✨｜sdxl

[west breach]

which branch do i git?

[late marsh]

I didn't know! 😂

And it's very slow

[uncut steeple]

Not sure what the green thing is

[soft zealot]

errrrm depends how patient you are , takes around 220 seconds from clicking go to ending up with a 2048x2048

which in the grand scheme of thinsg aint an eternity

[late marsh]

Of course not

[soft zealot]

out of curiosity what time is it taking you?

[ionic dragon]

An Avacado

[uncut steeple]

takes 100s for me with my workflow on a 4090
guess my workflow is kinda slow

[soft zealot]

looks more like a kiwi fruit or a lime

[dusty turtle]

LOL nice

[soft zealot]

100S on a 4090 !!!!!

Trty my workflow , be interested ina time comparison

[late marsh]

Around 150 seconds.. of course I don't mind the generation time if the concept works and it doesn't come out complete mess

[urban fjord]

Not sure what the green thing is
The arc-reactor? You are Ironman now.

[peak dove]

Discord SDXL1.0

[soft zealot]

sounda bout right , shows the old 1080ri may be old but still holding up lol

[late marsh]

Yeah it's a powerhouse

[uncut steeple]

(erotic boutique photography)

[soft zealot]

its a thing lol

[ionic dragon]

the black spots made me think its an avacado

[trim orbit]

i have a 4080 and i'm using sytan's newest layout getting 25 - 30 second gens

[west breach]

wouldn't want the cold rainwater getting into the hot bath

[uncut steeple]

my workflow is kinda complicated and takes longer

[soft zealot]

my workflow is based on Syatans, and then heavily tweaked lol

[uncut steeple]

Why dont you try mine

[trim orbit]

that's a two stage upscale ot 2048. and i've modified mine, added a face detailer chain onto the end

[uncut steeple]

I tried the face detailer twice now and I hated the results

not with SDXL right now afaik

[ionic dragon]

which controlnet do they use?

depth?

[uncut steeple]

Wrong channel
#🎥｜animation

[trim orbit]

i just use it as a simple beautifying pass. without controlnet the detailing pass is going to have to be very light. .55 denoise tops.

[uncut steeple]

Yeah but it always destroys the face more than it does make it better

[upbeat summit]

those burger look good

[heady vale]

even the android is hungry

[trim orbit]

don't know what you're doing wrong. i'm not getting such results with it. if i was i'd fix it.

don't know what you're doing to push 100 second generation times from your 4090, but if you're getting results you're happy with then that's what its all about

[uncut steeple]

got a workflow I can look at?

Also when do you run the detailer? At the end?

[vagrant wasp]

why sdxl ?

[trim orbit]

just using sytan's with facedetailer from impact pack. let me get the workflow together for it

[uncut steeple]

Also you can check my current workflow from my recent images to see how I get to 100s

[uncut steeple]

Why not

[vivid tide]

I got a 3060 let me test that out

[vagrant wasp]

im using SD a1111

[soft zealot]

just pointed all the LORAs at the default offset so I could quickly run it (used Avocado Man) and I can tell it is going to take forever, I see you're using Ultimate SD Upscale for one thing lol

[uncut steeple]

I use a mix of normal upscaling and ultimate

[soft zealot]

I just use the method Sytans used, works for me 🙂

[vivid tide]

Made with an rtx 3060:

[trim orbit]

@uncut steeple

[soft zealot]

ohh thats a nice version

[vivid tide]

Yea. I like it too

[vivid tide]

Whats that

[trim orbit]

face detailer node from impact pact rekajiggered onto sytan's workflow

[soft zealot]

BTW I fully accept my workflow may not be to everyones tatses or the easiet to read though but it works for me as a daily driver view 🙂

[uncut steeple]

The detailer node is wonky when I load it

[spring fulcrum]

Can I get a copy of that fancy plate of pasta... I'd like to try your workflow and see how it goes

[vivid tide]

Love to hear that!

[soft zealot]

its embeeded in the images I posted back up there

[soft zealot]

just open in bowser , download, savem, dragndrop into comfyui

[trim orbit]

yours looks way different from mine. i dont see a lot of those values on my layout

[uncut steeple]

you can drop images into comfy directly from the browser without downloading

[uncut steeple]

[soft zealot]

zzzzzzzzzzzzz still running lol

[uncut steeple]

[trim orbit]

before and after the facedetailer pass. just a simple prompt for arnold so it doesn't look much like him, but i would say that it doesn't "destroy" the face

interface looks like this on mine

[soft zealot]

Pass , I always DL first

[trim orbit]

seems we've got conflicting versions of impact pack

[uncut steeple]

Im up to date I think

[trim orbit]

i'm probably not

[soft zealot]

didnt i see somehwere that for the latest Impact Pack you also need to UG ComfyUI?

[trim orbit]

maybe? i'm always keeping my ui up to date

[uncut steeple]

Ill just copy your settings

[soft zealot]

750 seconds to get this

[soft zealot]

note no avacado

[trim orbit]

might work. mine are mostly the default the node came with.

could just delete it and rewire it

[uncut steeple]

I mean its not bad but damn LMAO
over 12 mins

[wet nacelle]

Yeah but do you really want some shit like this??

Great work on their part but the tech just isn't there yet for it to look good.

All of what I just said is subjective and should be taken as opinion.

[soft zealot]

as opposed to 187 seconds to get this with my normal daily driver workflow

[wet nacelle]

Jim

[uncut steeple]

Cursed

[soft zealot]

This is Stable Diffusion, This is The Way

[spring fulcrum]

my system did that image in 98 seconds

[soft zealot]

I wasnt surprised at the time it took, its why I dont bither with Ultimate (which is where most o fthe time was) .

Would I like a better GPU? Yes

Do I need a btter GPU (given I'm just tinkering) ? No

[uncut steeple]

4090 I guess

[spring fulcrum]

What are you running with for a GPU?

[spring fulcrum]

yes

[soft zealot]

1080ti

cost <£200 with an ek waterblock fitted and the OEM cooler in the box

[uncut steeple]

Yeah my workflow isnt low end friendly

[trim orbit]

i feel like it's the vae baking the image that takes most time. ultimate being tiled would have a lot of that

when i watch the process move through comfy nodes, inference seems to be ultra quick and it's always come the vae that things slow right down

[spring fulcrum]

How do I run your workflow? Can I get a copy of it?

[soft zealot]

aye, Im happy with the results I get from my daily driver though 🙂

[trim orbit]

same! no complaints. i wasn't able to generate any images a year ago

[dense chasm]

lora training for sdxl total steps around 3K-5K

[soft zealot]

sigh

opn image in bowser
download
~load in comfyui

job done

[vivid tide]

what gpu u have

[soft zealot]

1080ti

[spring fulcrum]

That didnt work with boto's image... For some reason some people don't have their workflows embedded in their images... either not using ComfyUI or the have it disabled

[trim orbit]

chrome window to chrome window, you can just drag the picture directly from one window into the other

[soft zealot]

thats how I got it

[vivid tide]

I have a 3060. if you got immersive images i can test them for you lster

[spring fulcrum]

It seems to be hit or miss for me

[trim orbit]

updating to the newest impact pack killed my layout lol. i'll fix it tonite

[soft zealot]

ta but Im fine 🙂

By all means grab one of mine and play with it but no need for testing per se ;o)

[trim orbit]

imo if custom node designers are going to break the way their node works with layouts, they should roll it out better. but i'm just an old back end guy with no idea of how this rapid development world works

[uncut steeple]

drag and drop an image i made

[spring fulcrum]

I tried and for some reason nothing happens

[spring fulcrum]

I tried dragging this one and it does nothing

[uncut steeple]

this one should work

[spring fulcrum]

that worked

[uncut steeple]

Ill upload a final version when its ready, still testing a lot so take it with a grain of salt

[spring fulcrum]

this one worked too

[vagrant wasp]

where to download sdxl ?

[spring fulcrum]

where do you get your LoRAs?

[uncut steeple]

You mean the nodes or?

Top left are 3 loaders

[spring fulcrum]

I'm just using the LoRAs that were dropped with 1.0 but you had something else in there.

[uncut steeple]

Oh those were from civit.ai just for testing if it works

[spring fulcrum]

I feel like since the drop of SDXL 1.0 I can't get anything with anime to look good. I feel like SDXL 0.9 was less restricted and had more variety in some way.

[next ruin]

Dreamshaper xl 1.0 with standard refiner

[fresh path]

there are already some finetuned anime models

https://cdn.discordapp.com/attachments/1019446913268973689/1136053044669329550/Anime_SDXL_Comparison_v1.png

[spring fulcrum]

Can you send a link for that model?

