always finds new ways to annoy us, huh? they never fail to disappoint me.

[soft zealot]

A photo of a floating clock melting in a serene desert landscape, in the style of Salvador Dalí and René Magritte.

[vital ermine]

On Linux we are stuck with 535 or 525 on ubuntu

[ionic gulch]

you could use step size, but i think you can overwrite that manually - so if you want to prevent it you would make sure that the output is odd in your own code i guess. the min/max values will help for the positive values and actually prevents the user from inputing negative values.

[hardy cipher]

well you don't one to do step size 2

step size 2 + 1

you hit 0 once and it bumps back to 2. then you're stuck on odd

[south frigate]

hi, fast question: control net working with SDXL in A1111? (not comfy) or not yet?

[ionic gulch]

if you set defaults and min max values to odd values, a step size of 2 will result in odd numbers only - unless the user overwrites the value manually

[hardy cipher]

no it won't

try it

I mean

maybe it will, but not in any way I've been able to figure out

[ionic gulch]

what happens if you set default:1 min:1 max:9 step: 2?

[hardy cipher]

it'll stay odd as long as you don't push max or min

[ionic gulch]

if it doesn't work and skip to even anyway, i think you will have to make the output odd in your own code.

[hardy cipher]

so if you go to 0 on the slider

it'll instantly push you back to 2

[ionic gulch]

thats why you set min 1

[indigo carbon]

now I get OOM errors 💀

[hardy cipher]

I mean, it's not a big deal, I have it set to correctin the code

but just curious

[ionic gulch]

that's the downside of the old memory management. the new one prevents oom, but makes 20% of your vram unused. with the old one the application has to avoid to use too much vram itself (which a lot of things used in the ai world don't do). but if you get oom, it shows your application needs too much vram, which would have reduced speed to 10% on the new drivers anyway. problem with the new memory management is that nvidia optimized that for gaming use cases, where shared vram is much less problematic and doesn't result in such bad performance. they know about the problem and said they fix it in a future update.

[hardy cipher]

check this

working on a canny node, lol

.1 seconds to render that though. vs 5 seconds with the bottom one

gaussian ksize there

supposed to always be odd

[ionic gulch]

so you can select 0 if you set min value to 1?

[hardy cipher]

it'll move there and then step up 2

like the slider doesn't just skip the other numbers, it goes over them

[ionic gulch]

ok, in that case that's how it was coded in comfy. not much you can do other than preventing it in your own code.

[hardy cipher]

yeah, works fine on my end. but was thinking of just adding the math to the widget info and then having it behave as it normally does. if it's not going to display exactly what I want anyway no use in even doing much to change it in the display itself

[indigo carbon]

@visual glade merging a commit from a week ago and updating reverting to the ones from before I tried to downgrade fixed the issue =\

[visual glade]

the comfy.model_management.free_memory(1000000000000, device) didn't fix it?

[indigo carbon]

I say you add a mechanism that handles memory like the old commits did when detects a 4000 series GPU under 4080 @visual glade

[indigo carbon]

nope, it didn't have anything to do with the node

it was the memory management. it is not working properly on 4000 series cards under 4080

[upbeat summit]

maybe related (also 4000 series card) - I got extreme slowdowns yesterday in a workflow and it looked liked the vram was just being filled all the time. only a restart of ComfyUI fixed it for 1 image. than it happened again.

[vital ermine]

UGH

[indigo carbon]

.. so it's not an issue on my end.. the fix was merging with a commit from a week ago.

[vital ermine]

4k cards having so many issues. 😦

[upbeat summit]

yeah I think I ran that workflow 1 or 2 weeks ago without this happening. but I would need to do more testing - so it could still be an issue on my end

[vital ermine]

Did you see my V2 of Segal? Came out great and flexible. I can have it in line and not called and never know but called and there he is. First time since 1.5 for me.

[upbeat summit]

but it they work they are pretty fast 🙂

[vital ermine]

A 4090 is still slower than a 3090 for training at least. 😦

[indigo carbon]

Nvidia are soooo smart, they were the geniuses that came up with the brilliant idea of using memory busses that use technology from a decade ago on all 4000 series under 4080; genius! best company!