[fresh path]

which one

[next ruin]

try counterfeit - I had some good anime-like results with it https://civitai.com/models/118406?modelVersionId=128439

[fresh path]

was about to link that

[spring fulcrum]

What is clip skip???

[static prawn]

dunno why but in my workflow my upscale and highresfix always removes a lot of detail

[warm hazel]

Clip skip ignores the last attention layers of a model. It was a trick used in NovelAI based models back in 1.5 days.

[spring fulcrum]

So its not very relevant to SDXL 1.0 usage?

[warm hazel]

Not really.

Some people swear by it but I basically never use it.

[uncut steeple]

Id tell you to try mine, but ive learned its not very performance friendly

[static prawn]

idk whats wrong with the sytan one, in general its perfect, just deleting details in the upscale

thats the only downside

[static prawn]

i could try

[spring fulcrum]

What are all of the relevant peices of a great image rendering process these days for SDXL 1.0?

[uncut steeple]

It runs ages on lower end GPUs

[worn quartz]

welp I have a new anxiety now - dragging pictures I like into comfy and seeing if the metadata is in it

[warm hazel]

I'm barely able to keep up with the relevant processes myself. VAEs, clip skip, refiner steps, it's gotten a bit complicated...

[mellow dome]

me too ... and comfy ui have this amount of different way to generate a image

[static prawn]

idk why sytans config deleting the details, in general its a really popular one, i allready set the steps on 30 for the highresfix so it should do the lowest possible changing of the pic.

and i set the contrast fix down too

but it still smooths the img in an extreme way

[uncut steeple]

I didnt really use it myself, only glanced over it
I prefer to work stuff out myself

[static prawn]

im just super unexperienced with creating nodes and all that stuff.

in general dont like sdxl so far, bec i feel its super overwhelmed with complex prompts where sd 1.5 shines imo

[trim orbit]

fixed it with newest versions of the nodes

[uncut steeple]

imo it listens a lot better to the prompts

[static prawn]

oh okok i have to try

[static prawn]

i just want a "nuclear worker in front of a nuclear explosion"

even that ends always in a mess for me

[trim orbit]

should say i wasn't suggesting i fixed sytan's workflow here. it works well for me. what i fixed was my hacky tagged on facedetailer custom node

[static prawn]

oh okok

dunno i didnt have luck so far with sdxl

[placid basin]

1024x8192 using SDXL (rundiffusionXL)

[spring fulcrum]

I like Sytan's new workflow but for some reason it struggles to let me upscale to 4096x4096

[uncut steeple]

Probably not what you want

[static prawn]

kinda goes in the right direction

i just have extremly problems adding multiple elements in the prompts

[trim orbit]

i tried taking it to 4096 too, with an 8x upscaler, but it kept failing on the vae for me. 16gb

[static prawn]

at some point he is just cutting out what i said in the prompt

[spring fulcrum]

it fails with 24gb too

[trim orbit]

think ultimate upscaler style tiled approaches will still benefit a lot for us home gpu boys

and girls

[soft zealot]

now lets try the same seed with different stles applied

[spring fulcrum]

agreed... But how do I frankenstein what he made to add the ultimate upscaler node to it? I'm not as good with this spaghetti as winston is.

[trim orbit]

the ultimate manual is willsmitheatingsghetti.avi

[soft zealot]

I like this but I lost themushroom , just got the white flash before the mushroom lol

[trim orbit]

i haven't seen a lot of dual negative prompts yet. neat setup

[soft zealot]

here we go @static prawn this what you're after ? 🙂

[uncut steeple]

#🤝｜tech-support

[soft zealot]

I take 2 Prompt stylers which are linked up as follows

Random/Linguistic -> Style1 pos -> pos base textG
Fundamental Negative = > Style1 neg -> neg base textG

Supplementary -> Style2 pos -> pos base textL
Fundamental Negative -> Style2 neg -> neg base textL

Supplementary -> Style2 pos -> pos refiner text (& pre con)
Fundamental Negative -> Style2 neg -> neg refiner text (& pre con)

[trim orbit]

nicely made. looks juicy

[soft zealot]

took a bit of fiddling to get it where I like it lol

[spice skiff]

What is meant by "portable" in the description for the ComfyUI download on GitHub: "ComfyUI Standalone Portable Windows Build"?

[soft zealot]

download, unzip nove folder where ever you want, run

[urban fjord]

Nice, Auto1111 has gotten refiner support through extensions 😄

I just hope it is efficient enough to be worth running

[toxic vortex]

surprise shirtless Robin

[uncut steeple]

I cant do anything against you posting this but ill say, I hate it

[static prawn]

yeah thats really nice 😄

[toxic vortex]

https://tenor.com/view/okily-dokily-flanders-the-simpsons-skiing-jokes-gif-19512585

[static prawn]

still think sd 1.5 models do something like that way better, but it would be unfair comparing trained 1.5 models with a base Model

[vivid tide]

How to add the prmpt styler?

[soft zealot]

slightly less "nuclear" but still , that's a nice explosion @static prawn

[soft zealot]

https://github.com/twri/sdxl_prompt_styler

[spice skiff]

Thank you.

[wet nacelle]

morning all

[toxic vortex]

Really though, I've found that "watercolor" token for one reason or another produces a lot more alluring/exposing subjects in SDXL

[soft zealot]

afternoon, evening, thats another day over

[wet nacelle]

https://tenor.com/view/bee-movie-lordsheen-gif-22664599

[wise edge]

Maybe watercolor painters are an especially horny lot

[soft zealot]

@static prawn

[autumn forum]

Yo where did you get those Lora loaders with an on off switch??

[uncut steeple]

ComfyRoll I think

Really handy

[soft zealot]

@static prawn this one inbound looks promising 🙂

[autumn forum]

Yep that’s it, thanks!

[soft zealot]

doh I didnt even notice the on/off switches lol

[uncut steeple]

I found those yesterday and had to use them

[ionic dragon]

how do you craft your prompts?

[soft zealot]

i let my fingers do the walking in the keyboard :p

[wet nacelle]

DOGS!

[toxic vortex]

"watercolor painting of David Duchovny"

"oil painting of David Duchovny"

Watercolor definitely more horny

[soft zealot]

but seriously @ionic dragon theres no special fomula. I may just type in a short pithy phrase and play mix and match with the stylers , or I may and try and describe something or If i get bored I turn on the random option and see what pops out lol

[ionic dragon]

whats the diff between prompt g and prompt l?

[soft zealot]

now thats a rabbit hole

https://www.reddit.com/r/StableDiffusion/comments/15c2n0q/sdxl_two_text_encoders_two_text_prompts/

[spring fulcrum]

lmao this is what I got with my experiment... Workflow is missing something for sure... ComfyUI spaghetti has a tall learning curve example... This was intended to be one image

[soft zealot]

am i using them as designed? Probably not, do I care? No

[ionic dragon]

thankyou, will take a deep dive into it

[midnight shuttle]

The best way to understand it is simply that you have 2 different text encoders to use, and the weights from both of them get concatenated. They don't really have different purposes. Just different tools and then the results are combined.

[crisp owl]

If you squint your eyes and look at it from a distance, you can see the background image haha

[midnight shuttle]

You can see how this works if you look in the Comfy code. G and L weights are just concatenated with a call to torch.cat().

[void skiff]

I use this prompt "a comical old dark giant of Ymir with a horrifying scream breaking loose from the dark icey void of ginnungagap in the mist of universe" and even a I change it a bit and do several generation the composition is always the same, when I search on Google for "ginnungagap" I can see what images it was inspired from, but what can I do to make the Bot here make larger variations?

[midnight shuttle]

I don't think the model is well trained on "ginnungagap". You would need to build a LoRA or hypernetwork (when/if these ever work in SDXL) to train the model what that word means.

[void skiff]

I though AI was born out of the darkest parts of Norse Mythology, but I had hope it had looked at "giant", "cosmical" and so on to make variations, I like to use the bot here for it do give better result than same prompt ran under Automatic1111 Web-ui.

[midnight shuttle]

If it is only trained on one or two examples of "ginnungagap" it wil always make pictures that look like those examples. I don't know the details of LoRA but I think hypernetwork can train the model on more examples of this. But hypernetwork is not yet working in A1111 with SDXL.

[soft zealot]

is this what you were getting?

[void skiff]

Yes, center placed, arms out, ice or water in foreground, sometime a cave.

[soft zealot]

took your prompt & added some terms and wrapped in styles

(that sounds like the definiton of emerging lol)

[void skiff]

Never any "breaking loose from the ice".

Or "Emerging from the ice" that I used first, just a gestalt standing center frame, arms on side with a ocean/cave/ice (with human) in foreground.

But darn that was a cool image

"trapped under the ice, the giant Ymir is breaking loose from the ice in the void of ginnungagap in the mist of universe, screaming in rage"

[fleet harness]

does anyone know how to use Wildcards in ComfyUI?

[void skiff]

"trapped under the ice is the giant Ymir, trying to break loose from the ice, in the dark ocean void of ginnungagap", even with a description that some is stuck under ice, under water give same resut if.

[wet nacelle]

https://tenor.com/view/me-when-fat-man-when-the-gif-22060334

[hoary saddle]

thoughts/ideas? best gpu with 24g vram for best price

[soft zealot]

how many fingers is this mythical dude supposed to have lol

[crisp owl]

at least 1

[soft zealot]

ptobably a 3090

[crisp owl]

ok so I've been running a batch of these also with different weights in the prompt and a couple different CFG's, i'll toss em all here in a sec. But this current one is hilarious 😆

[hoary saddle]

zotac, pny, evga, asus, gigabyte? any preferred brand to gravitate towards?

[autumn forum]

3090 easy

[void skiff]

"trapped under the ice is the giant Bob Ross trying to break loose from the ice, in the dark ocean void of ginnungagap". I is ginnungagap that lock the style, not even Bob Ross could help 😄

[midnight shuttle]

Asus is premium. EVGA no longer makes GPUs. Gigagyte and Zotac are cheap.

[crisp owl]

including one Orc as I forgot to change the supporting terms and that seemingly meant Orc. lol

[crisp owl]

woof

[wet nacelle]

some models are great for upscaling while others just smooth it out.

[alpine pine]

Both my motherboard and GPU from Asus were missing thermal pads. Kind of a worrying QA oversight.
I'd recommend MSI or PNY

[wet nacelle]

Upscaling with runDissfusionXL often smooths it out sadly.

[hardy cipher]

how does that even happen?

[alpine pine]

I have absolutely no idea

[hardy cipher]

how do they even function? seems like it'd cause things to pop

[ionic dragon]

can you please share your workflow, if you are comfortable sharing

[crisp owl]

just drag his image to your comfyui page

[visual plover]

is it possible to do transfer learning with my own images?

[hardy cipher]

I had a cpu that I had to remove for some reason, don't even recall now, but anyway, didn't distribute the thermal paste properly and that thing did not do well

[ionic dragon]

oh yeah, it has the workflow

[fleet harness]

DOes anyone have a working WIldCard Workflow I can take a look at?

for comfy

[alpine pine]

The GPU's main die had paste, but not some other components, so itd thermal throttle after an hour of use.

if you look up "ASUS motherboard thermal pad", youll find a many of posts where the gen 4(?) chipset is just barebones strapped to the cooler

[hardy cipher]

maybe someone does

[hardy cipher]

that's nuts

[spring fulcrum]

This has wildcard support it seems... But its not my workflow.. I'm not taking any credit for this

[ionic dragon]

@soft zealot are these your custom nodes?

[hardy cipher]

one of my gpu fans started acting funky in this stupid msi laptop. and it's pretty much impossible to find any real information or schematics for the internals of these things

[hardy cipher]

so had to take out about 50 screws to remove a tiny little fan. had no idea what I was doing, lol

[soft zealot]

I have no self authired custom nodes

[ionic dragon]

which custom nodes pack are you using for these

[soft zealot]

all th links are in the credits & notes section in the workflow

[ionic dragon]

oh, lemme check

thank you

[hardy cipher]

these node packs are out of control

[soft zealot]

apart from Comfyrool which is a new one Ive ust included a node from

[crisp owl]

No different than the copious amount of extensions for A1111

[visual plover]

you seem to use the refiner model. use the base model. then you can send to img2img and render with the refiner model there

[soft zealot]

exactly but I get frustarted by people who cant read 🙂

[hardy cipher]

I accidentally ran the refiner model for base as well recently. cloned the checkpoint loader and forgot to change it

[vivid tide]

Thanks

[hardy cipher]

images did turn out sort of. but no good at all

[soft zealot]

although fair does I havent added COmfyroll

[untold goblet]

Good results with chatGPT + SDXL + Sytan's workflow

Not enough detail though, how can I get a less "smooth" result?

[vivid tide]

How many diffrent things you have goddam

[hardy cipher]

higher cfg

[vivid tide]

i want that too

[vivid tide]

Are you using the refinder?

[soft zealot]

only the same as A1111

[untold goblet]

Yeah

[vivid tide]

hmmm

[hardy cipher]

how heavy are you using it?

you have to find the right balance

[vivid tide]

Nah i need that like u got it. You got a json that i can just load in comfyUI?

[hardy cipher]

also, noise offset might help

[soft zealot]

just load my images from here

[vivid tide]

How. Im new to all of this sorry for all the asking

[hardy cipher]

load it into comfyui

[vivid tide]

Yea but what.

[hardy cipher]

I don't know what that means

[vivid tide]

Is he talking about the images?

[hardy cipher]

yes

it has the workflow in the metadata of the image

so if you load it into comfyui the workflow appears

[vivid tide]

OHHHH. But how do i load in the image. IM so lost lmao

[hardy cipher]

you click the button that says "load"

[soft zealot]

https://www.google.com/search?sxsrf=AB5stBgnEW__iQHF0hmLr6CkaEXVE8iSmQ:1691081070150&q=tutorial+on+comfyui&tbm=vid

[vivid tide]

life saver

[hardy cipher]

believe me, I spent a few days perplexed by the tutorials because I didn't realize I could do that

[soft zealot]

BTW its nothing personal but every 2 minutes the same bloody question hgtes asked, does noo one do research before starting with a new program or bother scrolling back?

[hardy cipher]

just thought they were super vague

[vivid tide]

I did. But youre just on another level. You cant expect new users to be ultra smart and already knowing everything.

[vivid tide]

ahjsfouagsuhsag

i get it

When loading the graph, the following node types were not found:
Text Multiline
Logic Boolean
Text Random Prompt
Text String
Float
Text Input Switch
Text Parse Tokens
Seed With Text
Multiply
Evaluate Integers
CR Load LoRA
Text Add Token by Input
Text Concatenate
Samples Passthrough (Stat System)
Text to Console
Save Text File
Nodes that have failed to load will show as red on the graph.

[crisp owl]

#✨｜sdxl message

[cyan crown]

what sampler do you prefer in comfyUI?

[hardy cipher]

I don't mind questions. some pople are just SOOO needy, and they seem to have contempt for the people trying to help them if it doesn't work out

[soft zealot]

#✨｜sdxl message

[soft zealot]

hehehehe

[vivid tide]

Im ashemed by how stupid i currently am. Sorry to bother you god dayum

[hardy cipher]

get mad at someone answering their question because they don't like the answer.

[hardy cipher]

well I wasn't referring to you, lol

[soft zealot]

RTFM seems to have bypassed a lot of people these days

[hardy cipher]

I'm new to this ui myself, but I normally don't really like asking too many noob questions so I just experiment and google. but sometimes things aren't easily figured out without asking another person that knows more

it's hard to know exactly what you don't know

[vivid tide]

Normaly do that myself to. But google doesnt give me Winton Woof omega complex shit

[hardy cipher]

I used to ask the LLMs back when they were still helpful. but that didn't last long

[crisp owl]

complexity comes after understanding just the bare basics.

[vivid tide]

lol

[olive flare]

Hi! Do you know where/how I can find the list of tokens used by SDXL ?

[full acorn]

any one tried to run sdxl 1.0 on Google colab free tier? got stuck on vram issues

[hardy cipher]

I don't know if free tier is possible. maybe if you went no gpu?

[full acorn]

system memory is less than vram

[vivid tide]

Do i drop them all in \ComfyUI\ComfyUI\custom_nodes because there an install.bat

[full acorn]

and i tried gpu first

[soft zealot]

follow the install instructions on the individual pages

[full acorn]

maybe I should use #🤝｜tech-support instead?

[vivid tide]

For one there is none.

[hardy cipher]

my comfui install might need to be put out of it's misery tbh

[wet nacelle]

ages

[vivid tide]

It doesnt make sense

[hardy cipher]

they'd be better suited to help

[wet nacelle]

lol

[vivid tide]

It says that i want to install it on a callob wich i dont want to do

never mind

[soft zealot]

the thing is this is all bleeding edge tech orientated and not point & click for 12 o'clock flashers.