[vital ermine]

Yeah. Grrrrr. Well, they hate us all anyway.

Did you know training 2.1, not XL, but 2.1 in Kohya on Windows is slower on a 4090 than Colab's T4? Yep.

was a test I did

[indigo carbon]

also not to mention they became more sensible to make the GPUs at a more sane price of double the old price! thanks nvidia! love you guys

[vital ermine]

now if I go back to pre XL version of the scripts and use Ubuntu too it is faster

I never thought I would pay 1700 for a video card ever.

that hurt a lot

[indigo carbon]

not to mention the GPU you bought uses the BRILLIANT memory busses from a decade ago! hooray!

[vital ermine]

5k cards will be even more expensive with same ram. I am done unless this blows up for a decade. I was a 3 cycle person anyway so 6 years but this beast cost far too much to think about a new card even if it had 128GB. Now 128GB for 429 I'd do it.

[indigo carbon]

I hope I'll see the day when AMD puts Nvidia in their place. now, amen.

[willow bane]

hlo

[vital ermine]

Under Lisa never

[willow bane]

can anyone help me with sd

[mossy canopy]

sure

[thorny frost]

Really? so i nailed it buying a 3090ti for 1100€

[willow bane]

it is showing so much time

and my system is not so bad

i am new to sd

thats why donot know much about settings

[mossy canopy]

lmao

[mossy canopy]

what are you trying to do?

[willow bane]

i just write some prompts

of hitting a good cricketing shots

is it always take so much time

@mossy canopy .

[mossy canopy]

how many frames?

[willow bane]

300

maximum?

[thorny frost]

something is very wrong there. try the simplest workflow in comfyUI

[willow bane]

can anyone send me settings link for youtube shorts

[mossy canopy]

lower steps to around 16 and use dpmpp 2m karras

[willow bane]

i just want for yt

is tthere anyway i can get those settings?

and just drag my image and get result

[mossy canopy]

not sure png info works for deforum

[indigo carbon]

@visual glade actually, if all 4000 series under 4080 uses the same memory bus, I might be able to provide a value the highest VRAM can go without overflowing and dividing speed by 10.

[willow bane]

hey what should i do now

i stoped that

[mossy canopy]

change sampler to dmpp 2m karras and reduce steps to 13

[willow bane]

what should be width and height

[mossy canopy]

keeping the aspect ratio set the lowest res to 512 that should be optimal..

[mossy canopy]

which gpu?

[vital ermine]

in comfyui if we want to use more than 1 lora just daisy chain them? I am asking because I have never managed to get more than one working so I guess it is something I am not doing right.

[shy kelp]

my laptop having a hard breath when rendering

[willow bane]

Rtx 3060 6gb laptop

Is anyone made Instagram reel or YouTube shorts using SD deforum?

[indigo carbon]

anyways, back to the crazy speed I got before I pulled =]

[upbeat summit]

beautiful!

[vital ermine]

I am still afraid to touch ait as so much work

[mossy canopy]

LOL! put your laptop in fridger before hitting generate.. if one frame took min 12s then 300 frames should take around an hour of continuous hell

[willow bane]

Bro so what should I do

[indigo carbon]

if your GPU is at least 3000 series it's as simple as loading a workflow and clicking "install missing nodes"

[vital ermine]

I did install something ait

just not sure how to compile and use it

[mossy canopy]

ait should auto download compiled files

[indigo carbon]

the AIT workflow I made is accepted as the official one for AIT. just load it and install missing nodes, if you already have the nodes used update them. works for all gen params.

[mossy canopy]

or you can find them in hugging face and put them in modules

[vital ermine]

what workflow?

[thorny frost]

what is AIT for?

[mossy canopy]

200% speed boost

[thorny frost]

nice

[indigo carbon]

the workflow linked in the repo: https://github.com/FizzleDorf/AIT

[vital ermine]

https://civitai.com/models/133818

[indigo carbon]

yep =]

[indigo carbon]

that's actually the only thing 4000 series are good for. 3000 series gets a lower boost from AIT

[thorny frost]

😦

[vital ermine]

a missing node

manager couldn't find it.

[indigo carbon]

your manager isn't up to date then.

[vital ermine]

only 2 days old

[indigo carbon]

it's the advanced text encode; what makes that workflow not only fast; also incredibly coherent.

[vital ermine]

Well, I guess I shouldn't have hit clear and reload it as it now doesn't come up at all

[indigo carbon]

if you need to do it manually:
https://github.com/BlenderNeko/ComfyUI_ADV_CLIP_emb

[vital ermine]

Let me try and restarting comfy

It didn't complain this time

[south frigate]

looking for some ComfyUI workflow but with it https://github.com/FizzleDorf/AIT - any1 have good?

[indigo carbon]

there is one linked in that repo

[vital ermine]

I just need that upscaler

[indigo carbon]

those are models you don't have.. I don't think I need to explain

[indigo carbon]

oh, hold up, I have it somewhere

[vital ermine]

kk

[indigo carbon]

https://github.com/Phhofm/models/raw/main/4xLSDIRplus/4xLSDIRplus.pth

[vital ermine]

thanks

[indigo carbon]

easily my personal favorite upscaler for literally anything

[vital ermine]

I never upscale as that was a sytan thing not mine

[indigo carbon]

wait, why do you not have the refiner in your inference?

that's half of SDXL, very important

[vital ermine]

I have no idea this is installed for windows

Oh, wait nvm. I don't like refiner

[indigo carbon]

I mean, you do you; the refiner is a model that's specializing on fixing SDXL's common mistakes.

[vital ermine]

Sytan has an explanation about the refiner and he was saying people are stopping the use of it

[indigo carbon]

likely because of high parameter count, but it's not crucial when using AIT

I don't see any other reason not to use the refiner

[vital ermine]

Had to do with upscaling and stuff.

[heady vale]

Unless I am using a finetune model I use refiner

[vital ermine]

Sytan is anti refiner. I am neither pro nor con for it but I just don't use it

[south frigate]

@indigo carbon When I installed ur workflow - why? I used menager to install all custom content and still...

[indigo carbon]

you can try it once; the AIT workflow is intended to be used with it. and as you can tell; the images generated with it are very high quality. but you should be fine using 2 SDXL base models, if not one on stage 2

[indigo carbon]

if you restarted ComfyUI and this still happens: it's because ComfyUI isn't up to date.

[vital ermine]

Once I get everything install.

which folder does that upscaler go in?

[royal python]

Do you guys fill in the Optimizer extra arguments part when training lora?

[indigo carbon]

models\upscale_models

[south frigate]

after restart its the same and comfy said to me its already last version... hmm

[vital ermine]

ahhh, ty. Used to Auto

[indigo carbon]

look at CMD, it might be saying something like "IMPORT FAILED: AIT", that could be the case.

[south frigate]

found smth like that:

[indigo carbon]

yep, that's an issue with ComfyUI- not the node.

my fix was to take the latest commit, them merge it with the last one before the major update

[vital ermine]

11.48it/s

[indigo carbon]

how many were you getting before?

[vital ermine]

about half

[indigo carbon]

https://tenor.com/view/noice-nice-click-gif-8843762

this means it worked

[vital ermine]

Well, does it do a batch?

[indigo carbon]

@vital ermine you can also try doing batch size, 2/4

yep

just recently I updated the code to do so

[vital ermine]

ait doesn't work with ddim?

[indigo carbon]

I don't know, I don't see any reason for it not to

[vital ermine]

threw a lot of errors on me

[indigo carbon]

for me it works =\

DDiM with karras, right?

[indigo carbon]

pytorch nightly.... possible the issue

[upbeat summit]

the AIT node needs to download some files once,. do you have a firewall blocking access?

[vital ermine]

nope. It worked until I switched to ddim

[indigo carbon]

it did work for him when not using DDiM

[vital ermine]

now nothing is working

[indigo carbon]

?

try to restore to the workflow when unaltered, that 'shouldn't be the case.

[vital ermine]

I shut it down

damn

[south frigate]

@indigo carbon any idea how to fix it?

[indigo carbon]

tf, what did you delete? it worked just a second ago..