Thats not yuor fault and Im noy having a pop but the target demographic is not an American Iphone user

[vivid tide]

Language my guy. Write properly

[hardy cipher]

it's been a bit of a learning curve for me

[soft zealot]

you're presuming Im not dyslexic or have other impediments? tut tut lol

[hardy cipher]

I started using stable diffusion back in october or november, but only just started interacting with anyone else that uses it. felt like I'd get ridiculed for not knowing enough, lol

[soft zealot]

may I suggest you start with something like @high skiff s workflow which only uses OOTB ComfyUI nodes before loading other peoples more complex workdlows and also research into how to install from github

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

[ionic dragon]

@soft zealot please add this custom node link in your workflow
https://github.com/RockOfFire/ComfyUI_Comfyroll_CustomNodes

[hardy cipher]

I'm very perplexed by the fact that comfy won't load that sytan workflow

for me

[vivid tide]

Whatever. Just dont instantly answer my questions. I just scrolled down the page a little to far so i wasnt able to find an install guide.

[hardy cipher]

I really have no idea what it might be. might just reinstall from the ground up eventually

[soft zealot]

As it happens I have added it but guess what? This isnt a github or civitai supported package so all the images downnloaded wont get mgicvally updated

[ionic dragon]

oh ok

[oblique trellis]

still midjourney makes better than stable diffusion

[ionic dragon]

wait, that one isnt from that pack

[hardy cipher]

I make better than midjourney

[ionic dragon]

so nvm

[crisp owl]

Then use Midjourney?

[oblique trellis]

paid😢 😈

[alpine pine]

then use SDXL

[hardy cipher]

this is a complainer, everyone

[oblique trellis]

tru

[crisp owl]

Then learn how to use SDXL and make better images than midjourney and own your own images for free

[spark bear]

Sai will probably need to work on a base confy workflow with the meat and potatoes already built in and custom nodes bundled, new people trying to deal with our random noodle soup workflows will get confusing quick.

[vivid tide]

On the Reccommended Installation it says Navigate to your was-node-suite-comfyui folder
Portable/venv:
Run path/to/ComfUI/python_embeded/python.exe -s -m pip install -r requirements.txt. With what do i have to run it? cmd? and do i have to do it inside the root folder of it? Im not understanding it here

[soft zealot]

I was told it was somehwere back up there by @uncut steeple , I actually just copied and pasted it from his workflow

[wet nacelle]

Do you think that maybe a video tutorial could help you more than you guys talking here?

I'm just wondering

[ionic dragon]

i didnt get you

[vivid tide]

Indeed. Only problem. There is none

[soft zealot]

You can add a switchable node for lora
https://github.com/RockOfFire/ComfyUI_Comfyroll_CustomNodes

[vivid tide]

Atleast i didnt find it

[ionic dragon]

thank you

[wet nacelle]

Okay I'm lost in what you guys are talking about then. Can you briefly describe what the issue is?

I might be able to find something for you.

[soft zealot]

NPs I just used the discord search facility to find the answer

[ionic dragon]

switching my unrecognized node with a node from here?

[soft zealot]

probably

[vivid tide]

I think i fixed it.

[wet nacelle]

Nocie

Noice

[ionic dragon]

just to confirm, is it this one?

[soft zealot]

looks familiar

[ionic gulch]

how do "on trigger" and "on event" work for reroute nodes in comfyui?
https://i.imgur.com/QLC3geu.png

[vivid tide]

@soft zealot This one goes on you, where is the install guide huh?https://github.com/Derfuu/Derfuu_ComfyUI_ModdedNodes

Just paste in custom_nodes?

[soft zealot]

dont ask me , Im not the page maintainer, raise a support case 🙂

[royal fern]

install guide? just put it into the custom_nodes directory

[soft zealot]

or follow the link to their civitai page which has instructions

[crisp owl]

I see a link in that github to a very popular hosting page

[royal fern]

from the thumb it looks like he's holding a turd

[vocal rapids]

😆

[soft zealot]

thats too easy for some people

[ionic dragon]

install from comfyui manager

[wet nacelle]

https://tenor.com/view/jerma-pay-attention-to-me-sus-scream-scared-gif-26079473

[ionic dragon]

https://tenor.com/view/big-spoon-with-a-big-spoon-large-spoon-gif-12086211

[hardy cipher]

I like how it has only one taillight on the left

[crisp owl]

reverse ram air intake

[hardy cipher]

I like how it looks like its'f rom a 90s sitcom

[vocal rapids]

multiple characters is a mess , lol

[crisp owl]

is this a stickup?

[hardy cipher]

and I like it when I'm trying to connect a node on comfy and it 's like "I see you're trying to hit that 4th input down there, how about I go ahead put you up top?"

[sour obsidian]

gotta love the old school film look

[royal fern]

I TeleJunkies

[lilac wren]

Re guys,
How is it going for LoRA with the Refiner? Because with the Bases, it is enough to make a passthrought of the clip, but for the refiner??

[hardy cipher]

https://tenor.com/view/wat-what-gif-5360186

[cyan crown]

I'm working to a new showcase video... I think it will be ready next week

[hardy cipher]

decent. those guys look mean

[lilac wren]

I'm asking, how to (and if we are supposed to) use LoRAs with Refiner models??

[crisp owl]

nah, just base.
The refiner isn't gonna make any major changes, just very minor fine details

[hardy cipher]

ahh. well I normally don't, but maybe you're supposed to? not quite sure

[lilac wren]

Yeah, because Modifications made by a LoRA on a gen with a Base model are probably ruined by the Refiner afterwards... This is nonsense

[hardy cipher]

unless you're running the refiner way too hard it won't have that impact

[crisp owl]

refiner is doing just what it says, just refining. You shouldn't be having the refiner making substantial changes. If you do, probably to heavy with it

[upbeat summit]

that's awesome 🙂 great 80s look. those "blankets" on the armchair makes it even more believable 😉

[lilac wren]

That's what I thought, but the doubt is still there 🙂

[crisp owl]

That may change in the future, I know from the papers it's mentioned the ultimate plan would be to have it pared down to a single model instead of a base/refiner model

[hardy cipher]

the refiner really doesn't have a major impact at the default settings. and it's not very good at rendering stuff on it's own .at least in my experience

[upbeat summit]

so if you don't run the refiner with 50% of your generation it shouldn't really have a really big impact. but I've not tested using loras on the refiner in detail yet.

faces on the other hand...

[hardy cipher]

honestly the images are decent without the refiner, it just gives them that extra bit of refinement, as the name implies

k

[upbeat summit]

yes, your images are pretty decent 😄

[vale eagle]

With Sytan workflow, the upscale part is used with a base model which could apply the lora. This part would bring the lora likeness back after the refiner. The workflow would like base+lora->refiner->base+lora upscale

[lilac wren]

Another stupid question,
I saw very simple Workflows about img2img, but which did not suit me because we could not manage the denoising str, which on A1111 was very effective, and allowed to get as close as possible to the image of base, or not. How do I reproduce this on Comfy?

[broken hare]

me too. This is the one thing i prefer auto for

[hardy cipher]

I just get on weird tangents., lol

[vale eagle]

It has denoise in ksampler which does the same thing

[runic lark]

I remember I was able to right click on example checkpoint and add a Lora in between sampler and checkpoint. After reinstallation i don't have this any longer. I guess it was an addon, does anyone know what addon it might have been?

in comfyui*

[hardy cipher]

I didn't even know you could do that

maybe an addon of some kind? I don't know

[lilac wren]

To complete my question, this img2img is totatlly useless, we can't manage denoising str

[vale eagle]

Just double click and type lora, it should show lora related nodes, including load lora

[broken hare]

is there a workflow that has a way to send a generated image over to image to image

[strange mist]

you need to use advanced ksampler

[hardy cipher]

so many samplers. I never know which one to use

[strange mist]

there's only two? ksampler and ksampler (advanced)? 🤨
unless you have installed too many nodes

[hardy cipher]

what's too many?. more than you?

[strange mist]

nah, I've installed too few... but it will change... soon 😅

[broken hare]

thanks, i dont know how but ive missed this....im on it

[hardy cipher]

I probably did install too many tbh. but I wanted to explore things and see what was worthwhile. eventually I'll wipe this install and start anew and have a better understanding of what I wnat

[strange mist]

I've done the same couple of times 😄

[hardy cipher]

well coming from a1111 I barely understood what a lot of the concepts were. certainly didn't understand how they flowed

[strange mist]

same

[runic lark]

I mean that when you inserted a new empty lora by right clicking on the checkpoint loader and checkpoint was already hooked up to sampler it automatically was hooked up in between these. no need to realign the lines again

[hardy cipher]

stll haven't figured out how worthwhile all these extra things are. if there's really a tangible benefit to splitting up g and l

[vale eagle]

Not sure for that. Sometime when I removed the lora, the sampler and checkpoint was hooked automatically.