[indigo carbon]

latest Comfy version, then merge with a commit from a week ago

[vital ermine]

it blew everything out

now my port is not being released so I can't even get back in

[south frigate]

i have last version of Comfy, how to merge it? its easy?

[uncut fiber]

@vital ermine seems we are in same boat 🙂 in my case it probably as well stops working using DDIM

[indigo carbon]

try to delete and reinstall the node; I don't have any other explanation for what happened there

[vital ermine]

Yeah, it ripped me a new asshole

[uncut fiber]

but reinstall and nothing working for me, i mean AIT even clean install of Comfui, so it must be my settings somehow related, i mean PC, AV, FW...

[shy kelp]

does ait now works on the latest version?

[vital ermine]

Yep, same for me it tore everything out @uncut fiber

[shy kelp]

i keep getting this, already turn off preview

[thorny frost]

seems problematic. Wil avoid it for now (since apparently it is not so good for 3000 series)

[uncut fiber]

i have 50% boost

[thorny frost]

3000 or 4000?

[uncut fiber]

3070

[indigo carbon]

if you have files in models\vae_approx, delete them.

[vital ermine]

I wonder why the AIT stuff is in red when even the manager shows it is installed? Reinstall, shut down, restart comfy, and poof still in red.

[uncut fiber]

i must say myself it still isnt lossless 😄
or as some writer said about grappes he cant get "And yet they are sour"

have you both end connected?

[shy kelp]

still the same, it is related to gpu or something?

[vital ermine]

ddim did something bad

[upbeat summit]

just fyi - the high quality previews using taesdxl_decoder.pth have been working for at least 2 weeks. it was fixed on the AIT repo. I have not updated ComfyUI since 2 days though.

[vital ermine]

Nothing I do will bring those back

[south frigate]

i have the same problem, red nodes dis 3

[vital ermine]

It worked fine until I tried ddim and BOOM

[indigo carbon]

shouldn't be. it works just fine for me.. this is odd, seems like an issue with ComfyUI rather than node.

[shy kelp]

im using the default

[south frigate]

i have it when i just updated everything xD

[vital ermine]

I just did a git pull then restarted and it is back

[indigo carbon]

does the node work now?

[thin oasis]

is anyone aware of a way to pull off a somewhat workable inpainting using SDXL? Trying to remove someone from a photo with it its failing. (I could obviously go use an earlier version of SD, im just trying to fiddle and learn here)

[indigo carbon]

Nice!

[south frigate]

@indigo carbon can u tell me how to make a "merge" to avoid my errors? (like on screen before)

[vital ermine]

Oh, boy BS didn't work

[shy kelp]

what bs stands for

[soft zealot]

Bull Shit lol

[indigo carbon]

You download a commit as zip, place in ComfyUI folder, then override.

[vital ermine]

highest bs I can do is 4 with ait

very slow too

[indigo carbon]

AIT is slow? You must've changed something..

[uncut fiber]

i now realized i got some red warning, and after reconnecting it gone.

[vital ermine]

BS4 it was way too slow

took about as long as ddim BS10

[indigo carbon]

Uh, can you specify the speed?

BS2 should also work

[uncut fiber]

what is difference pure unipc and unipc bh2

[shy kelp]

do you need to see the error? ill share it to you

[vital ermine]

too bad this is capped at bs4

[indigo carbon]

BS4 is the highest I managed to compile it on. But it should work also on BS2, and the speed boost should be the same 200% you were describing on BS1.

[vital ermine]

Well, how do we get a native 1024x1024 instead of upscaling?

[indigo carbon]

Add preview image to the part before going to the upscale node

[vital ermine]

Smaller

[indigo carbon]

Or you could just remove the upscale node

[vital ermine]

1600x1600 comes out as 1038x1038

I did

[thorny frost]

my firsrt workflow attempt. It uses ultimate upscaler 3 times, face detailer, xl and 1.5 checkpoints, prompt randomizer, and controlnet. The interesting part is thet the second upsacler dont upscale anyhting, only add noise, so the third upscaler can extract extra detail of it.