[hardy cipher]

I know there's an impact, just not sure how well I can control it

[cold mica]

Starting sooooon! Join us at #1029055412764422214

https://discord.gg/tR2jCs8E?event=1135975612587184239

[hot rover]

Trying out some more things with my intermediary face refinement process, I hadn't tried a lora yet, this is the William Eggleston one, I really liked the results I was getting.

[upbeat summit]

these look great

[hardy cipher]

true

[boreal bough]

wait a minute. is comfy missing from the center stage about comfy

[soft zealot]

Please someon start talking (oh ok I could just mute the music lol)

[cold mica]

he's waiting for you guys to drop by and get comfy first

[spring fulcrum]

lol

[crisp owl]

oh keep stalling, my client meeting is close to being over

[vivid tide]

@soft zealot Installed everything but im stilling missing things. DId i forget something?

[upbeat summit]

you will need those nodes (it's the CR nodes): https://github.com/RockOfFire/ComfyUI_Comfyroll_CustomNodes

[vivid tide]

thanks!

did i install anything wrong

[upbeat summit]

yeah I already looked through my nodes to figure out which ones those are... the names aren't that descriptive - we need a better system to identify those heh

it's not efficiency nodes or WAS

[vivid tide]

damm. If you find a fix id really appreciate it

[upbeat summit]

yeah I'm searching 🙂

[vivid tide]

Thanks man.

[upbeat summit]

I think it's these:
https://github.com/Derfuu/Derfuu_ComfyUI_ModdedNodes

it does have a lot of math nodes and has a "Float" and "Multiply" node

[civic sigil]

Crazy there arent default math nodes so there doesnt have to be 100 different implementations

[upbeat summit]

I can multiply with multiply and lora load with a lora load 😉

[vivid tide]

I will reinstall

[echo pine]

decided to share my SDXL config for comfyUI 🙂
give it a try and feel free to make suggestions
https://civitai.com/models/122046/sdxl-10-comfy-config

[trim orbit]

https://huggingface.co/SargeZT boom. Praise be SargeZT i'd @ but there are many

[upbeat summit]

Thanks for sharing it!

[trim orbit]

suppose this isn't apparant. these are controlnet models for xl

[cold mica]

We're all getting comfy over at #1005545148211527800 & #1029055412764422214 if you aren't there yet!

[wet nacelle]

this is for 1.0?

ah sorry

I can read now

[trim orbit]

they're very developmental

[wet nacelle]

I really just want to get a pix2pix workflow.

[trim orbit]

looks like we've got depth and seg

[wet nacelle]

I do see that.

Pix2Pix will be great though

[trim orbit]

we already got reference node released by comfy too

[wet nacelle]

I'll have to look into that

[hardy cipher]

have any of you guys used an interrogator in comfy?

I see a couple options I'll try out. just wondering if anyone had recommendations

[soft zealot]

i have no idea

Theyre both part of Derfuu_ComfyUI_ModdedNodes which I mention in the notes

@vivid tide I am taking a guess that its these 2 BTW as unfortunatekly a screen shot of the nodes in red wasnt provided, just the error message.

[upbeat summit]

yeah I referred the custom node to him

[thick glen]

does anyone know if you can use https://stablediffusionapi.com/ with controlNet ?

[soft bone]

I finally got face loras scaling decently. with a huge mix of the training params people have been talking about around here (no one's params came close to working without major tweaking).
I'm tired boss

[lean garden]

SDXL - Deforum 🙂

[soft zealot]

@visual glade upgraded earlier and now the nodes search box is like this when I dbl click on the canvas 😦

or like this if I dbl clcik to left of canvas lol

[inland vessel]

howd u get those straight lines?

[static prawn]

is someone using auto 1111 with the refiner extension? i always get grey pictures out of it

[soft zealot]

https://github.com/failfa-st/failfast-comfyui-extensions

[soft bone]

trained subjects

[upbeat summit]

I don't know how you did it, but these are great

[trim orbit]

theres a total beast of a layout in the metadata

[queen raptor]

What clip refiner uses?

[upbeat summit]

yeah - that's very cool! the likeness is also so good. @fleet harness prompted well and won the seed lottery 😉

[fleet harness]

https://tenor.com/view/xmen-beast-covid19-covid-vaccine-gif-19720140

when you load the metadata

enjoy it! its my favorite right now

[inland vessel]

how to autofix the errors?

also damn thats crazy

thanks

!

[trim orbit]

install all those custom nodes, the reload the page and bam it's automatically installed

[fleet harness]

whoops

[fleet harness]

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

there

get that

[thorny frost]

wow, i am experimenting with several samplers and random prompts, Heun, precisely the one i never used in SD1 seems to work wonderfully now. Very consistent and beuatiful images.

[crisp owl]

testing out some img2img, getting closer to figuring it out
image input on right, generated image on left

[fleet harness]

make sure to read my notes. they are special. 😉

[upbeat summit]

https://tenor.com/view/computer-graphics-retro-graphics-computer-1980s-1980s-aesthetic-gif-18479588

me taking it apart

[fleet harness]

please do

[inland vessel]

i will it's interesting to see how it works lol

this way i can actually understand step by step what's happening

[fleet harness]

its made from a few workflows ive come across so I cant take credit for everything myself.

joins us. shall you?

[thorny frost]

To keep original image shape and sense you may try lower denoising strenght and do it in more iterations

[inland vessel]

is it normal to download tons of shit for the new custom nodes? xd

[crisp owl]

I'm doing it using the preconditioner/base/refiner/hiresfix method, so I'm starting the precon not at 0 steps, but a bit further. Still testing

[lilac wren]

Amouranth SDXL

[spark bear]

the film grain overlay is kind of cute...I had some 65mm grain off some paid site somewhere, should try to dig it up.

[inland vessel]

like gigabytes

[fleet harness]

https://www.filmcomposite.com/free-film-assets

its dangerous to go alone. take this

[inland vessel]

could u share what models you used?

[inland vessel]

to get this

[fleet harness]

just the vae fix models
https://github.com/Stability-AI/generative-models

[inland vessel]

no loras?

[fleet harness]

its not needed, you can disconnect the LoRAs if you want. but this is sometimes used:
https://civitai.com/models/119667

[inland vessel]

thanks!

cus i saw it was connected on the one you genrated haha

[fleet harness]

https://civitai.com/models/117060/photorealistic-slider-sdxl-lora this too

on the other end

[urban breach]

why does the KSampler Advanced not have a denoise value?

[fleet harness]

https://tenor.com/view/magic-space-magic-gif-14255718

idk lol

[crisp owl]

denoising basically reduces steps, and the advanced gives you that option instead

[inland vessel]

i put these in checpoint folder right?

[abstract comet]

i think the way it works is you use start_at step instead, for example start at 8 and total of 16 steps = 0.5 denoise, start at 4 would be 0.75

[crisp owl]

yup basically

[urban breach]

ah interesting, thanks

[inland vessel]

sorry i'm all new to this lol

[lusty wolf]

Me too, also search within model list cuts off, cannot search , have to scroll...

[inland vessel]

thankyou

[crisp owl]

I personally like the NMKD-Siax_200k

But it's an open road, test out a bunch if ya want
https://huggingface.co/uwg/upscaler/tree/main/ESRGAN

[soft zealot]

glad its not just me

[soft zealot]

have pointe dit out over at https://app.element.io/#/room/#comfyui:matrix.org as well

[fleet harness]

SDXL in Pika LAbs

[tribal jackal]

Any idea what I'm doing wrong in this test? Been at it for a bit, not sure what's causing it. Workflow in mind is to do a 20/80 split between stage 1 and 2

unless that's what's causing it and I need to scale differently lol

Stage two is intended to have multiple refinements inspired by stage one (eventually, debugging rn)

[inland vessel]

last error

[abstract comet]

try not touching end at step, leave i at 1000 I think, just do start_at + steps

[tribal jackal]

That would generate an image that is 1 pixel wide and 64 pixels high

[inland vessel]

oh xD

i thought

that

but wasnt sure cus that was how it loaded from the settings from the generated image

[tribal jackal]

will do, ty

wait end at step is already static at 1000...?

oh for stage one

[queen adder]

If i'm training a LoRA using SDXL, should my data set images to train on only be these resolutions?

[alpine pine]

any resolution that multiplies to 1 megapixel

[queen adder]

does it help if i use a variety of aspect ratios? instead of only 1024 x 1024

[tribal jackal]

time to test preview image and make sure stage 2 is still having an effect

[alpine pine]

I'm not 100% on this, I have found my 1024x1024 character training worked fine at other aspect ratios. Along with my 21:9 char training.

So you can essentially do what you want. Maybe more ratios can help varry the composition? But I am not 100% on that

[tribal jackal]

thanks again bud

Great success!

[thorny frost]

So i generated a 30 image batch in A1111 using Rundiffusion XL and a random prompt extension called "one button prompt" and every single image is wonderful! What a time to be alive!

[high skiff]

howdy hello people

[crisp owl]

howdy

[autumn forum]

Howdy partner

[high skiff]

I am trying to make a character in oobabooga to convert ideas into split prompts for SDXL, and I was wondering if anybody knows how to save the session tags in ooba, cause having to redo all of my command line flags every time I launch is very annoying

[upbeat summit]

did the node quick search palette explode in width for anyone else? otherwise it must be a custom_node doing this

[soft zealot]

nit something ive played with and you're right , google searching isnt very helpful.

Best I could come up with was
https://github.com/oobabooga/text-generation-webui/issues/1309

[soft zealot]

yes but only since I upgraded earlier this evening

[high skiff]

yeah, unfortunately thats not at all what I am looking for. I am not sure why you can't just check a box or something to save your settings, rather than having to restart the whole thing several times as you turn your tags on

[soft zealot]

as in upgarded comfyui

[upbeat summit]

yeah it's because all upscale models, checkpoints and loras are now a oneliner string in the filter list. that makes the the palette element explode. I mean it should be easily fixable with CSS.

but I guess the string shouldn't be in there in that format

let's see if it's a custom node

I haven't tried it in vanilla comfyui

[soft zealot]

I was just about to grab a clean install myself

[inland vessel]

thanks

[royal python]

Nice gun Brad

[soft zealot]

is fine in vanilla

[fleet harness]

https://tenor.com/view/youre-welcome-gif-25058970

[fleet harness]

now we fight lol

[inland vessel]

loooool

[fleet harness]

NO ONE EVER EXPECTS THE MITCH

:p

[inland vessel]

bro my gpu so slow tho im used to just render 512 x 512

[fleet harness]

https://tenor.com/view/its-about-time-to-upgrade-our-baby-boy-joseph-saratan-gg-sphere-oras-na-ito-na-gif-17423528

[inland vessel]

next year man haha

[fleet harness]

awesome!

[tribal jackal]

https://tenor.com/view/so-so-good-james-franco-smile-gif-12479863

[hardy cipher]

what a good looking man. wow

[hardy cipher]

woo, big dog gpu

[fleet harness]

not bad honestly

I only use a QRTX4000

[crisp owl]

hey now I run a 2060S

[inland vessel]

it takes me 8 mins to gen

[fleet harness]

hmm weird, takes me like 3 max on this rig

[hardy cipher]

I think I have the worst 3060 they made

[fleet harness]

LMAO

[inland vessel]

hmmmm its in a laptop tho for me

[midnight shuttle]

You mean the 8GB model? Yeah that's bad. Only the 6 GB mobile version (I think?) is worse.

[fleet harness]

makes sense now lol

[midnight shuttle]

Almost all major slowdowns are the result of not enough VRAM or bad VRAM management by the UI. Example: Comfy and A1111 run at the same speed on a 24 GB card. But Comfy is faster on 8 GB because of better VRAM management. Actually A1111 will be slightly faster on that card because it keeps everything in VRAM all the time instead of swapping out.

[hardy cipher]

is this good? should I be getting red text that says this?

File "C:\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\custom_nodes\was-node-suite-comfyui-main\WAS_Node_Suite.py", line 8421, in blip_caption_image
subprocess.check_call([sys.executable, '-s', '-m', 'pip', '-q', '--upgrade', '--force-reinstall', 'transformers==4.26.1'])
File "C:\Python311\Lib\subprocess.py", line 413, in check_call
raise CalledProcessError(retcode, cmd)

[midnight shuttle]

It's not good but you didn't show enough to troubleshoot.

[hardy cipher]

I can paste the rest, just one moment

[inland vessel]

how do i stop it when its fucked xd

[midnight shuttle]

Most important part of Python errors is the very end, not the beginning.

[hardy cipher]

that's why I posted the end

[soft zealot]

@upbeat summit its seargeSDXL is the problem one @primal vault

been bringing over the custom nodes one by one and she go boom on seargeSDXL

[fleet harness]

click on "See Que" on the settings block

[hardy cipher]

well that's just from the error that pooped up. console could say something different

[inland vessel]

beautifull

[upbeat summit]

Thank you! I just updated comfyui again and it updated the filterlist, but that wasn't it.

[fleet harness]

always use proper saftey practices

I think he's gonna shoot the console?

[primal vault]

What kind of problem are you dealing with?

[hardy cipher]

planning on looking at the code to see what I can figure out. but not really a programmer so kind of hard for me to discern what's going on

"no such option: --upgrade
!!! Exception during processing !!!
Traceback (most recent call last):
File "C:\comfyui\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\execution.py", line 144, in recursive_execute
output_data, output_ui = get_output_data(obj, input_data_all)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfyui\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\execution.py", line 74, in get_output_data
return_values = map_node_over_list(obj, input_data_all, obj.FUNCTION, allow_interrupt=True)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfyui\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\execution.py", line 67, in map_node_over_list
results.append(getattr(obj, func)(**slice_dict(input_data_all, i)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\comfyui\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\custom_nodes\was-node-suite-comfyui-main\WAS_Node_Suite.py", line 8421, in blip_caption_image
subprocess.check_call([sys.executable, '-s', '-m', 'pip', '-q', '--upgrade', '--force-reinstall', 'transformers==4.26.1'])
File "C:\Python311\Lib\subprocess.py", line 413, in check_call
raise CalledProcessError(retcode, cmd)
subprocess.CalledProcessError: Command '['C:\comfyui\ComfyUI_windows_portable_nvidia_cu118_or_cpu\ComfyUI_windows_portable\ComfyUI\venv\Scripts\python.exe', '-s', '-m', 'pip', '-q', '--upgrade', '--force-reinstall', 'transformers==4.26.1']' returned non-zero exit status 2."

[soft zealot]

#✨｜sdxl message

[hardy cipher]

I was curious about the aesthetic score option in your sdxl text encoders. I guess it's a thing in sdxl? do you know how much of an impact it has? is there any reason I shouldn't just max out the possible aesthetic score?

[soft zealot]

or #✨｜sdxl message @primal vault

Ive done a clean install of COmfyUI and readded nodes one by one until it went boom. Apologies but it was yours she went boom on

[midnight shuttle]

Looks like it's having a problem trying to reinstall the transformers module. I would have to look up the syntax of pip but it seems to be missing the main command. Did you upgrade python or make other changes since it last worked?

[fleet harness]

"You want to look at my UAPs? Come a little closer boy, I'll read you up on a little legislation."

[primal vault]

I can repro it here, didn't notice it because I never use the search box 🙂

[soft zealot]

🙂

[hardy cipher]

I haven't attempted to use it before now. I figured out what was triggering the error, but not really a solution as much as a band aid.

[soft zealot]

in fairness its only been an issue since I (& presumably others) upgraded core ComfyUI today

[hardy cipher]

sending the string data to a text node for whatever reason triggers the error

[soft zealot]

send it to a String to Text node first ?

[midnight shuttle]

That is very odd that it would trigger a pip reinstall command. You should report this as a bug.

[heady vale]

I missed the stage event earlier. anything cool presented?

[upbeat summit]

Thanks for all your work, Searge! Really appreciated.

[midnight shuttle]

I have send a string primitive to a CLIP encoder node, but I use the manual install version not the pre-built so it's different.

[hardy cipher]

yeah, I'll investigate a bit more. just seemed really odd to me, but I'd like to try to zero in on what the issue is. I think tantamount to anything else my installation of comfyui was sloppy initially so it's leading to all sorts of issues now

[primal vault]

hmm, need to update, been on the same comfy version for a week now

[soft zealot]

@hardy cipher ah there we go its part of WAS Nodes

[hardy cipher]

ahhh

[soft zealot]

worth trying maybe

[hardy cipher]

maybe that'll work then? it's quite confusing at times. because I'll do things that the ui tricks me into thinking will work. but then some subprocess or whatever has different plans

[upbeat summit]

that node hunting really needs to stop heh. we are spending so much time looking for a node with a generic name and no way to quickly find out in the interface which package it is part of, even when it is missing. I guess the package name could be part of the workflow metadata

[soft zealot]

A string field has a limit of 255 characters, whereas a text field has a character limit of 30,000 characters

[hardy cipher]

good to know!. I actually didn't really know the difference other than they are different things

[soft zealot]

that would make life so much easier

[hardy cipher]

what's really frustrating is there aern't any tags or anything, at least that I'm aware of, that signify which node pack they're from

well unless the creator added them

[soft zealot]

I mean technically a string can fit inside a text but who knows what the node is expecting

[upbeat summit]

some custom node developers have started to name their nodes with a tag (ttn, bnk etc) in the node's name. which at least helps a ton if you know what the tag means and you right-click a missing node you can at least find its origin out

I guess a package name in the properties of the node could make sense. the node properties could be used for many more cool things

[vernal cloak]

Can anyone tell me how i copy a color that i see in a image to use in the Sketch Inpainter? (A1111)

[hardy cipher]

ugh, got that same error, lol. not a huge deal. just trying to find an interrogator that works in comfy. no luck thus far

[copper rose]

This is from a Marc Simonetti LoRa I just finished

[hardy cipher]

nice. I need to look up who that is

[soft zealot]

that feels like your VAE is pooched

[azure oxide]

#1029055412764422214 message

[hardy cipher]

yeah, I was going to say. looks like excessive noise offset or something

[azure oxide]

The answer is the same reason some models like clip skip 2 or 3 from what I understand. The pictures are just more biased at the very end.