[indigo carbon]

You didn't remove the downscale node that goes after the upscale node to reduce file size

[thorny frost]

it is based on some old Searge workflow

[vital ermine]

that was it

[indigo carbon]

That's the preview error. Or you are using an older version of the node

[vital ermine]

BS2 DDIM

[shy kelp]

so i should do 'git pull'? i dont have any experience in git

[indigo carbon]

I don't know for sure, but it's worth trying at least

[indigo carbon]

Why 5 refiner steps?

[sand pier]

was the ksampler updated for better performance? 0.0

[shy kelp]

it shows this

[indigo carbon]

Just reinstall the entire node. Fixing this is a simple python task that I'm not the only person here capable of resolving

[shy kelp]

ok, thank you

[soft zealot]

A photo of a mature woman surrounded by vibrant floral arrangements, reminiscent of Frida Kahlo's bold and surreal self-portraits blended with Annie Leibovitz's intimate and visually striking compositions.

[thorny frost]

AIT works at first attempt for me. About a 50% faster. 3090ti

[south frigate]

can u give melink to good zip?

[thorny frost]

curious, when i load comfy using --cuda-device 1 to use my 1080ti it cdoesnt recognize ait nodes

is AIT not compatible with 1000 series and dont load the nodes?

[indigo carbon]

Yeah, that's about what you expect from 3000 series. This is the only case of 4000 being better

[indigo carbon]

1000 series don't have TC, so there is nothing to boost

[thorny frost]

but it should load the nodes the same, shouldnt it?

[uncut fiber]

mysterious exit

[indigo carbon]

No, when loading checkpoints using AIT it applies the module that boosts TC. 1000 series don't support that.

[thorny frost]

ok

[noble shoal]

AIT on my 4070 average 4.3 it/s. Without AIT 1.5 it/s

[uncut fiber]

you must have very complicate workflow

[noble shoal]

Not really. Just a bunch of notes cluttered together. (if you meant me with your message)

[uncut fiber]

yes

[noble shoal]

To the AIT professionals here: Is there a Lora workflow for AIT?

[indigo carbon]

You can take the official one and put the base through load LoRA none before going to load AIT

[noble shoal]

Ohhhhh. Thank you. Turns out i had the AIT Loader in front of the Lora.

[round walrus]

Anyone knows when Will be the reléase of 1.6?

[indigo vine]

sorry for the dumb question but I forgot: how do you move a Comfy node that has been "frozen"?

[noble shoal]

Rightclick and then "Pin"?

[indigo vine]

ty. "Unpin" would be a better ui text

[round walrus]

There is anyone on YouTube that explains how to do the 'git' thing to load 1.6 from 1.5? I can't find it.

[thorny frost]

50% faster is pretty noticeable when rendering big images using ultiamte upscaler. So adding AIT to my workflow...

[round walrus]

What is AIT?

[thorny frost]

https://github.com/FizzleDorf/AIT

it makes things daster

faster

[molten gull]

and it doesnt work with FaceDetailer/Restore 😦

[sand pier]

i tried installing it myself. ModuleNotFoundError: No module named 'comfy.controlnet'

[thorny frost]

https://github.com/ltdrdata/ComfyUI-Manager

that will find the nodes for you

[sand pier]

hmnn ill try

[dapper mural]

man when I generate hot anime girls I have to wait ten minutes and then when it comes out, the art kinda sucks. I guess I need ten thousand keywords to make a really good chick or what

[sand pier]

problem its not missing nodes, its missing some libs the node is trying to use

[sand pier]

update fixed it.

11 IT/s 0-0

jesus christ this thing is fast

[noble shoal]

What card do you have?

[sand pier]

4090

[south frigate]

@sand pier which workflow for comfy u using?

[sand pier]

Tdg8uU's, the one linked on the github

[south frigate]

and u dont have any errors with it?

hm

[sand pier]

nop

[south frigate]

and comfy updated to last version?

[sand pier]

yes

didnt work before i updated

[south frigate]

well... I updated and have a lot of errors

i will reinstall everything and check again

[sand pier]

what error do you have?

[dense chasm]

video training costs high extreamly

[south frigate]

this error

[sand pier]

it looks similar to what i had issues with before i updated

was not exactly that though

[south frigate]

yeah, but all addons are updated and comfy also

[sand pier]

The latest installer i used for comfy was this https://civitai.com/models/119528/sdxl-comfyui-ultimate-workflow

Adds a lot of extra stuff, and easy to use.