[soft zealot]

@copper rose ???

[hardy cipher]

I'd like to see some images from the lora with a bit more contrast

[copper rose]

I can send you it if you want to mess around

[hardy cipher]

sure, why not?

[soft zealot]

or he could do like what I did and download the image lol

[hardy cipher]

orly?

[copper rose]

sent

[soft zealot]

@copper rose have they chnaged it ?
Thought you had to use specific SDXL Conditioners that take a text input

[copper rose]

I use auto1111 sorry no comfy knowledge from me

[soft zealot]

tes really

ComfyUI 101 🙂

[soft zealot]

in which case I suggeste you start with @high skiff workflow if youre using SDXL in COmfy

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

[copper rose]

No need at the moment, I generated those images in auto1111

[soft zealot]

so if you generated them in A1111 how come I was able to load the workflow graph in ComfyUI after downloadingth emiddle image ??

[fleet harness]

nice lol

[inland vessel]

i love this

aight time to sleep

goodnight

[fleet harness]

good night!

[copper rose]

🙂 I like this a ton, I might need to switch to comfy

wanna mess around with my LoRa?

[soft zealot]

but your image contained a Comfy workflow (or you using a comfy plugin in A1111?

[soft zealot]

no

[copper rose]

no idea why it contains a comfy workflow....

[soft zealot]

well it does

[hardy cipher]

comfy will load a1111 workflows

I think I'm slow in the brain because I can't figure out where that lora was sent or how I would download it

[copper pasture]

can i use sdxl with 6 gb of ram?

vram*

[midnight shuttle]

It will be very slow if it works at all.

[crisp owl]

Can try, probably tiled vae for most of the encoding/decoding

[hardy cipher]

I have 6gb vram. but mine might be extra good vram or something

[soft zealot]

Well there we go

[hardy cipher]

what's the deal with node requirement.txt files? are they not supposed to be installed by the manager when you click to install them in the ui of comfy?

or requirements.txt rather

[heady vale]

you can use that file in a cmd to update the requirements

[hardy cipher]

I know that. but I'm saying that it seems like the manager should automate that if you install a group of nodes

[heady vale]

manager only handles custom node stuff not core requirements I think

[midnight shuttle]

Comfy works better with lower VRAM because it moves models back and forth between CPU and GPU. A1111 only does that when using --medvram or --lowvram.

[hardy cipher]

so anyway, ran " forfiles /s /m requirements.txt /c "cmd /c pip install -r @path" " because I live life on the edge

[midnight shuttle]

And even then Comfy does it better.

[hardy cipher]

I did not realize so many things weren't installed, lol. oops

[midnight shuttle]

Shouldn't need to do that at all with the portable version. And with the manual version should only need to do that with the main requirements.txt.

[hardy cipher]

wasn't getting errors for all of them though. well not on load at least

well I don't knowl. it's in venv so if iit all blows up not the end of the world

[west breach]

Morning, I just tried renaming my nodes with the NODE_DISPLAY_NAME_MAPPINGS, but it didn't do anything. If I rename them in the NODE_CLASS_MAPPINGS, it breaks the workflows

[hardy cipher]

that doesn't sound like what you want

[sweet abyss]

was anyone successful creating visual effects (2d anime) with sdxl ?

[west breach]

figured it out

Sweet, this means I can get rid of the V3 etc for the latest version of the node

[upbeat summit]

@fleet harness I like your new avatar image 👨‍🦰

[heady vale]

so much detail. really nice

[hardy cipher]

true

[upbeat summit]

that's really interesting to inspect!

[spark bear]

the lettuce ears really sell it

[tribal jackal]

Can't stop won't stop!

[urban fjord]

Is that base/refiner or LoRA?

[spark bear]

just base/refiner

[upbeat summit]

nice! go for it. are you familiar with string theory ;-)?

[tribal jackal]

lol passingly. I'm more of a multiverse guy.

[upbeat summit]

two spoilers for very different reasons

[trim vigil]

lol

[west breach]

how do you unspoil something? 🧼

[upbeat summit]

refresh the channel

[heady vale]

why does all the latent upscale methods make the image look worse?

[vital ermine]

I am still trying to figure out why latent is 2048x2048

[alpine pine]

Latent res for the CLIP?

[vital ermine]

No idea I see it used a lot while everything else is 1024x1024

[heady vale]

yeah fairly sure its CLIP. I have mine on 4096

[vital ermine]

why so big?

[heady vale]

it was mentioned a few days ago that its slighly better

[vital ermine]

the real latent is 128x128 now

[alpine pine]

which latent are you refering to 2048 then?

[heady vale]

latent resolution and latent clip sizes are different

[vital ermine]

yeah, it changed it

[autumn forum]

having the sdxl clip at 2048 or 4096 is slightly better when generating an image. dont know why but its the way it is lmao

[alpine pine]

When people set the CLIP resolution to 2-4x resolution, I believe it means clip is looking at higher resolution example images to base itself off of.
Kinda similar to choosing a size on google search. Higher resolution images tend to be better images.

(Or I am completely wrong)

[autumn forum]

ooo i mean that sounds right and interesting. would make sense

[vital ermine]

See that woman in my zombie pics above? How do I make her NOT a zombie? Tried all kinds of stuff.

[fossil thorn]

Probably something like regional prompting or inpainting.

Traits bleeding through is just a huge problem in general.

[vital ermine]

inpainting in comfy is no way no how for me. Needs some major love so I hope something comes along.

[heady vale]

woman being hunted down by zombies?

[alpine pine]

is this SDXL or 1.5 (or 2)?

[vital ermine]

tried that she became a zombie

XL

[spark bear]

maybe try prompting woman on one clip and zombies on other to try to seperate them a bit

[alpine pine]

why are you generating at 768?

[fossil thorn]

Deep Floyd was better about that, I think, but I never used it. Pity it died.

[vital ermine]

because I want to, and besides I have a 1060 so not wanting to spend hours per gen

getting a new card soon then it is on

[fossil thorn]

But I think there are ways to do regional prompting with SDXL and Comfy. Probably a pain, though.

[split magnet]

it's better to start low isn't it?

[vital ermine]

regional prompting is probably what is needed. In 2.x I would inpaint easy peasy

[alpine pine]

The SAI staff are pretty adamant about never generating below 1024, so it may just be a resolution issue with the best way around it being regional prompting

[vital ermine]

Let them buy me a card then, but until then 768 it is until I can afford one of the over priced things, lol.

[urban fjord]

Inpainting works well for me...

[alpine pine]

~~overpriced ~~expensive

[urban fjord]

And you can train SDXL to do 768x768 better.

[vital ermine]

Nah, AMD, and Nvidia are both overpriced by 10-30% depending on the model.

[spice skiff]

Anyone know if there is a way to adjust your mouse's scroll wheel sensitivity within the ComyUI, or is it all OS based?

[nocturne trail]

OS based, you can change how many "lines at a time" each scroll wheel click moves

by default it's set to like 5 which is nuts

[alpine pine]

Def with some cards (4060/80), but XX90 is just expensive

[vital ermine]

separating them into G and L didn't help

[nocturne trail]

debatable; two used 3090s is a better deal than one 4090 in a lot of ways

but if you're comparing new to new then yeah the 4090 is alright value

4060Ti 16GB gets a pass solely because it's 16GB

[alpine pine]

You generally compare new to new, yea 👍

[vital ermine]

See, I disagree. They are all over priced and the 4090 is a fantastic card that should have been no more than 1399 MSRP, and I mean MSRP being the old way of higher than street pricing.

Like buying a stock car MSRP 30k get to the dealer and the stock pricing is 45k. Not how it works (I said stock no addons).