(includes a bat file which just installs everything)

[south frigate]

will test it then

[sand pier]

just click update in the manager after installing and you should be good to go

(in theory)

[south frigate]

yeah I see how to install, let's see

[strong field]

150 steps you mad man!

[south frigate]

update and then download workflow by @indigo carbon y?

[sand pier]

was to benchmark

[sand pier]

update, install the AIT plugin, then load the workflow

[south frigate]

ah, so ait is not by default, ok

[sand pier]

its not no

[azure oxide]

Don't stop at 150, the limit is like 1k in comfy 🤪

[sand pier]

Lies

its 999

[azure oxide]

Same same

[south frigate]

@sand pier thanks, did everything like you and its working, so good!
but as a newbie in this comfy can i turn off refiner somehow? just using only one model

[sand pier]

just delete the refiner nodes and remap

[south frigate]

how to remap?

[sand pier]

you just drag the lines between the nodes

[strong field]

You can also bypass the node with Ctrl+B

[sand pier]

you can also just select the same model twice with this one

but learning how to create basic configs with comfy is quite easy very useful if your using comfyui

[native knot]

Good looking out, but sadly, that's not the one causing the issue on my system. 😦 So it seems there's at least some other add-on that's doing something similar.

[south frigate]

@sand pier can u make fast a version of this workflow without refiner? cuz errors everywhere

[sand pier]

like i said you can just use the same model twice, but here you go

you'll have to change the checkpoint and upscaler

(make sure you actually use a XL checkpoint)

[south frigate]

yeah its working, ty

what is best upscaler for sdxl?

[sand pier]

probably no best, i just use 4xUltrasharp

[south frigate]

u mean, the same like for sd 1.5?

or its some custom version

for sdxl

[sand pier]

just the same

[strong field]

Not really a best, many different solutions in different situations. Comfy has a latent upscale workflow posted now.

#✨｜sdxl message

My personal fav is NMKD Siax

https://nmkd.de/?esrgan

[soft zealot]

I went though my custom nodes folder moving each one one at a time to an archive folder, restarting, checking, control z if sill broken, rinse and repeat

[native knot]

Yeah, I was fearing having to do the same.

[soft zealot]

If its any help you can exclude these

and also these (which it runed out I want actually using so left them out)

[twilit knoll]

Would this work out-of-the-box with Google Colab? I see that it also includes some 'extractor.exe' which I'm not sure if notebook would allow running?

[sand pier]

the exe you dont have to think about

[twilit knoll]

I was using this one https://colab.research.google.com/github/comfyanonymous/ComfyUI/blob/master/notebooks/comfyui_colab.ipynb & I'm not sure how to integrate this ultimate-workflow to that

[sand pier]

google colab is linux only right? so you have to rewrite the bat to a .sh

not used it

anyways that there is just a shortcut to easily fetch repositories

tbf, rewriting it is very easy. Almost exlusively commands cross platform

[twilit knoll]

I see, file locations might be a bit of hussle I guess. But I'll try. Thanks for your help

[indigo vine]

Albert configuring ComfyUI

[pure crystal]

hahahah

[twilit nexus]

i thought that is unreal engine for a while

[uncut fiber]

Comfy, can i post error iam getting running comfyUI on startup with Cannot import AIT?

[sand pier]

if unreal used 3 minutes to render a frame

[native knot]

Impossible cube, but hey...can't ask for everything.

[pure crystal]

luxury haunted pirate ship

[ionic gulch]

is there a way to redirect sampler and scheduler in comfyui? that's the only thing that prevents me from getting a clean layout

https://i.imgur.com/9EdFz9W.png

[native knot]

Can't you just use a reroute node?

[ionic gulch]

@soul afaik reroute node doesn't work for sampler and scheduler.

[native knot]

Do you have a node that is able to output a sampler_name and scheduler? Because I can connect an input from a KSampler sampler_name and a KSampler scheduler to a reroute on the receiving end.

[pure crystal]

this is a 70% great giraffe

[ionic gulch]

@native knot do you know a node that fetches the samplers and schedulers from the directory on the disk and offers a dropdown menu (or some other way) to select them? i couldn't find one - so i have to use a primitive node and that doesn't work as input for the 4+5 reroute inputs in your screenshot

[native knot]

I do not. That would be nice, though. I looked through what I had loaded and everything that has that seems to just be the sampler nodes themselves.

So they don't pass it on, they use it.

[soft zealot]

https://github.com/giriss/comfy-image-saver has sampler/scheduler selectors

[native knot]

Ayyy...there you go.

[soft zealot]

which was the node cause my double click search box to go wonky

[native knot]

Right

[south frigate]

idk, trying to render smth in SDXL @sand pier with workflow by @indigo carbon and its looks bad I think, tryed dreamshaper and copaxTimless

what should i change for better quality?

[ionic gulch]

thanks. i will take a look.

[native knot]

Copax Timeless XL is good. I'd probably go with a resolution that's closer to the 1 megapixel size, though. Although you can run 1024x768, you'll probably find better results overall by running 1344x768.

[pure crystal]

CFG @ 10 is too high, try 6-7

[native knot]

I run cfg at 10 a lot without issue...just kinda depends. But that's absolutely something to play with until you get closer to what you're looking for.

Basically, the trick is that there is no trick.

You have to tweak things.

The best thing to do is learn what everything does; what each setting means. Then use that knowledge to make adjustments.

[pure crystal]

learn XY prompt tools, try stuff

[native knot]

XY is awesome for finding something you like

[soft zealot]

so fed your OG prompt "one white woman hold flowers on top of roof, cyberpunk world, lots of neon lights" through my ChatGPT Prompt Assist to get *"A photo of a young Caucasian woman with long flowing hair, dressed in futuristic attire, standing confidently on top of a rooftop. She holds a vibrant bouquet of flowers, surrounded by a sprawling cityscape illuminated with neon lights, showcasing a cyberpunk world." * as an alternitvely written prompt then wrapped it inside some styles to get this

Models used were ProtoVision as Primary & DynaVision as secondary

[pure crystal]

nice

when people ask for help then walk away is my pet peeve

[native knot]

I use ChatGPT occasionally as well to spice up prompts or to give me ideas. It's awesome.

[pure crystal]

image quality is nuts on this

[native knot]

My god doesn't that suck? Like, why did you ask?

[cedar tusk]

Anyone working with ip-adapter? I'm looking to merge two image inputs/clipvisions but just using the model merge node in Comfy doesn't seem to work. (Basically only first adapter has any influence). I was thinking it might be possible to get the diff b/t the adapted model and the base from each ip-adapter output and then merge those and merge with base. Thoughts?

[native knot]

That's what I ran into when I tried it. Haven't spent any time beyond that yet. It's great with text influence, but I couldn't get 2 images to merge concepts.

[cedar tusk]

My guess is that the layer added by the adapter is sooo small compared to the overall model that having 2x0.1% is effectively the same as 1x0.1% more data

Or the adapter overrides the same part of the model each time

[native knot]

I tried cranking up weights and making heavy adjustments and it basically just had zero effect.

[native knot]

This is what I had hypothesized as well.

[soft zealot]

As you may have seen I've now built it into my workflow, even helps with "refining" hand written prompts

so I started with the simple prompt of "a futuristic pop up tent in a forest" which my "Assistant" rewrote as "A photo of a futuristic pop-up tent in a dense forest at twilight. Vibrant LED lights illuminate the sleek, semi-transparent structure, blending harmoniously with nature's greenery."

[native knot]

Yeah, I saw that. I don't use it often enough to keep on my screen, but at some point, I may toy around with implementing it into one of my workflows.

[eternal fog]

Yeah you can't merge them. Every way I've tried it will just use the last node node you put through.

[cedar tusk]

I'm going to try to merge the clip visions and feed that into the ip adatper

might accomplish the same thing

[native knot]

It's super annoying that it doesn't work.

[eternal fog]

I don't think it's designed to use more than 1 image. If you get it to work let me know, but I don't think it's supposed to work.