[nocturne trail]

true, but that's also not a realistic/fair comparison - it's entirely reasonable for someone who's looking at buying GPU(s) for AI to compare a new 4090 with two used 3090

nobody's paying new prices for a 3090 rn that would be absurd

[alpine pine]

Yes, but BMW (to take General's point) doesn't compete with a used BMW. They compete with their competitors at the new price.

[vital ermine]

3090 is so expensive still because it can be NVLINK'd which was removed from the 4k series on purpose. Two 3090 for a LLM > a single 4090.

[alpine pine]

If someone wants a used product, thats fine. But you loose out on the length of warranty and (sometimes) return policy. The used price is assuming that along with any wear and tear

[nocturne trail]

NVLink is mostly pointless for inference, even for LLMs - the performance difference just isn't that big (with vaguely well written software)

two 3090s wins because more VRAM always wins

[vital ermine]

This is the problem with GPUs there are no competitors. 😦

[vital ermine]

my point precisely.

[alpine pine]

AMD/Apple/Intel. Intel nor AMD compete on the Top end. Apple competes with hardware accel in specific industries

[vital ermine]

48GB > 24GB

[nocturne trail]

the actual performance delta from NVLink is low single-digit percent on a PCIe Gen4 system

[vital ermine]

Never had an Apple, nor will I ever have an Apple and I have been saying that since the Apple ][.

[nocturne trail]

even in training on a pair of A40s it's 6-8%

source: have those, did a wide range of tests (also matches up with benchmarks you can find online)

and that's in a Gen3 system, Gen4 would be even less

[alpine pine]

so when it comes to, say, raw it/s in SDXL generations. Would a single 4090 out perform 2x3090 with nvlink?

[nocturne trail]

yes because you can't really split a unet across multiple GPUs in an effective fashion

[alpine pine]

If you dont like a competitor, that is ok, but other people in the world do and Nvidia has to calculate that fact. If claiming "There are no competitors" that doesnt mean "there are no competitors I like"

[nocturne trail]

and the 4090 does have approximately 2x the raw fp16 mixed precision compute power of a 3090

[visual glade]

for hires images it should go 2x faster if you split the cond and uncond across 2 GPUs

[nocturne trail]

that would be interesting to bench

[vital ermine]

You are not getting that Apple holds very little marketshare outside their bubble of influence. They are no competitor against Nvidia.

[nocturne trail]

there wouldn't be much in the way of cross-GPU data transfer either

[visual glade]

after a certain resolution batching doesn't increase speed

[nocturne trail]

you'd have to do some pretty major code refactoring to do that, no?

since it's basically instantiating two copies of the model

[somber walrus]

You have to realize that AI is a bubble, and NVIDIA is doing well inside the bubble

[visual glade]

it wouldn't be that complicated to add to comfyui

[somber walrus]

But if you go outside it, it's either more bubbles that NVIDIA does in, or AMD beats NVIDIA in terms of performance for price

Unless you go super high end

[alpine pine]

So, using comfy and neggles data:

2x3090 is about the price of one new 4090

• 2x the vram

• anywhere from half to the same amount of performance
• less warranty
• more wear-n-tear

[vital ermine]

Thank you for confirming what I said. No competition and no way is Apple a competitor against Nvidia.

[nocturne trail]

wear and tear is not really a significant consideration with ampere GPUs for the most part, warranty will vary depending on where you are in the world, but yeah

[somber walrus]

If you only mean within the AI space, then ig yeah NVIDIA is looking pretty strong

[nocturne trail]

Apple are in no way a competitor in the AI space (yet)

[alpine pine]

100%, I dont think Nvidia is pricing their GPUs on the AI community tho

[vital ermine]

I do since they are no longer caring about gamers only AI.

[brazen patrol]

As a rule, if it doesn't die or break in the first 3 months, it won't in the first year

[somber walrus]

There's actually some Apple computers that might outcompete NVIDIA at specific price points, but it's hard to calc out

[nocturne trail]

plus you can buy used EVGA 3090s and you can't buy EVGA 4090s

[alpine pine]

Warrantied my 3090 1.5 years in. The issue was invisible for most of the time and was a manufacturing fault

[spark bear]

4090 would use a lot less power, its surprisingly efficient and loses very little perf even at 60% power, 2x3090 would be prety painful if your electric rates are high

[nocturne trail]

yeah the power limits on 4090s are a huge meme

[vital ermine]

I can't buy used I have no faith in humans ESPECIALLY when money is involved so new for me only.

[brazen patrol]

I hear you. I warrantied my 2080ti almost four years in. It was not even being manufactured anymore. EVGA honored it though

[visual glade]

yeah 3090 TI heats up my room very fast if I don't power limit it

[boreal bough]

just double checking here, but vit-big-g needs 29gb vram, right?

[nocturne trail]

absolutely not

[somber walrus]

I buy it used, since I know that business will dump out these professional cards very frequently for reasons unrelated to the card performance itself

[nocturne trail]

350W is at most 2-3% slower than 495W on a 4090 and even dropping to 300 is 3-5%

[somber walrus]

But if I wasn't buying an used professional card line from NVIDIA

[vital ermine]

I don't even like the remaining brands for Nvidia as I was always (since 2005) EVGA for Nvidia and Sapphire for AMD.

[somber walrus]

I'd more likely go new

[alpine pine]

Whats the price dif for pro cards used? sounds very interesting tbh

[somber walrus]

50% - 55%

[nocturne trail]

$2000 for a used A40 or $6000 for a new one

[somber walrus]

They are overpriced

I got my A4000 for about 600€

A new one would cost about 1.2k€

[nocturne trail]

used A4000s are a pretty good deal rn yeah

[somber walrus]

And if this wasn't my first PC I was building

[nocturne trail]

lots of enterprise workstations being decommissioned

[somber walrus]

I would look up places where they frequently dump the GPUs when they go bankrupt or need to clear them out for new GPUs

[vital ermine]

that was weird how I had to do it

[somber walrus]

But because it was my first PC I kinda couldn't really wait

[vital ermine]

I had to add zombie in the neg

[somber walrus]

The A4000 has about 3070 performance, but 16GB VRAM

And very low electricity consumption compared to others

[nocturne trail]

most of that stuff gets immediately snapped up by ecycling companies and flipped on eBay

that's just the power limit

nothing stops you setting a 3070Ti to the same power

[alpine pine]

are a4000s not binned?

[somber walrus]

If you undervolt too much it's kinda

[nocturne trail]

nope

[heady vale]

I just realised you can right click a node and collapse it. Today I learnt 👌

[somber walrus]

Wdym with binned in this context?

[nocturne trail]

not undervolt just dropping power target

[somber walrus]

You can left click the circle

To also collapse it

[vital ermine]

A6000 I was looking at but dang they are expensive

[somber walrus]

If only you could buy them on ebay :sadlaif:

[nocturne trail]

you can get them for about $2.5k if you're patient

[alpine pine]

With 4090 your min power limit varries depending on silicone lottery. So i'd assume the power limited a4000 would be 3070s that are binned to run at a lower wattage (tho I am probably using the word wrong)

[spark bear]

would be nice to pick up old gen a6000 for training

[nocturne trail]

this isn't true

[spark bear]

but the new ones dont have nvlink so the old gen might keep value better

[nocturne trail]

you can lower the power limit to about 50W or so

[somber walrus]

Oh wait

Wtf why is the A6000 so expensive

[nocturne trail]

48GB

and it has a fan

people are afraid of A40s because you need to BYO fan

[vital ermine]

Or waterblock DIY

[somber walrus]

Tbh I'm using stock GPU fan

[nocturne trail]

they're otherwise the same exact card except A6000 has faster RAM so it's about 6-7% quicker

[somber walrus]

I sometimes get thermal throttled but like

[nocturne trail]

oh also fun fact

my 4090 is actually slower at 495W than 450W

[vital ermine]

I just can't take that screaming Lear Jet next to my head blower fan the A40 wants

[nocturne trail]

but yeah A40 and A6000 share a PCB so you can throw a waterblock on no trouble they're just not cheap

[somber walrus]

Oh teh A40 is ALSO a GPU

[alpine pine]

I believe this is why people recommend limiting below the set power limit. A lot of people gain performance as they new 320w

[somber walrus]

Jesus

[nocturne trail]

and you need an active backplate 'cause back-side RAM

[somber walrus]

That naming convention

[nocturne trail]

they've been fairly consistent since ampere at least

[vital ermine]

Waterblocks have become stupidly expensive. 4090 one is around 500-600 USD now

[nocturne trail]

but yeah 350W seems to be the sweet spot for 4090 and 3090 (300-350 for 3090/Ti)

[vital ermine]

Jay from Jayz2Cents had a double one for those backside ram and 600 USD